Integrating Remote Sensing, Multi-Objective Optimization, and Agent-Based Modelling for Nature-Based Flood Mitigation

Abstract

By 2030, pluvial flooding is estimated to cause nearly $500 billion in global damage, affecting 54 million people (UNDRR, 2015) each year. By 2050, metropolitan assets worth around $158 trillion could be at risk of floods. This rising risk has sparked an increased interest in nature-based solutions (NbS), as traditional hydraulic infrastructure may lack the flexibility required to manage more complex flood hazards. This study looks into the use of a data-driven approach to identify and assess the effectiveness of upstream afforestation in preventing pluvial floods. We constructed a thorough computational framework for the Girona/Empordà region of Catalonia, Spain. This framework combines cloud-based remote sensing via Google Earth Engine, multi-criteria decision analysis (MCDA), multi-objective evolutionary optimisation via NSGA-II, and agent-based flood simulation, which is linked to a reinforcement learning agent based on Proximal Policy Optimisation. The framework functions on two stacked spatial scales. At a regional scale (20 x 20 km), it finds and selects suitable afforestation patches. At a meso scale (1 x 1 km), it simulates flood responses to various afforestation schemes. The findings illustrate the technological viability of combining different methods into a single computing process. Preliminary reinforcement-learning results indicate a directional improvement in simulated flood outcomes.