A simple automated dynamic threshold extraction ...

International Journal of Remote Sensing

ISSN: 0143-1161 (Print) 1366-5901 (Online) Journal homepage: http://www.tandfonline.com/loi/tres20

A simple automated dynamic threshold extraction method for the classification of large water bodies from landsat-8 OLI water index images Fangfang Zhang, Junsheng Li, Bing Zhang, Qian Shen, Huping Ye, Shenglei Wang & Zhaoyi Lu To cite this article: Fangfang Zhang, Junsheng Li, Bing Zhang, Qian Shen, Huping Ye, Shenglei Wang & Zhaoyi Lu (2018) A simple automated dynamic threshold extraction method for the classification of large water bodies from landsat-8 OLI water index images, International Journal of Remote Sensing, 39:11, 3429-3451, DOI: 10.1080/01431161.2018.1444292 To link to this article: https://doi.org/10.1080/01431161.2018.1444292

Published online: 27 Feb 2018.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=tres20

INTERNATIONAL JOURNAL OF REMOTE SENSING, 2018 VOL. 39, NO. 11, 3429–3451 https://doi.org/10.1080/01431161.2018.1444292

A simple automated dynamic threshold extraction method for the classification of large water bodies from landsat-8 OLI water index images Fangfang Zhanga, Junsheng Lia, Bing Zhanga,b, Qian Shena, Huping Yea, Shenglei Wanga,b and Zhaoyi Lua,b a

Key Laboratory of Digital Earth Science, Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing, China; bUniversity of Chinese Academy of Sciences, Beijing, China ABSTRACT

ARTICLE HISTORY

Traditional manual methods of extracting water bodies from remote sensing images cannot satisfy the requirements for mass processing of remote sensing data, and new automated methods are complicated and require a large amount of auxiliary data. The histogram bimodal method is a frequently used objective tool for threshold selection in image segmentation. However, automatically calculating the threshold is difficult because of complex surfaces and image noise, which lead to imperfect twin peaks. To overcome these difficulties, we developed an operational automated water extraction method. This method does not require the identification of twin histogram peaks but instead seeks minimum values in the threshold range to achieve an automated dynamic threshold. We calibrated the method for 18 lakes in China using Landsat 8 Operational Land Imager images, for which the relative error (RE) and coefficient of determination (R2) for threshold accuracy were 2.1% and 0.96, respectively. The RE of area accuracy was 0.59%. The advantages of the method lie in its simplicity and minimal requirements for auxiliary data while still achieving an accuracy comparable to that of other automatic water extraction methods. It can be applied to mass remote sensing data to calculate water thresholds and automatically extract large water bodies.

Received 31 July 2017 Accepted 10 February 2018

1. Introduction Surface water is a critical resource that is impacted by changes in land cover and use, climate, and the environment, on a global scale. Changes in surface water can cause droughts, floods, and pollution. Monitoring surface water distribution using remote sensing is critical, and has the added advantages of having low costs, high speed and frequency, wide coverage, and low interference from surface conditions. Surface water distributions extracted using remote sensing data have been used in water resources assessments, coastal management, and studies of environmental change (Feng et al. 2016; Feyisa et al. 2014). CONTACT Junsheng Li [email protected] Key Laboratory of Digital Earth Science, Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing, China © 2018 Informa UK Limited, trading as Taylor & Francis Group

3430

F. ZHANG ET AL.

Water extraction using remote sensing techniques first began forty years ago (Work and Gilmer 1976). Since then, numerous studies have developed indices and algorithms of water extraction. Water surfaces appear dark in near-infrared radiation (NIR) because of strong absorption by water. In contrast, land appears bright because NIR is strongly reflected by terrestrial vegetation and dry soil. Therefore, initially, single-band methods based on NIR were used for water extraction (Work and Gilmer 1976; White 1978; Rundquist et al. 1987). However, surface terrain can be complex and other surface features can behave in a manner similar to water, resulting in errors. Subsequently, the normalised difference vegetation index (NDVI) was used to enhance water features by introducing red band (R) data to NIR techniques (Boland 1976; Townshend and Justice 1986; Tucker and Sellers 1986). While NDVI can delineate more effectively between water and non-water objects, and especially vegetation, it only suppresses non-water features rather than eliminating them completely (McFeeters 1996). Therefore, other studies introduced the green band (G) instead of the red band, resulting in the formation of the normalised difference water index (NDWI) (McFeeters 1996). The NDWI aims to enhance water features using the high reflectance of water in the green band, and low reflectance in the NIR band. However, the NDWI cannot efficiently suppress signals from built-up areas; moreover, water features are still interspersed with noise from other features (Xu 2006). To rectify this problem, the NDWI was altered to form the modification of normalised difference water index (MNDWI) (Xu 2006), which utilizes shortwave infrared radiation (SWIR-5) that has a lower reflectance from water than NIR. A new water index (NDWI-DB) for Landsat 8 Operational Land Imager (OLI) sensor utilized dark blue and SWIR-7 band was developed (Li et al. 2016). Despite improvements (Fisher, Flood, and Danaher 2016), all of these indices require a threshold to discriminate water from land, and subjective and static threshold values can lead to over- or underestimation of surface water area (Frazier and Page 2000). Threshold values usually vary with viewing angle, satellite attitude, illumination, atmospheric conditions, and environmental noise in the images obtained (Jain et al. 2005; Ji, Zhang, and Wylie 2009). Therefore, identifying an appropriate threshold is challenging and time consuming (Feyisa et al. 2014). For this reason, automated water extraction methods have increasingly been considered. Feyisa et al. (2014) introduced a multiple-band index (the automated water extraction index; AWEI) to increase the accuracy and robustness of water extraction. Although the suggested threshold value of 0.0 is more robust than NDWI and MNDWI, it changes in different images, which hinders automated water extraction. Both supervised and unsupervised classification procedures have frequently been applied to identify and classify water features in remote sensing images (Verpoorter, Kutser, and Tranvik 2012). These methods include maximum-likelihood classification (Kingsford et al. 1997), decision trees (Subramaniam, Babu, and Roy 2011; Leckie et al. 2005), artificial neural networks (Bishop 1999), support vector machines (Huang et al. 2011), the fuzzy clustering method (FCM) (Yang et al. 2015), isodata (Reis and Yilmaz 2008), and k-means clustering (Lu and Weng 2007). However, these methods require significant prior knowledge. Further, they are not stable because surface features vary in different images, preventing the application of these methods over large areas (Frazier and Page 2000). Computer graphics technology has also been used to extract water distributions, including tasselled cap transformation (Gao et al. 2016) and LBV transformation (Zhang et al. 2017), boundary tracing (Ouma and Tateishi 2006), principal component

