Explainable, Uncertainty‐Aware Machine Learning for Maize Land Suitability and Climate Resilience in Lorestan Province, Western Iran

Abstract

Maize ( Zea mays L.) is strategically important in Iran, yet its cultivation in the semi‐arid Zagros is constrained by limited and variable water, heat stress and heterogeneous soils, while conventional suitability assessments are deterministic, rarely validated spatially and provide no measure of confidence. The aim of this study was to develop an explainable, uncertainty‐aware machine‐learning framework for maize land‐suitability mapping that is interpretable, spatially robust and climate‐aware and to demonstrate it in Lorestan Province (≈28,300 km 2 ), western Iran. The experimental material comprised harmonised multi‐source geospatial predictors—soil properties (SoilGrids 2.0), climate variables (Climate Hazards Group InfraRed Precipitation with Station data, CHIRPS; TerraClimate), moderate resolution imaging spectroradiometer (MODIS) vegetation indices and shuttle radar topography mission (SRTM) terrain—together with 520 reference sites labelled by applying the Food and Agriculture Organization (FAO) maximum‐limitation land‐evaluation procedure to explicitly documented land qualities. As the method of investigation, six classifiers (Elastic‐Net, support vector machine, multilayer perceptron, random forest, LightGBM and XGBoost) were trained and evaluated under spatial block cross‐validation; predictions were interpreted with SHapley Additive exPlanations (SHAP) and uncertainty was quantified using predictive entropy and between‐model ensemble variance and then validated against observed classification error. Data collection drew on openly available gridded products resampled to a 250 m grid, with Coupled Model Intercomparison Project Phase 6 (CMIP6) projections under two shared socioeconomic pathways (SSPs) used for climate scenarios. Gradient‐boosting models performed best, achieving a macro‐ F 1 ‐score of 0.79 under spatial block cross‐validation against 0.87 under random cross‐validation, confirming that random partitioning inflates apparent skill. Growing‐season precipitation, aridity, heat‐stress frequency, vegetation condition and root‐zone soil properties dominated, with pronounced thresholds; predictive entropy was inversely related to classification accuracy, confirming that the uncertainty surface is diagnostic of error rather than merely descriptive and the most suitable land was not necessarily the most climate‐resilient. We conclude that the framework offers a transparent, probabilistic and transferable alternative to deterministic suitability mapping for climate‐smart dryland agricultural planning.