Mapping rice crops in cloud-covered mountainous areas using Sentinel-1 radar and random forest algorithm: a case study in Lao Cai, Vietnam

Abstract

Rice mapping in mountainous regions remains challenging because persistent monsoon cloud cover severely limits the availability of optical satellite imagery. In Lao Cai Province, northwestern Vietnam, rice monitoring is further complicated by steep terrain, fragmented agricultural fields, and heterogeneous cultivation systems dominated by smallholder terraced rice. This study developed a Sentinel-1 synthetic aperture radar (SAR) time-series framework within the Google Earth Engine platform to identify the most effective variables for operational rice mapping under cloud-prone and topographically complex conditions. A total of 22 candidate variables, including temporal backscatter features, phenology-based growth indicators, polarimetric indices, and topographic variables, were evaluated using a Random Forest classifier trained with 1639 reference samples. The classification achieved an Overall Accuracy of 91.4% and a Kappa coefficient of 0.83. The estimated rice area was 30 752 ha, corresponding to a relative error of 6.8% compared with official agricultural statistics. Notably, temporal Vertical–Horizontal backscatter dynamics during the tillering and harvest stages provided the strongest discriminative capability for rice detection, substantially outperforming conventional polarimetric indices. Furthermore, temporal SAR phenology is highly effective for detecting fragmented rice fields in mountainous environments where optical remote sensing is frequently obstructed by clouds.