INTERNATIONAL JOURNAL OF REMOTE SENSING

3431

analysis (Lira 2006; Rokni et al. 2014), morphological segmentation (Verpoorter, Kutser, and Tranvik 2012; Bagli and Soille 2004; Jiang et al. 2014; Xie et al. 2016), and end member extraction (Zhang et al. 2013), amongst others (Jiang et al. 2014). However, these methods have common complex weaknesses that lead to instabilities and difficulties in mass data processing. Recently, several methods have been used to generate large-scale water distribution products. Lehner and Döll (2004) provided the Global Lake and Wetland Database (GLWD), which combines multiple data sources such as inventories and archives. The Moderate Resolution Imaging Spectroradiometer (MODIS) water mask is a global, 250 m resolution representation of global inland surface water bodies in 2000 (Carroll et al. 2009; USGS 2012). Feng et al. (2016) produced a global, inland surface water dataset (GIW) of 30 m resolution with an automated algorithm using several datasets. Verpoorter, Kutser, and Tranvik (2012) developed an approach called GWEM (GeoCoverTM Water bodies Extraction Method) that combines remote sensing and Geographic Information System (GIS) to produce global lake maps with a spatial resolution of 14.25 m. Pekel et al. (2016) jointly used expert systems, visual analytics, and evidential reasoning to perform water detection via Landsat 7 Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+) and OLI data and then mapped global surface waters (GSW) from 1984 to 2015. GLWD and MODIS are low-resolution products, whereas Landsat water products have spatial resolutions of ≤30 m. However, the methods of obtaining Landsat water products are complicated and sophisticated, and require significant auxiliary data. If any of this auxiliary data is missing, the methods may not work. Simple, automated extraction of water bodies remains a challenge. Therefore, we developed an automated water extraction method (modified histogram bimodal method: MHBM) for simple and fast extraction of water bodies, and applied the method to Landsat 8 OLI images to evaluate its performance.

2. Data and study area Landsat 8 is the latest in the Landsat series of satellites that have been providing multispectral images of the Earth continuously since the early 1970s. A unique data record of the Earth’s land surface spanning over 40 years now exists (Zanter 2015). This unique retrospective portrait of the Earth’s surface has often been used for water management (McFeeters 1996; Xu 2006; Feyisa et al. 2014). Automated water extraction from time-series OLI mass data is difficult, but necessary. Therefore, Landsat 8 OLI images were used to test thresholds, accuracy, and the robustness of the MHBM. We downloaded Landsat L1T data from the open spatial data sharing system of the Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences (URL: http://ids.ceode.ac.cn/). The L1T data have already been subjected to radiometric and geometric correction from ground control points, and terrain correction by Digital Elevation Model (DEM). The study areas include several lakes, rivers, and reservoirs in different environmental conditions, including cities, plains, hills, mountains, and highlands (Figure 1). The data cover ten provinces of China, including Beijing, Tianjin, Hebei, Henan, Jiangsu, Shanghai, Anhui, Jiangxi, Hunan, Yunnan, Qinghai, and Tibet. Water bodies include turbid rivers, clear lakes, ephemeral lakes, eutrophic lakes, clear reservoirs, and eutrophic reservoirs. The total numbers of water bodies and images are 18 and 152, respectively.

3432

F. ZHANG ET AL.

Figure 1. Shorelines ((a)–(r)) and location map (centre) of 18 water bodies extracted using the modified histogram bimodal method (MHBM). Blue lines denote shorelines. Backgrounds are true colour images where red, green, and blue bands represent bands 4, 3 and 2, respectively, of the Landsat 8 Operational Land Imager (OLI) sensor.

3. Methods 3.1. Modified histogram bimodal method (MHBM) The histogram bimodal method (HBM) is often used for threshold selection in image segmentation of land and water bodies (Prewitt and Mendelsohn 1966; Li, Sheng, and Luo 2011; Khandelwal et al. 2017). In this method, histograms are created using grey images. Then, the valley values between the two peaks in a histogram are used as the threshold for image segmentation (Figure 2). However, it is difficult to automatically find two peaks using a computer, because complex surfaces and image noise usually result in

INTERNATIONAL JOURNAL OF REMOTE SENSING

3433

Figure 2. Diagram of the modified histogram bimodal method (MHBM). (a) Scene captured by the Landsat 8 Operational Land Imager (OLI). (b) Normalised difference water index (NDWI) image. (c) Histogram prepared using NDWI data from within the red line in (b). The blue line in (b) denotes the lake shoreline, and the red line is drawn as a buffer to the blue line. PLand and PWater denote the peaks of land and water, respectively. Tl and Tr denote the left and right edges of the threshold range, respectively, and Ti denotes the initial threshold. Rt denotes the length of the threshold range.

single or multi-peak retrieval rather than perfect twin peaks, and consequently in uncertainty of the threshold. In view of these problems, we modified the HBM to automatically calculate the threshold through an automated water extraction method (the MHBM). In the MHBM, a large number of histogram valley values formed the threshold range (Rt), and the mean valley value was defined as the initial threshold (Ti). Furthermore, we seek the minimum value in the threshold range rather than twin peaks and valleys, in order to avoid uncertainty. To obtain the best results, we restricted the data in our histograms to those from the areas surrounding water bodies and not from the full image. First, we extracted the rough distribution of a water body using Ti, and then the area was expanded to include

3434

F. ZHANG ET AL.

a buffer (generally 1.5 times the rough water body area), forming the statistical area of the histogram (Figure 2(b)). The expanded area was approximately two times that of the water area, allowing the formation of perfect twin peaks and a histogram of the water index (or single band reflectance) for the expanded area (Figure 2(c)). A Rayleigh scattering correction was required to ensure stable threshold results. Using this approach, we were able to automatically calculate the dynamic threshold, which verified that the MHBM could be used for automated water extraction.

