A Novel Object-Based Supervised Classification Method with ... - MDPI

remote sensing Article

A Novel Object-Based Supervised Classification Method with Active Learning and Random Forest for PolSAR Imagery Wensong Liu 1 1

2

*

ID

, Jie Yang 1, *, Pingxiang Li 1 , Yue Han 2 , Jinqi Zhao 1

ID

and Hongtao Shi 1

State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China; [email protected] (W.L.); [email protected] (P.L.); [email protected] (J.Z.); [email protected] (H.S.) School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN 47907, USA; [email protected] Correspondence: [email protected]; Tel.: +86-027-6877-8526

Received: 21 May 2018; Accepted: 3 July 2018; Published: 9 July 2018







Abstract: Most of the traditional supervised classification methods using full-polarimetric synthetic aperture radar (PolSAR) imagery are dependent on sufficient training samples, whereas the results of pixel-based supervised classification methods show a high false alarm rate due to the influence of speckle noise. In this paper, to solve these problems, an object-based supervised classification method with an active learning (AL) method and random forest (RF) classifier is presented, which can enhance the classification performance for PolSAR imagery. The first step of the proposed method is used to reduce the influence of speckle noise through the generalized statistical region merging (GSRM) algorithm. A reliable training set is then selected from the different polarimetric features of the PolSAR imagery by the AL method. Finally, the RF classifier is applied to identify the different types of land cover in the three PolSAR images acquired by different sensors. The experimental results demonstrate that the proposed method can not only better suppress the influence of speckle noise, but can also significantly improve the overall accuracy and Kappa coefficient of the classification results, when compared with the traditional supervised classification methods. Keywords: PolSAR imagery; object-based classification; generalized statistical region merging (GSRM); active learning (AL); random forest (RF)

1. Introduction Land-cover classification of remote sensing imagery is becoming more and more important for local and regional planning [1,2], environmental impact assessment [3,4], agriculture monitoring [5], etc. Meanwhile, classification methods using optical remote sensing images have been extensively developed in recent decades because of the widespread accessibility of optical imagery. However, optical remote sensing images are subject to the influence of weather and illumination [6]. Fortunately, PolSAR data can not only overcome the disadvantages of optical remote sensing images, but can also allow improved object interpretation through making full use of the polarimetric information of PolSAR imagery [7]. Several supervised classification methods for PolSAR imagery have been developed, such as the Wishart supervised classifier [1], convolutional neural networks (CNN)-based classifier [8], the support vector machine (SVM) classifier [9], and the random forest (RF) classifier [10], and they have shown good performances when the training samples are sufficient. However, the collection of training sample sets is complicated and expensive; moreover, the process of traditional classification is time-consuming

Remote Sens. 2018, 10, 1092; doi:10.3390/rs10071092

www.mdpi.com/journal/remotesensing

Remote Sens. 2018, 10, 1092

2 of 25

due to the redundant training samples. Fortunately, the active learning (AL) method [11,12] has been developed in recent years, which can obtain a better classification performance than the traditional supervised classification methods when the training samples are limited. Although numerous methods using AL have been successfully applied to hyperspectral remote sensing images [11–13], the AL method has rarely been applied to full-polarimetric SAR images, due to the influence of the complex imaging mechanism, such as foreshortening, shadow, and overlay. At the same time, PolSAR images are subject to the influence of inherent speckle noise and still have finely divided spots despite speckle filtering, so the results of the traditional pixel-based supervised classification methods show a high false alarm rate. Therefore, supervised classification using full-polarimetric SAR data remains a challenge. In this regard, it is necessary to design a supervised classification method for PolSAR imagery that uses as few labeled samples as possible to obtain better precision. In this paper, focusing on the aforementioned problems of the traditional supervised classification methods when using PolSAR imagery, we propose an object-based classification method named GSRM_MBT_RF to enhance the classification performance. The proposed method should not be regarded as a combination of many different methods, but rather as a fully unified algorithm with different steps. The GSRM method is used to reduce the impact of inherent speckle noise; the AL algorithm is applied to select reliable training samples from the different polarimetric features of the PolSAR imagery; and the RF classifier is used as the classifier to identify the different types of land cover in the PolSAR images. The rest of this paper is organized as follows. The proposed object-based supervised classification method with AL and RF is described in Section 2. Section 3 details the results of the proposed approach on different PolSAR images from different sensors. Section 4 discusses the results of the case study. Finally, our conclusions are drawn in Section 5. 2. Materials and Methods In this paper, a novel object-based supervised classification method which integrates the advantages of three different algorithms (the GSRM algorithm, the AL method, and the RF classifier) is proposed for the classification of PolSAR imagery. The procedure of the proposed method is introduced in this section. 2.1. Polarimetric Feature Extraction Generally speaking, PolSAR imagery includes the backscattering coefficients of the four polarimetric channels (HH/HV/VH/VV) [1]. For the orthogonal polarization basis, the scattering information of ground objects can be represented by the following Sinclair matrix S [14]: " S=

Shh Svh

Shv Svv

# (1)

The different elements in matrix S represent the backscattering coefficients of the different polarimetric channels. For multi-look conditions, the covariance matrix C of the PolSAR imagery obeys a complex Wishart distribution under the reciprocity hypothesis in mono-station case. D

C = S · S∗ T

E

*

=

|Shh |2  ∗  Shv Shh ∗ Svv Shh 

∗ Shh Shv |Shv |2 ∗ Svv Shv

 ∗ + Shh Svv  ∗ Shv Svv  |Svv |2

(2)

To better achieve object interpretation, it is necessary to acquire the parameters of the polarimetric 0 features based on a physical scattering mechanism [15]. The backscattering coefficients (σHH = |Shh |2 , 2 2 0 0 σHV = |Shv | and σVV = |Svv | ) are the basic polarimetric features of PolSAR imagery. Meanwhile,

Remote Sens. 2018, 10, 1092

3 of 25

the copolarized correlation coefficient (|ρ HHVV |) can represent the attenuation and randomness of different natural objects. The theorems of polarimetric decomposition, which are related to the properties of different objects, can also be used to extract the different polarimetric features of PolSAR imagery. In recent years, several polarimetric decomposition methods have been developed and applied. In this study, we extracted the polarimetric features using different methods of polarimetric decomposition. A summary of the polarimetric parameters used in this study is shown in Table 1, including Cloude polarimetric decomposition [15], H/A/Alpha polarimetric decomposition [16], Freeman–Durden three-component decomposition [17], van Zyl decomposition [18], Yamaguchi decomposition [19], TSVM decomposition [20] and Arii decomposition [21]. The polarimetric features were extracted using the Polarimetric SAR Data Processing and Educational Tool (PolSARpro). Table 1. A summary of the polarimetric parameters used in this study. Polarimetric Parameters 0 , σHH

0 , σ0 σHV VV

|ρ HHVV |

Physical Description 0 ), HV (σ0 ), and VV (σ0 ) The backscattering coefficients of HH (σHH HV VV

The copolarized phase difference between HH and VV

C3

The nine elements of covariance matrix C3

T3

The nine elements of coherence matrix T3

H/A/Alpha

The entropy (H), anisotropy (A), and alpha angle (Alpha) obtained from Cloude decomposition

Freeman_Odd Freeman_Dbl Freeman_Vol

The parameters of surface scattering, double scattering, and volume scattering obtained from Freeman–Durden three-component decomposition

VanZyl3_Odd VanZyl3_Dbl VanZyl3_Vol

The parameters of surface scattering, double scattering, and volume scattering obtained from van Zyl decomposition

Yamaguchi4_Odd Yamaguchi4_Dbl Yamaguchi4_Vol Yamaguchi4_Hlx

The parameters of surface scattering, double scattering, volume scattering, and helix scattering obtained from Yamaguchi four-component decomposition

TSVM_psi TSVM_tau TSVM_phi TSVM_alpha Arii3_NNED_Odd Arii3_NNED_Dbl Arii3_NNED_Vol

The sixteen decomposition parameters from TSVM decomposition

The parameters of surface scattering, double scattering, and volume scattering obtained from Arii decomposition

2.2. Generalized Statistical Region Merging (GSRM) PolSAR imagery is usually subjected to the impact of inherent speckle noise, which leads to a poor performance for the traditional pixel-based supervised classification methods [22]. Fortunately, object-based classification algorithms have the potential to suppress the influence of speckle noise, and we can also use spatial feature information when using a supervised classification method [23]. Segmentation is one of the most important aspects of the object-based classification algorithms, and a number of algorithms based on segmentation of PolSAR imagery have been developed in recent years, such as the GSRM algorithm [24], the mean shift segmentation algorithm [25], and the normalized cut (Ncut) segmentation algorithm [26]. The GSRM algorithm has a number of advantages, such as rapid computation and the fact that it is independent of the data distribution, when compared with some of the traditional segmentation methods using PolSAR data [24]. Therefore, the GSRM algorithm has been widely applied in the

Remote Sens. 2018, 10, 1092

4 of 25

segmentation of PolSAR imagery [27–29]. Two essential components are defined when using the GSRM algorithm: (1) the merging predicate, which is used to confirm whether adjacent regions are merged or not; and (2) the merging order, which is followed to test the merging of regions. 2.2.1. Merging Predicate The merging predicate of the GSRM algorithm can be derived from the theory of probability [30]. If (R, R’) is a pair of adjacent areas of an observed PolSAR imagery I, for any δ (0 < δ ≤ 1), a corollary can be represented according to the Nock and Nielsen model [30]: v     u
2 0  2    u R E 2 E R u B    ln 2  Pr R − R0 − E R − R0 ≥ t + ≤δ 2Q δ | R| | R0 | 

(3)

where Q is   the maximum value of the PolSAR imagery, and is uncertain. |·| stands for cardinal and 0 E R − R is the expectation over all corresponding statistical pixels. Parameter B is set as 2 in this paper [24]. As a result, a merging predicate PG ( R, R0 ) can be obtained as follows: ( 0

PG ( R, R ) = where bG

true. f alse.

k R0 − Rk1 ≤ bG ( R, R0 ) otherwise

v     u
u 2 k E R0 k 2 2
kE R k ∞ uB 2 0 ∞   ln R, R = t + 2Q δ | R| | R0 |

(4)

(5)

2.2.2. Merging Order The merging order of the GSRM algorithm is not a stepwise optimization tactic, but is rather a pre-ordering strategy. Under the guidance of pre-ordering, the first step of the merging order calculates the gradient from the gradient function f ( p, p0 ): f ( p, p0 ) = k

p − p0 k p + p0 1

(6)

where p and p0 is a pair of adjacent pixels in the PolSAR imagery. The second step of the merging order sorts the order of f ( p, p0 ) in increasing order and merge regions of p and p’ represent. 2.3. Active Learning (AL) Most of the supervised classification methods for PolSAR imagery are limited by the availability of effective training samples. Moreover, the process of classification is time-consuming due to the redundant information of training samples. To address these issues, the AL method has been introduced in recent years [13]. AL method can be used to better select a reliable training set from unlabeled data, and has been successfully applied in the supervised classification of optical imagery [31]. An AL algorithm is aimed at iteratively enlarging the training samples by human–machine interaction, and labels samples with the maximum information from a set of unlabeled features for each class [32]. The basic function of AL can be represented as follows [33]: A = (C, Q, S, L, U )

(7)

Remote Sens. 2018, 10, 1092

5 of 25

where component C represents the classifier used in the AL method; component L is a group of labeled training sets for each class; component Q stands for the query function, which is used to query the informative training samples; component U represents the unlabeled features of the PolSAR data; and component S stands for the unlabeled set to be labeled. The general process of AL is shown in Algorithm 1. Algorithm 1. The general process of AL. Input: labeled sample set L, unlabeled sample set SU , the number of training sample sets N 1. Initialize training sample set D L based on random selection from labeled sample set L 2. While size of training sample set n < N do 3. Learn a model by classifier C according to training sample set D L 4. Select the most informative samples based on query function Q 5. Label the most informative samples from unlabeled sample set U 6. Update unlabeled sample set SU and new labeled sample set S L 7. n ← n + 1 8. End while Output: the training sample set S L of the most informative samples

