Pattern recognition of soil classes across a semi-arid landscape of western Iran using knowledge-based fuzzy approach

Background and objectives: Modeling soil variation plays an essential role in sustainable management of the resource. However, discontinuous models used for decades, do not describe soil variation enough as that required in modern agriculture, since lead to basic shortcomings in spatial predictability value of maps. To conquer the problem, new statistical algorithms known as “machine learning” tools are increasingly used to construct and improve soil maps, digitally. As a means of machine learning, SoLIM employs knowledge-based fuzzy approach to realize soil-landscape relations and predict soil pattern in a continuous way. In this work, SoLIM used to predict soil distribution pattern in a 2300 ha area of Miandarband region of Kermanshah province.
Materials and methods: Maps of slope gradient and aspect, planform and profile curvature, and wetness index derived from a 10m-resolution digital elevation model (DEM), and along with geological map used in the study as most effective environmental covariates of soil diversity over the area. Based on physiographic analysis, 26 pedons were described and classified in 7 subgroups of Soil Taxonomy (ST) and 16 RSGs of WRB at second-level, respectively. To train the algorithm to recognize relations between covariates and classified soils in both systems, required fuzzy rules defined in SoLIM environment. Following inference, a fuzzy distribution map for each subgroup and RSG constructed. After combining the fuzzy outputs, a non-fuzzy map of predicted soil distribution pattern over the study area obtained for each classification system.
Results: Though results confirmed good learning ability of the algorithm, outputs where different for two the classification systems. As a reflection of its hierarchical structure, map of ST great groups was more contiguous than that of WRB. However, patchy appearance of WRB map interpreted as a sign of better spatial predictability, because of its more flexible two-leveled structure. Thus, probably WRB-based inference leads to more realistic predictions. This indicates how the results are affected by logical structure of soil classification system. To evaluate model performance, results of 25 more pedons aligned in 4 transects and 5 purposive points so that capture most soil variability over the study area, compared to SoLIM predictions. Based on overall map accuracy (OA) and Kappa agreement index (K), SoLIM predictions at ST subgroup level, were correct by 78 and 64 percent, respectively. Same values for WRB were 67 and 62. Inference at family level led to poor results. However, considering transects, correct predictions were 78.3 and 65.2 percent for ST and WRB, but for the random points was 75 for both. Results confirmed good predictions by SoLIM in the study area. At lower categories of ST with hierarchical structure, the model showed a poor ability to identify various soils.
Conclusion: No doubt, increasing sample points is the most effective factor on improving predictability of maps either in traditional or modern soil mapping techniques. However, such viewpoint seems unfeasible and not conforms to economic considerations of DSM. Probably, adopting some other strategies such as identifying most effective environmental covariates, increasing algorithms sensitivity, and better sampling designs to obtain optimal number and distribution of observations over the study area.