3.2. Water extraction procedure based on MHBM To achieve automated water extraction, we designed a workflow based on MHBM (Figure 3), which includes Rayleigh scattering correction, threshold statistics, preliminary water extraction, water mask buffer, and accurate water distribution extraction. To ensure a stable threshold, accurate atmospheric correction of the remote sensing images was required; however, accurate atmospheric correction is challenging (Wang 2007). Reflectance values after Rayleigh scattering correction (ρrc) satisfy water extraction and are often used in water colour remote sensing (Hu 2009; Hu

Figure 3. Flow chart showing the water extraction procedure based on the modified histogram bimodal method (MHBM).

INTERNATIONAL JOURNAL OF REMOTE SENSING

3435

et al. 2010). Therefore, we used ρrc to calculate water indices (i.e., NDWI, MNDWI, and AWEI; Table 1) and replaced surface reflectance (ρo) through accurate atmospheric correction. Reflectance from the top of the atmosphere (ρTOA) can be expressed as the linear sum of the contribution from Rayleigh scattering (ρr), aerosol scattering in the atmosphere (ρa), and diffuse transmittance to the sensor (ρo) using Equation (1) (Zanter 2015): ρTOA ¼ ρr þ ρa þ t  ρo

(1)

where t is the diffuse transmittance of the atmosphere (considering only gas molecules). We simulated ρr using the second simulation of a satellite signal in the solar spectrumvector (6SV) to obtain reflectance after the Rayleigh scattering correction, which included aerosol scattering reflectance and surface reflectance: ρrc ¼ ρTOA  ρr

(2)

From correcting Rayleigh scattering and converting to reflectance, to delineation of water distribution, MHBM follows four steps: (1) Preliminary water extraction: We extracted water bodies using Ti to obtain a rough water distribution mask. The purpose of this step was to identify the general location and distribution of water bodies, and not to represent accurate shorelines. (2) Water area buffer: To obtain well-defined twin peaks in the histograms, we expanded the preliminary water masks by 1.5 times to obtain work areas. Water index images were then masked to isolate the work areas. In the masked water index images (work areas), the number of water pixels and non-water pixels were equal, and the histogram peaks of water and non-water, along with the valley between the two peaks, were distinct. (3) Accurate water extraction: The minimum values of threshold ranges in the histograms are easily processed and represent precise thresholds for accurate water extraction. The use of threshold ranges ensures the stability and correctness of the selected thresholds, even in non-ideal histograms. For instance, if several minima or a range of minima occur within the threshold range, the median position minimum was selected as the precise threshold. For every water body within an image, an exclusive threshold (dynamic threshold: Td) for image segmentation was obtained using the above described method, and automated water extraction was successfully accomplished. Table 1. Water index expressions. Water indexa NDWI MNDWI AWEI a

Expression

Ref.

(ρrc,563 − ρrc,865)/(ρrc,563 + ρrc,865) (ρrc,563 − ρrc,1610)/(ρrc,563 + ρrc,1610) 4 × (ρrc,563 − ρrc,1610) − (0.25 × ρrc,865 + 2.75 × ρrc,2200)

McFeeters 1996 Xu 2006 Feyisa et al. 2014

NDWI: normalised difference water index. MNDWI: modification of normalised difference water index. AWEI: Automated water extraction index. ρrc,563, ρrc,865, ρrc,1610, ρrc,2200 are the reflectance values after Rayleigh scattering correction at the Landsat 8 OLI wavelength of 563 nm, 865 nm, 1610 nm, and 2200 nm.

3436

F. ZHANG ET AL.

Threshold statistics are necessary when calibrating the threshold range and initial threshold for data of a new sensor (Figure 3). However, this step can be skipped when the threshold range and initial threshold are already known for the sensor of interest. Optimal thresholds were defined as the minimum valley value occurring between the twin peaks of a given histogram, which was based on image statistics derived from water areas with a 1.5 times buffer. For adequate representativeness, threshold statistics were calculated using mass images that included many typical water body types. The minimum and maximum valley values were defined as the left and right edges of the threshold (Tl and Tr), respectively. The length between Tl and Tr was the threshold range (Rt), and the mean of the valley values was defined as the initial threshold (Ti) for preliminary water extraction. We obtained thresholds of NDWI, MNDWI, and AWEI for 18 water bodies within each of the 152 Landsat 8 OLI images. The 152 thresholds of each water index formed a threshold range (Table 2). The statistical threshold results of this study could be used by other researchers for delineating other water bodies.

3.3. Accuracy assessment We evaluated the accuracy of the automated MHBM dynamic threshold calculation using relative error (RE) and the squared correlation coefficient (R2); furthermore, we evaluated the area accuracy of the extracted area of water using the MHBM in comparison to the initial threshold and a static threshold (0.0) using RE and the root-mean-square error (RMSE): RE ¼

n 1X jXi Yi j 100% N i¼1 Z

(3)

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi n 1X ðXi Yi Þ2 RMSE ¼ N i¼1

(4)

For threshold accuracy, X is the dynamic threshold of the MHBM (Td), Y is the optimal threshold, Z the interval of the threshold ranges (Rt) of NDWI, MNDWI, and AWEI, respectively, and N is the number of images and water bodies. The dynamic threshold, X, is derived through automatic calculation by computer using the MHBM method. The optimal threshold, Y, is the minimum valley value between the two peaks (i.e., water and non-water peaks) of the histogram derived from statistics of preliminary water masks expanded by a buffer 1.5 times the area of the water body and defined by visual interpretation. For area accuracy, X is the water area using different thresholds (Td, Ti, and T0), Y and Z are equal and represent the water area using the optimal threshold, and N Table 2. Statistics of water index expressionsa. Water indexb NDWI MNDWI AWEI

Tl

Tr

Rt

SD

Ti

−0.184 0.104 −0.295

0.228 0.427 0.083

0.412 0.323 0.378

0.099 0.084 0.098

0.020 0.277 −0.099

Tl, Tr, and Ti: left edge, right edge, and initial threshold ranges, respectively; Rt: threshold range between Tl and Tr; SD: standard deviation. b NDWI: normalised difference water index; MNDWI: modification of normalised difference water index; AWEI: Automated water extraction index. a

INTERNATIONAL JOURNAL OF REMOTE SENSING

3437

is the number of images and water bodies. To obtain reference data for validation, visual interpretation was used by the image itself, along with high-resolution satellite and Google Earth imagery. All 152 images of the 18 water bodies were used for the accuracy assessment.