The query function plays a key role in the AL method. Different schemes of sample selection have been developed in recent years, such as membership query synthesis, which includes stream-based selective sampling and pool-based selective sampling [32]. The pool-based query methods are some of the most popular methods, and they can make the best use of the posterior probabilities for each class [33]. Four different sampling schemes based on polarimetric and contextual information were implemented in this study: (1) random selection (RS), which selects the training samples from the unlabeled candidate pool using a random method; (2) a mutual information (MI)-based criterion; (3) the breaking ties (BT) algorithm; and (4) a modified breaking ties (MBT) scheme [11,13]. 2.3.1. The Mutual Information (MI)-Based Criterion The MI-based criterion [11] of AL method obtains the training samples by maximizing the mutual information between the classifier and class labels, which can select samples from the most complicated region. We suppose that xˆiMI stands for the MI between the classifier and the class labels, and a new feature vector xi is selected. The MI-based criterion of AL method can then be represented as follows: xˆiMI = arg max I (ω; yi | xi )

(8)

MI H 1 I (ω; yi | xi ) = log( ) 2 H

(9)

xi ,i ∈SU

where

where H is the Hessian matrix, and H MI stands for the Hessian matrix after including the new polarimetric feature xi . H and H MI can be represented as follows: H ≡ ∇2 (−logp(ωˆ | DL )) H MI = H + (diag( pi (ωˆ | DL )) − pi (ωˆ | DL ) pi (ωˆ | DL ) T ) ⊗ h( xi )h( xi ) T

(10)

where pi (ωˆ | DL ) represents the posterior probabilities of class labels, which we can obtain by Bayes’ theorem. Component h( xi ) stands for the input polarimetric feature. ⊗ is an operation of the Kronecker product. We obtain the MI by combining Equations (8)–(10). I (ω; yi | xi ) =

1 K 1 log(1 + ∏ xiT H −1 xi ) 2 K k =1

where K represents the number of categories.

(11)

Remote Sens. 2018, 10, 1092

6 of 25

2.3.2. Breaking Ties (BT) Algorithm The BT algorithm [32] selects the training samples by minimizing the distance of the two most probable classes, and it can select samples from the boundary region between the two most probable classes. The BT algorithm for AL method can be represented as follows: xˆiBT = arg min {max p(yi = k| xi , ωˆ ) − xi ,i ∈SU

where

k∈ L

max p(yi = k| xi , ωˆ )}

k ∈ L\{k+ }

(12)

max p(yi = k | xi , ωˆ ) stands for the most probable class for feature vector xi , which is the

k ∈ L\{k+ }

maximum of the posterior probabilities of class labels; and max p(yi = k| xi , ωˆ ) represents the second k∈ L

most probable class for feature vector xi , which can also be easily obtained by Bayes’ theorem [11,13]. 2.3.3. The Modified Breaking Ties (MBT) Algorithm Even though the MI-based criterion and BT algorithm can obtain better performances than RS, they do have some disadvantages: (1) the MI-based criterion focuses on the most complicated area, so it is prone to confused among different classes and less accurate; and (2) the BT algorithm focuses on the boundary area between the two most probable classes, so it is prone to having only one boundary. Fortunately, the MBT scheme was put forward, which can promote more diversity than the other two methods [13]. The goal of MBT scheme is to obtain the training samples by calculating the minimum probability between the largest classes in each individual class. The MBT scheme includes two main steps: the first step of MBT is to select samples from the unlabeled pool according to the samples with the same maximum a posteriori (MAP) estimation; and the second step of MBT is used to iteratively select samples from the most complicated region according to the following Formula (13): xˆiMBT = arg where

max

xi ,i ∈SU ,k ∈ L\{s}

max

xi ,i ∈SUs ,k ∈ L\{s}

p(yi = k| xi , ωˆ )

(13)

p(yi = k| xi , ωˆ ) stands for the most possibility of each class for feature vector xi .

2.4. The Procedure of the Proposed Method The basic processing procedure of the proposed method (as shown on Figure 1) consists of: (1) preprocessing of the PolSAR imagery; (2) polarimetric decomposition and feature extraction; (3) segmentation of PolSAR imagery, i.e., delineation of the land parcels using the GSRM algorithm; (4) selection of an effective training set; and (5) image classification. In this proposed method, it is essential to preprocess the PolSAR imagery because of the complexity of the PolSAR imaging procedure and the characteristics of phase coherence processing. The preprocessing of PolSAR data includes: (1) radiation correction; and (2) filtering. In this study, the radiation correction was implemented using PolSARpro software and Next ESA SAR Toolbox (NEST) software. The Lee Sigma filter has the advantages of being able to retain better details and a reduced time cost when compared with other filter methods [7], so we chose the 7 × 7 Lee Sigma filter to suppress the speckle noise. The different polarimetric features were extracted using PolSARpro software, and the GSRM algorithm was used to reduce the impact of the inherent speckle noise in the next step. The AL algorithm was then applied to select reliable training samples from the different polarimetric features of the PolSAR imagery. Finally, the RF classifier was used as the classifier to identify the different types of land cover in the PolSAR images.

Remote Sens. 2018, 10, 1092

7 of 25

Remote Sens. 2018, 10, x FOR PEER REVIEW

7 of 25

Figure Figure 1. 1. The The process process flow flow of of the the proposed proposed method.

In this proposed method, it is essential to preprocess the PolSAR imagery because of the 3. Experiments and Results complexity of the PolSAR imaging procedure and the characteristics of phase coherence processing. Three PolSARofimage datasets were used the evaluation of the method our The preprocessing PolSAR data includes: (1)for radiation correction; andproposed (2) filtering. In thisand study, method was implemented using MATLAB using language. To evaluate the effectiveness of SAR the proposed the radiation correction was implemented PolSARpro software and Next ESA Toolbox method, we compared the overall accuracy (OA) and Kappa coefficient (Kappa) of the (NEST) software. The Lee Sigma filter has the advantages of being able to retain better experimental details and a results with other classifierswith thatother selectfilter data methods points either at random via7another reduced timethose cost of when compared [7], so we choseorthe × 7 Lee related Sigma sample selection strategies. filter to suppress the speckle noise. The different polarimetric features were extracted using PolSARpro software, and the GSRM algorithm was used to reduce the impact of the inherent speckle 3.1. Description of the Image Datasets noise in the next step. The AL algorithm was then applied to select reliable training samples from In the experiments, first full-polarimetric SAR dataset used Airborne the different polarimetricthe features of the PolSAR imagery. Finally, thewas RF obtained classifier by wasthe used as the Synthetic Aperture Radar (AIRSAR) sensor. The imagery is L-band data acquired by NASA/JPL from classifier to identify the different types of land cover in the PolSAR images. Flevoland in the Netherlands in 1989. The PolSAR imagery size is 750 × 1024 pixels and the spatial resolution is 6.60and m× 12.10 m. The Pauli-RGB imagery is shown in Figure 2a and the ground truth is 3. Experiments Results shown in Figure 2b. The land-cover types of this area are mainly crops, and 15 different land-cover Three PolSAR image datasets were used for the evaluation of the proposed method and our categories are listed in Figure 2c. method was implemented using MATLAB language. To evaluate the effectiveness of the proposed The second full-polarimetric SAR dataset was acquired by the 38th Research Institute of the China method, we compared the overall accuracy (OA) and Kappa coefficient (Kappa) of the experimental Electronics Technology Group Corporation (CETC38) in 2012. This imagery is an Uninhabited Aerial results with those of other classifiers that select data points either at random or via another related Vehicle Synthetic Aperture Radar (UAVSAR) imagery (X-band) of an area near the city of Lingshui in sample selection strategies. China. The Pauli-RGB image is shown in Figure 2d. The PolSAR image size is 2220 × 2333 pixels and the spatial resolution is 0.5 m × 0.5 m. The ground truth of the PolSAR image is shown in Figure 2e, 3.1. Description of the Image Datasets which was established by on-the-spot field investigation. Five different growth stages of paddy fields In the experiments, are identified in Figure 2f.the first full-polarimetric SAR dataset used was obtained by the Airborne Synthetic Aperture Radar (AIRSAR) sensor. was The acquired imagery is data acquired by NASA/JPL The third full-polarimetric SAR dataset byL-band the Radarsat-2 on 7 December 2011. from Flevoland in the Netherlands in 1989. The PolSAR imagery size is 750 × 1024 pixels and the This imagery is a C-band of an area near the city of Wuhan, in China. The Pauli-RGB image is spatial m ×PolSAR 12.10 m. The Pauli-RGB imagery shownand in Figure 2a andresolution the ground shown resolution in Figure is 2g.6.60 The image size is 1200 × 1000is pixels the spatial is truth in Figure Theofland-cover ofisthis areainare mainly different 8.0 m is × shown 8.0 m. The ground2b. truth the PolSARtypes image shown Figure 2h, crops, which and was 15 obtained by land-cover categories are listed in Figure 2c. visual interpretation of the optical imagery corresponding to the time of Radarsat-2 imagery. Four The land-cover second full-polarimetric was different categories are SAR listeddataset in Figure 2i. acquired by the 38th Research Institute of the China Electronics Technology Group Corporation (CETC38) in 2012. This imagery is an Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) imagery (X-band) of an area near the city of Lingshui in China. The Pauli-RGB image is shown in Figure 2d. The PolSAR image size is 2220 ×

stages of paddy fields are identified in Figure 2f. The third full-polarimetric SAR dataset was acquired by the Radarsat-2 on 7 December 2011. This imagery is a C-band of an area near the city of Wuhan, in China. The Pauli-RGB image is shown in Figure 2g. The PolSAR image size is 1200 × 1000 pixels and the spatial resolution is 8.0 m × 8.0 m. The ground truth of the PolSAR image is shown in Figure 2h, which was obtained by visual Remote Sens. 2018, 10, 1092 8 of 25 interpretation of the optical imagery corresponding to the time of Radarsat-2 imagery. Four different land-cover categories are listed in Figure 2i.

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

Figure 2. Pauli-RGB images of the three experimental datasets and the ground-truth maps: (a)

Figure 2. Pauli-RGB images of the three experimental datasets and the ground-truth maps: Pauli-RGB of AIRSAR imagery; (b) Ground-truth map for the AIRSAR imagery;(c) Land-cover (a) Pauli-RGB of AIRSAR imagery; (b) Ground-truth map for the AIRSAR imagery; (c) Land-cover categories of AIRSAR imagery;(d) Pauli-RGB of UAVSAR imagery; (e) Ground-truth map for the categories of AIRSAR Pauli-RGB (e) Pauli-RGB Ground-truth map for the UAVSAR imagery; imagery;(d) (f) Land-cover categories of ofUAVSAR UAVSAR imagery; imagery; (g) of RadarSat-2 UAVSAR imagery; (f) Land-cover of UAVSAR imagery; (g) Pauli-RGB of RadarSat-2 imagery; imagery; (h) Ground-truth mapcategories for the RadarSat-2 imagery; (i) Land-cover categories of RadarSat-2 (h) Ground-truth map for the RadarSat-2 imagery; (i) Land-cover categories of RadarSat-2 imagery. imagery.

