Improved FedProx-TinyViT Model for Urban Flood Depth Reconstruction under Crowdsourced Images and Sensor Readings

Abstract

Urban flood depth reconstruction needs to quickly infer street-level information from dispersed observations by citizens, road sensors and municipal departments. This paper introduces an improved FedProx-TinyViT model for flood depth estimation from crowdsourced images and synchronous sensor data without transferring raw visual data to a central server. A compact TinyViT backbone is employed to add sensor-conditioned prompt tokens, reliability-weighted local learning is used to suppress noisy image-sensor pairs, and a proximal federated objective controls client drift under heterogeneous rainfall, camera and drainage conditions. A municipal-edge benchmark was prepared with 18,420 images, 126,000 sensor records, 5,860 corrected depth labels and 24 distributed clients for evaluation. The proposed model had an RMSE of 7.8cm and an MAE of 5.9cm, with 82.6% of the predictions within a 10cm absolute error. FedAvg-TinyViT was outperformed; it had a lower RMSE of 31.4%, a smaller median client drift of 0.129, and reduced communication volume by 45.6% compared with raw-data upload. Based on the above results, the compact federated multimodal transformer supports privacy-preserving flood intelligence with good practical accuracy, calibration and edge deployment efficiency.