Automated rice mapping under diverse cropping patterns and establishment methods by integrating phenological knowledge and synergy of optical and SAR imagery

Abstract

Accurate and timely information on rice cultivation areas and cropping intensity is essential for precision crop management, food security and environmental sustainability. However, the generation of high-quality rice products is often hindered by the lack of ground truth samples, particularly in the regions with complex cropping patterns. While most rice mapping methods rely on the use of remotely sensed flooding signal from the transplanting period to distinguish rice from other crops, they face significant challenges from the increasingly prevalent direct-seeded rice for which the flooding signal at the beginning of season is weak due to the unique management measures. To address these issues and achieve the comprehensive extraction of rice under diverse cropping patterns and establishment methods, this study proposed the phenological knowledge-guided automatic rice mapping approach using optical and synthetic aperture radar (SAR) data (PHAROS). This method first determined the cropping intensity of rice and its concurrent crops using harmonized multi-resource Normalized Difference Vegetation Index (NDVI) time series and phenological knowledge. Subsequently, the Double Phase SAR Index (DPSI) was constructed to extract candidate rice samples by synthesizing the early and peak growth phases retrieved from NDVI time series and the corresponding SAR polarization features. Consequently, training data were generated automatically and fed into a machine learning classifier for rice mapping. The effectiveness of the PHAROS was evaluated over the Middle and Lower Reaches of the Yangtze River (MLRYR) of China and four other major rice production regions in Asia. Furthermore, the earliest timing of early, middle and late rice in the MLRYR was also quantified via the PHAROS and cross-year model transfer.

The results demonstrated that the PHAROS could identify rice of diverse cropping pattern with an overall accuracy (OA) from 0.965 to 0.976 in the MLRYR from 2019 to 2023. The classification maps exhibited well-delineated rice parcels and clear separation between single- and double-cropping rice. The PHAROS also yielded OA values of 0.902–0.979 and similar distribution pattern with the reference rice products in Suihua of China, Sataka of Japan, An Giang of Vietnam and Punjab of Pakistan, which represent diverse climatic conditions and cropping patterns across Asia. Compared to its counterpart methods, PHAROS demonstrated significant improvements by 0.022–0.235 in OA in rice planting area extraction and cropping intensity detection. The early, middle and late rice could be identified as early as at tillering, tillering-jointing and sowing/transplanting stages, respectively. This study reveals the necessity of handling the overlooked weak flooding signal from direct-seeded rice and offers a viable solution for large-scale rice cropping intensity detection and mapping under diverse establishment methods.