ARTIFICIAL INTELLIGENCE AND REMOTE SENSING IN URBAN HYDROLOGICAL APPLICATIONS

Abstract

Urban flooding is a big concern because it causes material, economic, environmental losses, and in the worst situations results in the death of living beings. To deal with these issues, preventive approaches must be implemented to minimize such impacts. Although there are researches seeking to solve the issue of flooding in urban areas, there are few works related to Deep Learning (DL) techniques for monitoring water resources. Due to this problem, this paper investigates and proposes DL-based methods for water monitoring. First, the performance of the SegNet semantic segmentation model in delineating water bodies in RGB images was analyzed, presenting an accuracy above 97%, showing that the model is suitable for water segmentation. Next, an automated approach was introduced in water level measurements combining DL and photogrammetry, showing correlations between reference measurements and the proposed approach of 93%. We also analyzed different configurations for SegNet evaluating the performance of the models for generalization tasks in segmenting different water surfaces, showing that techniques such as transfer learning and fine-tuning improved the results. Furthermore, it was shown that there is a reduction in the segmentation accuracy when the number of labeled images used in the network training is reduced. Finally, the performance of the Space-Time Correspondence Network (STCN) model in the segmentation of water based on video structures was analyzed, which the results show that the model is accurate in delimiting the contours of a body of water in different situations. The major contribution of this study is the optimization of information concerning a body of water using techniques different from traditional measurement systems.