Landslide potential mapping applying maximum entropy to continuous change maps

Landslide mapping inventories are crucial for disaster prevention and risk mitigation. Remote sensing uses remote sensors that record data from the Earth's surface encoded in digital images distributed in electromagnetic spectrum ranges, allowing us access to various types of information. This, in conjunction with appropriate spatial analysis and modeling techniques, allows us to monitor the phenomena, such as landslides, that put man-nature coupled systems at risk. This paper presents a practical alternative for integrating landslide inventories in the central area of the state of Guerrero in Mexico by using the maximum entropy model (MaxEnt), a machine learning algorithm oriented to the potential prediction of patterns using continuous change (CC) maps as input. These maps were obtained using the unsupervised change detection methods linear regression and difference applied to transformed images, the normalized difference vegetation index (NDVI), and principal component analysis (PCA). The selection of supplementary input data was made by using the jackknife test to assess the contribution of the main determinant factors of slope stability: lithology (L), angular slopes (AS), and terrain orientation (TO). Ground truth landslide samples were used for the algorithm training (2/3) and the accuracy assessment of the final inventory map (1/3). The landslide inventory map derived by combining the MaxEnt model, the thresholding by the secant method, and the discrimination of pixels with slope values less than 5 degrees reveals a high accuracy and visual concordance with reality, reaching 3.0% and 3.5% in commission and omission errors, a Kappa concordance index of 93.37%, and an AUC of 0.75, indicating MaxEnt is a practical and efficient tool that allows for the rapid and accurate generation of reliable maps for the detection of landslides.