Spatiotemporal Rainfall Variability and Integrated Multi-Model Forecasting in the Greater Upper Nile Region of South Sudan

Rainfall volatility heavily drives hydrological risk and agricultural productivity in South Sudan's Greater Upper Nile Region, yet severe data scarcity hampers quantitative forecasting. To overcome these barriers, this study introduces an integrated multi-model deep learning framework utilizing a Long Short-Term Memory (LSTM) network that achieves a dramatic performance boost over traditional linear benchmarks—delivering validation R2 scores up to 0.912 and significantly lowering baseline errors relative to standard SARIMA and SARIMAX approaches. This framework was built by extracting high-resolution NASA POWER reanalysis and CHRS PERSIANN satellite data across twelve localized stations. To ensure mathematical stability, multi-station Variance Inflation Factor (VIF) diagnostics were executed to eliminate multicollinearity, and Principal Component Analysis (PCA) was used to compress the remaining features into four orthogonal components capturing 85.7% of cumulative atmospheric variance. Unsupervised K-means clustering within the PCA space divided the twelve target stations into three macro-climatic clusters, verified by a 5-fold cross-validation Silhouette score of 0.3628. Four paradigms—Principal Component Regression (PCR), SARIMA, SARIMAX, and the deep learning LSTM network—were evaluated using a leakage-free 5-fold time-series forward-chaining protocol. The multi-variable LSTM network outperformed all other frameworks, consistently capturing highly non-linear, chaotic atmospheric interactions to yield the lowest overall error metrics (RMSE = 7.84 mm). A Diebold–Mariano test confirmed that the LSTM’s performance edge over traditional models was statistically significant (p < 0.01). Forward-looking projections (2026–2035), bounded by ±95% bootstrap prediction intervals, indicate a stable continuation of macro-monsoonal envelopes alongside substantial localized spatial volatility. These findings provide granular, data-driven climate intelligence essential for optimizing agricultural calendars, water infrastructure design, and early-warning systems in climate-vulnerable zones.