Morphometrical analysis of torrential flows-prone catchments in tropical and mountainous terrain of the Colombian Andes by machine learning techniques

Assessment of susceptibility to torrential flows at the catchment scale is a necessary step for prioritising zones that can be identified by hazard and risk maps. In the Andean region of Colombia, where these phenomena are common and highly destructive, few studies have been carried out on the subject. Considering that the effects of torrential flows will likely increase in the future due to extreme weather events and urbanisation of low prone areas, effective land management plans should include the assessment of susceptibility to this type of event. In this study, machine learning techniques were applied with morphometric parameters of mountainous catchments in the northern Colombian Andes to analyse torrential susceptibility in tropical environments. Several feature selection methods were implemented using statistics to evaluate the relationship between morphometric parameters and torrential susceptibility. The parameters that showed the strongest relationship with torrential catchments were selected for application of machine learning techniques, along with evaluation of performance and prediction ability. The analysis was carried out on morphometric parameters regarding the catchment’s drainage network, basin geometry, drainage texture, and relief characteristics. According to the results, the catchments characteristics related to relief and drainage texture are the most important for the study of torrential susceptibility. Within these groups of predictor parameters, relief ratio and constant of channel maintenance are the highest ranked variables. Although the Melton index shows a good prediction capacity, thresholds proposed in the literature are not valid in the Colombian Andes.