Federated Learning for Simulating Land-Use Change and Carbon Storage in Sri Lanka’s Coastal Zone: A Decentralized GeoAI Approach Without Exposing Local Government Planning Data

ABSTRACT: Multi-jurisdictional carbon modelling for coastal zone management and rural development planning is constrained by data sovereignty barriers that prevent collaborative analysis across administrative boundaries. A novel FedProx-PLUS-InVEST framework was developed to enable privacy-preserving land-use simulations and carbon storage assessments using federated learning. The framework was validated in the Batticaloa Lagoon watershed, Eastern Sri Lanka, where three scenarios were simulated: Business as Usual, Rapid Rural Economic Expansion, and Ecological Protection. Model accuracy approaching centralised training performance was achieved while maintaining complete data locality across the participating jurisdictions. Carbon storage projections revealed that carbon neutrality targets require stringent mangrove protection measures, with the Ecological Protection scenario achieving a net carbon gain relative to the baseline (+2.4%), demonstrating the highest carbon sequestration potential among the three scenarios evaluated. This framework represents the first coupling of federated learning with PLUS-InVEST modelling, addressing critical gaps in collaborative regional planning where sensitive land-use data cannot be centralised. This approach enables multi-stakeholder environmental planning for coastal zone management and rural development in Sri Lanka and other data-scarce regions of the Global South without exposing proprietary datasets, facilitating coordinated climate adaptation strategies across institutional and political boundaries in Sri Lanka. The methodology provides a replicable template for privacy-preserving ecosystem service modeling in data-sensitive contexts globally.