4548
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 7, NO. 11, NOVEMBER 2014
A Decision-Tree Classifier for Extracting Transparent Plastic-Mulched Landcover from Landsat-5 TM Images Lizhen Lu, Liping Di, Senior Member, IEEE, and Yanmei Ye
Abstract—The area of plastic-mulched landcover (PML), an important agricultural landscape, is increasing rapidly at a rate of 20% per year globally. However, the spatial and temporal distributions of PML have been poorly understood because of lack of effective technology to extract PML for large geographic areas. This paper presents a decision-tree classifier for extracting the transparent PML information from Landsat-5 TM images. The classifier was built with rules obtained from analyzing the spectral characteristics of transparent PML on Landsat-5 TM images, covering the study area of in Xinjiang, the largest PML-based cotton plantation provinces in China. Then, the classifier was applied at the study area for years 1998, 2007, and 2011. Results indicate that the classifier successfully extracted the PML from Landsat-5 TM images at overall accuracies of 97.82%, 85.27%, and 95.00% and Kappa coefficients of 0.9782, 0.80, and 0.93 for years 2011, 2007, and 1998, respectively. The results also imply that the decision-tree classifier is temporally stable and can be applied in different years. Visual comparison of the results with the high-spatial resolution images on Google Earth also shows that detected locations of PML are correct. The study shows that the classifier is an effective method for extracting PML for large geographic areas from Landsat-5 TM. Because of the long history of global coverages as well as free availability of Landsat-5 TM images, it is feasible to map the spatio–temporal dynamics of PML over large geographic areas with the technologies presented in this paper. Index Terms—Decision-tree classifier, Landsat TM, plasticmulched landcover (PML), Xinjiang cotton region.
I. INTRODUCTION INCE the first use of plastic film in agriculture was in 1948 [1], plastic covering has been used extensively in the cultivation of vegetables (e.g., tomato and beans), fruits (e.g., strawberries), and crops (e.g., cotton, wheat, and rice) [2]. The purpose of using plastics cover in agriculture is to increase the productivity by mitigating the threats of coldness, heat, drought, wind, insects, and crop diseases [3]. Plasticulture, in a broad
S
Manuscript received November 19, 2013; revised January 19, 2014; accepted May 13, 2014. Date of publication June 16, 2014; date of current version January 06, 2015. This work was supported in part by China’s National Science and Technology Support Program under Grant 2008BAB38B05. (Corresponding author: Liping Di.) L. Lu is with the Department of Earth Science, Zhejiang University, Hangzhou 310027, China (e-mail:
[email protected]). L. Di is with the Center for Spatial Information Science and Systems, George Mason University, Fairfax, VA 22030 USA (e-mail:
[email protected]). Y. Ye is with the Institute of Land Science and Real Estate, Zhejiang University, Hangzhou 310027, China (e-mail:
[email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JSTARS.2014.2327226
sense, is defined as the use of plastics in agriculture, including both plant and animal production [3]–[6]. The main types of plastic landcover include greenhouse or walk-in tunnels, small tunnels, and plastic mulch (Fig. 1). There are mainly two types of plastic film used as mulch: 1) the transparent and 2) the black. This study concentrated on plastic mulch with transparent film, the dominate type of plastic landcover. Plasticultural area has been rapidly expanding worldwide in the past 2 decades and now represents an important cultivated landscape. It was reported that plasticultural area has been expanded at the rate of about 20% per year on average globally over the last decade [2]. In China, this is especially true due to the large-scale shift in land use from cereal crops to more profitable vegetables and cash crops in the last 2 decades [7]. The plasticultural area in China has grown from 4200 ha in 1981, 11 966 873 ha in 2003, to 28 000 000 ha in 2010, with 99% of them transparent plastic mulch [8]–[10]. The large-scale landcover change introduced by plasticulture must have impacts on climate, ecosystem, and environment regionally and globally because it alters the energy balance and water cycles on the land surfaces, deteriorates the soil structure, and reduces the biodiversity. For example, transparent plastic-mulch film allows visible lights to penetrate, but blocks outgoing longwave radiation and thus causes the greenhouse effect [11], whereas black plastic-mulch film has opposite effects. Plastic-mulch film also prevents the evaporation of water from the soil and reduces energy consumption. The plastic film wastes in the soil harm the plant root system development [12], [13]. Plasticulture also changes pollination of plants and distribution of insects and bird and has a negative effect on biodiversity [14]. Therefore, mapping and monitoring the plasticulture on a large geographic area are important both scientifically and socio-economically. Remote sensing is the only feasible approach for monitoring plasticulture and understanding its impacts on climate and ecoenvironment in a large geographic area (e.g., whole China or whole East Asia). In recent years, there have been a limited number of studies on remote sensing of plasticulture by using high (at meter level) spatial resolution images. Carvajal et al. [15] proposed an artificial intelligence neural network to detect greenhouse using 2.44 m-resolution QuickBird imagery. With analysis of data collected by ground spectrometer, Levin et al. [16] claimed that white and transparent plastic-mulch films have three absorptions centered at 1218, 1732, and 2313 nm that are not affected by dust, rinse, and surface factors. They classified 1 m-resolution AISA-ES hyper-spectral imagery to reach a
1939-1404 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
LU et al.: DECISION-TREE CLASSIFIER FOR EXTRACTING TRANSPARENT PML FROM LANDSAT-5 TM IMAGES
Fig. 1. Plasticulture landscapes in Asia. (a) Small tunnel covers Pujiang, Chengdu, Sichuan, China. (b) Plastic mulched landPujiang, Chengdu, Sichuan, China. cover Ulsan Airport, (c) Greenhouse or walk-in tunnels Bukgu, Ulsan, South Korea. (d) Greenhouse or walk-in tunnels Sakurai City, Nara Prefecture, Japan. (e) Plastic mulched , Shawan, Yili, Xinjiang China. landcover , Shouguang, (f) Greenhouse or walk-in tunnels Weifang, Shandong, China.
