Satellite-based Crop Discrimination with Machine Learning on GEE Platform: Insights from Udham Singh Nagar

Abstract

Crop discrimination is crucial for environmental monitoring, agricultural planning, and sustainable development. This study assessed the performance of optical (Sentinel-2, 10 m, atmospherically corrected to surface reflectance) and microwave (Sentinel-1, 10 m, preprocessed with radiometric calibration and speckle filtering) remote sensing data for crop classification in Udham Singh Nagar district, Uttarakhand, India, during the June–October 2023 kharif season. Ground truth data for major crops, namely rice (815 samples) and sugarcane (62 samples), were collected through field surveys and split into 70% training and 30% validation subsets to ensure robust model evaluation. Five machine-learning classifiers Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Classification and Regression Tree (CART), and Gradient Boosted Machine (GBM) were applied to individual and fused datasets. GBM consistently achieved the highest classification accuracy at the monthly scale, likely due to its sequential error-correction mechanism that effectively exploits distinct phenological patterns captured in monthly temporal composites, while RF produced the highest overall accuracy (89.21%) for the season-long fused optical and microwave dataset. SVM and KNN showed comparatively lower performance, especially during transitional crop growth stages. The results highlight the effectiveness of ensemble learning methods and demonstrate the benefit of multi-sensor data fusion for accurate and reliable crop discrimination and land use/land cover mapping.