3.2. Experiments Results 3.2. Experiments andand Results

3.2.1. Experiments with AIRSAR Image In this set of experiments, 15 land-cover categories were considered for the classification. Polarimetric decomposition was implemented to extract the polarimetric parameters from the AIRSAR data and the GSRM algorithm was applied to segment the AIRSAR imagery. We used the AL algorithm to select an effective training set with high representation quality and low redundancy; the RF classifier was then used as the classifier to identify the different types of land cover in the AIRSAR images; and, finally, we evaluated the classification accuracy of the different methods. As shown in Figure 3a, PolSAR imagery is usually affected by speckle noise, which leads to the AIRSAR imagery to still have finely divided spots despite the speckle filtering in Figure 3b. For this reason, the GSRM algorithm is used to segment original AIRSAR image. To avoid the phenomenon

AIRSAR data and the GSRM algorithm was applied to segment the AIRSAR imagery. We used the AL algorithm to select an effective training set with high representation quality and low redundancy; the RF classifier was then used as the classifier to identify the different types of land cover in the AIRSAR images; and, finally, we evaluated the classification accuracy of the different methods. Remote Sens. 2018, 10,in 1092 of 25 As shown Figure 3a, PolSAR imagery is usually affected by speckle noise, which leads to9the AIRSAR imagery to still have finely divided spots despite the speckle filtering in Figure 3b. For this reason, the GSRM algorithm is used to segment original AIRSAR image. To avoid the phenomenon of over-segmentation or under-segmentation, the segmentation parameters are set during the GSRM of over-segmentation or under-segmentation, the segmentation parameters are set during the segmentation in this paper and we found the segmentation parameters which the scale parameter GSRM segmentation in this paper and we found the segmentation parameters which the scale Q is 32 and the gradient threshold ∆ is 0.5 can better retain details and preserve the shape of small parameter Q is 32 and the gradient threshold Δ is 0.5 can better retain details and preserve the land parcels than other scale parameters or gradient thresholds by visual assessment. The number of shape of small land parcels than other scale parameters or gradient thresholds by visual assessment. regions is 2235 and average number of pixels per region is 343, where the red line is the boundary of The number of regions is 2235 and average number of pixels per region is 343, where the red line is homogeneous area in Figure 3c. The result shows that GSRM segmentation can better suppress the the boundary of homogeneous area in Figure 3c. The result shows that GSRM segmentation can influence of speckle thanoftraditional filtering better suppress the noise influence speckle noise than methods. traditional filtering methods.

(a)

(b)

(c)

Figure 3. The Pauli-RGB imagery of AIRSAR after speckle filtering and GSRM segmentation: (a) The Figure 3. The Pauli-RGB imagery of AIRSAR after speckle filtering and GSRM segmentation: (a) The Pauli-RGB imagery of original AIRSAR imagery; (b) The Pauli-RGB imagery of AIRSAR after Pauli-RGB imagery of original AIRSAR imagery; (b) The Pauli-RGB imagery of AIRSAR after speckle speckle filtering; (c) The Pauli-RGB imagery of AIRSAR after GSRM segmentation. filtering; (c) The Pauli-RGB imagery of AIRSAR after GSRM segmentation.

We designed several contrasting experiments to evaluate the effectiveness of the proposed We designed severalsample contrasting experiments to (RS, evaluate the effectiveness of the proposedfirst. method: method: Four different selection strategies MI, BT, and MBT) were considered In Four different sample selection strategies (RS, MI, BT, andper MBT) considered first. In detail, wewas first detail, we first randomly selected five training samples classwere as the initial training set, which randomly selected training samples class as the initial training set, and which was spatially done spatially by five randomly selecting theper segmented objects within a class that all done polarimetric byparameters randomly were selecting the segmented objects a class andfive thatsamples all polarimetric were considered in this step; in within the second step, per classparameters were selected considered in this step; insample the second step,strategies five samples wereMBT) selected using the fourwith different using the four different selection (RS,per MI,class BT, and at each iteration, the stopping criterion of the iteration toand 10 times; the RF classifier wasthe finally applied to identify sample selection strategies (RS, MI,set BT, MBT)and at each iteration, with stopping criterion of the the different of land cover the AIRSAR imagery. In addition, two contrasting iteration set to types 10 times; and the RF in classifier was finally applied to identify theother different types of experiments were designed: (1) a pixel-based method using MBT strategies and RF land cover in the AIRSAR imagery. In addition,classification two other contrasting experiments were designed: classifier (namedclassification MBT_pixel); method and (2) an object-based classification MBTMBT_pixel); strategies (1) a pixel-based using MBT strategies and RFcombined classifierwith (named and(2)RF butclassification it only usescombined the full with T3 matrix (named MBT_T3). Figure but 4 shows and an classifier, object-based MBT strategies and RF classifier, it only the uses classification results obtained withFigure the AIRSAR when 55results samples in each category are the full T3 matrix (named MBT_T3). 4 shows imagery the classification obtained with the AIRSAR picked using differentinstrategies. It is noted that the training samplesstrategies. are used to train model imagery when the 55 samples each category are picked using the different It is noted that of the classification and test samples are used to evaluate the performance of the classification, so training training samples are used to train model of classification and test samples are used to evaluate the samples used assessment independent of those used for algorithms performance of for theclassification classification,accuracy so training samplesare used for classification accuracy assessment are training in this paper. independent of those used for algorithms training in this paper.

Remote Sens. 2018, 10, 1092

10 of 25

Remote Sens. 2018, 10, x FOR PEER REVIEW

10 of 25

(a)

(b)

(c)

(d)

(e)

(f)

Figure 4. Classification results obtained for theforAIRSAR imagery with the different strategies: (a) Figure 4. Classification results obtained the AIRSAR imagery with the different strategies: (e) BT; (f) MBT. MBT_pixel; (b) MBT_T3; (c) RS; (d) MI; (a) MBT_pixel; (b) MBT_T3; (c) RS; (d) MI; (e) BT; (f) MBT.

Table 2 lists the classification accuracies of each category, OA and Kappa of the classification Table 2 lists the classification accuracies of each category, OA and Kappa of the classification results obtained for the AIRSAR imagery with the different strategies when 55 samples in each results obtained for the AIRSAR imagery with the different strategies when 55 samples in each category are picked, which can be used to quantitatively compare the classification performances of category are picked, which can be used to quantitatively compare the classification performances of different strategies. different strategies.

Remote Sens. 2018, 10, 1092

11 of 25

Table 2. The classification results obtained for the AIRSAR imagery with different strategies. Categories Remote Sens.MBT_pixel 2018, 10, x FOR PEER MBT_T3 REVIEW

RS

MI

BT

MBT 11 of 25

Rape seed 0.7891 ± 0.0183 0.9325 ± 0.0071 0.9658 ± 0.0025 0.9997 ± 0.0002 0.9608 ± 0.0034 0.9985 ± 0.0007 Stem beams 0.9441 ± 0.0095 0.9607 ± 0.0053 0.9350 ± 0.0047 0.7747 ± 0.0214 0.9949 ± 0.0005 0.9958 ± 0.0012 Table ± 2. The classification the AIRSAR imagery with different Bare soil 0.2694 0.0212 0.7284 ±results 0.0115 obtained 0.9116 for ± 0.0102 0.9998 ± 0.0001 0.9910 ±strategies. 0.0010 0.9998 ± 0.0001 Water 0.9636 ± 0.0064 0.9637 ± 0.0087 0.9876 ± 0.0034 0.9997 ± 0.0002 0.9916 ± 0.0004 0.9996 ± 0.0001 MBT_pixel MBT_T3 RS MI BT Categories Forest 0.8947 ± 0.0120 0.9572 ± 0.0046 0.8219 ± 0.0072 0.9969 ± 0.0024 0.9791 ± 0.0011 MBT 0.9974 ± 0.0010 Rape seed 0.7891 ± 0.0183 0.9325 ± 0.0071 0.9926 0.9658±± 0.0012 0.0025 0.9997 ± 0.0002 ± 0.0034 0.9985 ±0.9989 0.0007 ± 0.0004 Wheat C 0.8975 ± 0.0074 0.9802 ± 0.0027 0.9996 ± 0.00010.9608 0.9988 ± 0.0002 0.9441 ± 0.0095 0.9607 ± 0.0053 0.9993 0.9350±± 0.0007 0.0047 0.7747 ± 0.0214 ± 0.0005 0.9958 ±0.9997 0.0012 ± 0.0002 LucerneStem beams 0.9101 ± 0.0038 0.9917 ± 0.0013 0.8615 ± 0.02160.9949 0.9991 ± 0.0005 soil 0.2694 ± 0.0212 0.7284 ± 0.0115 0.9116 ± 0.0102 0.9998 ± 0.0001 0.9910 ± 0.0010 0.9998 ± 0.0001 Wheat A Bare 0.8876 ± 0.0167 0.9909 ± 0.0011 0.9954 ± 0.0014 0.9698 ± 0.0117 0.9976 ± 0.0011 0.9827 ± 0.0040 Water 0.9636 ± 0.0064 0.9637 ± 0.0087 0.9876 ± 0.0034 0.9997 ± 0.0002 0.9916 ± 0.0004 0.9996 ± 0.0001 Peas 0.9601 ± 0.0078 0.9925 ± 0.0004 0.9908 ± 0.0003 0.9998 ± 0.0001 0.9814 ± 0.0018 0.9720 ± 0.0037 Forest 0.8947 ± 0.0120 0.9572 ± 0.0046 0.8219 ± 0.0072 0.9969 ± 0.0024 0.9791 ± 0.0011 0.9974 ± 0.0010 Wheat B 0.7901 ± 0.0147 0.9169 ± 0.0029 0.9493 ± 0.0041 0.9964 ± 0.0013 0.9879 ± 0.0025 0.9902 ± 0.0010 Wheat C 0.8975 ± 0.0074 0.9802 ± 0.0027 0.9926 ± 0.0012 0.9996 ± 0.0001 0.9988 ± 0.0002 0.9989 ± 0.0004 Beet 0.9308 ± 0.0201 0.4568 ± 0.0146 0.9597 ± 0.0023 0.9884 ± 0.0024 0.8275 ± 0.0033 0.9969 ± 0.0002 Lucerne 0.9101 ± 0.0038 0.9917 ± 0.0013 0.9993 ± 0.0007 0.8615 ± 0.0216 0.9991 ± 0.0005 0.9997 ± 0.0002 Potatoes 0.9124 ± 0.0094 0.9997 ± 0.0001 0.9775 ± 0.0017 0.9738 ± 0.0011 0.9168 ± 0.0101 0.9891 ± 0.0011 Wheat A 0.8876 ± 0.0167 0.9909 ± 0.0011 0.9954 ± 0.0014 0.9698 ± 0.0117 0.9976 ± 0.0011 0.9827 ± 0.0040 Barely 0.8654 ± 0.0182 0.9891 ± 0.0016 0.9692 ± 0.0051 0.9989 ± 0.0007 0.9983 ± 0.0020 0.9997 ± 0.0002 Peas 0.9601 ± 0.0078 0.9925 ± 0.0004 0.9908 ± 0.0003 0.9998 ± 0.0001 0.9814 ± 0.0018 0.9720 ± 0.0037 Building 0.9943 ± 0.0024 0.9989 ± 0.0007 0.9624 ± 0.0011 0.9981 ± 0.0003 0.8273 ± 0.0007 0.9981 ± 0.0004 Wheat B 0.7901 ± 0.0147 0.9169 ± 0.0029 0.9493 ± 0.0041 0.9964 ± 0.0013 0.9879 ± 0.0025 0.9902 ± 0.0010 Grass 0.8484 ± 0.0102 0.9662 ± 0.0035 0.9814 ± 0.0014 0.9888 ± 0.0010 0.9935 ± 0.0004 0.9917 ± 0.0017 Beet 0.9308 ± 0.0201 0.4568 ± 0.0146 0.9597 ± 0.0023 0.9884 ± 0.0024 0.8275 ± 0.0033 0.9969 ± 0.0002 OA 0.9014 ± 0.0111 0.9442 ± 0.0052 0.9816 ± 0.0016 0.9862 ± 0.0022 0.9924 ± 0.0017 0.9974 ± 0.0012 Potatoes 0.9124 ± 0.0094 0.9997 ± 0.0001 0.9775 ± 0.0017 0.9738 ± 0.0011 0.9168 ± 0.0101 0.9891 ± 0.0011 Kappa 0.8915 ± 0.0181 0.9388 ± 0.0057 0.9798 ± 0.0015 0.9848 ± 0.0024 0.9916 ± 0.0019 0.9971 ± 0.0015