detection accuracy of 90% for the transparent plastic-mulch film and 70% for the black one. Agüera et al. [17] used the maximumlikelihood method to extract greenhouse locations from Quickbird and IKONOS imageries. Result show that although both Quickbird and IKONOS images have satisfactory detection accuracy, Quickbird has a higher accuracy and the use of texture information does not raise the detection accuracy. In a follow-on study, Agüera and Liu [18] proposed an algorithm to detect greenhouses based on Quickbird and IKONOS imagery. The algorithm detected greenhouses by using the maximumlikelihood classification method and then auto-vectorized the raster greenhouses to irregular polygons whose outline is transformed to straight lines based on Hough Transformation. In that case study, the detection accuracy of Quickbird imagery is 66.7%, whereas that of IKONOS is 49%. Tarantino and Figorito [19] mapped rural areas with widespread plastic covered vineyards using very high spatial resolution true color aerial images. Both spectral information and spatial texture are used in their classification. The type of plasticulture they dealt with is the tunnel/greenhouse, interlacing with uncovered farmlands and man-made structures. This type of plasticulture is significantly
4549
different from the plasticulture this paper deals with, the transparent plastic-mulch, which has much large application scale which forms more uniformed landscape. A limited number of studies based on median (at tens meter level) resolution imagery have also been conducted. The researches yielded mixed results. Colby and Keating [20] employed the parallelepiped method in combination with the maximum-likelihood and the minimum distance for landcover classification of Landsat TM imagery, but failed to obtain the expected results. Thunnissen and Wit [21] found, during the creation of the Netherland LGN3 database using Landsat TM and SPOT imagery, that remote sensing imagery solely is not able to provide precise classification information of mulched fields. Picuno et al. [14] employed the parallelepiped method to extract mulched fields from Landsat TM imagery, and then used SAR imagery to validate detection accuracy. They then simulated the scenery of plastic-mulched fields using a geographic information system (GIS) and 3-D landcover modeling. The above studies show that high spatial resolution (at meter level) remote sensing imagery can be used to extract plasticulture landcover effectively. Those studies have mainly concentrated on extraction of greenhouse/tunnel type of plasticultural landcover in the mixed heterogeneous landscape. However, plastic mulch is by far the largest type of plasticulture in term of the area it covers. In China, 95% of plasticulture exists in mulch form [9] and 99% of them use the transparent plastic film (TPF), which is directly laid on the ground. Therefore, mapping plastic-mulched landcover (PML) is more important than mapping the tunnel/greenhouse type of plasticultural landcover in term of understanding the impacts of large scale plasticulture on the climate and environment. Plasticulture mulch is typically applied on field crops (such as cotton and corn) in large area continuously, forming a much more uniformed landscape [e.g., Fig. 1(e)] than the tunnel type of plasticulture. Therefore, it is no doubt that PML can be easily extracted from the high spatial resolution imagery. However, the use of high spatial resolution images is not feasible for mapping PML in a large geographic area because of high data cost and infrequent temporal coverage. Therefore, because of its availability and low or no cost, moderate resolution images, such as Landsat TM, are the only option for monitoring PML for large geographic areas if effective algorithms can be developed and validated, although the effectiveness of using such data may depend on the spatial scale of plasticulture in comparison with the spatial resolution of images. The large-scale and unformed landscape of PML makes the development of such algorithms possible. In this paper, we concentrated our study on the method for mapping the transparent PML over large geographic areas. An effective decision-tree classifier has been proposed to extract PML for large geographic areas from Landsat TM data. The classifier has been tested and validated for mapping the PML of cotton fields in a study area located in Xinjiang Province, China. II. STUDY AREA AND DATA SETS A. Study Area Xinjiang Province is the largest cotton producer in China, accounting for 40% of national annual production. In 2007, there were 1 484 100 ha of cotton fields and 100% of them were
4550
