Mapping paddy rice planting area in wheat-rice double-cropped areas through integration of Landsat-8 OLI, MODIS and PALSAR images

Jie Wang, Xiangming Xiao, Yuanwei Qin, Jinwei Dong, Geli Zhang, Weili Kou, Cui Jin, Yuting Zhou, Yao Zhang
2015 Scientific Reports  
As farmland systems vary over space and time (season and year), accurate and updated maps of paddy rice are needed for studies of food security and environmental problems. We selected a wheat-rice double-cropped area from fragmented landscapes along the rural-urban complex (Jiangsu Province, China) and explored the potential utility of integrating time series optical images (Landsat-8, MODIS) and radar images (PALSAR) in mapping paddy rice planting areas. We first identified several main types
more » ... f non-cropland land cover and then identified paddy rice fields by selecting pixels that were inundated only during paddy rice flooding periods. These key temporal windows were determined based on MODIS Land Surface Temperature and vegetation indices. The resultant paddy rice map was evaluated using regions of interest (ROIs) drawn from multiple highresolution images, Google Earth, and in-situ cropland photos. The estimated overall accuracy and Kappa coefficient were 89.8% and 0.79, respectively. In comparison with the National Land Cover Data (China) from 2010, the resultant map better detected changes in the paddy rice fields and revealed more details about their distribution. These results demonstrate the efficacy of using images from multiple sources to generate paddy rice maps for two-crop rotation systems. Studies on paddy rice fields aim to provide direct or indirect information for researches on food security, water resource management, and environmental sustainability. Paddy rice fields provide one of the main staple foods for more than half of the world's population with 11% of cultivated land 1 . In Asia, the majority of rice agriculture relies on irrigation, accounting for 70% of global fresh water withdrawals 2 . Determining the area of paddy rice fields is an important component of obtaining more accurate information about agricultural water use to effectively manage fresh water resources 3 . In addition, as a kind of cultivated wetland, seasonally flooded paddy fields contribute 10-13% of the atmosphere's anthropogenic methane 4 . Meanwhile, paddy rice fields are changing at a breakneck pace due to dramatic encroachment by expanding cities 5 and the potentially reduced availability of water resources caused by climate change 6 . Therefore, it is urgently necessary to update and refine information about paddy rice planting areas in order to efficiently and accurately estimate crop production 7 , manage water resources 8,9 and monitor greenhouse gas emissions 10 . At the global and regional scales, several early studies of ecosystems and land cover have involved mapping paddy rice fields based on agricultural census data. In the late 1980s and early 1990s, several paddy rice datasets with coarse spatial resolution were produced to analyze global climate and greenhouse gas emission 11,12 . In the years following, two global cropland datasets representing crop patterns in the early 1990s and in the year 2000 were created at a spatial resolution of 5 arc minutes (~10 km) 13,14 . At the regional scale, a rice dataset for Asia was developed at the beginning of the 1980s 15 . In recent years, several studies on paddy rice planting areas were conducted by combining agricultural census data at the national scale 16, 17 . Although most of the crop datasets were produced with input from multiple sources, these datasets were developed mainly by relying on statistics with coarse spatial resolution that were supplied by administrative units. Given the limitations of this spatial and temporal information, it is a challenge to apply these datasets to finer spatial research and to update them year to year. Remote sensing is an efficient technique to acquire temporal and spatial cropland information repeatedly and consistently 18 . Historically, two main kinds of satellites have been used to map paddy rice fields: microwave and optical. Microwave satellites can penetrate through clouds and are thus superior for mapping paddy rice in regions dominated by long-term cloudy and rainy weather 19-22 , but available synthetic aperture radar (SAR) imagery is limited 19 or expensive 23 . Commonly used optical sensors are the Multispectral Scanner System/Thematic Mapper/Enhanced Thematic Mapper Plus (MSS/TM/ ETM+), the Moderate Resolution Imaging Spectroradiometer (MODIS) 24-29 , NOAA's Advanced Very High Resolution Radiometer (AVHRR) 30,31 , and SPOT High Resolution Geometrical/High Resolution Visible Infrared/VEGETATION (HRG/HRVIR/VGT) 32-34 . Owing to continuous archiving and free to acquire, MODIS and Landsat have been the prevalent data sources of mapping paddy rice fields over the last several years. Optical image classification techniques often use either individual image(s) or individual pixel(s) of time series data as input, here namely image-based methods and pixel-based methods. Image-based methods quantify the relationships (such as similarities or differences in spectra or texture) among all the pixels in an image for classification or object detection. These image-based methods have been applied to paddy rice mapping using single or multi-temporal optical images (e.g., MODIS and Landsat) at regional scales 23,35-37 . For image-based methods, collection of training samples (pixels) from ground reference data in each corresponding year remains a challenge 38 . Pixel-based methods rely primarily on the time series data for a pixel. These methods track the seasonal dynamics of a type of land cover and provide phenology information about the land surface. Several pixel-based algorithms have been developed to classify cropland using various optical images (e.g., MODIS, Landsat, SPOT, and FORMOSAT-2) 32,38-40 . To take into account the phenology of paddy rice, a pixel-and phenology-based algorithm using time series data of vegetation indices has been proposed. This was successfully applied to VGT and MODIS data for southern China and South and Southeast Asia 28,29 . Due to the high temporal resolution and continuous observation, MODIS data have frequently been combined with pixel-and phenology-based algorithms to track paddy rice phenology in order to map paddy rice areas or detect the intensity of paddy rice fields 24, 26, 27, [41] [42] [43] . However, MODIS-based phenology information cannot capture the sub-pixel dynamics of small paddy rice fields in heterogeneous and fragmented agricultural landscapes 26,44 . This could be improved by using Landsat images with 30-m spatial resolution. Nevertheless, more research is needed to document the combination of Landsat images with pixel-and phenology-based algorithms for paddy rice mapping, especially in the case of Landsat 8 OLI images. Reduced Landsat data availability caused by cloud cover or other problems may result in the failure of Landsat time series with 16-day intervals to distinguish the crops and trees. This problem is obvious in regions with double or multiple rotation agricultural systems, where cropland tends to be covered by plants year round, resulting in unavoidable confusion with natural evergreen forest. Fortunately, this problem can be solved by the Phased Array Type L-band Synthetic Aperture Radar (PALSAR) onboard the Advanced Land Observing Satellite (ALOS), given its capacity for forest detection 45, 46 . The current study, which involves mapping paddy rice planting areas at a 30-m spatial resolution, has two aims: (1) to develop a pixel-and phenology-based algorithm by integrating time series optical images (Landsat-8, MODIS) and radar images (PALSAR) to map paddy rice planting areas in wheat-rice agricultural systems; and (2) to evaluate the potential utility of Landsat-8 OLI data in identifying fragmented paddy rice fields in complex agricultural landscapes. The case study area is located at the Yangzi-Huaihe Plain, China, which is characterized by a two-crop rotation (wheat and rice) agricultural system and intermixture of rural and urban landscapes ( Figure S1 ). Accuracy assessment of resultant maps. We used ground truth data (field photos ( Figure S5 )), Google Earth (GE), and high-resolution images to locate and digitize ROIs. Google Earth displays high-resolution images, which have been used to validate land cover classification in several studies 57-59 . However, GE images were not enough to visually interpret ROIs as it lacked images within key time windows. We also ordered multiple high-resolution images from 2012 and 2013 from the NASA Goddard Space Flight Center, including WorldView-2 (WV2), OrbView5 (OV5), and QuickBird2 (QB2). According to the reference information, we generated a series of random sampling points and interpreted them into ROIs. In total, 15,751 Landsat-8 pixels were acquired, including 7,388 paddy rice pixels (173 ROIs) and 8,363 non-paddy rice pixels (427 ROIs) ( Figure S6 ). The accuracy of the paddy rice map produced in this study was assessed by using the "Ground Truth ROIs" method in ENVI software. We obtained a confusion matrix between the paddy rice map and the ROI data, and producer's accuracy, user's accuracy, overall accuracy, and the Kappa coefficient. We estimated the accuracy of a paddy rice map using only good quality observations during the flooding/ transplanting period. Finally, according to the cloud/cloud shadow masks from June 24 and July 10, we selected 3,610 paddy rice pixels (199 ROIs) and 3,113 non-paddy rice pixels (85 ROIs) located in good observation regions that were then used to validate the final paddy rice map. Comparison with other available datasets of paddy rice fields. We compared our results with NLCD2010 to analyze their differences and the dynamics of paddy rice fields. The paddy rice map was compared with NLCD2010 at two scales. At the regional level, we analyzed the variations of paddy rice fields in spatial distribution and planting areas. At the pixel level, correlation analysis was used to compare these two datasets in regard to the estimation of paddy rice planting areas.
doi:10.1038/srep10088 pmid:25965027 pmcid:PMC4428029 fatcat:vtqgystconbgjcl4iukgrggi7u