Machine-learning modelling of regional landslide susceptibility in eThekwini, South Africa

The occurrence of landslides, particularly within urban landscapes, has gained significant research interest because of their devastating impacts on infrastructure, resulting in casualties, displacement, and economic losses. This study aimed to examine the probability of landslide events and landslide-prone areas within the eThekwini Metropolitan Municipality in KwaZulu-Natal Province, South Africa. To achieve this, robust machine learning models, including Random Forest (RF) and Support Vector Machines (SVM), combined with geospatial tools and key landslide-triggering factors, were considered in the analysis and characterised based on susceptibility indexes ranging from very low to very high based on various carefully chosen thresholds. The results indicate that land use, elevation, NDVI, and lithology are important landslide-influencing factors. In contrast, land use parameters, such as cultivated land and built-up areas, were associated with high susceptibility zones. Furthermore, RF yielded slightly higher overall accuracy scores than SVM, with accuracy scores of 99.45% and 98.63%, respectively. The SVM susceptibility output was further analysed based on the spatial distribution of formal and informal housing types. The results confirmed that only a small percentage of each housing type fell within the high-susceptibility zones, calculated as 1.33% and 2.41% for formal and informal housing, respectively. Although SVM is one of the strongest models, RF shows higher accuracy in landslide susceptibility modelling. Furthermore, its slightly higher recall score compared to its precision score is deemed most applicable for identifying landslide-prone areas. The study suggested that human-engineered land-use activities, such as road and property development, contribute significantly to increased landslide susceptibility within the eThekwini area.