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 7, NO. 11, NOVEMBER 2014
PHENOLOGICAL CALENDAR
OF
TABLE I MAJOR PLANTATION
IN
NORTHERN XINJIANG [22]
TABLE II LANDSAT TM LEVEL 1T DATA SOURCE PARAMETERS (PATH 144, ROW 29)
mulched by TPF [12]. Because of large PML area, it is ideal to select a cotton planting region in the province as the study area. The region is selected through visually inspecting the meter-level resolution remote sensing images available on Google Earth. According to visual inspection of the GeoEye images acquired on April 27, 2011 and shown on Google Earth, a large region bounded with spatial coordinates of ( ), ( ), ( ), and ( ) contains large percentage of PML [see Fig. 1(e)]. We chose this region as the study area. As shown in Fig. 2, the study area is located in the northern Xinjiang. It borders with the Tian Shan Mountain on the south and the Gurbantunggut Desert on the north and contains Shawan, Shihezi Shi, and Manasi Counties. The cotton fields in the study area mainly belong to the Eighth Division of the Xinjiang Production and Construction Corps. Geographically, the study area is located in inland China with mountain ranges in north and south sides blocking moist air from oceans. It has typical temperate continental arid climate with cold winter, hot summer, and a large diurnal temperature range. The study area has abundant sunshine (2600 h annually) and receives the most in July. The annual average temperature is between 6.0 C and 7.5 C, the accumulated temperature above 0 C about 3900 C–4100 C, the annual precipitation around 140–200 mm, and the annual evaporation 1237–1770 mm. The study area is irrigated mainly with water from the Manas River fed by precipitation and snowmelt. Soil types in the study area include gray desert soil, aquic soil, and saline meadow soil. Soil textures include gravel soil, sand soil, clay soil, etc. In the plantation area, long-term irrigation leads to formation of irrigated desert sierozem and irrigated meadow soil. Salinization and sandstorm are two major natural hazards in the study area. The major crops in the study area include cotton, winter wheat, spring corn, tomato, and pepper, and cotton accounts for over 70% of the total planted area [22]. B. Data Sets Table I shows the phenological calendar of major crops, including cotton, winter wheat, and spring corn, in the study area. This table shows that cotton is planted around middle and late April and
emerges in early May, whereas corn is planted in the late April and early May. Both cotton and corn fields are mulched with transparent PML. During the planting and emerging period, the Normal Difference Vegetation Index (NDVI) values of both cotton and corn fields are very low, whereas the NDVI values of the winter wheat and other vegetated fields are high due to the difference in the phonological stages (i.e., winter wheat is at the jointing and heading stage). Such a phenological difference makes PML easy to be extracted from satellite remote sensing imagery. In addition, this paper concerns only on extracting the PML, not on further distinguishing between PML on cotton and PML on corn. Therefore, this study selects high-quality Landsat TM images mainly in May and some in late July and early August of 1998, 2007, and 2011 as the source data. The use of Landsat images acquired in July and August is for distinguishing the Bare land and Fallow land. The Landsat TM data were downloaded from the International Science Data Services Platform of China (http://datamirror.csdb. cn/index.jsp) and the U.S. Geological Survey (USGS) official website (http://earthexplorer.usgs.gov/). Table II lists the Landsat TM data used in this study. In addition to the Landsat TM images, statistics data from other references, such as the Xinjiang Yearbook of 1998, 2007, and 2011, are also used in this study. C. Data Preprocessing The original Landsat scenes include a part of the Tian Shan Mountain in the south and the Gurbantunggut desert in the north. Both are not the region of interest (ROI) for this study. In order to reduce data processing time, a subsetting operation was performed to extract a rectangle region [Fig. 2(b)] in the middle of the Landsat scenes, which excludes both mountainous area in the south and the desert in the north, as our study area. The ROI has a size of (including the blank area) with the land area of 1 402 403 ha. Before PML extraction is carried out, the DNs of the Landsat images in ROI are converted to normalized at-the-satellite spectral radiance
LU et al.: DECISION-TREE CLASSIFIER FOR EXTRACTING TRANSPARENT PML FROM LANDSAT-5 TM IMAGES
4551
Fig. 2. Study area in Xinjiang, China.
