Optimizing Flood Monitoring Networks Using eXplainable AI and Physical Informed approaches: A Case Study of the Piracicaba River Basin in Brazil

Effective monitoring and forecasting of flood events are key aspects of early warning systems, especially in areas susceptible to frequent floods. In this context, Artificial Intelligence (AI) techniques have proven to be a strong tool for enhancing such systems because they can capture non-linear processes of flood genesis. AI models make accurate predictions with minimum processing time, thus providing a strong alternative to nowcasting. However, the quality and quantity of monitoring stations and the black-box nature of machine learning (ML) models hamper the development of efficient and adaptable Early Warning Systems (EWS). Accordingly, this study aims to investigate the impact of data quality and quantity on the performance of data-driven flood forecasting models built upon eXplainable AI (XAI) and Physically Informed (PI) approaches. Intending to develop a predictive analysis of stage level in this flood-hit city of Piracicaba, the Piracicaba River catchment had 18 gauging stations with 10-minute time-step rainfall and stage monitoring used for model resiliency checks based on the MLP network. This included defining the structure of the model and the input variables determined by previous studies on flood wave propagation times prior to training and testing the model. This approach considered such hydrological aspects as incorporation into the machine learning framework. A deterioration algorithm was developed to simulate the gradual introduction of gaps in the stage and rainfall time series (10% to 100%), designed to assess the impacts of failures and monitoring errors on model prediction. This methodology provides a way to assess how the ML model would be able to handle misinformation and failures in flood predictions, while at the same time, drawing a line of priority regarding the stations to maintain the effectiveness of EWS. XAI enabled us to assess the hydrological aspects behind the models’ performance and the stations’ importance, which are crucial pieces of information for planning maintenance campaigns and allocating budget. Moreover, improvement in the monitoring network is possible by defining places for installing new sensors based on the physical aspects behind runoff onset and flood propagation. Therefore, PI and XAI are central to enhancing EWS under changing climate because they incorporate knowledge of hydrological dynamics into station selection and ML model development.