4. Results We extracted all 18 water bodies (Figure 3), and confirmed good performance of MHBM. Despite differences in environmental conditions and water qualities, accurate and clear shorelines could be extracted. We also calculated the threshold accuracy and area accuracy, and compared it with other methods.

4.1. Accuracy of the MHBM 4.1.1. Threshold accuracy The MHBM yields the greatest accuracy when the buffer size is 150% (Figure 4). Most dynamic thresholds were equal to the optimal threshold, and fitted trends had high R2 and slopes close to 1.0. Furthermore, the RE was low (under 3%) for all three water indices. These findings confirm that the MHBM is able to establish optimal thresholds in most situations. The MNDWI has the lowest RE and highest R2. The threshold accuracy of the NDWI is slightly lower than that of the MNDWI, while that of the AWEI is slightly lower than that of the NDWI. However, even the AWEI attained high accuracy. These results show that the MHBM has good threshold accuracy when using any of the three water indices, but particularly when using MNDWI. 4.1.2. Area accuracy Table 3 presents the area accuracies of water extraction using three different thresholds: dynamic threshold (Td) of the MHBM, static threshold 0.0 (T0), and initial threshold (Ti). For RMSE, the water extraction area accuracy by T0 was lowest, with RMSE up to 40 km2 (Table 3). Water extraction area accuracy by Ti was higher than T0, with RMSE more than

Figure 4. Threshold accuracy of the modified histogram bimodal method (MHBM) with 150% buffer for (a) the modification of normalised difference water index (MNDWI), (b) the normalised difference water index (NDWI), and (c) the automated water extraction index (AWEI).

3438

F. ZHANG ET AL.

Table 3. Accuracy of modified histogram bimodal method (MHBM) Accuracy index

b

Threshold

c

T0 Ti Td T0 Ti Td

RE (%) RMSE (km2)

a.

MNDWI

NDWI

AWEI

2.82 2.74 0.59 44.5 28.1 7.4

3.54 1.91 0.61 24.5 25.2 11.9

4.80 2.38 0.61 44.4 23.6 7.2

a

NDWI: normalised difference water index; MNDWI: modification of normalised difference water index; AWEI: Automated water extraction index. b RE: relative accuracy; RMSE: root mean square error. c Td: dynamic threshold; T0: static threshold (0.0); Ti: initial threshold.

20 km2; however, the MHBM using Td had the highest area accuracy compared with Ti and T0, with a minimum RMSE of just 7 km2. The trend was similar for RE, with that obtained by the MHBM using Td much lower than that when using Ti and T0, having a mean RE value about 0.6%. These results show that the MHBM using MNDWI and the dynamic threshold has the highest accuracy (i.e., RE of the threshold and area accuracies as low as 2.1% and 0.59%, respectively).

4.2. Comparisons with other methods To verify the performance of the MHBM, we selected two automatic water extraction methods: MMFCM (Yang et al. 2015) and k-means (Lu and Weng 2007). MMFCM is the combination of a modified fuzzy clustering method (MFCM) algorithm and the modified normalised difference water index (MNDWI) (Yang et al. 2015). k-means is a conventional classification method (Lu and Weng 2007). An additional ten images were used for independent accuracy assessment. We performed water extraction for the ten images using MHBM, MMFCM and k-means, separately. Comparison between the three water extraction methods indicates that the relative error of MHBM is the lowest, followed by MMFCM, then k-means (Table 4). We selected several examples to analyse the performance of the MHBM, MMFCM, and k-means water extraction methods (Figure 5). For example, Figure 5 shows three typical water bodies. Guanting Reservoir is an area with shallow water where the underwater ground surface is similar to the surrounding soil. The k-means method extracted portions of this shallow water as ‘non-water,’ causing many errors and punctuated shorelines on the water map. The shorelines produced by the MHBM and MMFCM were smooth, with the MHBM results being more accurate (Figure 5(b) and (b′)). All three methods yielded accurate results for Miyun Reservoir, which possesses high water quality (Figure 5(c) and (c′)). At Yuqiao Reservoir, grass and algae were Table 4. Comparison of MHBM with other classification methods MMFCM and k-means Accuracy index RE (%) RMSE (km2) a

b

a.

MHBM

MMFCM

k-means

0.11 0.09

2.04 1.59

8.56 8.09

MHBM: modified histogram bimodal method; MMFCM: An automatic water extraction method combining a modified fuzzy clustering method (MFCM) algorithm with the modified normalised difference water index (MNDWI); k-means: k-means clustering method. b RE: relative accuracy; RMSE: root mean square error.

INTERNATIONAL JOURNAL OF REMOTE SENSING

3439

Figure 5. Examples of water body extraction results within (a) a sample Landsat 8 OLI true colour image, and at sub-regions (b) Guanting, (c) Miyun, and (d) Yuqiao Reservoirs, using the modified histogram bimodal method (MHBM; red line), an automatic water extraction method combining a modified fuzzy clustering method (MFCM) algorithm with the modified normalised difference water index (MMFCM; blue line), and k-means (green line) methods. (b′), (c′), and (d′) are more detailed local views of each sub-region, displayed in false colour (RGB: 5, 4, and 3).

present at the edge of the water body. As a result, the k-means method again produced water distributions with complex spots and curves. The shoreline produced by MMFCM shows a shift from land to water, while MHBM produced the most accurate shoreline results (Figure 5(d) and (d′)). Compared with the two classification methods, MHBM shows greater accuracy and more stable performance throughout the delineation of different typical water body types. Feng et al. (2016) developed an automated method for mapping inland surface water bodies, which has been applied to the roughly 9000 Landsat scenes of the GLS 2000 data collection to produce a global, 30 m resolution inland surface water body dataset (GIW) for circa 2000. This GIW dataset has been adopted by the Global Land Cover Facility (GLCF). In this study, we modified the MHBM to adapt to Landsat 7 ETM+ images and applied the modified method to three ETM+ images (11 August 1999: path 122, row 32; 1 July 1999: path 123, row 32; and 7 May 2000: path 124, row 32). The results of MHBM and GIW from the same images were then compared

