An Integrated Optical and SAR Multi-Index Approach for Crop Discrimination Using Optimized Random Forest

Accurately distinguishing multiple crop types in diverse agricultural landscapes is essential for effective crop monitoring, food security assessments, and sustainable land management. However, this task remains challenging in mixed and smallholder farming systems, where crops often share overlapping phenological stages, and optical observations are frequently obstructed by persistent cloud cover. To address these challenges, this study developed a unified optical–SAR multi-index framework that integrates complementary multispectral and radar features with statistical separability analysis and an optimized Random Forest (RF) classifier for robust multi-crop mapping. Multi-temporal Sentinel-2 surface reflectance data were used to derive vegetation greenness, red-edge, moisture, stress, and disturbance indices that capture crop physiological conditions and phenological dynamics. Concurrently, Sentinel-1 SAR VV and VH backscatter coefficients and polarization-based indices were extracted to quantify the canopy structure, surface roughness, and moisture variability under cloud-prone conditions. These optical and SAR predictors were fused into a high-dimensional feature stack and evaluated using the Bhattacharyya and Jeffries–Matusita distance metrics to quantify pairwise inter-class separability and identify informative, non-redundant features, particularly for spectrally similar crops. Multi-crop classification was performed using an RF classifier with hyperparameter optimization and feature importance–driven top-N predictor selection to reduce dimensionality and improve model generalization. Post-classification refinement using majority filtering and connected pixel cleaning enhanced the spatial coherence of the output maps. The framework was implemented in the coastal agricultural region of Cuddalore District, Tamil Nadu, India, during the Rabi season (2024–2025). The optimized model achieved an overall classification accuracy of approximately 0.81 and produced realistic crop area estimates that were consistent with regional agricultural patterns. The results indicate that Sentinel-2 red-edge and greenness indices were the dominant predictors, whereas Sentinel-1 SAR features provided complementary structural and moisture information that strengthened the discrimination among spectrally ambiguous crops. The framework successfully mapped rice, cotton, maize, groundnut, sugarcane, coconut, cashew, and casuarina, demonstrating a scalable and operational solution for multi-crop mapping in complex and cloud-prone agricultural landscapes.