MULTI-SENSOR INTEGRATION OF ENMAP HYPERSPECTRAL, MULTISPECTRAL, AND SAR DATA FOR CROP MONITORING IN IRRIGATED AGRICULTURE: A MACHINE LEARNING COMPARATIVE STUDY

Accurate crop monitoring in Central Asia requires overcoming challenges posed by cloud cover, spectral similarity between crops, and the computational constraints of processing large satellite archives. This study investigates a multi-sensor approach combining EnMAP hyperspectral imagery, Sentinel-2 multispectral time series, and Sentinel-1 SAR data for cotton and wheat classification in the Tashkent region, Uzbekistan. Machine learning classifiers including Support Vector Machine (SVM), Random Forest (RF), CART, and Gradient Boosting were applied to each sensor type independently and in fusion. Adding climate variables (CHIRPS precipitation and ERA5 evapotranspiration) to multispectral inputs increased overall accuracy by 8–10%, achieving OA > 90% and Kappa > 0.85. Monthly NDVI time series for 2023 confirmed phenological separability between crops: cotton peaks in July–August and wheat in April–May. SAR-only classification (Sentinel-1 VV and VH) achieved OA = 63% with Random Forest, demonstrating cloud-independent monitoring capability. EnMAP data yielded the highest classification accuracy (OA = 95%, K = 0.91) across all sensor configurations, confirming the superiority of narrow-band hyperspectral data for fine-grained crop discrimination. The integrated multi-sensor approach raised overall accuracy above 90% across all classifiers.