Most precision-agriculture systems that rely on soil and climate sensor data feed those raw readings straight into a machine learning model, leaving the model to work out on its own how temperature, moisture and nutrient levels combine into something resembling environmental stress. We take a different approach here. Working with the standard multimodal inputs used in soil health monitoring — nitrogen, phosphorus, potassium, temperature, humidity, pH and rainfall — we introduce an engineered feature we call the Adaptive Climate Stress Index (ACSI), computed before any classification takes place. ACSI folds four climate-sensitive variables together with an NPK-balance term into a single stress score bounded between 0 and 1, using a deviation-from-ideal normalization scheme. That score, combined with the raw sensor features, feeds into a continuous 0–100 Soil Health Score and a three-tier Climate Risk Level, both of which we predict using Random Forest and Decision Tree classifiers and then interpret with SHAP (SHapley Additive exPlanations). Tested on a public crop-recommendation dataset of 2,200 samples, the pipeline reaches up to 97.95% accuracy for Soil Health classification and 97.73% for Climate Risk classification, and SHAP analysis confirms that ACSI is, by a wide margin, the most influential feature driving the climate risk predictions. What this shows is that a fairly simple, domain-motivated index can turn climate stress into something explicit, learnable and explainable — without needing any sensors beyond what standard agricultural datasets already provide.