Using machine learning to build safer cities

This study models regional landslide susceptibility within the eThekwini Metropolitan Municipality, KwaZulu-Natal, South Africa, following the April 2022 flood event. Using a combination of machine learning techniques—Random Forest (RF) and Support Vector Machines (SVM)—the authors analysed the probability of landslide events and identified landslide-prone areas. The analysis integrated a detailed landslide inventory from satellite imagery and open data sources collected between April and May 2022, along with key conditioning factors (such as land use, elevation, NDVI, lithology, and proximity to infrastructure and hydrological features). Data was processed and split for model training and testing, with 10,000 samples used (evenly divided between landslide and non-landslide points). Accuracy, kappa, precision, recall, and F1 metrics were computed, and spatial analyses included formal and informal housing overlays.

The study found that land use, elevation, NDVI, and lithology are the most influential factors for landslide susceptibility in eThekwini. High susceptibility was associated with built-up areas, cultivated land, sparse vegetation, low elevations, and specific lithological units such as the Maputaland Group. The RF model slightly outperformed SVM, yielding overall accuracies of 99.45% and 98.63%, respectively, though both showed robust performance. Most housing types (formal and informal) are located in areas of low susceptibility; only 1.33% of formal and 2.41% of informal housing overlapped with high-susceptibility zones. The authors recommend further investigation into human-induced land use changes and call for higher resolution terrain and imagery data, as well as in-field verification, to improve landslide risk assessment for urban planning and risk reduction in the region.