Urban flood exposure assessment through citizen science, 2D hydraulic modeling, and artificial intelligence.

Addressing and predicting urban flooding remains a significant challenge. This study combines citizen observations, two-dimensional modelling, and machine learning (ML) to model, calibrate, validate, and forecast flooding in an urban area of central Mexico with limited runoff and rain gauge data. Citizen observations via social media and newspapers identified flood events and locations. Two events were modeled using a hydraulic model (FLO-2D), which were calibrated and validated using water depths estimated from citizen observations. Error metrics were calculated using mean squared error (MSE), mean absolute error (MAE), and root mean square error (RMSE), and statistical differences between estimated and modeled water depths were assessed using the Mann-Whitney U test. Flood prediction and its contributing factors were assessed with ML and evaluated through the area under the curve (AUC), as well as accuracy, precision, cross-validation (k-fold), and agreement with validation points. Results indicated 24 flood events and 297 citizen observations between 2010 and 2021, with a strong relationship (r = 0.91; p < 0.05). The FLO-2D model was successfully developed, calibrated, and validated to replicate floods (MSE<2.91 %, MAE <19.21 %, RMSE <26.97, and p > 0.05). Additionally, the ML model effectively predicted flood and non-flood zones (with 0.87, 84.9 %, 88.5 %, 82.76 %, and 85.3 % for AUC, accuracy, precision, k-fold, and validation) and a high probability (>60 %) of flooding, with urban density (0.060), citizen observation frequency (0.046), Manning (0.045), and rainfall (0.044) as key factors in flood prediction. These findings confirm the effectiveness of this approach for urban flood modelling and forecasting, underscoring the importance of citizen participation.