Spatial Modeling of Flood Susceptibility Using Machine Learning Algorithms
This study has been conducted during my Data Science internship at the Moroccan Association for Regional Sciences laboratory and its resuslts have been published in the Arabian Journal of Geosciences.
1. Introduction
Floods constitute one of the most devastating and destructive natural forces in the world. They have a considerable impact on the economy and can result in significant loss of life. The frequency and severity of flooding in Morocco are likely to increase due to global warming, urbanization, poor watershed management, deforestation, and land-use changes. It is therefore incumbent upon the government and researchers to identify and implement mitigation strategies to reduce the damage caused by flooding.
Several strategies, including studies by advanced data analysis methods, have been adopted to curb this phenomenon and ultimately limit the accompanying damage. In this study, four supervised models based on machine learning (ML) algorithms were used to map food vulnerability in the Souss watershed located in southern Morocco, where numerous devastating floods have been recorded over the past decade. They include random forest, x-gradient boost, k-nearest neighbors, and artificial neural networks.
2. Data and flood predisposing factors
Access to data is considered the most difficult aspect of this type of study. However, thanks to the Agence du Bassin Hydraulique du Souss-Massa (ABHSM), the data for the inventory of flooding points in the watershed was made available. In total, 87 historical food points were provided by ABHSM. Another 87 non-food points were chosen randomly from areas with slopes greater than 50%. The assumption is that if the slope is greater than 50%, they are non-food points, as all 87 food points are on a slope less than 50%.
Many flood predisposing factors have been selected, looking into the literature. Those identified in this study were the digital elevation model (DEM), aspect, curvature, distance to rivers, drainage density, flow accumulation, flow direction, geology, land use, precipitation, slope, soil type, and topographic moisture index.
3. Results
Evaluating the predictions of the models were based on the AUC value of the ROC curve. The ML models adopted in this study performed reasonably well. Indeed, they had AUC values greater than 80%. Nevertheless, differences in model accuracy and overall performance remain, as was the case in this study. As such, it is imperative to compare model results to identify points of
agreement and disagreement. Areas where the same classes of food susceptibility were predicted, were characterized by good models agreement. Conversely, the models disagreed where different classes of food susceptibility were predicted. The uncertainty of the results is often low where all models agree and high where they disagree. The susceptibility maps were overlaid to make spatial comparisons of the model results. The four models of each model type (KNN, ANN, RF, XGB) were overlaid to compare their results.
The susceptibility maps were classified into five groups including very low (0–0.2), low (0.2–0.4), moderate (0.4–0.6), high (0.6–0.8), and very high (0.8–1) based on the probability of food risk before comparing the model results.
In addition, all the models were overlaid to identify areas of agreement and disagreement.
4. Conclusion
Floods are one of the most destructive natural hazards occurring in the world. In this study, the models based on ML algorithms including KNN, ANN, RF, and XGB were compared for their performance in mapping food susceptibility in the Souss watershed. Using these models, the relationships between flooding and conditioning factors were adequately characterized.
The results of this study are proof that food susceptibility mapping projects in Morocco have the potential to be accomplished using ML models. Indeed, susceptibility maps could serve as a useful tool for food management and for mitigating the damage caused by these events in vulnerable areas.
