Predicting badland occurrence in Catalonia by applying random forest techniques and a multi-scale approach

Abstract

Badlands are highly erosive landforms of dissected morphology, which can be found on soft rocks and unconsolidated sediments, with little or no vegetation, that are useless for agriculture. The erosion rates of these areas are high, causing important environmental and economic problems. For that reason, detecting the main conditioning factors that control badland occurrence and identifying susceptible areas is higly important to prevent soil erosion phenomena and their negative consequences.This work attempts to assess badland susceptibility and their governing factors at a multi-scale level, using a random forest (RF) modelling approach. Previous RF-based research have demonstrated that RF modelling is a powerful tool for making predictions in the same spatial context and scale where the model has been trained. However, upscaling RF-modelling results to obtain accurate predictions in other, more extensive spatial contexts than that used for model training, remains an important challenge.For that, the Upper Llobregat River Basin (ULRB, 504.8 km2) and Catalonia region (CAT, 32000 km2) have been selected as study areas. We have evaluated the viability of training a RF model for the analysis of badland suceptibility in the small spatial context of the ULRB, and further testing it to the more extensive spatial context of CAT. Revealing the most important factors that control badland distribution in the territory has been another goal in the present study. Eleven governing factors and two badland inventories developed for these study areas have been used for model training and testing. Model performance has been analyzed through validation tests and three different evaluation metrics: AUC, Kappa coefficient and accuracy. The outcomes of this work manifest that the two variables that have the most important relevance for badland occurrence are lithology and NDVI. In addition, our results showed that upscaling RF model results defined in the ULRB to the more extensive spatial context of CAT in order to predict badland occurrence, it’s possible but not ideal. Last, badland susceptibility maps of ULRB and Catalonia have been obtained with a very high accuracy (96% and 97% respectively), confirming the feasibility and uselfuness of RF approach for badland susceptibility assessment.