دانشگاه محقق اردبیلی مدل سازی و مدیریت آب و خاک 2783-2546 6 2 2026 05 22 Optimizing machine learning to reduce crop classification uncertainty in semi-arid Canal command areas Optimizing machine learning to reduce crop classification uncertainty in semi-arid Canal command areas 303 325 4546 10.22098/mmws.2026.18944.1745 FA Mohansing Rajaput Ph.D. Scholar, Department of Water Resources and Ocean Engineering, National Institute of Technology Karnataka, Surathkal, Mangaluru-575025, Karnataka, India Abhilash Ramadasa Scientist ‘C’, National Institute of Hydrology, Hard Rock Regional Centre, Visvesvaraya Nagar, Belagavi-590019, Karnataka, India 0000-0002-7505-5263 Basavanand M. Dodamani Professor, Department of Water Resources and Ocean Engineering, National Institute of Technology Karnataka, Surathkal, Mangaluru-575025, Karnataka, India Journal Article 2025 12 05 Effective water resource management and canal performance analysis in semi-arid regions are fundamentally reliant on precise estimates of crop water demand, which are typically derived from accurate, up-to-date crop inventories. For command areas in semi-arid India, this vital information is a primary input for agro-hydrological models used to assess irrigation efficiency and plan water allocation. However, the inherent complexity of these landscapes characterized by small, fragmented landholdings introduces substantial uncertainty into remote sensing based crop classification, threatening the reliability of subsequent management decisions. This study systematically addresses this input uncertainty by performing a comprehensive, multi-factorial sensitivity analysis using multi-temporal Sentinel-1 (SAR) and Sentinel-2 (optical) data. We investigated the combined effects of four multi-sensor data fusion strategies, six Machine Learning (ML) classifiers, three feature selection techniques, and five training/testing data split ratios. The findings of the study provide crucial operational insights for modelers and managers. The synergistic fusion of Sentinel-1 and Sentinel-2 data was identified as the single most critical factor for achieving high accuracy. Furthermore, classification performance showed high sensitivity to training data volume, with an optimal threshold observed at an 80/20 train/test split. The Extreme Gradient Boosting (XGBoost) classifier, coupled with Backward Elimination feature selection, emerged as the superior strategy, achieving a maximum overall accuracy of 98.4%. By identifying this optimized workflow, this research provides a robust and scalable method for generating highly reliable spatial input data, thereby minimizing uncertainty in crop water requirement calculations and significantly enhancing the predictive capacity and practical utility of agro-hydrological models for sustainable water management. Effective water resource management and canal performance analysis in semi-arid regions are fundamentally reliant on precise estimates of crop water demand, which are typically derived from accurate, up-to-date crop inventories. For command areas in semi-arid India, this vital information is a primary input for agro-hydrological models used to assess irrigation efficiency and plan water allocation. However, the inherent complexity of these landscapes characterized by small, fragmented landholdings introduces substantial uncertainty into remote sensing based crop classification, threatening the reliability of subsequent management decisions. This study systematically addresses this input uncertainty by performing a comprehensive, multi-factorial sensitivity analysis using multi-temporal Sentinel-1 (SAR) and Sentinel-2 (optical) data. We investigated the combined effects of four multi-sensor data fusion strategies, six Machine Learning (ML) classifiers, three feature selection techniques, and five training/testing data split ratios. The findings of the study provide crucial operational insights for modelers and managers. The synergistic fusion of Sentinel-1 and Sentinel-2 data was identified as the single most critical factor for achieving high accuracy. Furthermore, classification performance showed high sensitivity to training data volume, with an optimal threshold observed at an 80/20 train/test split. The Extreme Gradient Boosting (XGBoost) classifier, coupled with Backward Elimination feature selection, emerged as the superior strategy, achieving a maximum overall accuracy of 98.4%. By identifying this optimized workflow, this research provides a robust and scalable method for generating highly reliable spatial input data, thereby minimizing uncertainty in crop water requirement calculations and significantly enhancing the predictive capacity and practical utility of agro-hydrological models for sustainable water management. https://mmws.uma.ac.ir/article_4546_eafde0e1d6610d18ea992a70862ed7a6.pdf