An enhanced pixel-based phenological feature for accurate paddy rice mapping with Sentinel-2 imagery in Google Earth Engine

Abstract

Accurate paddy rice mapping with remote sensing at a regional scale plays critical roles in agriculture and ecology. Previous studies mainly employed a single key phenological period (i.e., transplanting) for paddy rice mapping. However, the prominentpoor spectral separability between paddy rice and others (e.g., wetland vegetation) exists in this period. To this end, we developed an enhanced pixel-based phenological feature composite method (Eppf-CM). Subsequently, the feature derived from Eppf-CM was served as the input data to a one-class classifier (One-Class Support Vector Machine, OCSVM). Eppf-CM includes two steps: (1) four distinctive phenological periods, specifically designed for rice mapping, were identified by time-series analysis of Sentinel-2 imagery. (2) We strived to choose one or two vegetation indices for each phenological period, and then stacking all the indices together. The new developed paddy rice mapping method with Eppf-CM and OCSVM is low costs and high precision. To fully demonstrate the outstanding precision of Eppf-CM based paddy rice map (Eppf map) in this study, three different sources of reference data were employed for comparison purposes. Compared with the field survey data, Eppf map achieved an overall accuracy higher than 0.98. The paddy rice area in Northeast China from Eppf map is only 1.86% less than that of the National Bureau of Statistics in 2019. Compared with a latest paddy rice map at the same spatial resolution (10-m), Eppf map significantly reduced commission and omission errors. To the best of our knowledge, the Eppf-CM has obtained one of the highest accuracy rice maps in Northeast China up-to-date. As a whole, we expect that: (1) Eppf-CM will advance the phenology-based agricultural remote sensing mapping method. (2) The paddy rice map will provide a new baseline data for the study of agriculture and ecology.