FedGIS-Water: A Federated Learning-Enhanced GIS Framework for Circular Water Infrastructure Assessment in Railway Stations

We introduce FedGIS-Water, a federated learning-enhanced GIS framework designed for assessing circular water infrastructure across railway stations in emerging regions. As sustainable water management becomes increasingly vital in decentralised environments, solutions that promote collaboration while preserving privacy are needed. Existing methods often depend on centralised data collection, which isn't feasible with fragmented infrastructures. Our approach combines decentralised spatio-temporal modelling with federated parameter sharing, allowing stations to optimise water use, reuse, and quality monitoring without sharing raw data. Each node uses a GIS-Enhanced Spatio-Temporal Model (GESTM), leveraging graph neural networks for spatial relationships and temporal fusion transformers for dynamic patterns, processing data like geographic location, consumption rates, and water quality. Parameters are aggregated through a resilient federated learning protocol, FedProx, which manages data heterogeneity and reduces bandwidth use. The system also connects with traditional water infrastructure via automated input replacement and output integration, replacing manual logging with sensor-based predictions to control valves, pumps, and filters. Running on edge devices with Kubernetes-managed federated servers, FedGIS-Water offers scalability and privacy through hybrid spatial-temporal federated learning, a unique contribution for emerging regions. Experimental results confirm its success in minimising water waste and enhancing quality alerts, demonstrating its potential for large-scale use in resource-limited environments.