3440

F. ZHANG ET AL.

(Figure 6). The relative error between MHBM and GIW is 5.03%, which indicates that the two products produce broadly similar shorelines (Figure 6(a), (b), (b′), (c), (c′), (d), (d′), (d′′), and (e)). However, there are small differences. The total area of inland surface water bodies in the MHBM product is greater than that of GIW (Figure 6(a′), 6 (a′′) and (e′)). Field investigation and spectral analysis indicate that these additional areas are water with water grass growth. Therefore, a primary difference is that MHBM tends to distinguish water with water grass growth as water, while GIW tends to distinguish it as land. MHBM is a useful potential water extraction method. However, it needs to be tested in more water bodies. The GSW dataset over the past 32 years at a 30 m resolution was produced by Pekel et al. (2016) and shows different facets of the spatial and temporal distribution of surface water over this period. We selected the water occurrence product for the validation of the MHBM. The frequency with which water was present on the surface was captured in a single product known as surface water occurrence (SWO). To

Figure 6. Comparison of the modified histogram bimodal method (MHBM) and global inland water body dataset (GIW) applied to several water bodies. The red and blue lines are the shorelines extracted by MHBM and GIW. (a), (b), (c), (d), and (e) are Landsat 7 ETM+ true colour images (RGB: 3, 2, and 1) of Guanting, Miyun, and Yuqiao Reservoirs. (a′), (a′′), (b′), (c′), (d′), (d′′), and (e′) provide more detailed local views of each reservoir, displayed in false colour (RGB: 4, 3, 2). The image acquisition date of (a), (b), and (c) is 1 July 1999 (path 123, row 32); of (d) is 7 May 2000 (path 124, row 32); and of (e) is 11 August 1999 (path 122, row 32).

INTERNATIONAL JOURNAL OF REMOTE SENSING

3441

compute SWO, the water detections (WDs) and valid observations (VOs) were summed, and SWO was computed by ∑WD/∑VO. The Guanting Reservoir in North China was selected as the study area, and 616 Landsat images (Landsat 5 TM: 283; Landsat 7 ETM+: 277; Landsat 8 OLI: 56) from 1984 to 2015 were used to extract water distribution. The comparison of SWO as revealed by GSW and the MHBM is shown in Figure 7. Over the past 32 years, the water distribution of the Guanting Reservoir has changed greatly, especially in the northeast and southwest. Both GSW and the MHBM could detect these changes. The permanent water surfaces (with nearly 100% occurrence over 32 years) were distinct and highly similar on the map. The details of changes in seasonal water surfaces were also highly similar. GSW achieved high-accuracy comprehensive apply expert systems, visual analytics, and evidential reasoning (Pekel et al. 2016). The MHBM achieved similar results by only using the image itself and minimal auxiliary data. The advantages of the MHBM include simplicity and robustness for application to other areas and sensors.

5. Discussion 5.1. Dynamic thresholds The main indicator of good performance of the MHBM using the dynamic threshold is that different thresholds for different water bodies are observed within single images. For example, the Guanting, Yuqiao, and Miyun Reservoirs yielded different thresholds, which reflects their differing surface environments, water qualities, atmospheric conditions, and image-forming conditions (Figure 8). Threshold values for Guanting and Miyun Reservoirs are much lower than the initial threshold (i.e., 0.277), while for Yuqiao Reservoir, the values are similar. Moreover, the dynamic threshold was calculated

Figure 7. Comparison of surface water occurrence (1984–2015) by the modified histogram bimodal method (MHBM) and global surface water dataset (GSW) applied to the Guanting Reservoir. (a), (a′), (a′′) represent the surface water occurrence of GSW and detailed local views; (b), (b′), (b′′) represent the surface water occurrence of the MHBM and detailed local views.

3442

F. ZHANG ET AL.

Figure 8. High performance image of the modified histogram bimodal method (MHBM). Different identified water bodies correspond to different thresholds.

automatically, removing the effects of human intervention and thereby improving accuracy and objectivity in water extraction.

5.2. Buffer size Buffer size determines histogram shape. In the MNDWI histogram, the peak on the left represents non-water and that on the right represents water bodies. Balance between the two peaks results in more accurate automatic threshold calculation. The water peaks for five buffer sizes were invariant (Figure 9(c)); however, the peaks for non-water gradually increased with increasing buffer sizes (50–250%), until their heights at 100% and 150% buffer were close to that of the water peaks. Therefore, in complex surface environments, multiple surface features can cause non-water peaks to be flatter than a single object, and the height of two peaks at 150% buffer may be much more balanced. We investigated the threshold accuracy at five buffer sizes (Figure 10) and found that while RE initially decreased with increasing buffers (50–150%), it subsequently increased with increasing buffers (150–250%). In contrast, while R2 increased as buffers increased from 50% to 150%, it decreased between 150% and 250% buffer. Therefore, 150% is the most appropriate buffer size for use with MHBM.

INTERNATIONAL JOURNAL OF REMOTE SENSING

3443

Figure 9. (a) True colour, and (b) modified histogram bimodal method (MHBM) images of an identified water body. (c) Histograms for the five buffer sizes (50%–250%) of MHBM.

5.3. Optimal water index MNDWI has the smallest threshold range (Rt) and standard deviation (SD) when compared with AWEI and NDWI (Table 2). Low Rt values provide a narrow threshold range that allows for the precise identification of thresholds, while the small SD shows that thresholds are close to the initial threshold. Therefore, MNDWI is the optimal water index for automated water extraction using OLI images because the stable thresholds generated ensure low error in preliminary water extraction.

3444

F. ZHANG ET AL.

Figure 10. Threshold accuracy of five buffer sizes (50%–250%) for the modified histogram bimodal method (MHBM) buffer size in terms of (a) relative error (RE), and (b) R2.