NUMBER
OF
TRAINING PIXELS
FOR
TABLE III DIFFERENT LANDCOVER TYPES AFTER PURIFICATION
where is the at-the-satellite spectral radiance of a specific pixel with its pixel value to be DN, is the at-satellite normalized spectral radiance, is band-specific spectral radiance scaled to the maximum DN value (i.e., 255), and is band-specific spectral radiance scaled to the minimum DN value (i.e., 0). Both and are from the USGS Landsat 5 post-launch calibration table (http://landsat.usgs.gov/documents/L5TM_ postcal.pdf) which has considered the sensor degradation. In this study, we use the normalized at-the-satellite spectral radiance to approximate the surface reflectance since atmospheric conditions were not available for the study area. The geometric correction is carried out by selecting the Ground Control Points (GCPs) through Google Earth and using the Image to Map function of ENVI. D. Ground Truth The ground truth for this study was obtained through visually interpreting the high spatial resolution GeoEye image of April 27, 2011 for the study area, since the plasticultural area can be easily identified on GeoEye images visually. The visually interpreted landcover information was overlaid on the Landsat scenes of the same year to obtain the ground truth at the Landsat TM resolution. Then, randomly a half of the ground-truth pixels on the Landsat TM images were used for deriving rules for the decision tree and another half for validating results. Because of unavailability of high spatial resolution images, the ground truth for years 1998 and 2007 was obtained through direct visual interpretation of Landsat TM images based on the spectral and spatial patterns learned from the images of year 2011. The Landsat TM training pixels obtained from the ground truth in 1998, 2007, and 2011 were purified by using n-Dimensional Visualizer in ENVI 4.7 to obtain the representative pixels for each landcover type. Table III summarizes the number of training pixels for different landcover types after the purification.
III. ALGORITHM This section describes the proposed PML algorithm. Note that the term of the surface reflectance is used to describe the algorithm. In the implementation, we use normalized spectral radiance to approximate it since surface reflectance is not available for this study. A. Detectable Features of PML In order to obtain the rules for the decision-tree classifier, we have to discover the detectable features of PML in Landsat TM images. To facilitate the discovery, we synthesized both the true and false color composite images and then visually inspected the composites. A true color composite image was synthesized by using Landsat TM images band 3 ( μ ), band 2 ( μ ), and band 1 ( μ ). In addition, two false color composite images were synthesized by using band 7 ( μ ), band 5 ( μ ), and band 3, as well as band 7, band 4 ( μ ), and band 3. Fig. 3 shows the Landsat color composite images of May 10, 2011 for a part of the study area. From Fig. 3, we can find that the composite of band 7: R, band 4: G, and band 3: B (RGB:band 743) preserves ground geographic features and surface colors better than other composites. Therefore, RGB:band 743 composite is selected as the base imagery for future inspection. As shown in Fig. 3(b), the PML can be easily seen in the RGB:band 743 composite. Because of almost all plastic mulch used in the research area is transparent, dependent on the crop growing condition, soil types, and soil moisture, the PML shows either grayish blue or grayish green colors in the composite image. It can be easily distinguished visually from other landcover types (see the visually interpreted features of typical land use/cover types in Table IV). The areas with grayish blue color are mainly distributed along rivers and low-lying land.
4552
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 7, NO. 11, NOVEMBER 2014
Fig. 3. Landsat TM color composites using different bands (date: May 10, 2011). (a) RGB:band 321. (b) RGB:band 743. (c) RGB:band 753.
VISUAL INTERPRETATION FEATURES
OF
TABLE IV TYPICAL LAND USE/COVER TYPES
IN THE
STUDY AREA
Bare land 1 is gobi of quartz sands mixed with small gravel and Saline land is the bare land whose top soil has high salt content.
LU et al.: DECISION-TREE CLASSIFIER FOR EXTRACTING TRANSPARENT PML FROM LANDSAT-5 TM IMAGES
4553
Fig. 4. Spectral signatures of typical landcover types in study area ( is the mean radiance of the training samples, and is the maximum radiance). (a) The eight typical landcover types in May 2011. (b) The eight typical landcover types in May 2007. (c) The eight typical landcover types in May 1998. (d) PML1 and PML2 in May 2011, 2007, and 1998, respectively.
