A hybrid ensemble deep learning model for advanced time series rainfall forecasting using satellite data and climate variability analysis

Accurate rainfall prediction is important for climate adaptation, managing water resources, and planning for farming in dry areas and places where data is difficult to obtain. By collecting long-term data on precipitation from satellite sources Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) between the years of 2000 and 2024, the study was able to create an uncertainty-aware ensemble deep learning framework for predicting monthly rainfall amounts. The study focused on Northern Iraq, which is a semi-arid and climate-sensitive area that has very few ground observation stations. The ensemble framework combines three different types of recurrent neural networks (LSTM, GRU, and hybrid bidirectional LSTM/GRU) using a stacking methodology with a ridge regression algorithm as the meta-learner. The incorporation of advanced feature engineering techniques (e.g., rolling statistics, anomaly detection, and seasonal encoding) enables capturing the dynamic information contained in rainfall on both short- and long-term timescales. Results from experiments using time-series cross-validation support the claim that stacking ensemble models perform consistently better than individual models (with highly accurate predictions [R2 ≈ 0.99] and minimal errors [MAE ≈ 2.3 mm]). Probabilistic forecasts for the period of 2025-2028 have been developed through the application of the Monte Carlo Dropout technique, allowing for explicit quantification of uncertainty associated with projections of rainfall decline and increased uncertainty. Thus, the proposed approach will provide an effective, reproducible, and scalable method for satellite-based rainfall forecasting and support decisions on climate adaptation and environmental policies in arid regions with insufficient data to develop traditional, ground-level forecasting techniques.