The distance between the water peak and non-water peak (Dw-nw) determines the possibility of distinguishing water from other objects (Figure 11). The Dw-nw of MNDWI is the greatest among the three water indices, showing that water pixels are most easily distinguished using MNDWI. Compared with NDWI and AWEI, the curve in the MNDWI histogram is flatter and the value between Tl and Tr is smallest; therefore, the accuracy of water extraction is high for any value within the threshold range. These results indicate that MNDWI is a highly accurate and stable index; however, for sensors with absent SWIR bands, NDWI would also produce accurate results when applied to MHBM. All water indices (including NDWI, AWEI, and even MNDWI) have inherent defects, such as classification errors caused by ice and snow, cloud shadows, topographic shadows, and building shadows. The MHBM implements an automatic threshold value to extract water bodies from common water indices. It does not attempt to modify the indices to avoid their inherent defects. Therefore, in this study, MNDWI fits MHBM; however, its inherent defects are still present within the results. We have further tested the accuracy of MHBM by employing MNDWI in a full scene Landsat 8 OLI dataset (Figure 12). We distributed 229 points in order to visually interpret water bodies and calculate accuracies. Total accuracy is 98.7%, kappa coefficient (κκ is 95.0%, mapping accuracy is 91.9%, and user accuracy is 100%. In this image, there are several typical confounding factors: clouds and cloud shadows (Figure 12(b)), building shadows (Figure 12(c)), small water bodies (Figure 12(d)), and topographic shadows (Figure 12(e)). The water under cloud shadow can be extracted very well (Figure 12(b)). There are few effects of building shadows (Figure 12(c)) and topographic shadows (Figure 12(e)). However, the small water bodies were ignored by MHBM (Figure 12(d)). Therefore, the dynamic threshold of MHBM employing MNDWI can adapt to several common confounding factors, except very small water bodies.

5.4. Initial threshold and optimal threshold The initial threshold was derived from the statistics of finite samples. In exceptional cases, the initial threshold can be too high to extract a valid water area, or so low that

INTERNATIONAL JOURNAL OF REMOTE SENSING

3445

Figure 11. True colour water body image and associated images and histograms derived using: (b) and (e) the normalised difference water index (NDWI); (c) and (f) the modification of normalised difference water index (MNDWI); and (d) and (g) the automated water extraction index (AWEI). Tl, Tr, and Ti denote the left edge, right edge, and initial threshold range, respectively. The peaks on the right of the histograms are water peaks and those on the left are non-water peaks.

land is extracted as water. Only optimal preliminary water masks can provide good twin peaks and valleys as precise as dynamic thresholds. In an extreme case, MHBM could not be applied if the initial threshold failed to extract valid water bodies completely. Therefore, vector shorelines of water distribution could be used to generate preliminary water masks instead of applying those extracted using the initial threshold. Furthermore, vector shorelines could provide the locations and approximate distributions of water bodies. However, the vector approach could only be used for extracting known water bodies. For extracting unknown water bodies, preliminary water masks must be extracted by the initial threshold approach. Additionally, the dynamic threshold may not be optimal at the limits of the threshold range; however, when it is limited within a specific threshold range, it can be used to extract water to within an acceptable level of accuracy.

3446

F. ZHANG ET AL.

Figure 12. Examples of water body extraction results in a full scene Landsat 8 OLI dataset (acquisition date: 15 January 2017): (a) false colour image (RGB: 5, 4, 3); and sub-regions of (b), clouds and cloud shadows, (c) building shadows, (d) small water bodies, and (e) topographic shadows.

5.5. Water area Drawing the histogram is an important step of MHBM. The histogram requires a certain number of pixels to achieve well-defined twin peaks. We simulated the histograms of different sized areas of pixels (from 100 to 5000), based on the Gaussian function (Figure 13). Areas of 100–500 pixels produce histogram curves that are uneven and jagged. However, the twin peaks are clear, which indicates that the threshold (valley value) could be found and MHBM would be successful for areas ≥100 pixels. Areas of ≥1000 pixels resulted in smooth histogram curves and perfect twin peaks. A Landsat 8 OLI image pixel is 900 m2 (30 m × 30 m). Therefore, MHBM is effective on water bodies that have an area of greater than 100 pixels, or 90 000 m2 (0.09 km2) in Landsat 8 OLI images, and the optimal area is ≥1000 pixels, or 900 000 m2 (0.9 km2). Uncertainty will arise associated with small water bodies such as small ponds and narrow streams. MHBM could be used for small water bodies using high resolution data, such as that of Sentinel-2A/B, which has a spatial resolution of

INTERNATIONAL JOURNAL OF REMOTE SENSING

3447

Figure 13. Histograms derived by Gaussian function simulation. Modification of normalised difference water index (MNDWI) histograms change with pixel number, showing different curves and twin peaks.

10 m, where it could be applied to ≥100 pixels area (10 000 m2 or 0.01 km2), with an optimal area of ≥1000 pixels (100 000 m2 or 0.1 km2).

5.6. Extensive application for other sensors MHBM is an automated method of water threshold calculation from water index images. Therefore, MHBM could be used in any water index image or single band grey image (such as the near infrared band). In addition, multi-spectral satellite sensors, such as Landsat-5 TM, Landsat-7 ETM+, and Sentinel-2A/B could be employed to calculate water indexes and used with MHBM. However, threshold statistics are necessary for new sensor data and a new water index, in order to calibrate the threshold range and initial threshold. Fortunately, the threshold range and initial threshold of a water index derived for a particular sensor could be directly applied by other users, reducing duplication of effort.

6. Conclusions In this study, we developed a modified histogram bimodal method (MHBM) for automatic extraction of water bodies, focusing on the automatic calculation of a dynamic threshold by following several steps: Rayleigh scattering correction, preliminary water extraction, water mask buffer, and accurate water distribution extraction. MHBM features two important modifications making it automatic and simple: 1) Rayleigh scattering correction and water area buffer generate suitable twin peaks, which are the foundation of automation; and 2) the method of dynamic threshold calculation seeks minimum values as precise thresholds within threshold ranges instead of seeking valleys between twin peaks, making the method simple to use. MHBM requires inputs of threshold range, initial threshold (or vector), and buffer size. Along with changes in sensor and water indexes, threshold range and initial threshold

3448

F. ZHANG ET AL.

statistics are required. We considered 152 scenes of Landsat 8 OLI image statistics to show that the most suitable water index is MNDWI, with a threshold range of 0.104– 0.427, initial threshold of 0.277, and optimum buffer size of 150%. We tested the MHBM method based on MNDWI using Landsat 8 OLI images of 18 water bodies. The RE and R2 for threshold accuracy were 2.1% and 0.962, respectively and the RE for area accuracy was 0.59%. The advantages of the method lie in its simplicity and minimal requirements for auxiliary data to achieve accuracy comparable to that of other automatic water extraction methods. Therefore, MHBM is applicable to mass time-series remote sensing data for calculating water thresholds from water index images and automatically extracting the spatial distribution of large water bodies.