In reference to the high spatial resolution imagery in Google Earth and our prior knowledge of the landcover in the region, we recognize eight typical land use/cover types in the ROI, including PML1, PML2, Saline land, Bare land 1, Bare land 2, Vegetation cover (mainly winter wheat), Fallow land, and water body, from images taken in May of 1998, 2007, and 2011, respectively. After sample purification by n-Dimensional Visualizer in ENVI 4.7 software, spectral signatures of the eight typical landcover types in the ROI are generated. Fig. 4(a) shows the spectral signatures for 2011, Fig. 4(b) for 2008, and Fig. 4(c) for 1998. Fig. 4(a) shows that the differences between the normalized radiance ( ) of PML1 and PML2 on different bands are consistently around 0.1, whereas their spectral signatures are quite similar in shape. Meanwhile, the overall spectral signatures of PML1 and PML2 resemble Bare land and Fallow land, except that the latter have a much larger difference in the normalized spectral radiance between band 5 and band 3. In addition, the spectral signatures of PML1 and PML2 are significantly different from those of Saline land, Vegetation cover, and water bodies at band 3, band 4, and band 5, respectively. In summary, PML1 and PML2 can be identified and extracted from TM images based on their reflective features on band 3, 4, and 5. Fig. 4(d) shows the spectral patterns of PML1 and PML2 in May of years 2011, 2007, and 1998, respectively. From Fig. 4(d),
we can find that the spectral patterns of PML are very similar in different years. Therefore, the same spectral patterns of PML1 and PML2 can be used to extract the PML in different years. Based on the above analysis, we employed the following indices for the PML extraction from Landsat TM images. 1) Normalized Difference Water Index (NDWI): NDWI is an index that uses the reflected near-infrared and visible green radiations to delineate open water features and enhance their presence in remotely-sensed digital imagery [23]. NDWI can be calculated as follows:
where Green is the surface reflectance in the green band (band 2 in this study) and NIR is the surface reflectance in the nearinfrared band (band 4 in this study). In this study, we use the normalized radiance to approximate the surface reflectance. 2) Normalized Difference Vegetation Index: NDVI is an index that is widely used for evaluating vegetation condition over the land surface [24]–[27]. NDVI can be calculated as follows:
4554
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 7, NO. 11, NOVEMBER 2014
where Red is the surface reflectance of the red band (band 3 in this study) and NIR is the surface reflectance in the nearinfrared band (band 4 in this study). Similar to NWDI, we use normalized spectral radiance instead of surface reflectance in this study. 3) PML Index (PMLI): Landsat band 3 (b3) locates at chlorophyll absorption band. Bare land has high reflectance in this band, whereas heavily vegetated area has low reflectance in this band. Saline land barely has any vegetation. Therefore, it has high reflectance in this band. In the early growing season, transparent PML has high reflectance in this band too since the newly plowed land with cotton seedlings has little vegetation. On other hand, Bare land and Fallow land also grow some weeds. Therefore, there is no big difference in reflectance among PML, Bare land, and Fallow land in b3. The b5 locates at the water absorption band ( μ ). Therefore, it is sensitive to the water content on the surface layer of soil. Saline land has high surface reflectance on b5 since its surface usually is very dry. Since the water contents in the surface layers of PML, Bare land, Fallow land have no much difference, their reflectances in b5 also have no significant difference [see Fig. 4(a)]. Therefore, we proposed a PMLI by using the reflectances of Landsat TM band 3 (b3) and band 5 (b5):
By examining the PMLI, we found that at index value of 0.28, PML and Saline land classes can be separated from Bare land, Fallow land, and Vegetation classes. Further separation between PML and Saline land classes can be done by using the property that Saline land has much higher reflectance in b5 than PML. B. Construction of the Decision-Tree Classifier Decision tree is a nonparametric classifier involving a recursive partitioning of the feature space, based on a set of rules learned through analyzing the training set [28]. A decision tree, known as a top-down classification approach, is composed of a root node (containing all data), a set of internal nodes (splits), and a set of terminal nodes (leaves) [29], [30]. Due to their relatively simple, explicit, and intuitive classification structure [29], the decision-tree classifiers have been successfully used for extraction of landcover information from remote sensing data [30]. For instance, Hansen et al. [31] compared the performance of a decision-tree classifier with that of a maximum-likelihood classifier by using a 1 by 1 global data set and found the classification accuracy was comparable. Simard et al. [32] constructed a decision tree without assuming a particular probability density distribution of the input data and applied the tree to classify SAR images. The results showed that the tree is adaptive for landcover classification. As shown in Table IV and Fig. 4, there are significant differences in the visual interpretation features and the spectral curves between PML1 and PML2 and between Bare land 1 and Bare land 2 due to the difference in underlying soils. However, this paper does not care much on the types of the underlying surface. Therefore, PML1 and PML2 are merged into one PML
Fig. 5. Decision-tree classifier (note: T stands for True; F, False; and NDVI8, the NDVI value in the peak of a growing season e.g., August. All other variables in the decision tree are derived from the TM image in the early growing season e.g., middle to late May in the study area).
class and Bare land 1 and Bare land 2 into Bare land class. Therefore, there are six landcover types in this study. Based on the discussion above, a decision-tree classifier has been constructed. The tree’s decision nodes, all calculated from May Landsat Images except for those explicitly stated, are described below. 1) : If “ ” is true, it is a water body. Otherwise, it is other types of landcovers, such as PML, Saline land, Vegetation cover, Bare land, and Fallow land. 2) > : If “ > ” is true, it is the Vegetation cover; otherwise, it is PML, Saline land, Bare land, or Fallow land. 3) : If “ ” is false, it is either PML or Saline land; otherwise, it is either Bare land or Fallow land. 4) > : Bare land and Fallow land have similar spectral signatures in May, but can be differentiated by the Landsat images in July/August (peak of the growing season) since Fallow land grows a lot of weeds but Bare land do not. NDVI value in the peak of growing season can easily distinguish the two. In this study, we use NDVI8 to represent the NDVI in the peak of a growing season (i.e., July/August). If “ > ” is true, it is Fallow land; otherwise, it is Bare land, including Bare land 1 and Bare land 2. 5) > : If “ > ” is false, it is PML, including PML1 and PML2; otherwise, it is Saline land. In above, b5 stands for the reflectance of TM band 5.
