Using machine learning to build safer cities

Using machine learning to build safer cities Swati Mestri Scientific Editor Andrew Zinin Chief Editor In eThekwini, researchers have harnessed machine learning and satellite data to spot the city's most landslide-prone zones. A new study published in Scientific Reports has found that urban development activities, including road construction, property development and vegetation loss, are significantly increasing landslide susceptibility across parts of eThekwini. The research comes just four years after the devastating April 2022 floods and landslides that claimed hundreds of lives, displaced more than 40,000 residents and caused an estimated R36 billion in damage to infrastructure and businesses. Using advanced machine learning models, researchers mapped areas most likely to experience landslides and identified the factors driving risk. The study found that land use, elevation, vegetation cover and geological conditions play a critical role in determining where landslides are most likely to occur. One of the strongest findings was the link between human-modified landscapes and increased landslide susceptibility. Areas characterized by built-up development and cultivated land were more likely to experience landslides than less disturbed environments. According to the researchers, removing vegetation can reduce slope stability, while construction activities may alter drainage patterns and put additional pressure on vulnerable landscapes. "We wanted to understand better the factors driving landslide occurrence in eThekwini and identify areas that may be vulnerable to future events," said Dr. Laven Naidoo (Gauteng City-Region Observatory, GCRO, senior researcher). "As extreme weather events become more frequent, this information can help support preventive planning and risk reduction." Harnessing artificial intelligence for disaster risk reduction To conduct the study, researchers combined geographic information systems (GIS), satellite imagery and machine learning techniques to analyze landslide-prone areas across the municipality. Two models, random forest and support vector machines, achieved accuracy levels above 98%, allowing researchers to identify potential landslide hot spots with high confidence. The resulting maps showed that areas along the coastal belt of eThekwini are particularly susceptible to landslides, while mountainous inland areas generally face lower levels of risk. Researchers say these tools could help municipalities and planners better understand where future development should be carefully managed and where disaster preparedness efforts should be prioritized. The study also demonstrates the potential of artificial intelligence and Earth observation technologies to support evidence-based planning and disaster preparedness. "Traditional landslide assessments can be labor-intensive and time-consuming," the study notes. "Combining machine learning with satellite-derived data allows for more efficient and accurate identification of at-risk areas." Lessons from the 2022 disaster The findings build on lessons from the April 2022 floods, widely regarded as one of the most destructive natural disasters in KwaZulu-Natal's history. Over four days, parts of the province received more than 300 mm (11.8 inches) of rainfall, triggering widespread flooding and slope failures. The impacts affected both informal and formal settlements, highlighting how landslide risk extends beyond traditionally vulnerable communities. While poor households in informal settlements often face heightened exposure to environmental hazards, the study also notes that formal residential developments were affected. In areas such as Umdloti, infrastructure damage raised questions about development practices, vegetation removal and the suitability of building on geologically sensitive terrain. The research suggests that landslides in eThekwini are not driven solely by natural factors such as heavy rainfall and geology. Instead, they emerge from a complex interaction between environmental conditions and human decisions about where and how development occurs. A growing challenge in a changing climate As climate change increases the likelihood of extreme rainfall events, understanding landslide risk is becoming increasingly important for South African cities. The researchers argue that advanced technologies such as machine learning, remote sensing and satellite-based monitoring can provide valuable decision-support tools, particularly in resource-constrained contexts where traditional hazard assessments are often costly and time-intensive. By identifying areas most vulnerable to future landslides, the study gives municipalities information to support safer urban planning, infrastructure development and disaster risk reduction efforts. Ultimately, the findings point to an important lesson from the 2022 floods: Reducing landslide risk will require more than responding to extreme weather events. It will also depend on how cities manage growth, protect natural landscapes and plan future development amid climate uncertainty. Publication details Cher Petersen et al., Machine-learning modelling of regional landslide susceptibility in eThekwini, South Africa, Scientific Reports (2026). DOI: 10.1038/s41598-026-73316-x Journal information: Scientific Reports Provided by Wits University