Variance-based sensitivity analysis of climate variability impact on crop yield using machine learning: A case study in Jordan

Abstract

Climate variability poses a significant threat to crop production in arid and semi-arid regions, where droughts are becoming more frequent and intense. This study employs a variance-based sensitivity analysis combined with machine learning to assess the impact of climate variability on crop yield in Jordan, a water-scarce country with a declining agricultural sector. Using meteorological, environmental, and demographic datasets, we predict the yields of four major crops – wheat, barley, date palm, and olive – and evaluate the relative importance of input variables, including drought indices, using the stratified first-order Sobol’ index. Machine learning models, particularly eXtreme Gradient Boosting, outperformed traditional methods, achieving out-of-sample $R^2$ values of 0.79 for wheat, 0.92 for date palm, 0.83 for olive, and 0.48 for barley yield prediction. Our sensitivity analysis reveals that barley exhibits greater resilience to climate variability, with climate-related variables explaining only 20% of its yield variance. In contrast, wheat is highly vulnerable to prolonged, low-intensity droughts, with a long-term precipitation index accounting for 36% to its yield variance, while short-term climate variables explaining 49% of the remaining variability. Date palm and olive yields are more sensitive to short-term, high-magnitude droughts, with short-term precipitation indices explaining 35% and 44% of their variance, respectively. These findings can help inform policies that optimize water allocation, prioritize drought-resilient crops, and implement targeted strategies to enhance agricultural resilience in Jordan. By leveraging public remote sensing data and advanced sensitivity analysis methods, this approach can be adapted to other data-scarce regions to support food security and sustainable agricultural management.