In June 2017, The Chittagong Hill Tracts, Bangladesh, experienced torrential rainfall followed by massive landslides. According to UNOCHA, 160 people were killed, and 187 were injured at the foothills, with 6,000 houses destroyed. Once a landslide starts, it is difficult to stop the damage due to its fast-onset nature.
When you see a colorful risk map—a simulated 100-year flood inundation map in a city or red-colored road sections indicating a high risk of landslide, for example—have you ever asked yourself how this risk map was prepared and how reliable it is? This is a crucial question for decision-makers who prioritize investments based on the risk maps.
There are several ways to increase the reliability of risk maps. These include utilizing onsite monitoring data for model calibration and validation, leveraging local knowledge and information for spot checks, and conducting robust “eyes on the ground” quality control with different experts. But
Our recently published working paper, Machine learning and sensitivity analysis approach to quantify uncertainty in landslide susceptibility mapping, tackles the uncertainty question using machine learning technology for landslide risk mapping.
Why does landslide mapping matter?
Using this map, decision-makers and practitioners can take necessary steps to reduce vulnerability. Actions can include restricting new development in highly susceptible areas and stabilizing slopes along roads, schools, and other critical infrastructure to prevent landslides and minimize damages and losses.
How do we develop landslide susceptibility mapping, and what are the challenges?
In a common method of landslide susceptibility mapping for roads, we first identify factors that contribute to the occurrence of landslides, such as rainfall, land cover, distance to roads, lithology, and slope. Second, we determine the factor weights either statistically—based on comparing historical landslide data and derived factors—or simply by experts’ judgment or literature values. But the lack of data coverage or accuracy and reliance on experts’ subjective judgment leaves room for an unpredictable error margin.
How can we reduce uncertainties?
Building on a past landslide risk study conducted in the Chittagong Hill Tracts and Sylhet regions in Bangladesh, the study applied the Random Forest Classification Machine Learning Algorithm for training the model based on historical landslide and non-landslide locations. Through the process, in combination with large-scale sensitivity analysis, the reasonable range of each factor weight was identified (Figure 1), resulting in reducing the “uncertainty” around the weights.
Figure 1: Box plots of the uncertainty in the weights of factors contributing to landslide susceptibility. The ends of the boxes in panel (b) represent the upper and lower quartiles, the continuous vertical lines inside the boxes mark the medians, the circles show the outlier points, and the whiskers extend to the maximum and minimum values, excluding the outliers.
How can this tool support decision-making?
Depending on whether you look for the most conservative or the least conservative scenario, this study highlighted the highest and lowest road susceptibility to landslides and the associated combination of the weights (Figure 3). All other simulations fall between the two scenarios. Figure 4 translated this information into maps and visualized the landslide susceptibility of roads. Decision-makers can select a suitable weight combination within the ranges depending on the resources available to implement mitigation measures.
Figure 3: Parallel coordinate plot of the sensitivity analysis of factor weight values within the uncertainty range determined using machine learning.
Figure 4: A large part of the roads is colored red (extremely high susceptibility) on the most conservative map (right), while most of the roads on the least conservative map (left) are colored yellow (high susceptibility). If the government has enough budget, they may select the right map to guide their investment.
The uncertainty tool can also be used in combination with and to supplement expert judgment. Collecting more accurate, detailed (classified by scale), wider coverage of historical landslide data or higher resolution model input (slope) can also help improve the quality of the model.
The authors would like to thank GFDRR-EU which funded the South Asia Region Building Resilience to Landslide and Geohazard Risk in the South Asia Region Program and Africa Fellowship Program for funding the work.