Saavedra Navarro, Jorge Andrés (2026) Improving Flood Susceptibility Mapping Using Machine Learning Techniques and Geomorphic Features. [Tesi di dottorato]

Documento PDF Thesis_SaavedraNavarro_JorgeAndres_XXXVIII_cycle.pdf Visibile a [TBR] Amministratori dell'archivio Download (10MB) | Richiedi una copia

Abstract

Flood susceptibility mapping is increasingly supported by simplified modelling approaches that enable large-scale, data-efficient analysis. However, their reliability remains constrained by challenges related to temporal generalisation, parameter uncertainty, and the physical representation of flood processes. Machine learning (ML) models often struggle to maintain predictive performance beyond their calibration period, while geomorphic approaches such as the Geomorphic Flood Index (GFI) remain sensitive to parameter selection. This thesis aims to improve flood susceptibility mapping along two complementary fronts: enhancing the temporal transferability of ML models, and reducing the parameter uncertainty of geomorphic approaches such as the GFI. For the ML approach, two models and a multi-source national flood inventory (1980–2026) were used to evaluate temporal generalisation under varying climatic and environmental conditions, along with a systematic framework to define non-flooded areas and to account for overfitting, domain shift, and threshold sensitivity. The results reveal a strong dependence on both the inventory type and the sampling configuration, while generalisation capacity—both temporal and spatial—is limited, with a systematic and reproducible decline in performance that increases the frequency of low probabilities and suggests that future predictions are prone to underestimation. In parallel, the GFI was examined across multiple catchments representing diverse geomorphological settings and benchmarked against a Global Flood Map, with GFI 2.0 outperforming the original version. The findings reveal significant spatial variability in the scaling exponent n, indicating that a single universal parameter is unsuitable across differing topographic environments. To mitigate this limitation, a streamlined theoretical formulation was proposed, establishing a connection between parameter variability, contributing area, and drainage network structure. Preliminary results of this spatially adaptive parameterisation enhance model stability, reduce uncertainty in parameter selection, and preserve computational efficiency. Overall, this thesis demonstrates that reliable flood susceptibility mapping requires both temporally robust machine learning models and spatially adaptive geomorphic parameterisation. Through these complementary contributions, the study provides practical strategies to improve the operational use of simplified flood models in large-scale, data-limited environments.

Downloads

Actions (login required)

Modifica documento