Abstract
Agricultural systems in semi-arid regions are increasingly exposed to climate variability, yet the drivers of seasonal vulnerability remain insufficiently understood. This study examines whether perceived vulnerability during winter and summer cropping seasons is shaped by distinct mechanisms—climatic exposure during the wet season and adaptive capacity during the dry season. Using Morocco as a representative case for North African agriculture, we integrate long-term Earth observation indicators of environmental variability (precipitation, temperature, and NDVI) with survey data from 3,591 smallholder farmers and apply machine learning classification models evaluated under spatially explicit cross-validation. Model performance was benchmarked against multinomial logistic regression and majority-class baselines. While standard cross-validation yielded optimistic estimates, spatial cross-validation produced more conservative and policy-relevant results. Under spatial validation, Random Forest and XGBoost consistently outperformed simpler models, with macro F1 scores ranging from approximately 0.53 for overall vulnerability to over 0.70 for seasonal outcomes. Comparative experiments showed that survey-based predictors explain a larger share of perceived vulnerability than Earth observation indicators alone, while their combination provides complementary gains, particularly for winter vulnerability. Explainable model analysis revealed clear seasonal contrasts: winter vulnerability was dominated by hydroclimatic variability and soil moisture conditions, whereas summer vulnerability was more strongly shaped by adaptive capacity, including groundwater access, irrigation practices, and climate information use. The proposed framework offers a transferable approach for assessing climate vulnerability and informing targeted adaptation strategies in semi-arid farming systems.