ENHANCING HYDROLOGICAL MONITORING THROUGH DEEP LEARNING AND PHOTOGRAMMETRY

Abstract

Observing components of the hydrological cycle can be challenging due to the escalation that occurs and the cost of sensors. Measuring the formation of surface runoff and flow is fundamental to understanding water dynamics, as it also influences human activities in order to keep natural ecosystems balanced. The main objective of this doctoral thesis is to propose deep learning approaches combined with photogrammetry to automatically measure surface teaching formation and flow. Our results suggest that considering class imbalance and label uncertainty when training deep learning to segment water pocket areas is more important than the network itself as well as ensembles. Area, number and connectivity of water pools and their comparison with the flow measurement, where different behaviors were found in relation to the generation of surface runoff. Regarding flow rate, our results demonstrated that both STCN and SAM using fixed points and SAM combined with Dino achieved overwhelming results for water segmentation, even with minimal or unannotated label dataset. Measurements of water levels using these masks resulted in a good fit with reference data, being able to capture changes in water flow, especially at higher water levels. In dynamic images, STCN and SAM Dino obtained similar results, however the choice of the first frame influenced the STCN results. The results found in this doctoral thesis open a new frontier for hydrologists and soil science practitioners with the possibility of directly measuring surface runoff formation and a cheaper and more scalable solution for flow measurement.