Deep Learning for Short-Term Precipitation Prediction in Four Major Indian Cities: A ConvLSTM Approach with Explainable AI

Explainable AI makes weather prediction more trustworthy: The paper demonstrates that deep learning models for precipitation forecasting don't have to remain "black boxes." By applying interpretability techniques like Grad-CAM and permutation importance, the researchers revealed which specific weather variables the model actually uses to make predictions, making the system more transparent for meteorologists and policymakers who need to trust AI recommendations.

Different cities require different AI approaches: The study found that optimal model performance varied significantly by location—Bengaluru's model used 32 convolutional filters while Kolkata's required 128—and prediction accuracy ranged from 0.21 mm/day error in Bengaluru to 1.80 mm/day in Kolkata. This highlights that responsible AI deployment requires customization rather than one-size-fits-all solutions, especially across diverse climate zones.

Prediction reliability depends on local climate patterns: The models achieved different prediction horizons for each city (one day ahead for Bengaluru versus five days for Kolkata), suggesting that AI systems must be calibrated to reflect real-world complexity. This finding is crucial for responsible deployment—users need to understand that the same model cannot provide equally reliable forecasts everywhere.

Interpretability techniques reveal model reasoning: By using multiple explanation methods (temporal occlusion, counterfactual perturbation, and activation mapping), the researchers identified city-specific patterns in how the model makes decisions. This multi-method approach to explainability provides stronger evidence that the AI is learning genuine meteorological relationships rather than spurious correlations.