Field-scale irrigation information is needed to quantify agricultural water use and support water accounting in sub-humid regions, where irrigation records are sparse and rainfall can obscure irrigation signals. This study developed a crop-specific, two-stage machine-learning framework to classify irrigated versus rainfed corn, cotton, and soybean fields in West Tennessee during the 2023 and 2024 growing seasons. Time-series predictors were derived from Harmonized Landsat and Sentinel-2 (HLS) optical surface reflectance and Landsat thermal infrared brightness temperature (TIRS1 and TIRS2). LightGBM (Light Gradient Boosting Machine) models were evaluated using five-fold cross-validation and an independent test set. To emulate an operational setting without an external crop map, crop type was first predicted and then used to select the corresponding crop-specific irrigation classifier. Test accuracies ranged from 86% to 93% (Cohen's kappa = 0.57–0.83), with consistently strong performance for corn. Soybean performance was weaker in 2023 but improved in 2024. Feature-importance analysis showed that near-infrared and shortwave infrared predictors were consistently influential across crops and years. Thermal predictors, particularly TIRS1 (with additional gains from TIRS2 in 2024), were among the most influential variables, indicating that irrigation signals were captured through combined effects on canopy vigor, moisture status, and land surface temperature. Overall, the two-stage workflow achieved up to 93% accuracy, supporting operational field-scale irrigation mapping in sub-humid regions.
Machekposhti et al. (Mon,) studied this question.