Barely 0.8654 ± 0.0182 0.9891 ± 0.0016 0.9692 ± 0.0051 0.9989 ± 0.0007 0.9983 ± 0.0020 0.9997 ± 0.0002 Building 0.9943 ± 0.0024 0.9989 ± 0.0007 0.9624 ± 0.0011 0.9981 ± 0.0003 0.8273 ± 0.0007 0.9981 ± 0.0004 Grass 0.8484 ± 0.0102 0.9662 ± 0.0035 0.9814 ± 0.0014 0.9888 ± 0.0010 0.9935 ± 0.0004 0.9917 ± 0.0017 As Figure Table± 0.0111 2 show, the± object-based supervised methods can suppress the OA4 and 0.9014 0.9442 0.0052 0.9816 ± 0.0016 0.9862 ±classification 0.0022 0.9924 ± 0.0017 0.9974 ± 0.0012 0.8915 ±(MBT_pixel, 0.0181 0.9388 ± OA 0.005790.14% 0.9798 ±and 0.0015 0.98480.8915) ± 0.0024 and 0.9916 ± 0.0019 pleasing 0.9971 ± 0.0015 influence of Kappa speckle noise Kappa obtain classification

performances when compared with the pixel-based supervised classification method. However, the As Figure 4 and Table 2 show, the object-based supervised classification methods can suppress performance with the strategy 98.16%OA and90.14% Kappa 0.9798) the worst amongpleasing the different the influence of RS speckle noise (OA (MBT_pixel, and Kappais0.8915) and obtain sample classification selection strategies when 55 samples in each category are picked, because the RS strategy performances when compared with the pixel-based supervised classification method. However, with thefrom RS strategy (OA 98.16% and Kappa is the worst randomly selects the theperformance training samples the ground-truth maps, and0.9798) the strategy doesamong not consider the different sample selection strategies when 55 The samples in each category becauseand the Kappa the information contained in the training samples. performances withare MIpicked, (OA 98.62% randomly selects the training samples from the ground-truth maps, and the strategy 0.9848) RS andstrategy BT (OA 99.24% and Kappa 0.9916) are better than the performance with RS strategy, does not consider the information contained in the training samples. The performances with MI as these methods adequately consider the information contained in the training samples. The MI (OA 98.62% and Kappa 0.9848) and BT (OA 99.24% and Kappa 0.9916) are better than the strategyperformance focuses onwith the RS most complicated area and the strategy of the algorithm focuses strategy, as these methods adequately consider the BT information contained in on the boundary between most probable As a result,area these methods are both the area training samples.the Thetwo MI strategy focuses onclasses. the most complicated and two the strategy of the algorithmand focuses boundary areathe between the two most classes. As a 0.9971) result, when prone toBT confusion are on lessthe accurate than MBT strategy (OA probable 99.74% and Kappa these is two methods are both prone to confusion are less accurate MBT strategy (OA the scenario complicated. At the same time, theand proposed methodthan canthe fully use the polarimetric 99.74% and Kappa 0.9971) when the scenario is complicated. At the same time, the proposed information and it can improve the performance when compared with the MBT_T3 (OA 94.42% and method can fully use the polarimetric information and it can improve the performance when Kappa 0.9338) method. compared with the MBT_T3 (OA 94.42% and Kappa 0.9338) method. Figure 5Figure shows the classification accuracies with oftraining training samples from the 5 shows the classification accuracies withdifferent different numbers numbers of samples from differentthe strategies. horizontal axis represents the number of training samples and the axis different The strategies. The horizontal axis represents the number of training samples andvertical the vertical axis represents canthe be seen that the classification performances represents the OA or Kappa.theIt OA canorbeKappa. seen It that classification performances becomebecome stable when stableinwhen samplesare in each more 50, and theofperformance of the MBT the samples eachthe category morecategory than 50,are and thethan performance the MBT strategy is the best strategy is the best among the different strategies of sample selection when the samples in each among the different strategies of sample selection when the samples in each category are more than 35. category are more than 35.

(a)

(b)

Figure 5. OA and Kappa curves with different numbers of training samples from different strategies: (a) OA; (b) Kappa.

Remote Sens. 2018, 10, x FOR PEER REVIEW

12 of 25

Figure 5. OA and Kappa curves with different numbers of training samples from different strategies: Remote Sens. 10, 1092 (a) OA; (b)2018, Kappa.

12 of 25

To quantitatively compare the classification results obtained by different classification To quantitatively the classification obtained by(1) different classification algorithms, four differentcompare classification algorithms results were considered: k-nearest neighbor algorithms, (KNN) four different classification algorithms were considered: (1) k-nearest neighbor (KNN) classifier; classifier; (2) Wishart classifier; (3) LOR-LBP classifier; and (4) RF classifier. The four different (2) Wishart classifier;classification (3) LOR-LBPalgorithms classifier; and (4) RF classifier. The four different classifiers classifiers are common of machine learning, which have been widely are common classification algorithms of machine learning, which have been widely applied for the applied for the classification of PolSAR images. The KNN classifier is one of the most common classification of PolSAR images. The KNN classifier is one of the most common classification algorithms classification algorithms in the remote sensing field [34]. The Wishart classifier is an important in the based remoteon sensing field [34]. The Wishart classifier is anofimportant classifier [1]. based the Wishart classifier the Wishart probability density function PolSAR imagery TheonLOR-LBP probability density function of PolSAR imagery [1]. The LOR-LBP classifier is a classifier classifier is a classifier based on Logistic regression via the loopy belief propagation based (LBP) on Logistic[13]. regression via the loopy belief propagation (LBP) algorithm [13]. Thetrees RF classifier is an algorithm The RF classifier is an algorithm used to integrate multiple decision by the idea algorithm used to integrate multiple decision trees by the idea of ensemble learning [4]. of ensemble learning [4]. In detail, randomly selected training samples as the initial training In detail, we we firstfirst randomly selected fivefive training samples per per classclass as the initial training set, set, which done spatially by randomly selecting segmented objects within a class all the which waswas done spatially by randomly selecting the the segmented objects within a class andand thatthat all the polarimetric parameters were considered in this step; in the second step, five samples per class were polarimetric parameters were considered in this step; in the second step, five samples per class were selected using the MBT strategy of sample selection at each iteration, with the stopping criterion selected using the MBT strategy of sample selection at each iteration, with the stopping criterion of of iteration to times; 10 times; different classification algorithms were finally applied the the iteration set set to 10 andand the the fourfour different classification algorithms were finally applied to to identify the different types of land cover in the AIRSAR imagery. Figure 6 shows the classification identify the different types of land cover in the AIRSAR imagery. Figure 6 shows the classification results obtained with AIRSAR imagery when 55 samples in each category picked using results obtained with the the AIRSAR imagery when 55 samples in each category are are picked using the the different classification algorithms. different classification algorithms.

(a)

(b)

(c)

(d)

Figure 6. Classification results obtained for the AIRSAR imagery with the different classification Figure 6. Classification results obtained for the AIRSAR imagery with the different classification algorithms: (a) KNN classifier; (b) Wishart classifier; (c) LOR-LBP classifier; (d) RF classifier. algorithms: (a) KNN classifier; (b) Wishart classifier; (c) LOR-LBP classifier; (d) RF classifier.

Remote Sens. 2018, 10, 1092

13 of 25

Table 3 lists the performances (OA and Kappa) obtained for the AIRSAR Remote Sens. 2018, 10, classification x FOR PEER REVIEW 13 imagery of 25 with the different classification algorithms when 55 samples in each category are picked, which can Table 3 lists the classification (OA and Kappa) the AIRSAR imagery be used to quantitatively compare performances the classification results of obtained differentforclassification algorithms. with the different classification algorithms when 55 samples in each category are picked, which can The performance with the RF classifier (OA 99.74% and Kappa 0.9771) is the best among the different be used to quantitatively compare the classification results of different classification algorithms. The classification algorithms, which is because the RF classifier has strong generalization ability and also performance with the RF classifier (OA 99.74% and Kappa 0.9771) is the best among the different carries out the selection of implied features in the process of classification. classification algorithms, which is because the RF classifier has strong generalization ability and also carries out the selection of implied features in the process of classification. Table 3. Classification results obtained for the AIRSAR imagery with the different classification algorithms. Table 3. Classification results obtained for the AIRSAR imagery with the different classification Classification Algorithm OA Kappa algorithms.

KNN Algorithm Classification Wishart KNN LOR-LBP Wishart RF LOR-LBP RF

0.8332OA ± 0.0143 0.8680 ± 0.0024 0.8332 ± 0.0143 0.9557 ± 0.0039 0.8680± ± 0.0024 0.9974 0.0012 0.9557 ± 0.0039 0.9974 ± 0.0012

0.8173 ± 0.0139 Kappa 0.8571 ± 0.0028 0.8173 ± 0.0139 0.9512 ± 0.0041 0.8571 0.0028 0.9971±± 0.0015 0.9512 ± 0.0041 0.9971 ± 0.0015

Figure 7 plots the classification accuracies of the different classification algorithms with different Figure 7 plots the classification accuracies of the differentthe classification algorithms different numbers of training samples. The horizontal axis represents number of training with samples, and the numbers of training samples. The horizontal axis represents the number of training samples, and the the vertical axis represents the OA or Kappa of the different classification algorithms. Analogously, vertical axis represents the OA orstable Kappawhen of thethe different classification algorithms. Analogously, classification performances become samples in each category are more than 40, the and the classification performances become stable when the samples in each category are more than 40, and performance with the RF classifier is the best among the different classification algorithms. the performance with the RF classifier is the best among the different classification algorithms.

(a)

(b)

Figure 7. OA and Kappa curves with different number of training samples for the classification

Figure 7. OA and Kappa curves with different number of training samples for the classification algorithms: (a) OA; (b) Kappa. algorithms: (a) OA; (b) Kappa.

Several conclusions can be made from the results in Figures 4 and 6. On the one hand, the proposed method (GSRM_MBT_RF) obtains performance (OA 99.74% 0.9971) Several conclusions can be made fromthe thebest results in Figures 4 andand 6. Kappa On the oneinhand, all experiments, i.e., it can better suppress the speckle noise and obtain a higher OA and Kappa the proposed method (GSRM_MBT_RF) obtains the best performance (OA 99.74% and Kappa by 0.9971) taking full advantage of the polarimetric and spatial information. On the other hand, the in all experiments, i.e., it can better suppress the speckle noise and obtain a higher OA and Kappa by performance using the MBT strategy is better than the other sample selection strategies, because the taking full advantage of the polarimetric and spatial information. On the other hand, the performance MBT strategy can more effectively select the labeled samples.

using the MBT strategy is better than the other sample selection strategies, because the MBT strategy can more select labeled samples. 3.2.2.effectively Experiments with the UAVSAR Image To furtherwith assess the effectiveness and feasibility of the proposed approach, a relatively 3.2.2. Experiments UAVSAR Image

complex situation (Paddy land from the city of Lingshui) was used to test the proposed approach.

To further assess the effectiveness and(as feasibility of Figure the proposed approach, in a relatively complex Five different growth stages of paddy shown in 8) are researched this situation, including stage of tillering, stem-elongation, panicle-exsertion, flowering and ripening. There are situation (Paddy land from the city of Lingshui) was used to test the proposed approach. Five different different heights and densities of paddies, so the performance of Radar cross-section (RCS) in of growth stages of paddy (as shown in Figure 8) are researched in this situation, including stage UAVSAR imagery is different and it also leads to different responses. tillering, stem-elongation, panicle-exsertion, flowering andpolarimetric ripening. There are different heights and densities of paddies, so the performance of Radar cross-section (RCS) in UAVSAR imagery is different and it also leads to different polarimetric responses.

Remote Sens. 2018, 10, 1092

14 of 25

Remote Sens. 2018, 10, x FOR PEER REVIEW Remote Sens. 2018, 10, x FOR PEER REVIEW

(a) (a)

14 of 25

14 of 25

(b)

(c)

(b)

(c)

(d)

(e)

(d)

(e)

Figure 8. Five different growth stages of paddy: (a) stage of tillering; (b) stage of stem-elongation; (c) Figure 8. Five different growth stages of paddy: (a) stage of tillering; (b) stage of stem-elongation; stage of8.panicle-exsertion; (d) stage of flowering; stage of ripening. Figure Five different growth stages of paddy:and (a) (e) stage of tillering; (b) stage of stem-elongation; (c) (c) stage of panicle-exsertion; (d) stage of flowering; and (e) stage of ripening. stage of panicle-exsertion; (d) stage of flowering; and (e) stage of ripening.

