Automatic sar-based rapeseed mapping in all terrain and weather conditions using dual-aspect Sentinel-1 time series

Abstract

Timely and reliable rapeseed mapping is crucial for vegetable oil supply and bioenergy industry. Synthetic Aperture Radar (SAR) remote sensing is able to track rapeseed phenology and map rapeseed fields in cloudy regions. However, SAR-based rapeseed mapping is challenging in mountainous areas due to the highly fragmented farming land and terrain-induced distortions on SAR signals. To address this challenge, this study proposed a novel SAR-based automatic rapeseed mapping (SARM) method for all terrain and weather conditions. SARM first composites high-quality dual-aspect Sentinel-1 time series by combining ascending and descending orbits and smoothing temporal noises. Second, SARM embeds a novel terrain-adjustment modeling to mitigate confounding terrain effects on the SAR intensity of sloped pixels. Third, SARM quantifies unique shape and intensity features of SAR signals during the leaf-flower-pod period to estimate the probability of rapeseed cultivation with the aid of automatically extracted local high-confidence rapeseed pixels. SARM was tested at three sites with varying topographic conditions, rapeseed phenology and cultivation systems. Results demonstrate that SARM achieved accurate rapeseed mapping with the overall accuracy 0.9 or higher, and F1 score 0.85 or higher at all three sites. Compared with the existing rapeseed mapping methods, SARM excelled in mapping fragmented rapeseed fields in both flat and sloped terrains. SARM utilizes unique and universal SAR time-series features of rapeseed growth without relying on any prior knowledge or pre-collected training samples, making it flexible and robust for cross-regional rapeseed mapping, especially for cloudy and mountainous regions where optical data is often contaminated by clouds during rapeseed growing stages.