Lower agricultural drought predictability in canal-irrigated districts of Pakistan

Agricultural drought threatens food security in irrigated and rainfed regions worldwide, and early warning is especially difficult in large irrigated basins, where canal water partly decouples crop condition from local climate. Machine learning forecasts often claim multi-month skill, yet few are tested over multiple years against simple benchmarks, and few examine irrigation's effect on predictability. Using Pakistan's Indus Basin as a test case, this study forecasts the monthly Vegetation Health Index (VHI) at 1-, 3-, and 6-month leads for 94 districts of Sindh, Punjab, and Khyber Pakhtunkhwa (KP). Zone-stratified stacking ensembles combined Random Forest, XGBoost, LightGBM, and CatBoost with a Ridge meta-learner. Predictors were restricted to information available at the forecast origin, and skill was evaluated by rolling-origin validation over seven test years (2016–2022). Forecasts outperformed climatology only one month ahead, with a mean squared error skill score (MSESS) of 0.234 (95% confidence interval 0.151–0.302). This skill was positive in every test year, including the 2022 compound drought and flood, while skill at three and six months did not differ from climatology. Removing all VHI-derived predictors halved one-month skill without eliminating it. Shapley attribution showed vegetation persistence dominating at one month and soil moisture gaining weight at longer leads. VHI responded to the three-month Standardized Precipitation Evapotranspiration Index (SPEI-3) within the same month in 80 of 94 districts, and propagation timescale did not constrain predictability. Districts with more canal-irrigated cropland were less predictable: a one standard deviation increase in canal-irrigated fraction was associated with a 0.046 lower one-month MSESS, about 20% of mean skill. This association remained robust after controlling for agro-climatic factors, province effects, within-province permutation, and spatial dependence, whereas tubewell irrigation had a smaller, non-significant effect. The canal association arose mainly in the thermal component of VHI. Conformal intervals calibrated on preceding years achieved 91% coverage at one month, and a deterministic warning gave the best balance between detection and false alarms (Peirce skill score 0.29). Predicted VHI was associated with rice but not wheat yield anomalies. The framework provides district drought managers with a skillful one-month forecast and calibrated uncertainty, and points to canal supply data as the priority input for extending warning lead times in irrigated basins.