As shown in Figure 9a, UAVSAR imagery is more affected by speckle noise due to its As 9a, UAVSAR imagery isis more affected speckle noise due high-resolution and leads to fewer scattering elements inmore the resolution unit of UAVSAR As shown shown in in Figure Figure 9a, UAVSAR imagery affected by by speckle noiseimagery. due toto its its high-resolution and leads to fewer scattering elements in the resolution unit of UAVSAR imagery. Therefore, the AIRSAR imagery still has finely divided spots despite the speckle filtering in Figure high-resolution and leads to fewer scattering elements in the resolution unit of UAVSAR imagery. Therefore, the imagery still has finely divided spots despite theisthe speckle inoriginal Figure 9b. 9b. To suppress the influence of speckle noise, the GSRM algorithm used tofiltering segment Therefore, theAIRSAR AIRSAR imagery still has finely divided spots despite speckle filtering in Figure UAVSAR imagery. The performance of segmentation is better when the scale parameter Q is 8 and the To suppress the influence of speckle noise, the GSRM algorithm is used to segment original UAVSAR 9b. To suppress the influence of speckle noise, the GSRM algorithm is used to segment original gradientThe threshold Δ is 4, by assessment. The number of regions is 11,836 and number imagery. performance of visual segmentation is better when the scale Q isaverage 8 andQthe UAVSAR imagery. The performance of segmentation is better whenparameter the scale parameter is gradient 8 and the of pixels ∆per region is 437.assessment. The result after GSRM segmentation is shown in average Figure 9cnumber and it shows threshold is 4, by visual The number of regions is 11,836 and pixels gradient threshold Δ is 4, by visual assessment. The number of regions is 11,836 and averageofnumber that GSRM can better suppress the influence of speckle noise than traditional filtering method. per regionper is 437. Theisresult GSRMafter segmentation is shown in is Figure 9c in and it shows thatitGSRM of pixels region 437. after The result GSRM segmentation shown Figure 9c and shows can better suppress the influence of speckle noise than traditional filtering method. that GSRM can better suppress the influence of speckle noise than traditional filtering method.

(a)

(b)

(c)

Figure 9. The Pauli-RGB imagery of UAVSAR after speckle filtering and GSRM segmentation: (a) (a) imagery of original UAVSAR imagery; (b) (c) The Pauli-RGB (b) The Pauli-RGB imagery of UAVSAR after speckle filtering; (c) The Pauli-RGB imagery of UAVSAR after GSRM segmentation. Figure9.9.The ThePauli-RGB Pauli-RGB imagery UAVSAR after speckle filtering GSRM segmentation: (a) Figure imagery of of UAVSAR after speckle filtering and and GSRM segmentation: (a) The The Pauli-RGB imagery of original UAVSAR imagery; (b) The Pauli-RGB imagery of UAVSAR after Pauli-RGB of originalsample UAVSAR imagery; (b) The Pauli-RGB UAVSAR after speckle Similarly,imagery four different selection strategies (RS, MI, imagery BT, andofMBT) were considered speckle filtering; (c) The Pauli-RGB imagery of UAVSAR after GSRM segmentation. filtering; (c) we Thefirst Pauli-RGB imagery of UAVSAR after samples GSRM segmentation. first. In detail, randomly selected five training in each category as the initial training

set in these experiments, which was done spatially by randomly selecting the segmented objects Similarly, four different sample selection strategies (RS, MI, BT, and MBT) were considered Similarly, sample selection strategies were (RS, MI, BT, andin MBT) considered within a classfour and different that all the polarimetric parameters considered this were step; in the secondfirst. first. In detail, we first randomly selected five training samples in each category as the initial training fivewe samples in each category were selected using the four sampleas selection strategies Instep, detail, first randomly selected five training samples indifferent each category the initial training set in these experiments, which was done spatially by randomly selecting the segmented objects at each iteration, where the stopping criterion of the iteration was set to 10 times; and, finally, the RF set in these experiments, which was done spatially by randomly selecting the segmented objects within a class and that all the polarimetric parameters were considered in this step; in the second classifier wasand applied to the identify the different types ofwere landconsidered cover in the PolSAR In within a class that all polarimetric parameters in this step;imagery. in the second step, five samples in each category were selected using the four different sample selection strategies addition, two other contrasting experiments wereusing designed: (1) adifferent pixel-based classification step, five samples in each category were selected the four sample selectionmethod strategies at each MBT iteration, where theRF stopping criterion of the iterationand was setantoobject-based 10 times; and, finally, the RF strategies classifier (named atusing each iteration, whereand the stopping criterion of MBT_pixel); the iteration was(2) set to 10 times; and,classification finally, the RF classifier was applied to identify the different types of land cover in the PolSAR imagery. In combined with MBT strategies classifier, but of it only the matrix (named MBT_T3). classifier was applied to identifyand theRF different types land uses cover infull theT3 PolSAR imagery. In addition, addition, two otherthe contrasting experiments designed: (1) a pixel-based method Figure shows classification results forwere the(1) UAVSAR imagery when 55classification samples in each two other10contrasting experiments were designed: a pixel-based classification method using MBT using MBT strategies and RF classifier (named MBT_pixel); and (2) an object-based classification category are picked from the different strategies. strategies and RF classifier (named MBT_pixel); and (2) an object-based classification combined with combined with MBT strategies and RF classifier, but it only uses the full T3 matrix (named MBT_T3). MBT strategies and RF classifier, but it only uses the full T3 matrix (named MBT_T3). Figure 10 shows Figure 10 shows the classification results for the UAVSAR imagery when 55 samples in each the classification results for the UAVSAR imagery when 55 samples in each category are picked from category are picked from the different strategies. the different strategies.

Remote Sens. 2018, 10, 1092

15 of 25

Remote Sens. 2018, 10, x FOR PEER REVIEW

15 of 25

(a)

(b)

(c)

(d)

(e)

(f)

Figure 10. Classification results for the UAVSAR imagery with different sample selection strategies:

Figure 10. Classification results(c) for the (f) MBT. with different sample selection strategies: (a) MBT_pixel; (b) MBT_T3; RS; (d)UAVSAR MI; (e) BT;imagery (a) MBT_pixel; (b) MBT_T3; (c) RS; (d) MI; (e) BT; (f) MBT. Table 4 lists classification accuracies of each category and the classification results (OA and Kappa) for the UAVSAR imagery with the different strategies when 55 samples in each category are Table 4 lists classification accuracies of each category and the classification results (OA and Kappa) picked, which can be used to quantitatively compare the classification performances of different for the UAVSAR strategies. imagery with the different strategies when 55 samples in each category are picked,

which can be used to quantitatively compare the classification performances of different strategies. Table 4. Classification results for the UAVSAR imagery with different strategies.

Table 4. Classification results for the UAVSAR imagery Categories MBT_pixel MBT_T3 RS MI with different BT strategies. MBT 0.9080 ± 0.9229 ± 0.9371 ± Paddy 1 MBT_pixel0.0075 MBT_T3 RS 0.0021 0.0020 0.8697 0.9229 ± 0.9021 ± 0.9371 0.9187 ± Paddy 1 Paddy 0.9080 ± 0.0075 ± 0.0021 ± 0.0020 2 0.0124 0.0030 0.0044 Paddy 2 0.8697 ± 0.0124 0.9021 ± 0.0030 0.9187 ± 0.0044 Paddy 3 0.8670 ± 0.9567 ± 0.9534 ± Paddy 3 0.8670 ± 0.0100 0.9567 ± 0.0017 0.9534 ± 0.0017

Categories

0.9257 ± MI 0.0025 0.8251±±0.0025 0.9257 0.0091 0.8251 ± 0.0091 0.9617 ± 0.9617 ± 0.0010

0.9072 ± 0.8730 ± 0.0031 BT 0.0023 MBT 0.8714 ± ± 0.0031 0.7836 ± 0.9072 0.8730 ± 0.0023 0.0020 0.0102 0.8714 ± 0.0020 0.7836 ± 0.0102 0.9835 ± 0.9731 ± 0.9835 ± 0.0004 0.9731 ± 0.0009

Paddy 4

0.5489 ± 0.0089

0.7855 ± 0.0041

0.7774 ± 0.0059

0.7897 ± 0.0028

0.8503 ± 0.0011

0.8145 ± 0.0012

Paddy 5

0.5989 ± 0.0102

0.7671 ± 0.0033

0.7606 ± 0.0043

0.8193 ± 0.0027

0.7333 ± 0.0102

0.8723 ± 0.0007

OA

0.7704 ± 0.0060

0.8725 ± 0.0041

0.8794 ± 0.0052

0.8856 ± 0.0043

0.8873 ± 0.0043

0.9093 ± 0.0012

Kappa

0.7516 ± 0.0063

0.8317 ± 0.0069

0.8207 ± 0.0070

0.8349 ± 0.0043

0.8372 ± 0.0042

0.8509 ± 0.0025

0.0089 0.5989 ± Paddy 5 0.0102 0.7704 ± OA 0.0060 0.7516 ± Remote Sens. 2018, 10, 1092 Kappa 0.0063

0.0041 0.7671 ± 0.0033 0.8725 ± 0.0041 0.8317 ± 0.0069

0.0059 0.7606 ± 0.0043 0.8794 ± 0.0052 0.8207 ± 0.0070

0.0028 0.8193 ± 0.0027 0.8856 ± 0.0043 0.8349 ± 0.0043

0.0011 0.7333 ± 0.0102 0.8873 ± 0.0043 0.8372 ± 0.0042

0.0012 0.8723 ± 0.0007 0.9093 ± 0.0012 0.8509 ±16 of 25 0.0025

Again, the the experimental experimental results results demonstrate that the Again, demonstrate that the object-based object-based supervised supervised classification classification methods can suppress the influence of speckle noise (MBT_pixel, OA 77.04% and Kappa Kappa 0.7516) 0.7516) methods can suppress the influence of speckle noise (MBT_pixel, OA 77.04% and and obtain obtain pleasing pleasing classification classification performances, performances, as as shown shown in in Figure Figure 10 10 and and Table Table 4. the and 4. Meanwhile, Meanwhile, the performance with the RS strategy (OA 87.94% and Kappa 0.8207) is the worst among the different performance with the RS strategy (OA 87.94% and Kappa 0.8207) is the worst among the different strategies. This because the RS randomly selects the training from the ground-truth strategies. Thisis is because thestrategy RS strategy randomly selects thesamples training samples from the maps, and the RS strategy does not consider the information contained in the training ground-truth maps, and the RS strategy does not consider the information contained in the samples. training The performances with MI (OA Kappa BT (OA 88.73% Kappa samples. The performances with 88.56% MI (OAand 88.56% and0.8349) Kappa and 0.8349) and BT (OA and 88.73% and 0.8372) Kappa are better thethan performance with RS,with which because former adequately consider 0.8372) arethan better the performance RS,iswhich is the because thestrategies former strategies adequately the information containedcontained in the training However,However, the MI strategy the most consider the information in thesamples. training samples. the MI focuses strategyon focuses on complicated area, and the BT algorithm strategy focuses on the boundary area between the two most the most complicated area, and the BT algorithm strategy focuses on the boundary area between the probable so these methods aremethods prone toare confusion and are less accurate the MBT than strategy two mostclasses, probable classes, so these prone to confusion and are than less accurate the (OA 90.93% and Kappa 0.8509). At the same time, the proposed method can fully use the polarimetric MBT strategy (OA 90.93% and Kappa 0.8509). At the same time, the proposed method can fully use information and itinformation can improve theitperformance compared with the compared MBT_T3 method (OA 87.25% the polarimetric and can improvewhen the performance when with the MBT_T3 and Kappa method (OA0.8317). 87.25% and Kappa 0.8317). Figure 11 shows shows the the classification classification accuracy accuracy with with different different numbers of training from Figure 11 numbers of training samples samples from the different differentstrategies. strategies.The The horizontal represents the number of training samples, the the horizontal axisaxis represents the number of training samples, and theand vertical vertical axis represents the OA or Kappa of the different sample selection strategies. The classification axis represents the OA or Kappa of the different sample selection strategies. The classification performances when the samples in eachincategory are moreare thanmore 50, and the performance performances become becomestable stable when the samples each category than 50, and the of the proposed method is themethod best among different performance of the proposed is thethe best amongstrategies. the different strategies.

