Synergistic active and passive remote sensing with multi-feature optimization for winter wheat identification in oasis agriculture of Arid Desert Regions.

Abstract

Accurate mapping of winter wheat is essential for food security, agricultural management, and climate adaptation strategies. We present a multi-source, phenology-aware approach for identifying winter wheat in the arid oasis agricultural landscape of Changji City (northern Xinjiang) using Sentinel-1 C-band SAR and Sentinel-2 surface reflectance imagery acquired during key growth stages in 2023. We constructed a comprehensive six-category feature set-polarimetric, textural, spectral, vegetation indices, normalized-difference indices, and temporal-difference features-and designed eleven feature-combination schemes for comparative evaluation. A Random Forest-based forward sequential feature selection (RF-FS) procedure was used to identify the most informative predictors. From an initial 171 features, RF-FS selected 27 key variables (an 84.4% reduction), dominated by SWIR bands (B11/B12), a phenology index (NDPI), and SAR polarimetric combinations; moisture-sensitive features comprised 40.7% of the selected subset. The optimized model achieved an overall accuracy (OA) of 94.60% and a Kappa coefficient of 0.9338 (producer's accuracy [PA] = 100%). Temporal analysis indicates that April and May contributed the most discriminative information, corresponding to rapid biomass accumulation and the jointing-heading stages. Applying the optimized classifier yielded an estimated 2023 winter wheat area of approximately 93.39 km2 for Changji, with a relative error of 7.79% compared to statistical records. The results validate a "multi-source synergy, phenology-driven, moisture-sensitive" framework for efficient winter wheat mapping in arid desert-oasis mosaics and provide an operational basis for crop structure monitoring and agricultural risk assessment.