PHTF Select Phenology Aware Hurdle Foundation Modelling for Uncertainty Guided Crop Screening

Abstract

District-level crop screening is difficult when cultivation occurrence, conditional productivity, spatial dependence, and prediction uncertainty are modelled as a single task. We introduce PHTF-Select, a phenology-aware two-stage framework that first estimates recorded cultivation viability and then predicts productivity conditional on positive cultivated area. The analysis used 4,606 valid crop records from all 64 districts of Bangladesh; 4,190 positive-area records were retained for productivity modelling. Outcome-derived variables were excluded, phenology-window climate features were constructed within each training fold, and all models were evaluated with five district-held-out folds. TabICLv2 achieved the strongest classification performance (AUROC 0.956, balanced accuracy 0.895, MCC 0.621) and the lowest regression error (RMSLE 0.409), reducing district-level RMSLE by 0.021 relative to XGBoost and outperforming it in 73% of districts. A leakage control that restored Area and Production reduced RMSLE to 0.104, demonstrating how outcome-derived predictors can create misleading gains. Overall 90% interval coverage reached 0.907, but coverage fell to 0.798 in the highest-productivity quartile. Retaining the 60% least-uncertain predictions reduced RMSLE from 0.409 to 0.321, at the cost of abstaining on 40% of cases. Phenology features improved RMSLE by approximately 3%, with post-transplant rainfall producing the largest ablation loss. PHTF-Select therefore provides a reproducible framework for retrospective crop screening, while its seasonal-transfer failures and omitted agronomic variables preclude autonomous planting recommendations.