(a)

(b)

Figure 11. OA and Kappa curves with different numbers of training samples from the different Figure 11. OA and Kappa curves with different numbers of training samples from the different strategies: (a) OA; (b) Kappa. strategies: (a) OA; (b) Kappa.

To quantitatively compare the classification results with different classification algorithms, four To quantitatively compare the classification resultsInwith different classification algorithms, different classification algorithms were also considered. detail, we first randomly selected five four different classification algorithms were also considered. In detail, we first randomly selected five training samples per class as the initial training set, which was done spatially by randomly selecting training samples per class as the initial training set, which was done spatially by randomly selecting the segmented objects within a class and that all the polarimetric parameters were considered in this the segmented objects within a class and the polarimetric parameters considered in step; in the second step, five samples perthat classallwere selected using the MBT were strategy of sample this step; at in each the second step, fivethe samples per criterion class were usingset thetoMBT strategy sample selection iteration, with stopping ofselected the iteration 10 times; andofthe four selection at each iteration, with the stopping criterion of the iteration set to 10 times; and the four different classification algorithms were finally applied to identify the different types of land cover in different classification algorithms were finally applied to identify the different types of land cover in the UAVSAR imagery. Figure 12 shows the classification results obtained with the UAVSAR imagery when 55 samples in each category are picked using the different classification algorithms.

Remote Sens. 2018, 10, x FOR PEER REVIEW

17 of 25

the UAVSAR imagery. Figure 12 shows the classification results obtained with the UAVSAR Remote Sens. 2018, 10,55 1092 17 of 25 imagery when samples in each category are picked using the different classification algorithms.

(a)

(b)

(c)

(d)

Figure 12. Classification results for the UAVSAR imagery with the different classification algorithms: Figure 12. Classification results for the UAVSAR imagery with the different classification algorithms: (a) KNN classifier; (b) Wishart classifier; (c) LOR-LBP classifier; (d) RF classifier. (a) KNN classifier; (b) Wishart classifier; (c) LOR-LBP classifier; (d) RF classifier.

Table 5 lists the classification results (OA and Kappa) for the UAVSAR imagery with the different Table 5 lists the classification results (OA andcategory Kappa) are for picked, the UAVSAR imagery withtothe classification algorithms when 55 samples in each which can be used different classification algorithms when 55 samples eachdifferent categoryclassification are picked, which can be used quantitatively compare the classification resultsin of algorithms. The to performance compare with RF the (OAclassification 90.93% and Kappa is the best among algorithms. the differentThe classification quantitatively results 0.8509) of different classification performance algorithms, which is because RF classifier hasamong strong the generalization ability and algorithms, also carries out with RF (OA 90.93% and Kappathe 0.8509) is the best different classification which the selection of implied features in the process of classification. is because the RF classifier has strong generalization ability and also carries out the selection of implied features in the process of classification. Table 5. Classification results for the UAVSAR imagery with different classification algorithms. Classification Algorithm

OA

Kappa

KNN Wishart LOR-LBP RF

0.8391 ± 0.0092 0.8492 ± 0.0051 0.8747 ± 0.0022 0.9093 ± 0.0012

0.7717 ± 0.0084 0.7890 ± 0.0039 0.8210 ± 0.0020 0.8509 ± 0.0025

Figure 13 shows the classification accuracies with different numbers of training samples from the different classification algorithms. The horizontal axis represents the number of training samples, and the vertical axis represents the OA or Kappa of the different classification algorithms. Analogously, the performance with the RF classifier remains the best among the different classification algorithms.

RF

0.9093 ± 0.0012

0.8509 ± 0.0025

Figure 13 shows the classification accuracies with different numbers of training samples from the different classification algorithms. The horizontal axis represents the number of training samples, and the vertical axis Remote Sens. 2018, 10, 1092represents the OA or Kappa of the different classification algorithms. Analogously, 18 of 25 the performance with the RF classifier remains the best among the different classification algorithms.

(a)

(b)

Figure 13. 13. OA Kappa curves curves with with different different numbers numbers of of training training samples samples from from the the different different Figure OA and and Kappa classification algorithms: algorithms: (a) (a) OA; OA; (b) (b) Kappa. Kappa. classification

Analogously, several conclusions can be made from the results in Figures 10 and 12. On the one Analogously, several conclusions can be made from the results in Figures 10 and 12. On the one hand, the proposed method obtains the best performance (OA 90.93% and Kappa 0.8509) in all the hand, the proposed method obtains the best performance (OA 90.93% and Kappa 0.8509) in all the experiments, i.e., it can better suppress the speckle noise and obtain a higher OA and Kappa by experiments, i.e., it can better suppress the speckle noise and obtain a higher OA and Kappa by taking taking full advantage of the polarimetric and spatial information. On the other hand, the full advantage of the polarimetric and spatial information. On the other hand, the performance with performance with the MBT strategy is better than the other sample selection strategies, because the the MBT strategy is better than the other sample selection strategies, because the MBT strategy can MBT strategy can more effectively select the labeled samples. more effectively select the labeled samples. 3.2.3. Experiments with RadarSat-2 Image 3.2.3. Experiments with RadarSat-2 Image To further assess the effectiveness and feasibility of the proposed approach, an actual urban To further assess the effectiveness and feasibility of the proposed approach, an actual urban scene (the city of Wuhan, China) was used to test the proposed approach. To avoid the influence of scene (the city of Wuhan, China) was used to test the proposed approach. To avoid the influence speckle noise, filtering method and GSRM algorithm were applied to original RadarSat-2 image. The of speckle noise, filtering method and GSRM algorithm were applied to original RadarSat-2 image. segmentation parameters were set during the GSRM segmentation and we found the segmentation The segmentation parameters were set during the GSRM segmentation and we found the segmentation parameters when the scale parameter Q is 16 and the gradient threshold Δ is 0.5 can better retain parameters when the scale parameter Q is 16 and the gradient threshold ∆ is 0.5 can better retain details and preserve the shape of small land parcels than other scale parameters or gradient details and preserve the shape of small land parcels than other scale parameters or gradient thresholds thresholds by visual assessment. The number of regions is 4135 and average number of pixels per by visual assessment. The number of regions is 4135 and average number of pixels per region is 283. region is 283. The result after GSRM segmentation (Figure 14c) shows that GSRM segmentation can The result after GSRM segmentation (Figure 14c) shows that GSRM segmentation can better suppress better suppress the influence of speckle noise than traditional filtering methods. the influence of speckle noise than traditional filtering methods.

Remote Sens. 2018, 10, 1092

19 of 25

RemoteSens. Sens.2018, 2018,1010 FORPEER PEERREVIEW REVIEW Remote , x, xFOR

(a)The ThePauli-RGB Pauli-RGBimagery imageryofof (a) original RadarSat-2 imagery original RadarSat-2 imagery

1919ofof2525

(b)The ThePauli-RGB Pauli-RGBimagery imageryofof (c)The ThePauli-RGB Pauli-RGBimagery imageryofof (b) (c) RadarSat-2 after speckle filtering RadarSat-2 after GSRM segmentation RadarSat-2 after speckle filtering RadarSat-2 after GSRM segmentation

Figure14. 14.The ThePauli-RGB Pauli-RGBimagery imageryofofUAVSAR UAVSARafter afterspeckle specklefiltering filteringand andGSRM GSRMsegmentation: segmentation:(a) (a) Figure Figure 14. The Pauli-RGB imagery of UAVSAR after speckle filtering and GSRM segmentation: (a) The ThePauli-RGB Pauli-RGBimagery imageryofoforiginal originalRadarSat-2 RadarSat-2imagery; imagery;(b) (b)The ThePauli-RGB Pauli-RGBimagery imageryofofRadarSat-2 RadarSat-2 The Pauli-RGB imagery of original RadarSat-2 imagery; (b) The Pauli-RGB imagery of RadarSat-2 after afterspeckle specklefiltering; filtering;(c) (c) ThePauli-RGB Pauli-RGBimagery imageryofofRadarSat-2 RadarSat-2 afterGSRM GSRMsegmentation. segmentation. after speckle filtering; (c) The The Pauli-RGB imagery of RadarSat-2 afterafter GSRM segmentation.

Similarly,four fourdifferent differentsample sampleselection selectionstrategies strategieswere wereconsidered consideredfirst. first.InIndetail, detail,we westill stillfirst first Similarly, Similarly, four different sample selection strategies were considered first. training In detail,set we in stillthese first randomly selected five training training samples each category category the initial initial randomly selected five samples inin each asas the training set in these randomly selected five training samples in category as the initial training setobjects in thesewithin experiments, experiments, which was donespatially spatially byeach randomly selecting thesegmented segmented class experiments, which was done by randomly selecting the objects within aaclass which was done spatially by randomly selecting the segmented objects within a class and allfive the and that that all all the the polarimetric polarimetric parameters parameters were were considered considered inin this this step; step;ininthe thesecond secondthat step, and step, five polarimetric parameters were considered in this step; in the second step, five samples in each category samplesinineach eachcategory categorywere wereselected selectedusing usingthe thefour fourdifferent differentsample sampleselection selectionstrategies strategiesatateach each samples were selected using the four different sample selection strategies at each iteration, where the stopping iteration,where wherethe thestopping stoppingcriterion criterionofofthe theiteration iterationwas wasset settoto10 10times; times;and, and,finally, finally,the theRF RF iteration, criterion was of theapplied iteration was set to the 10 times; and,types finally, the RF classifier identify the classifier was applied identify the different types ofof land cover ininwas theapplied PolSARtoimagery. imagery. classifier toto identify different land cover the PolSAR InIn differenttwo types of land cover in experiments the PolSAR imagery. In addition, other contrasting experiments addition, twoother other contrasting experiments weredesigned: designed: (1)aatwo pixel-based classification method addition, contrasting were (1) pixel-based classification method were designed: (1) a pixel-based classification method using MBT strategies and RF classifier (named usingMBT MBTstrategies strategiesand andRF RFclassifier classifier(named (namedMBT_pixel); MBT_pixel);and and(2) (2)an anobject-based object-basedclassification classification using MBT_pixel); and (2) an object-based classification combined with MBT strategies and RF classifier, combinedwith withMBT MBTstrategies strategiesand andRF RFclassifier, classifier,but butititonly onlyuses usesthe thefull fullT3 T3matrix matrix(named (namedMBT_T3). MBT_T3). combined but it only uses the full T3 matrix (named MBT_T3). Figure 15 shows the classification results the Figure 15 shows the classification results for the RadarSat-2 imagery when 55 samples each Figure 15 shows the classification results for the RadarSat-2 imagery when 55 samples ininfor each RadarSat-2 imagery when 55 samples in each category are picked from the different strategies. categoryare arepicked pickedfrom fromthe thedifferent differentstrategies. strategies. category

(a) (a)

(b) (b) Figure 15. Cont.

Remote Sens. 2018, 10, 1092

20 of 25

Remote Sens. 2018, 10, x FOR PEER REVIEW

20 of 25

(c)

(d)

(e)

(f)

Figure 15. Classification Classification results for the the RadarSat-2 RadarSat-2 imagery imagery with strategies: (a) Figure 15. results for with different different strategies: (a) MBT_pixel; MBT_pixel; (e) BT; (f) MBT . (b) MBT_T3; (c) RS; (d) MI; (b) MBT_T3; (c) RS; (d) MI; (e) BT; (f) MBT.

