Development and optimization of a morphodiversity model for mountainous areas using supervised classification and artificial neural networks

Abstract Morphodiversity assessments in mountainous landscapes represent an important tool for geoconservation. Particularly useful are the Raster Continuous Morphodiversity (RCM) models, which are based on continuous variables and measures of pixel variability within statistical zones. Traditional RCM models of the Aggregating Ratings (AR) type define morphodiversity as the sum of standardized partial criteria; however, their drawback lies in the redundant aggregation of variability, which leads to evaluation errors. To mitigate this issue, a new RCM model was developed, based on Supervised Classification (SC) and Artificial Neural Networks (ANN), which accounts for interdependencies between variables and eliminates low-informative ones. This study presents the results of RCM modeling for the Pieniny Mountains (southern Poland) using the RCM SC−ANN and the improved RCM SC−ANN−M models. Variable reduction was performed using the Global Sensitivity Analysis (GSA) method in its Backward Analysis (BA) and Backward Stepwise Analysis (BSA) variants, resulting in simplified models. Comparisons with the SC-ANN VSR−9−BS and RCM SDcm models showed that RCM SC−ANN−M produces more reliable evaluations and can be recommended for testing in other mountainous areas. The high correlation coefficients (0.96–0.98) between full and reduced sets of criteria confirm the effectiveness and efficiency of the BA and BSA procedures.