Acknowledgments We would like to thank Yongxue Liu and Yuhao Yang for providing the MMFCM program, the Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences and Geospatial Data Cloud for providing Landsat images, the Joint Research Centre for providing Global Surface Water (GSW) data, and the Global Land Cover Facility for providing Global Inland Water products (GIW).

Disclosure statement No potential conflict of interest was reported by the authors.

Funding This work was supported by the Director Foundation of the Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences under Grant number Y6SJ2100CX, High Resolution Earth Observation Systems of National Science and Technology Major Projects under Grant number 41-Y20A31-9003-15/17, Strategic Priority Research Program of the Chinese Academy of Sciences, Grant number XDA19080304, National Natural Science Foundation of China (41701402, 41471308, 41571361, 91638201), and Youth Innovation Promotion Association of Chinese Academy of Sciences (2015128)

References Bagli, S., and P. Soille. 2004. “Automatic Delineation of Shoreline and Lake Boundaries from Landsat Satellite Images.” In Proceedings of Initial ECO-IMAGINE GI and GIS for Integrated Coastal Management,Seville, Spain, May 13–15, p. 4. Bishop, M. P. 1999. “SPOT Panchromatic Imagery and Neural Networks for Information Extraction in a Complex Mountain Environment.” Geocarto International 14 (2): 19–28. doi:10.1080/ 10106049908542100. Boland, D. H. 1976. Trophic Classification of Lakes Using Landsat-1 (ERTS-1) Multispectral Scanner Data. US Environmental Protection Agency, Office of Research and Development, Corvallis Environmental Research Laboratory,Corvallis, Oregon. Carroll, M., J. R. Townshend, C. M. DiMiceli, P. Noojipady, and R. Sohlberg. 2009. “A New Global Raster Water Mask at 250 M Resolution.” International Journal of Digital Earth 2 (4): 291–308. doi:10.1080/17538940902951401.

INTERNATIONAL JOURNAL OF REMOTE SENSING

3449

Feng, M., J. O. Sexton, S. Channan, and J. R. Townshend. 2016. “A Global, High-Resolution (30-M) Inland Water Body Dataset for 2000: First Results of A Topographic–Spectral Classification Algorithm.” International Journal of Digital Earth 9 (2): 113–133. doi:10.1080/ 17538947.2015.1026420. Feyisa, G. L., H. Meilby, R. Fensholt, and S. R. Proud. 2014. “Automated Water Extraction Index: A New Technique for Surface Water Mapping Using Landsat Imagery.” Remote Sensing of Environment 140: 23–35. doi:10.1016/j.rse.2013.08.029. Fisher, A., N. Flood, and T. Danaher. 2016. “Comparing Landsat Water Index Methods for Automated Water Classification in Eastern Australia.” Remote Sensing of Environment 175: 167– 182. doi:10.1016/j.rse.2015.12.055. Frazier, P. S., and K. J. Page. 2000. “Water Body Detection and Delineation with Landsat TM Data.” Photogrammetric Engineering and Remote Sensing 66 (12): 1461–1468. Gao, H., L. Wang, L. Jing, and J. Xu. 2016. “An Effective Modified Water Extraction Method for Landsat-8 OLI Imagery of Mountainous Plateau Regions.” IOP Conference Series: Earth and Environmental Science. IOP Conference Series: Earth and Environmental Science 34: 012010. doi:10.1088/1755-1315/34/1/012010. Hu, C. 2009. “A Novel Ocean Color Index to Detect Floating Algae in the Global Oceans.” Remote Sensing of Environment 113 (10): 2118–2129. doi:10.1016/j.rse.2009.05.012. Hu, C., J. Cannizzaro, K. L. Carder, F. E. Muller-Karger, and R. Hardy. 2010. “Remote Detection of Trichodesmium Blooms in Optically Complex Coastal Waters: Examples with MODIS FullSpectral Data.” Remote Sensing of Environment 114 (9): 2048–2058. doi:10.1016/j.rse.2010.04.011. Huang, L., Z. Li, B. S. Tian, Q. Chen, J. L. Liu, and R. Zhang. 2011. “Classification and Snow Line Detection for Glacial Areas Using the Polarimetric Sar Image.” Remote Sensing of Environment 115 (7): 1721–1732. doi:10.1016/j.rse.2011.03.004. Jain, S. K., R. Singh, M. Jain, and A. Lohani. 2005. “Delineation of Flood-Prone Areas Using Remote Sensing Techniques.” Water Resources Management 19 (4): 333–347. doi:10.1007/s11269-0053281-5. Ji, L., L. Zhang, and B. Wylie. 2009. “Analysis of Dynamic Thresholds for the Normalized Difference Water Index.” Photogrammetric Engineering and Remote Sensing 75 (11): 1307–1317. doi:10.14358/PERS.75.11.1307. Jiang, H., M. Feng, Y. Zhu, N. Lu, J. Huang, and T. Xiao. 2014. “An Automated Method for Extracting Rivers and Lakes from Landsat Imagery.” Remote Sensing 6 (6): 5067–5089. doi:10.3390/ rs6065067. Khandelwal, A., A. Karpatne, M. Marlier, J. Kim, D. Lettenmaier, and V. Kumar. 2017. “An Approach for Global Monitoring of Surface Water Extent Variations Using Modis Data.” Remote Sensing of Environment (In Press, Available Online 21 June 2017). doi:10.1016/j.rse.2017.05.039. Kingsford, R., R. Thomas, P. Wong, and E. Knowles. 1997. GIS Database for Wetlands of the Murray Darling Basin, Final Report of the Murray Darling Basin Commission. Sydney, Australia: National Parks and Wildlife Service. Leckie, D. G., E. Cloney, C. Jay, and D. Paradine. 2005. “Automated Mapping of Stream Features with High-Resolution Multispectral Imagery.” Photogrammetric Engineering and Remote Sensing 71 (2): 145–155. doi:10.14358/PERS.71.2.145. Lehner, B., and P. Döll. 2004. “Development and Validation of a Global Database of Lakes, Reservoirs and Wetlands.” Journal of Hydrology 296 (1): 1–22. doi:10.1016/j.jhydrol.2004.03.028. Li, J., Y.-W. Sheng, and J. Luo. 2011. “Automatic Extraction of Himalayan Glacial Lakes with Remote Sensing.” Yaogan Xuebao- Journal of Remote Sensing 15 (1): 29–43. Li, Y., X. Gong, Z. Guo, K. Xu, D. Hu, and H. Zhou. 2016. “An Index and Approach for Water Extraction Using Landsat–OLI Data.” International Journal of Remote Sensing 37 (16): 3611–3635. doi:10.1080/01431161.2016.1201228. Lira, J. 2006. “Segmentation and Morphology of Open Water Bodies from Multispectral Images.” International Journal of Remote Sensing 27 (18): 4015–4038. doi:10.1080/01431160600702384. Lu, D., and Q. Weng. 2007. “A Survey of Image Classification Methods and Techniques for Improving Classification Performance.” International Journal of Remote Sensing 28 (5): 823– 870. doi:10.1080/01431160600746456.