LU et al.: DECISION-TREE CLASSIFIER FOR EXTRACTING TRANSPARENT PML FROM LANDSAT-5 TM IMAGES
4555
Fig. 7. Comparison of classification results and their corresponding colorcomposite images in 2011.
Fig. 6. Classification results from the decision-tree classifier. (a) 2011. (b) 2007. (c) 1998.
IV. RESULTS AND DISCUSSION ENVI 4.7 Decision-tree tool is used to implement the decisiontree classifier. The classifier is used to classify the Landsat TM images for the study area. The results are shown in Fig. 6. From Fig. 6, it is apparent that PML pixels mainly occur in the middle and south parts of the study area, whereas Bare land, another main type of landcover in the study area, in the north part of the study area. This pattern of landcover distribution matches well with the ground truths visually interpreted from the highresolution GeoEye images and the Landsat color composites. Statistics of the classification result show that PML accounts for 79.9%, 84.4%, and 80.4% of the total farmland, which includes PML, Fallow land, and Vegetation cover in years 2011, 2007, and 1998, respectively. The right-side images of Fig. 7 show the zoom-in of classification results for two small areas in the study area, whereas the left-side images show the corresponding Landsat color composites. From Fig. 7, it is clear that the decision-tree classifier proposed in this paper is very effective for extracting PML information.
Further analysis on the classification accuracy with ground truth is made by using the confusion matrix tool in ENVI 4.7 software, which generates the producer accuracy (PA), the user accuracy (UA), the overall accuracy (OA), and Kappa coefficients [33]. PA is a measure indicating the probability that the classifier has labeled an image pixel into Class A given that the ground truth is Class A. UA is a measure indicating the probability that a pixel is Class A given that the classifier has labeled the pixel into Class A. Table V shows the summary of classification accuracy assessment. From Table V, it can be found that OAs in 2011, 2007, and 1998 are all higher than 85%, whereas Kappa coefficients are greater than 0.8. For all landcover types, only the producer accuracies of Saline land in 2007 and 1998 and Vegetation cover in 2007 are lower than 80%, which indicates this landcover type is most easily misclassified into other classes, although those two classes are not the key classes in this study. In addition, UAs of all landcover types are higher than 74%. This means that the decisiontree classifier is an effective method for extracting not only PML, but also other types of landcover, except for Saline land. We attribute the high classification accuracy to the following reasons: 1) the transparent PML has very distinct spectral signatures from the other landcover types in the study area; 2) the study area has a large field size compared with the spatial resolution of Landsat TM images; 3) there is a single type of plastic film, TPF, used as the mulch; and 4) the large scale, spatially-continued application of TPF forms the uniformed landscape. With the multiple years of PML for the ROI, we can analyze the change of PML from 1998 to 2007 and from 2007 to 2011. Table VI shows the landcover change from 1998 to 2007 and Table VII from 2007 to 2011. From Table VI, we can find that PML from 1998 to 2007 increased 78.04%. Most of the increase comes from the conversion from the Vegetation (mainly winter
4556
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 7, NO. 11, NOVEMBER 2014
CONFUSION MATRIX
FOR THE
TABLE V DECISION-TREE CLASSIFIER USING THE LANDSAT TM IMAGES
IN
2011, 2007, AND 1998
PA, producer accuracy; UA, user accuracy.
TABLE VI 1998 TO 2007 IN
THE
STUDY AREA
TABLE VII 2007 TO 2011 IN
THE
STUDY AREA
LANDCOVER CHANGE
FROM
LANDCOVER CHANGE
FROM
wheat) fields, Fallow land, and Saline land. The large increase of PML in the study area during this period corresponded with the large increase of PML in entire China during the same period
when the Chinese government promoted PML as the major farming advance to increase crop productivity and farmer incomes [8]. As shown in Table VII, PML from 2007 to 2011
LU et al.: DECISION-TREE CLASSIFIER FOR EXTRACTING TRANSPARENT PML FROM LANDSAT-5 TM IMAGES