Table 6 lists classification accuracies of each category and the classification results (OA and Table 6 lists classification accuracies of each category and the classification results (OA and Kappa) Kappa) for the RadarSat-2 imagery with the different strategies when 55 samples in each category for the RadarSat-2 imagery with the different strategies when 55 samples in each category are picked, are picked, which can be used to quantitatively compare the classification performances of different which can be used to quantitatively compare the classification performances of different strategies. strategies. Table 6. Classification results for the RadarSat-2 imagery with different strategies. Table 6. Classification results for the RadarSat-2 imagery with different strategies. Categories Categories Building Building Forst Forst Water Water Soil Soil OA OA Kappa Kappa

MBT_pixel MBT_pixel 0.6045 ± 0.0164 0.6045 ± 0.0164 0.5928 ± 0.0138 0.5928 ± 0.0138 0.9286 ± 0.0019 0.9286 ± 0.0019 0.2997 ± 0.0022 0.2997 ± 0.0022 0.8226 ± 0.0031 0.8226 ± 0.0031 0.6959 ± 0.0064 0.6959 ± 0.0064

MBT _T3 MBT _T3 0.6096 ± ±0.0093 0.6096 0.0093 0.8294 ± ±0.0037 0.8294 0.0037 0.9450 ± ±0.0014 0.9450 0.0014 0.1501 ± 0.0077 0.1501 ± 0.0077 0.8486 ± 0.0038 0.8486 ± 0.0038 0.7398 ± 0.0054 0.7398 ± 0.0054

RS RS 0.6593 ±±0.0141 0.6593 0.0141 0.7825 ±±0.0133 0.7825 0.0133 0.9483 ±±0.0020 0.9483 0.0020 0.1309 ± 0.0083 0.1309 ± 0.0083 0.8179 ± 0.0113 0.8179 ± 0.0113 0.6936 ± 0.0154 0.6936 ± 0.0154

MI MI 0.7347 0.7347±± 0.0072 0.0072 0.6754 0.6754±± 0.0124 0.0124 0.9599 0.9599±± 0.0010 0.0010 0.1446 ± 0.0049 0.1446 ± 0.0049 0.8382 ± 0.0116 0.8382 ± 0.0116 0.7180 ± 0.0205 0.7180 ± 0.0205

BT BT 0.6606 0.0047 0.6606± ± 0.0047 0.8573 0.0028 0.8573 ± ± 0.0028 0.9287 0.0011 0.9287 ± ± 0.0011 0.1374 ± 0.0106 0.1374 ± 0.0106 0.8462 ± 0.0040 0.8462 ± 0.0040 0.7345 ± 0.0073 0.7345 ± 0.0073

MBT MBT 0.6537 0.0047 0.6537 ±±0.0047 0.8360 0.0021 0.8360 ±±0.0021 0.9583 0.0023 0.9583 ±±0.0023 0.2591 ± 0.0104 0.2591 ± 0.0104 0.8623 ± 0.0039 0.8623 ± 0.0039 0.7590 ± 0.0068 0.7590 ± 0.0068

Again, the experimental results demonstrate that the object-based supervised classification methods can suppress the influence of speckle noise (MBT_pixel, OA OA 82.26% 82.26% and Kappa Kappa 0.6959) 0.6959) and obtain pleasing classification performances, as shown in Figure 15 and Table 6. 6. Meanwhile, the

performance with the RS strategy (OA 81.79% and Kappa 0.6936) is the worst among the different sample selection strategies. This is because the RS strategy randomly selects the training samples

Remote Sens. 2018, 10, 1092

21 of 25

performance with the RS strategy (OA 81.79% and Kappa 0.6936) is the worst among the different Remote Sens. 2018, 10, x FOR PEER REVIEW 21 of 25 sample selection strategies. This is because the RS strategy randomly selects the training samples from the ground-truth maps, and theand RS strategy does notdoes consider the information contained in the training from the ground-truth maps, the RS strategy not consider the information contained in the samples. The performances with MI (OA 83.82% and Kappa 0.7180) and BT (OA 84.64% and Kappa training samples. The performances with MI (OA 83.82% and Kappa 0.7180) and BT (OA 84.64% 0.7345) are better than performance with RS, which because former strategies adequately and Kappa 0.7345) arethe better than the performance withisRS, whichthe is because the former strategies consider the consider information in the traininginsamples. However, the MI strategythe focuses on the adequately the contained information contained the training samples. However, MI strategy most complicated area, and the BT algorithm strategy focuses on the boundary area between the focuses on the most complicated area, and the BT algorithm strategy focuses on the boundary two area most probable classes, these methods confusion are less accurate than between the two mostsoprobable classes,aresoprone thesetomethods areand prone to confusion andthe areMBT less strategy (OA 86.23% and Kappa 0.7590). The proposed method fully use the polarimetric information accurate than the MBT strategy (OA 86.23% and Kappa 0.7590). The proposed method fully use the and it can improve the performance when compared with the when MBT_T3 methodwith (OA the 84.86% and polarimetric information and it can improve the performance compared MBT_T3 Kappa method0.7398). (OA 84.86% and Kappa 0.7398). Figure with different numbers of training samples fromfrom the Figure 16 16shows showsthe theclassification classificationaccuracy accuracy with different numbers of training samples different strategies. The horizontal axis represents the number of training samples, and the vertical axis the different strategies. The horizontal axis represents the number of training samples, and the vertical represents the OAthe or Kappa the different selection strategies. classification axis represents OA orofKappa of thesample different sample selectionThe strategies. The performances classification become stable when the samples in each category are more than 50, and the performance performances become stable when the samples in each category are more than 50, with and the the proposed method theproposed best among the different strategies. performance withisthe method is the best among the different strategies.

(a)

(b)

Figure 16. 16. OA OA and and Kappa Kappa curves curves with with different different numbers numbers of of training training samples samples from from the the different different Figure strategies: (a) OA; (b) Kappa. strategies: (a) OA; (b) Kappa.

To quantitatively compare the classification results with different classification algorithms, four To quantitatively compare the classification results with different classification algorithms, four different classification algorithms were also considered. In detail, we first randomly selected five different classification algorithms were also considered. In detail, we first randomly selected five training samples per class as the initial training set in these experiments, which was done spatially training samples per class as the initial training set in these experiments, which was done spatially by by randomly selecting the segmented objects within a class and that all the polarimetric parameters randomly selecting the segmented objects within a class and that all the polarimetric parameters were were considered in this step; in the second step, five samples per class were selected using the MBT considered in this step; in the second step, five samples per class were selected using the MBT strategy strategy of sample selection at each iteration, with the stopping criterion of the iteration set to 10 of sample selection at each iteration, with the stopping criterion of the iteration set to 10 times; and the times; and the four different classification algorithms were finally applied to identify the different four different classification algorithms were finally applied to identify the different types of land cover types of land cover in the RadarSat-2 imagery. Figure 17 shows the classification results obtained in the RadarSat-2 imagery. Figure 17 shows the classification results obtained with the RadarSat-2 with the RadarSat-2 imagery when 55 samples in each category are picked using the different imagery when 55 samples in each category are picked using the different classification algorithms. classification algorithms.

Remote Sens. 2018, 10, 1092

22 of 25

Remote Sens. 2018, 10, x FOR PEER REVIEW

22 of 25

(a)

(b)

(c)

(d)

Figure 17. Classification results for the RadarSat-2 imagery with the different classification

Figure 17. Classification results for the RadarSat-2 imagery with the different classification algorithms: algorithms: (a) KNN classifier; (b) Wishart classifier; (c) LOR-LBP classifier; (d) RF classifier. (a) KNN classifier; (b) Wishart classifier; (c) LOR-LBP classifier; (d) RF classifier.

Table 7 lists the classification results (OA and Kappa) for the RadarSat-2 imagery with the different classification algorithms when 55 (OA samples each category areRadarSat-2 picked, which can be used Table 7 lists the classification results andinKappa) for the imagery with the to quantitatively compare the classification results of different classification algorithms. different classification algorithms when 55 samples in each category are picked, which can beThe used to performance with RF 86.23% andresults Kappaof 0.7590) is the best among algorithms. the different The classification quantitatively compare the(OA classification different classification performance which is because the RF classifier has strong generalization ability and also carries out with algorithms, RF (OA 86.23% and Kappa 0.7590) is the best among the different classification algorithms, the selection of implied features in the process of classification. which is because the RF classifier has strong generalization ability and also carries out the selection of implied features process results of classification. Tablein 7. the Classification for RadarSat-2 imagery with different classification algorithms. Classification Algorithm OAwith different Kappa Table 7. Classification results for RadarSat-2 imagery classification algorithms. KNN 0.8109 ± 0.0149 0.6786 ± 0.0112 Wishart 0.8201OA ± 0.0103 0.6927Kappa ± 0.0191 Classification Algorithm LOR-LBP 0.8469 ± 0.0064 0.7351 ± 0.0101 KNN 0.8109 ± 0.0149 0.6786 ± 0.0112 RF 0.8623 ± 0.0039 0.7590 ± 0.0068 Wishart 0.8201 ± 0.0103 0.6927 ± 0.0191 LOR-LBP 0.8469 ± 0.0064 0.7351 ± 0.0101 Figure 18 shows the classification accuracies with±different training samples from RF 0.8623 0.0039 numbers 0.7590 ±of0.0068 the different classification algorithms. The horizontal axis represents the number of training samples, and the vertical axis represents the OA or Kappa of the different classification algorithms. Figure 18 shows classification accuracies with different numbers of training from the Analogously, the the performance with the RF classifier remains the best among samples the different classification algorithms. different classification algorithms. The horizontal axis represents the number of training samples, and

the vertical axis represents the OA or Kappa of the different classification algorithms. Analogously, the performance with the RF classifier remains the best among the different classification algorithms. Analogously, several conclusions can be made from the results in Figures 16 and 18. On the one hand, the proposed method obtains the best performance (OA 86.23% and Kappa 0.7590) in all the

Remote Sens. 2018, 10, 1092

23 of 25

experiments, i.e., it can better suppress the speckle noise and obtain a higher OA and Kappa by taking full advantage of the polarimetric and spatial information. On the other hand, the performance with the MBT strategy is better than the other sample selection strategies, because the MBT strategy can more effectively select the labeled samples. Remote Sens. 2018, 10 , x FOR PEER REVIEW 23 of 25

(a)

(b)

Figure 18. OA and Kappa curves with different numbers of training samples from the different Figure 18. OA and Kappa curves with different numbers of training samples from the different classification algorithms: (a) OA; (b) Kappa. classification algorithms: (a) OA; (b) Kappa.