3450

F. ZHANG ET AL.

McFeeters, S. K. 1996. “The Use of the Normalized Difference Water Index (NDWI) in the Delineation of Open Water Features.” International Journal of Remote Sensing 17 (7): 1425– 1432. doi:10.1080/01431169608948714. Ouma, Y. O., and R. Tateishi. 2006. “A Water Index for Rapid Mapping of Shoreline Changes of Five East African Rift Valley Lakes: An Empirical Analysis Using Landsat TM and ETM+ Data.” International Journal of Remote Sensing 27 (15): 3153–3181. doi:10.1080/ 01431160500309934. Pekel, J. F., A. Cottam, N. Gorelick, and A. S. Belward. 2016. “High-Resolution Mapping of Global Surface Water and Its Long-Term Changes.” Nature 540: 418–422. doi:10.1038/nature20584. Prewitt, J., and M. L. Mendelsohn. 1966. “The Analysis of Cell Images.” Annals of the New York Academy of Sciences 128 (1): 1035–1053. doi:10.1111/j.1749-6632.1965.tb11715.x. Reis, S., and H. M. Yilmaz. 2008. “Temporal Monitoring of Water Level Changes in Seyfe Lake Using Remote Sensing.” Hydrological Processes 22 (22): 4448–4454. doi:10.1002/hyp.7047. Rokni, K., A. Ahmad, A. Selamat, and S. Hazini. 2014. “Water Feature Extraction and Change Detection Using Multitemporal Landsat Imagery.” Remote Sensing 6 (5): 4173–4189. doi:10.3390/rs6054173. Rundquist, D. C., M. P. Lawson, L. P. Queen, and R. S. Cerveny. 1987. “The Relationship between Summer-Season Rainfall Events and Lake-Surface Area1.” JAWRA Journal of the American Water Resources Association 23 (3): 493–508. doi:10.1111/j.1752-1688.1987.tb00828.x. Subramaniam, S., A. S. Babu, and P. S. Roy. 2011. “Automated Water Spread Mapping Using ResourceSat-1 AWiFS Data for Water Bodies Information System.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 4 (1): 205–215. doi:10.1109/ JSTARS.2010.2085032. Townshend, J. R., and C. Justice. 1986. “Analysis of the Dynamics of African Vegetation Using the Normalized Difference Vegetation Index.” International Journal of Remote Sensing 7 (11): 1435– 1445. doi:10.1080/01431168608948946. Tucker, C., and P. Sellers. 1986. “Satellite Remote Sensing of Primary Production.” International Journal of Remote Sensing 7 (11): 1395–1416. doi:10.1080/01431168608948944. USGS. SRTM Water Body Dataset. 2012. Accessed June 17 2014. https://lta.cr.usgs.gov/srtm_water_ body_dataset Verpoorter, C., T. Kutser, and L. Tranvik. 2012. “Automated Mapping of Water Bodies Using Landsat Multispectral Data.” Limnol. Oceanogr. Methods 10: 1037–1050. doi:10.4319/lom.2012.10.1037. Wang, M. 2007. “Remote Sensing of the Ocean Contributions from Ultraviolet to Near-Infrared Using the Shortwave Infrared Bands: Simulations.” Applied Optics 46 (9): 1535–1547. doi:10.1364/ AO.46.001535. White, M. 1978. “Reservoir Surface-Area from Landsat Imagery.” Photogrammetric Engineering and Remote Sensing 44 (11): 1421–1426. Work, E. A. Jr, and D. S. Gilmer. 1976. “Utilization of Satellite Data for Inventorying Prairie Ponds and Lakes.” Photogrammetric Engineering and Remote Sensing 42 (5): 685–694. http://worldcat. org/issn/00991112. Xie, H., X. Luo, X. Xu, H. Pan, and X. Tong. 2016. “Evaluation of Landsat 8 OLI Imagery for Unsupervised Inland Water Extraction.” International Journal of Remote Sensing 37 (8): 1826– 1844. doi:10.1080/01431161.2016.1168948. Xu, H. 2006. “Modification of Normalised Difference Water Index (NDWI) to Enhance Open Water Features in Remotely Sensed Imagery.” International Journal of Remote Sensing 27 (14): 3025– 3033. doi:10.1080/01431160600589179. Yang, Y., Y. Liu, M. Zhou, S. Zhang, W. Zhan, C. Sun, and Y. Duan. 2015. “Landsat 8 OLI Image Based Terrestrial Water Extraction from Heterogeneous Backgrounds Using a Reflectance Homogenization Approach.” Remote Sensing of Environment 171: 14–32. doi:10.1016/j. rse.2015.10.005. Zanter, K. 2015. Landsat 8 (L8) Data Users Handbook. U.S. Geological Survey, National Oceanic and Atmospheric Administration. http://www.webcitation.org/6mu9r7riR.

INTERNATIONAL JOURNAL OF REMOTE SENSING

3451

Zhang, T., H. Ren, Q. Qin, C. Zhang, and Y. Sun. 2017. “Surface Water Extraction from Landsat 8 OLI Imagery Using the LBV Transformation.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 10 (10): 4417–4429. doi:10.1109/JSTARS.2017.2719029. Zhang, X., C. Zhu, J. Luo, and J. Yang. 2013. “A Method for Water Body Adaptive Extraction from Remote Sensing Imagery Based on Local End Member Spectral Characterization.” Spectroscopy and Spectral Analysis 33 (6): 1670–1677. doi:10.3964/j.issn.1000-0593(2013)06-1670-08.