decreased 7.25% in the study area, whereas Vegetation (mainly winter wheat) increased 74.07% and Fallow land 15.04%. This means that the area of PML in the study area reached the peak around 2007 and has been stable in recent years. The large increases in the Vegetation (mainly winter wheat) and Fallow landcovers reflected the government incentives on cereal crops and land abandoning due to migration of farmers to urban areas in recent years [9]. V. CONCLUSION The decision-tree classifier proposed in this paper can map the spatial and temporal dynamics of transparent PML over a large geographic region with Landsat TM images. Because of global spatial coverage, long temporal coverage, and no cost of Landsat TM images, it is feasible to map the spatio–temporal dynamics of PML over large geographic areas with the technologies presented in this paper. Such mapping will provide valid scientific information to support governmental agricultural decision-making. It also provides the fundamental information to study the impacts of large-scale PML on climate and environment. ACKNOWLEDGMENT The authors would like to thank Ms. X. Sha, a graduate student from Zheijiang University for the data preprocessing. They would also like to thank the reviewers for their valuable comments. REFERENCES [1] J. C. Garnaud, “Plasticulture magazine: A milestone for a history of progress in plasticulture,” Plasticulture, vol. 1, no. 119, pp. 30–43, 2000. [2] A. Espí, A. SalmerÓn, A. Fontecha, Y. García, and A. I. Real, “Plastic films for agricultural applications,” J. Plast. Flim Sheeting, vol. 22, no. 2, pp. 85–102, Apr. 2006. [3] T. Takakura and W. Fang. (2002). Climate Under Cover-Digital Dynamic Simulation in Plant Bio-Engineering, 2nd ed. [Online]. Available: http:// www.google.com.tw/books?id=QRM7QeW0yZUC&printsec=frontcover& hl=zh-CN& [4] P. Dubois, Plastics in Agriculture. London, U.K.: Applied Science Publishers, 1978, p. 176. [5] W. J. Lamont, “What are the components of a plasticulture vegetable system?” HortTechnology, vol. 6, no. 3, pp. 150–154, Sep. 1996. [6] E. E. Carey, L. Jett, W. J. Lamont, T. T. Nennich, M. D. Orzolek, and K. A. Williams, “Horticultural crop production in high tunnels in the United States: A snapshot,” HortTechnology, vol. 19, no. 1, pp. 37–43, 2009. [7] L. Di, “Landuse pattern and landuse change,” in Changing China: A Geographical Appraisal, C. Hsieh and M. Lu, Eds. Boulder, CO, USA: Westview, 2003, pp. 17–31. [8] MA/PRC. (2007, May). Abridged National Statistics of agriculture. P.R.C. Ministry of Agriculture ed. [Online]. Available: http://www.agri.gov.cn/ sjzl/nongyety.htm [9] MA/PRC. (2010, May). Statistical information: The Summary of National Agricultural Statistics. Ministry of Agriculture [Online]. Available: http:// www.agri.gov.cn/sjzl/nongyety.htm [10] G. Zhou, “Analysis of situations of China agro-film industry (2010) and countermeasures for its development,” China Plast., vol. 24, no. 8, pp. 9–12, Aug. 2010. [11] E. Waggoner, P. M. Miller, and H. C. Deroo, “Plastic mulching-principles and benefits,” Connecticut Agri. Exp. Station Bull., no. 634, pp. 5–44, 1960. [12] J. Li, “Economic analysis of agro-film pollution in Xinjiang region,” Border Econ. Culture, vol. 1, no. 1, pp. 16–17, Jan. 2008. [13] W. Q. He et al., “The use of plastic mulch film in typical cotton planting regions and the associated environmental pollution,” J. Agro-Environ. Sci., vol. 28, no. 8, pp. 1618–1622, Aug. 2009.
4557
[14] P. Picuno, A. Tortora, and R. L. Capobianco, “Analysis of plasticulture landscapes in southern Italy through remote sensing and solid modeling techniques,” Landscape Urban Plann., vol. 100, no. 1–2, pp. 45–56, Mar. 2011. [15] F. Carvajal, E. Crisanto, F. J. Aguilar, F. Aguera, and M. A. Aguilar, “Greenhouses detection using an artificial neural network with a very high resolution satellite image,” in Proc. ISPRS Tech. Comm. II Symp., Vienna, 2006, pp. 37–42. [16] N. Levin, R. Lugassi, U. Ramon, O. Braun, and E. Ben-Dor, “Remote sensing as a tool for monitoring plasticulture in agricultural landscapes,” Int. J. Remote Sens., vol. 28, no. 1, pp. 183–202, Jan. 2007. [17] F. Agüera, F. J. Aguilar, and M. A. Aguilar, “Using texture analysis to improve per-pixel classification of very high resolution images for mapping plastic greenhouses,” ISPRS J. Photogramm. Remote Sens., vol. 63, no. 6, pp. 635–646, Nov. 2008. [18] F. Agüera and J. G. Liu, “Automatic greenhouse delineation from QuickBird and IKONOS satellite images,” Comput. Electron. Agric., vol. 66, no. 2, pp. 191–200, May 2009. [19] E. Tarantino and B. Figorito, “Mapping rural areas with widespread plastic covered vineyards using true color aerial data,” Remote Sens., vol. 4, no. 7, pp. 1913–1928, Jul. 2012. [20] D. Colby and P. L. Keating, “Land cover classification using Landsat TM imagery in the tropical islands: The influence of anisotropic reflectance,” Int. J. Remote Sens., vol. 19, no. 8, pp. 1479–1500, May 1998. [21] H. Thunnissen and A. D. Wit, “The national land cover database of the Netherlands,” ISPRS J. Photogramm. Remote Sens., vol. 33, pp. 223–230, Jul. 2000. [22] W. B. Cao, J. D. Liu, and