Enhancing Streamflow Prediction in Vulnerable Regions through Probabilistic Deep Learning and Satellite-Derived Data

Accurate and timely prediction of streamflow is critical for managing the increasing risks associated with floods, particularly in developing countries where traditional in-situ monitoring systems are often sparse or non-existent. This study introduces a novel probabilistic multi-step ahead prediction model that leverages Graph Neural Networks (GNNs), self-attention mechanisms via the Informer network, and a distributional output layer to enhance the predictive accuracy and uncertainty quantification of streamflow time series. By integrating satellite-derived data, this approach addresses the acute data scarcity prevalent in regions most vulnerable to the impacts of climate change and hydrological extremes. The proposed model captures complex, non-linear spatiotemporal dependencies within multi-sensors data, offering significant improvements over conventional geo-spatiotemporal analysis. This approach is validated across multiple case studies, demonstrating superior performance in both accuracy and reliability enhanced accuracy and reliability over conventional neural network architectures such as Vanilla LSTM, CNN-LSTM, traditional Transformers, and Informers. The incorporation of probabilistic outputs alongside sophisticated self-attention mechanisms significantly improves the model's capability to forecast streamflow over extended sequences, addressing critical gaps in flood forecasting. The findings underscore its potential as a practical tool for enhancing disaster preparedness and optimizing water resource management strategies in data-scarce regions, thereby contributing significantly to the resilience of vulnerable communities against climate-induced threats.