Bayesian calibration of a 2D hydraulic model using a convolutional neural network emulator

Abstract

This study presents a Bayesian calibration framework for 2D hydraulic models using convolutional neural networks (CNNs) as surrogate emulators of TELEMAC-2D. Applied to the Lower Piura River Basin in northern Peru, the method estimates spatially distributed Manning's roughness coefficients while accounting for structural model error. A CNN trained on a simulation ensemble predicts flood depth under varying roughness scenarios, enabling substantial computational savings. The emulator is embedded in a Bayesian inference scheme with a Gaussian Process discrepancy model to capture systematic deviations. Validation with synthetic scenarios demonstrates accurate roughness retrieval in hydraulically sensitive areas. Additionally, a real-case validation was performed using PeruSAT-1, a high-resolution Earth observation satellite operated by the Peruvian Space Agency (CONIDA), acquired during the 04/10/2017 flood. This confirmed the framework's ability to reproduce observed depth patterns under data scarcity. The method provides a scalable solution for parameter inference in flood-prone regions where conventional validation approaches remain limited.