Analogously, several conclusions can be made from the results in Figures 16 and 18. On the one 4. hand, Discussion the proposed method obtains the best performance (OA 86.23% and Kappa 0.7590) in all the experiments, i.e., it can supervised better suppress the speckle noise and obtain a higher OAare andlimited Kappaby bythe Most state-of-the-art classification methods using PolSAR imagery taking fullof advantage of the samples, polarimetric spatial of information. On pixel-based the other hand, the availability effective training and and the results the traditional supervised performance with the MBT strategy better than the other sample selection because the of classification methods are subject to theisinfluence of speckle noise. In this paper,strategies, to solve the problems MBT strategy can more effectively select the labeled samples. existing supervised classification methods when using PolSAR data, we have proposed an object-based supervised classification method (GSRM_MBT_RF), which combines the advantages of the GSRM 4. Discussion algorithm, the AL method, and the RF classifier. The experimental results indicated that the proposed Most state-of-the-art classification methods arethe limited by approach achieves the bestsupervised performances with regard to OAusing and PolSAR Kappa. imagery Moreover, proposed the availability of effective training samples, and the results of the traditional pixel-based supervised method can not only better suppress the speckle noise, but can also perform well when the training classification methods are subject to the influence of speckle noise. In this paper, to solve the samples are limited. problems of existing classification methods when are using PolSARby data, we havemethod proposed The accuracies ofsupervised experiments from different sensors different proposed and an object-based supervised classification method (GSRM_MBT_RF), which combines the advantages the sources of the results mainly include that: (1) The resolution of PolSAR imagery is different. of the GSRM algorithm, the AL method, and the RF classifier. The experimental results indicated The UAVSAR is high-resolution and it leads to the fewer scattering elements in the resolution unit of that the proposed approach achieves the best performances with regard to OA and Kappa. UAVSAR imagery. Nevertheless, the AIRSAR imagery and RadarSat-2 imagery are medium resolution Moreover, the proposed method can not only better suppress the speckle noise, but can also and they are relatively simple. (2) The categories of PolSAR imagery are different. The AIRSAR perform well when the training samples are limited. imagery and UAVSAR imagery mainly focus on agricultural field. The categories of AIRSAR imagery The accuracies of experiments from different sensors are different by proposed method and the aresources entirely and it leads to different responses. categories of different the results mainly include that: (1) polarization The resolution of PolSARHowever, imagery isthe different. The of UAVSAR imagery are five different growth stages of paddy and it might bring about the confusion UAVSAR is high-resolution and it leads to the fewer scattering elements in the resolution unit of in UAVSAR classification procedures. The RadarSat-2 imagery is urban and the situation relatively imagery. Nevertheless, the AIRSAR imagery and scene RadarSat-2 imagery are ismedium complex. (3) The influence of topography is different. The city of Wuhan has an undulating relief and resolution and they are relatively simple. (2) The categories of PolSAR imagery are different. The it is prone to misclassification. AIRSAR imagery and UAVSAR imagery mainly focus on agricultural field. The categories of The proposed has different several advantages compared with the responses. traditionalHowever, supervised AIRSAR imagery method are entirely and it leadswhen to different polarization classification methods for PolSAR (1) different The proposed obtain higher the categories of UAVSAR imagerydata: are five growthmethod stages ofcan paddy andobviously it might bring about the confusion classification procedures. The RadarSat-2 imagery is urban methods, scene anddue the to classification accuracyin when compared with pixel-based supervised classification situation is relatively complex. (3) The influence of topography is different. TheThe cityproposed of Wuhanmethod has using the GSRM algorithm to better suppress the influence of speckle noise. (2) an undulating relief andinformation it is prone toofmisclassification. considers the maximum the training samples and more effectively selects the labeled methodThis has allows severalthe advantages compared withwell, the traditional supervised samplesThe by proposed the AL method. proposedwhen method to perform even when the training classification methods for PolSAR data: (1) The proposed method can obtain obviously samples are limited. (3) The RF classifier integrates a feature evaluation technique, so thathigher it is not classification accuracy when compared supervised classification dueoftothe necessary to carry out feature selection.with (4) pixel-based The proposed approach takes fullmethods, advantage using the GSRM algorithm to better suppress the influence of speckle noise. (2) The proposed method considers the maximum information of the training samples and more effectively selects the labeled samples by the AL method. This allows the proposed method to perform well, even when the training samples are limited. (3) The RF classifier integrates a feature evaluation technique, so that it is not necessary to carry out feature selection. (4) The proposed approach takes full

Remote Sens. 2018, 10, 1092

24 of 25

polarimetric information and spatial information of PolSAR imagery. As a result, it can obtain a better OA and Kappa when compared with the traditional supervised classification methods. However, the proposed method still has some disadvantages: (1) the redundancy of polarimetric features might pollute the classifier and it is necessary to analyze and select the most relevant features before implementation of proposed method; and (2) the algorithm structure is relatively complex and time-consuming when compared with the state-of-the-art supervised classification methods. 5. Conclusions A novel object-based supervised classification method for PolSAR imagery is proposed in this paper. The first step of the proposed method is aimed at reducing the speckle noise through the GSRM algorithm. A reliable training set is then selected from the different polarimetric features of the PolSAR imagery by the AL method. Finally, the RF classifier is applied to identify the different types of land cover. To validate the performance of the proposed method, three PolSAR datasets acquired by different sensors were considered, and the experimental results showed that the proposed method significantly improves the classification accuracy for PolSAR images. Meanwhile, four different sample selection strategies and four different classification algorithms were applied to assess the effectiveness and feasibility of the proposed method. The analysis showed that the proposed method can not only better suppress the speckle noise, but can also significantly improve the OA and Kappa, even when the training samples are limited. Nevertheless, further investigation is still necessary. For example, in our future work, it is necessary to suppress the effect of terrain for classification using relevant DEM data, and the proposed method should be optimized by considering feature optimization. Author Contributions: W.L. defined the research problem, proposed the methods, undertook most of the programming, and wrote the paper. J.Y. gave some key advice. P.L. gave useful advice and contributed to the paper writing. Y.H. contributed to the paper writing and provided the ground truth. J.Z. contributed to the paper writing. H.S. collected the data and processed the datasets. Funding: This research was funded by the National Natural Science Foundation of China (No. 91438203, No. 61371199, No. 41501382, and No. 41601355), the Hubei Provincial Natural Science Foundation (No. 2015CFB328 and No. 2016CFB246), the National Basic Technology Program of Surveying and Mapping (No. 2016KJ0103), and the Technology of Target Recognition Based on GF-3 Program (No. 03-Y20A10-9001-15/16). Conflicts of Interest: The authors declare no conflict of interest.

References 1. 2. 3. 4. 5.

6. 7. 8. 9.

Lee, J.S.; Grunes, M.R.; Kwok, R. Classification of multi-look polarimetric SAR imagery based on complex Wishart distribution. Int. J. Remote. Sens. 1994, 15, 2299–2311. [CrossRef] Martinuzzi, S.; Gould, W.A.; González, O.M.R. Land development, land use, and urban sprawl in Puerto Rico integrating remote sensing and population census data. Landsc. Urban Plan. 2007, 79, 288–297. [CrossRef] Zhao, L.; Yang, J.; Li, P. Damage assessment in urban areas using post-earthquake airborne PolSAR imagery. Int. J. Remote. Sens. 2013, 34, 8952–8966. [CrossRef] Sun, W.; Li, P.; Yang, J. Polarimetric SAR image classification using a wishart test statistic and a wishart dissimilarity measure. IEEE Geosci. Remote Sens. Lett. 2017, 14, 1–5. [CrossRef] Zhao, L.; Yang, J.; Li, P. Seasonal inundation monitoring and vegetation pattern mapping of the Erguna floodplain by means of a RADARSAT-2 fully polarimetric time series. Remote Sens. Environ. 2014, 152, 426–440. [CrossRef] Shi, L.; Zhang, L.; Yang, J. Supervised graph embedding for polarimetric SAR image classification. IEEE Geosci. Remote Sens. Lett. 2013, 10, 216–220. [CrossRef] Qi, Z.; Yeh, A.G.-O.; Li, X. A novel algorithm for land use and land cover classification using RADARSAT-2 polarimetric SAR data. Remote Sens. Environ. 2012, 118, 21–39. [CrossRef] Maggiori, E.; Tarabalka, Y.; Charpiat, G.; Alliez, P. Convolutional neural networks for large-scale remote-sensing image classification. IEEE Trans. Geosci. Remote Sens. 2017, 55, 645–657. [CrossRef] Samat, A.; Gamba, P.; Du, P. Active extreme learning machines for quad-polarimetric SAR imagery classification. Int. J. Appl. Earth Obs. Geoinf. 2015, 35, 305–319. [CrossRef]

Remote Sens. 2018, 10, 1092

10.

11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. 31. 32. 33. 34.

25 of 25

Mahdianpari, M.; Salehi, B.; Mohammadimanesh, F. Random forest wetland classification using ALOS-2 L-band, RADARSAT-2 C-band, and TerraSAR-X imagery. ISPRS J. Photogramm. Remote Sens. 2017, 130, 13–31. [CrossRef] Li, J.; Bioucas-Dias, J.M.; Plaza, A. Hyperspectral image segmentation using a new bayesian approach with active learning. IEEE Trans. Geosci. Remote Sens. 2011, 49, 3947–3960. [CrossRef] Rajan, S.; Ghosh, J.; Crawford, M.M. An active learning approach to hyperspectral data classification. IEEE Trans. Geosci. Remote Sens. 2008, 46, 1231–1242. [CrossRef] Li, J.; Bioucas-Dias, J.M.; Plaza, A. Spectral–spatial classification of hyperspectral data using loopy belief propagation and active learning. IEEE Trans. Geosci. Remote Sens. 2013, 51, 844–856. [CrossRef] Lee, J.S.; Pottier, E. Polarimetric Radar Imaging From Basics to Applications; CRC Press: Boca Raton, FL, USA, 2017. Cloude, S.R.; Pottier, E. A review of target decomposition theorems in radar polarimetry. IEEE Trans. Geosci. Remote Sens. 1996, 34, 498–518. [CrossRef] Cloude, S.R. Target decomposition theorems in radar scattering. Electron. Lett. 1985, 21, 22–24. [CrossRef] Freeman, A.; Durden, S.L. A three-component scattering model for polarimetric SAR data. IEEE Trans. Geosci. Remote Sens. 1998, 36, 963–973. [CrossRef] Van Zyl, J.J. Application of Cloude’s target decomposition theorem to polarimetric imaging radar data. In Radar Polarimetry; Bellingham: Washington, DC, USA, 1993; Volume 1748, pp. 184–192. Park, S.-E.; Yamaguchi, Y.; Kim, D.-J. Polarimetric SAR remote sensing of the 2011 Tohoku earthquake using ALOS/PALSAR. Remote Sens. Environ. 2013, 132, 212–220. [CrossRef] Touzi, R. Target scattering decomposition in terms of roll-invariant target parameters. IEEE Trans. Geosci. Remote Sens. 2007, 45, 73–84. [CrossRef] Arii, M.; van Zyl, J.J.; Kim, Y. Adaptive model-based decomposition of polarimetric SAR covariance matrices. IEEE Trans. Geosci. Remote Sens. 2011, 49, 1104–1113. [CrossRef] Qin, F.; Guo, J.; Sun, W. Object-oriented ensemble classification for polarimetric SAR imagery using restricted Boltzmann machines. Remote Sens. Lett. 2016, 8, 204–213. [CrossRef] Ma, L.; Li, M.; Ma, X. A review of supervised object-based land-cover image classification. ISPRS J. Photogramm. Remote Sens. 2017, 130, 277–293. [CrossRef] Lang, F.; Yang, J.; Li, D.; Zhao, L.; Shi, L. Polarimetric SAR image segmentation using statistical region merging. IEEE Geosci. Remote Sens. Lett. 2014, 11, 509–513. [CrossRef] Comaniciu, D.; Meer, P. Mean shift: A robust approach toward feature space analysis. IEEE Trans Pattern Anal. Mach. Intell. 2002, 24, 603–619. [CrossRef] Qin, F.; Guo, J.; Lang, F. Superpixel segmentation for polarimetric SAR imagery using local iterative clustering. IEEE Geosci. Remote Sens. Lett. 2017, 12, 13–17. Xiang, D.; Ban, Y.; Wang, W. Adaptive superpixel generation for polarimetric SAR images with local iterative clustering and SIRV model. IEEE Trans. Geosci. Remote Sens. 2017, 55, 3115–3131. [CrossRef] Liu, W.; Yang, J.; Zhao, J. A novel method of unsupervised change detection using multi-temporal PolSAR images. Remote Sens. 2017, 9, 1135. [CrossRef] Zhang, L.; Li, A.; Li, X. Remote sensing image segmentation based on an improved 2-D gradient histogram and MMAD model. IEEE Geosci. Remote Sens. Lett. 2015, 12, 58–62. [CrossRef] Nock, R.; Nielsen, F. Statistical region merging. IEEE Trans. Pattern Anal. Mach. Intell. 2004, 26, 1452. [CrossRef] [PubMed] Luo, T.; Kramer, K.; Goldgof, D.B.; Hall, L.O.; Samson, S.; Remsen, A.; Hopkins, T. Active learning to recognize multiple types of plankton. J. Mach. Learn. Res. 2005, 6, 589–613. Tuia, D.; Ratle, F.; Pacifici, F. Active learning methods for remote sensing image classification. IEEE Trans. Geosci. Remote Sens. 2009, 47, 2218–2232. [CrossRef] Tuia, D.; Volpi, M.; Copa, L.; Kanevski, M.; Munoz-Mari, J. A survey of active learning algorithms for supervised remote sensing image classification. IEEE J. Sel. Top. Signal Process. 2011, 5, 606–617. [CrossRef] Hastie, T.; Tibshirani, R. Discriminant adaptive nearest neighbor classification and regression. IEEE Trans. Pattern Anal. Mach. Intell. 1996, 18, 409–415. [CrossRef] © 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).