Climate-resilient Durum wheat selection using explainable machine learning and multi-environment trials

Introduction: Climate change poses a growing threat to durum wheat production in Mediterranean semi-arid regions, mainly through rising temperatures, frequent water shortages, and increased genotype × environment interactions. Methods: Over five consecutive growing seasons (2021-2025), ten durum wheat genotypes were evaluated across three rainfed environments in Algeria: Oued Smar, Sétif, and El Khroub. Agronomic performance, yield stability, and grain quality were assessed using ANOVA, AMMI, ASI, GGE biplots, drought tolerance indices (STI, DSI, and YSI), and supervised machine learning models coupled with SHAP-based interpretation. Results: Environment, genotype, and genotype × environment interaction had highly significant effects on grain yield. Yield differences exceeded 50% between sites. Waha and Skh/hau.heca-1 were identified as broadly adapted and stable genotypes, whereas Ammar-8, Da-6blak, and Lahaucan showed high yield potential but stronger environmental sensitivity. SHAP analysis indicated that precipitation and water deficit were the main drivers of yield variability, together explaining 50% of model importance, followed by grain number and soil plant-available water capacity. Discussion: The integration of stability analysis, drought tolerance indices, and explainable machine learning provides a robust framework for identifying climateresilient durum wheat genotypes and guiding targeted varietal deployment under Mediterranean rainfed conditions.