R. Ma, “Regional planning of Xinjiang cotton growing areas for monitoring and recognition using remote sensing,” Trans. CSAE, vol. 24, no. 4, pp. 172–176, Apr. 2008. [23] K. McFeeters, “The use of normalized difference water index (NDWI) in the delineation of open water features,” Int. J. Remote Sens., vol. 17, no. 7, pp. 1425–1432, May 1996. [24] J. D. Tarpley, S. R. Schneider, and R. L. Money, “Global vegetation indexes from the NOAA-7 meteorological satellite,” J. Clim. Appl. Meteorol., vol. 23, no. 3, pp. 491–494, Mar. 1984. [25] J. R. G. Townshend, T. E. Goff, and C. J. Tucker, “Multitemporal dimensionality of images of normalized difference vegetation index at continental scales,” IEEE Trans. Geosci. Remote Sens., vol. GE-23, no. 6, pp. 888–895, Nov. 1985. [26] W. D. Rosental, B. J. Blanchard, and A. J. Blanchard, “Visible/infrared microwave visible/infrared microwave agriculture classification, biomass, and plant height algorithms,” IEEE Trans. Geosci. Remote Sens., vol. GE-23, no. 2, pp. 84–90, Mar. 1985. [27] D. S. Bartlett et al., “Continental scale variability in vegetation reflectance and its relationship to canopy morphology,” Int. J. Remote Sens., vol. 9, no. 7, pp. 1223–1241, Jul. 1988. [28] U. Kumar, N. Kerle, M. Punia, and T. V. Ramchandra, “Mining land cover information using multiplayer perception and decision tree from MODIS data,” J. Indian Soc. Remote Sens., vol. 38, no. 4, pp. 592–602, Dec. 2010. [29] A. Friedl and C. E. Brodley, “Decision tree classification of land cover from remotely sensed data,” Remote Sens. Environ., vol. 61, no. 3, pp. 399–409, Sep. 1997. [30] M. Xu, P. Watanachaturaporn, P. K. Varshney, and M. K. Arora, “Decision tree regression for soft classification of remote sensing data,” Remote Sens. Environ., vol. 97, no. 3, pp. 322–336, Aug. 2005. [31] M. Hansen, R. Dubayah, and R. Defries, “Classification trees: An alternative to traditional land cover classifiers,” Int. J. Remote Sens., vol. 17, no. 5, pp. 1075–1081, Mar. 1996. [32] M. Simard, S. S. Saatchi, and G. D. Grandi, “The use of decision tree and multiscale texture for classification of JERS-1 SAR data over tropical forest,” IEEE Trans. Geosci. Remote Sens., vol. 38, no. 5, pp. 2310–2321, Sep. 2000. [33] J. Cohen, “A coefficient of agreement for nominal scales,” Educ. Psychol. Meas., vol. 20, no. 1, pp. 37–46, Apr. 1960. Lizhen Lu received the M.Sc. and Ph.D. degrees in cartography and geographic information systems from Zhejiang University, Hangzhou, China, in 1996 and 2005, respectively. She is currently an Associate Professor with the Department of Earth Science, Zhejiang University. Her research interests include remote sensing image processing, feature extraction and classification, and mechanism of land use/cover change.
4558
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 7, NO. 11, NOVEMBER 2014
Liping Di (M’01–SM’06) received the Ph.D. degree in remote sensing/GIS (geography) from the University of Nebraska–Lincoln, Lincoln, NE, USA, in 1991. He is a Professor and the Founding Director of the Center for Spatial Information Science and Systems (CSISS) and a Professor with the Department of Geography and Geoinformation Science, George Mason University, Fairfax, VA, USA. He has engaged in geoinformatics and remote sensing research for more than 25 years and has published over 300 publications. He has served as the Principal Investigator (PI) for more than $34 million in research grants and as Co-PI for more than $8 million in research grants/contracts awarded by U.S. federal agencies and international organizations. His research interests include remote sensing standards, web-based geospatial information and knowledge systems, and remote sensing applications. Dr. Di has actively participated in the activities of a number of professional societies and international organizations, such as IEEE GRSS, ISPRS, CEOS, ISO TC 211, OGC, INCITS, and GEO. He served as the Cochair of the Data Archiving and Distribution Technical Committee (DAD TC) of IEEE GRSS from 2002 to 2005 and the Chair of DAD TC from 2005 to 2009. He currently Chairs INCITS/L1, a U.S. National Committee responsible for setting U.S. National Standards on geographic information and representing the US at ISO Technical Committee 211 (ISO TC 211).
Yanmei Ye received the M.Sc. degree in hydrogeology from Chengdu University of Technology, Chengdu, China, in 1988, and the Ph.D. degree in land resource management from Zhejiang University, Hangzhou, China, in 2002. Currently, she is a Professor with the Department Of Land Resource Management, Zhejiang University. Her research interests include mechanism of land use/ cover change and land use planning and management.