A Machine Learning-Driven Framework for Comparative Multi-Algorithm Classification and Spatial Suitability Modelling of Saffron Cultivation in Pampore Karewas

Abstract. The study presents a comprehensive geospatial and machine learning-based framework for mapping, evaluating, and forecasting the spatial suitability of saffron cultivation in the Pampore Karewas of Jammu & Kashmir. Sentinel-2 imagery and ground-truth data were used to implement three supervised classifiers—Gradient Boosting, Random Forest, and Support Vector Machine (SVM)—to accurately detect saffron fields. Among these, Gradient Boosting demonstrated the highest classification performance, while Random Forest provided a balanced accuracy across classes. An Analytic Hierarchy Process (AHP)-based multi-criteria decision model, incorporating nine agro-environmental parameters, was developed to generate a saffron site suitability analysis for 2024. This was further integrated with classification outputs to delineate spatial priority zones for cultivation. To evaluate near-future viability, a climate-adjusted suitability model was constructed using ERA5 and CHIRPS datasets, incorporating saffron-specific temperature and rainfall thresholds through a rule-based penalty approach. Results indicate a projected 6.76% net reduction in highly suitable areas due to changing climatic conditions. While some traditional core production zones may face declining suitability, the analysis also highlights emerging potential in transitional zones, suggesting both challenges and new opportunities for adaptive saffron cultivation. The proposed methodology offers a scalable and replicable decision-support tool for climate-resilient agricultural planning and sustainable saffron farming under evolving environmental conditions.