Joint use of time series Sentinel-1 and Sentinel-2 imagery for cotton field mapping in heterogeneous cultivated areas of Xinjiang, China

Abstract

Cotton is an important crop playing a key role in both economy and regional environment. In recent years, remote sensing has become the most feasible tool of crop field mapping in large-scale. This study evaluates the feature fusion of time series Sentinel-1 (S1) and Sentienl-2 (S2) data for cotton filed mapping in heterogeneous smallholder agricultural systems in Xinjiang, China. A SHP (Statistically Homogeneous Pixel) algorithm originally used for identification of distributed scatterers in Interferometric Synthetic Aperture Radar (InSAR) applications was implemented in de-speckling of SAR intensities. A semi-automated approach based on Jeffries-Matusita (J-M) distance and Recursive Feature Elimination (RFE) algorithm was used to select optimal combination of SAR or/and optical features in the cotton field mapping to achieve highest accuracy. In experiments, we demonstrated that feature fusion of Sentinel-1&2 is able to improve the cotton mapping accuracy.