An efficient technique for hazard susceptibility mapping using AHP and ML based models

Landslides are one of the most frequently occurring natural disasters worldwide, often resulting in significant casualties, substantial loss of life, and extensive property damage. Certain areas are more susceptible to landslides due to various natural factors such as soil type, average rainfall, and land cover, as well as human factors like proximity to roads, mining activities, and other human settlements. Over the years, researchers have employed various methods to identify areas with a higher likelihood of landslides. One such method is the Analytical Hierarchy Process (AHP), a multicriteria decision-making process. The main advantage of AHP is its welldefined mathematical framework, which incorporates expert opinions to generate insights. Additionally, various machine learning models have been used to calculate the susceptibility of an area to landslides, with their performance depending on the number and quality of training samples available. Generating input data for both AHP and machine learning models is time- and effort-intensive. It usually requires obtaining data from multiple sources at different resolutions, resampling to a common resolution, and mosaicking the data to form a unified raster for each input feature. Furthermore, all input features must be resampled to a common resolution. In our current work, we propose a unique technique that combines both AHP and machine learning-based models. Our primary objective is to accurately detect past landslides in a given region while optimizing the time and effort required to generate input data. We evaluate our proposed model’s performance in terms of accuracy on landslide data from Italy and compare it with the performance of standalone AHP and machine learning models in terms of (i) accuracy and (ii) the time and effort required to generate input data for each model. The results are promising, with our proposed model surpassing the AHP model’s performance and being almost as accurate as the standalone machine learning model, while requiring less than half the time and effort to generate input data compared to the standalone AHP and machine learning models.