Stepwise integration of analytical hierarchy process with machine learning algorithms for landslide, gully erosion and flash flood susceptibility mapping over the North-Moungo perimeter, Cameroon

Background The Cameroon Volcanic Line (CVL) is an oceanic-continental megastructure prone to geo-hazards, including landslide/mudslide, gully erosion and flash floods targeted in this paper. Recent geospatial practices advocated a multi-hazard analysis approach supported by artificial intelligence. This study proposes the Multi-Geoenvironmental Hazards Susceptibility (MGHS) tool, by combining Analytical Hierarchy Process (AHP) with Machine Learning (ML) over the North-Moungo perimeter (Littoral Region, Cameroon). Methods Twenty-four factors were constructed from satellite imagery, global geodatabase and fieldwork data. Multicollinearity among these factors was quantified using the tolerance coefficient (TOL) and variance inflation factor (VIF). The AHP coefficients were used to weigh the factors and produce a preliminary map per Geoenvironmental hazard through weighted linear combination (WLC). The sampling was conducted based on events records and analyst knowledge to proceed with classification using Google Earth Engine (GEE) cloud computing interface. Classification and Regression Trees (CART), Random Forest (RF) and Gradient Boosting Regression Trees (GBRT), were used as basic learners of the stacked hazard factors, whereas, Support Vector Regression (SVR), was used for a meta-learning. Results The rainfall was ranked as the highest triggering factor for all Geoenvironmental hazards according to AHP, with a coefficient of 1 , while the after-learning importance assessment was more varied. The area under receiver operating characteristic (AUROC/AUC) was always more than 0.96 , and F 1 -score is between [ 0.86–0.88 ] for basic classifiers. Landslides, gully erosion and flash floods showed different spatial distributions, confirming then their probability of co-occurrence. MGHS outputs clearly displayed two and three simultaneous occurrences. Finally, the human vulnerability assessed with population layer and SVR outputs showed that high human concentrations are also the most exposed, using the example of Nkongsamba’s extract. Conclusions Combining AHP with single learners, then a meta-learner, was efficient in modelling MGHS and related human vulnerability. Interactions among geo-environmental hazards are the next step and city councils are recommended to integrate results in the planning process.