Exploiting Sentinel-1 data and machine learning–based random forest for collectively mapping rice fields in Taiwan

Abstract

Rice is the most important crop in Taiwan. Monitoring rice-growing areas is thus essential for crop management and food decision-making processes. This research aims to develop an approach for seasonally mapping rice areas from time-series Sentinel-1 data in Taiwan. The data were processed for 2019 and 2020 rice cropping seasons, following three main steps: (1) data pre-processing to construct smooth time-series satellite data, (2) rice area estimation using random forests (RF), and (3) accuracy assessment. The mapping results compared with the government’s reference data showed overall accuracy and kappa coefficient higher than 87.7% and 0.76, respectively. The rice area estimates at the county level well agreed with the official statistics, with the root mean square error (RMSE) in percentage smaller than 19.7%. An examination of changes in cropping areas between 2019 and 2020 showed a noticeable reduction of rice areas in 2020, mainly attributed to severe drought conditions.