A Digital Decision-Support Tool for Climate-Resilient Crop Planning: Design, Validation and Policy Implications

Abstract

Soil sensitive crop selection demands decision tools that not only relate soil conditions to the shifting temperature, humidity and rainfall, but also do not translate into agronomic prescriptions that lack support from the algorithms. The digital planning prototype for crops was designed and internally validated in this study based on a public benchmark, which has 2200 records, 7 soil–climate predictors and 22 balanced crop classes. Dummy, multinomial logistic regression, decision tree, random forest and extra trees classifiers were tested using a stratified 80:20 holdout and a 5-fold cross validation. Random Forest achieved the best cross-validated macro-F1 (0.994 ± 0.005), a test accuracy of 0.993, test macro-F1 of 0.993 and top three accuracy of 1.000. The most important factors were humidity, rainfall and potassium. Recommendation stability was then quantified by illustrative warming and drying and humidity-reduction perturbations. Combined stress resulted in 86.18% of top-1 recommendations being retained, and a mean top-3 overlap of 57.79%. The resulting architecture offers ranked recommendations, empirical-range warnings and scenario sensitivity as opposed to deterministic advice. The study provides a replicable process from benchmark modelling to decision support for policy and sets out clear guidelines for geospatial, temporal, economic and field validation prior to implementation.