Modeling and evaluating older landscape classifications with modern quantitative methods

Landscape classifications and mapping are important elements of numerous fields of research, e.g. geography, geomorphology, agriculture, climatology, ecology, and biology (e.g. Iwahashi, Pike 2007; van der Zanden 2016; Rivas-Martínez, Penas, & Díaz, 2009; Olson et al. 2001; Bohn et al. 2002/2003). Some classifications are officially used, even at European level (European environmental agency 2016). Thus they must be prepared transparently and as objectively as possible.
We presume that existing landscape classifications are well prepared, although they were made more or less subjectively. Thus we assume that they can be confirmed (reproduced) with different mathematical models. Since there are new geoinformation, statistical, and machine learning methods available we believe that older classifications can be quantitatively evaluated.
The objective of this paper is to model and evaluate existing natural landscape classifications of Slovenia with the latest quantitative methods that can process data faster than conventional methods. Slovenia has a very diverse environment (Ciglič, Perko 2013), because there different European natural geographical regions meet – the Alps, the Pannonian Basin, the Mediterranean, and the Dinaric Alps.
In our research we rasterized landscape classifications of Slovenia made by Perko (1998) and Špes et al. (2002) and collected more than 40 raster data layers (predictors). These data layers represent different environmental characteristics (e.g. elevation, slope, precipitation regime, bedrock). From the 500.000 cells database we then selected different sets of training cells and produced several modeled landscape classifications with selected predictors. For this purpose we used some well-known methods (e.g. kNN, decision trees, and maximum likelihood classifier). We were also able to apply Bayesian multinomial regression to a large landscape type dataset with all the 500.000 cells (observations) and all the predictors, which would not have been possible without an efficient custom implementation on the GPU (Graphical Processing Unit) using OpenCL (bayesCL package for R on CRAN).
In the final stage, we compared the computer based classifications (modeled classifications) to the original classifications to find areas of agreements and disagreements. The rate of agreement differed according to the methods and training samples used. However, modeling was possible and gave accurate and reasonable results, thus we can conclude that the original landscape classifications were prepared with great sense. Our results also showed which areas can be confirmed by computer models with high confidence and which areas should be reconsidered or cannot be easily classified, due to high uncertainty in computer model classification.
The methods we tested showed promising results and are of great help in the process of landscape classification. They can be used also for other areas anywhere in the world.