Supervised classification of morphodiversity using artificial neural networks on the example of the Pieniny Mts (Poland)

Probably for the second time, supervised classification (SC) using artificial intelligence - artificial neural networks (ANN) - was applied for the morphodiversity rating of the landscape, and in the Pieniny Mountains and adjacent areas, it was undoubtedly used for the first time. Vector and raster data describing the landscape characteristics in hexagonal grid cells with a size of 200 m (small diagonal) were utilized. Since high values of variables indicate high rating scores, in the traditional bonitation procedure (Aggregating Rating Method; AR), standardized values of variables are summed. However, supervised classification can associate individual bonitation degrees with classes, considering the interdependence between variables to a greater extent and allowing for the elimination of less informative variables. As choosing patterns for multiple classes seemed less credible, only two classes were defined: morphodiverse and non-morphodiverse. Subsequently, in several variants with different subsets of variables, activation functions of neuron in output layers assigned to the morphodiverse class were discretized into 2 and 5 categories, resulting in 2- and 5-class rating sets. The results of the 2-class ratings were positively verified using the Terrain Relief model (TR) reflecting the potential morphodiversity, while the results of the 5-class ratings were verified by comparison with the boundaries of the Pieniny National Park and Poprad Landscape Park. It was observed that high rating scores appear in mountainous areas or where rocks are present, indicating protected areas, excluding those with gentle landscapes and without klippen The latter areas were protected not for morphodiversity but for biodiversity. Due to the high variability of ratings obtained using SC-ANN and the AR methods, clustering was performed to facilitate their comparison. It was found that in the SC-ANN approach, areas where ratings are uncertain are significantly smaller than in the AR approach. This observation was confirmed by comparison with the TR model. This suggests that the SC-ANN metod operate more efficiently, allowing for a better assessment of the values of abiotic landscape components and, consequently, better protection of environmentally valuable areas. A significant advantage of the SC-ANN methods was the identification and elimination of variables carrying little information. This revealed the most valuable, informative variables, subsequently used for calculations. Their frequencies ranged from 16 to 25 %, and only in one case did they reach 60 % of the original variable count. The three best variables: the presence of klippen, slope steepness, and elevation, were repeated among different sets of informative features. The conducted research allows recommending the SC-ANN method for assessing morphodiversity.