Machine learning-based approach for flood susceptibility mapping in the Amazon River Basin

Código

I-EBHE0124

Autores

ALENA GONZALEZ BEVACQUA, Giha Lee

Tema

WG 1.12: Development & application of river basin simulators

Resumo

Flood events are among the most frequent and destructive natural disasters. Between 2000 and 2019, floods affected approximately 1.65 billion people globally. In the Amazon River basin, floods are recurrent and are anticipated to increase in both frequency and intensity. Despite this, the Amazon River basin still lacks comprehensive flood susceptibility maps. Therefore, the creation of accurate flood susceptibility maps is essential to enhance the management, monitoring, and prediction of these events. This study aims to evaluate the performance of Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) for flood susceptibility mapping in the Amazon River basin, as well as to identify the most critical flood conditioning factors. Flood extent data from the Global Flood Database were utilized to create a flood inventory comprising sixteen flood events from 2002 to 2018. Fourteen flood conditioning factors were selected, including aspect, curvature, curve number (CN), drainage density, distance to rivers, elevation, land use/land cover (LULC), lithology, rainfall, slope, soil, stream power index (SPI), topographic roughness index (TRI), and topographic wetness index (TWI). Statistical metrics such as Accuracy, Precision, Recall, F1-score, and Kappa score were used to evaluate the models' performance. Additionally, SHapley Additive exPlanations (SHAP) were employed to identify and explain the relationships of the most important conditioning factors in flood susceptibility mapping. The results demonstrate that both models performed well overall, with XGBoost slightly outperforming RF. XGBoost achieved an accuracy of 0.9188 and a Kappa score of 0.8375, while RF achieved an accuracy of 0.9120 and a Kappa score of 0.8240. However, in terms of spatial variability, both models tended to overestimate areas with a very high probability of floods, particularly in the southern Amazon. SHAP results revealed that the most important features for XGBoost were LULC, elevation, rainfall, drainage density, slope, CN, and soil. In conclusion, this study demonstrates that both RF and XGBoost techniques are effective for flood susceptibility mapping.