We present a machine learning framework for climate change attribution and hotspot identification, with emphasis on testing geographic transferability through spatial cross-validation. We evaluate eight ML models for explaining temperature anomalies in Sindh, Pakistan on 44 years of observations. We performed spatial transferability testing using Leave-One-District-Out Cross-Validation (LODO-CV) to assess whether learned relationships generalize to districts not seen during training. Gradient Boosting was the most successful algorithm with (\(R^2 = 0.914\pm 0.098\)) under LODO-CV, which indicates robust transferability across the region’s diverse climatic zones. The modest difference between LODO-CV and random cross-validation (\(\Delta R^2 \approx 0.02\)) is consistent with findings that temporal structure often dominates spatial structure in climate time series1; we thus interpret LODO-CV as a complementary robustness check rather than a correction for inflated performance. SHAP feature attribution showed that climate variables (37.6%), temporal trends (32.0%), and anthropogenic proxies (23.7%), are the most important predictors, although it is also important to note the caveat that the importance of proxies is only indicative of correlation, not causation, and must be carefully considered when applying to policy matters. A sensitivity analysis excluding all temperature-derived predictors confirms that the model achieves \(R^2 = 0.914\) using only temporal, geographic, and anthropogenic features, ruling out feature-target circularity. Temporally blocked cross-validation (train \(\le\)2015, test 2016–2024) yields \(R^2 = 0.445\), honestly demonstrating that temporal extrapolation is substantially more challenging than spatial interpolation—a well-documented limitation of statistical climate models. By using a dual-index model, which integrates the frequency of extreme events with average climate changes, we were able to pinpoint seven hotspots of climate change, concentrated in Karachi and Hyderabad urban areas, which are exposed to compound risk of urbanization, coastal exposure, and rising temperature extremes. We emphasize that our framework addresses attribution—understanding which factors correlate with observed temperature variability—rather than operational forecasting, which appropriately relies on numerical weather prediction models. The results demonstrate the value of spatially explicit validation procedures for confirming geographic transferability of climate ML models and offer practical suggestions to specific adaptation planning for the most climate-prone regions.
