Paddy rice field mapping using GF-1 images with SVM method

Paddy rice is one of the main grain crops in China. Accurate evaluation of rice planted area and spatial distribution are significant for the estimation ofgrain output and cropland allocation. Meanwhile, paddy rice is a main contributor to greenhouse gas emissions and therefore its cultivation and spatial distribution have significant implications for rising temperatures and global climate change. In southeast China, even though paddy rice is widely cultivated, the paddy rice field is highly fragmented by other land use/cover. The difficulty in mapping small and fragmented rice fields is further exacerbated by the cloudy weather conditions in this area at periods corresponding to the rice growing season which impedes the utility of optical images. As optical data possess important spectral information which cannot be provided by satellite sensors (radar) least affected by weather conditions, much attention should be paid to the optimal use of optical data. GF-1 is an optical satellite launched in China on April 26, 2013 and has been providing remotely-sensed images with multispectral data of relatively high spatial and temporal resolution. Despite its better resolution and freely available 16 m products, the application of GF-1 images in agricultural land-cover studies, especially in paddy rice mapping, is not as popular as other optical products such as Landsat, MODIS, SPOT and so on. In this study, we have employed multi-temporal images of GF-1 covering the 2016 rice growing season to map paddy rice cultivated fields in Tongxiang County, Zhejiang Province, China. The radiometric and geometric corrections were first carried out for multiple temporal GF-1 images. These corrected images were layer-stacked to enable time series analysis. Since the two-band enhanced vegetation (EVI2) is changing with paddy rice growth stages, we extracted and smoothed the EVI2 temporal series from GF-1 images using the Savitzky-Golay filter (S_G filter) approach to further reduce the influence of atmospheric effects from clouds and water vapor. The Support Vector Machine (SVM) classification algorithm was then introduced on the temporal and filtered GF-1 images to map paddy rice fields in the area investigated. Classification results show two major types of rice crop planting patterns in Tongxiang County: single-cropped rice fields and paddy rice-upland crop rotation fields. The former cropping pattern is more uniformly distributed in Tongxiang County, whereas the latter mainly occupies the western and northern parts of the County. From the error matrix, overall classification accuracy, user's accuracy and producer's accuracy are the approximately 87.68%, 88.8% and 90.3%, respectively. These results indicate the promising potential of GF-1 temporal images in extracting paddy rice cultivated areas with higher accuracy in irregular and highly fragmented paddy rice cultivated areas.