Application of multinomial logistic regression model in digital survey of soil classes in Kouhbanan region of Kerman

Soil digital mapping represents a set of mathematical computations to predict the distribution of soil classes in the landscape. . The digital identification of soils as a tool for creating soil spatial data provides ways to address the growing need for high-resolution soil maps. The use of digital soil mapping technique has been expanded considerably; therefore, new methods of mapping and preparing digital maps have been developed by researchers to eliminate the limitations of traditional methods. This approach relies on statistical relationships between measured soil observations and environmental covariates at the sampling locations. Digital soil data is increasing based on new processing tools and various digital data. The present study was conducted with the purpose of digital soil mapping in Kouhbanan region of Kerman based on a Multinomial logistic regression model.

The study area is located in southeastern Iran, northwest of Kerman city, in Kouhbanan distinct. This study covers a 2000 ha area. In this study, a Latin hypercube sampling design was applied and the sampling was done according to the difference in landforms (geomorphology map), topography (including digital elevation map) and geology. Finally, the geographic locations of 70 profiles were identified. Soil profiles were described according to U.S. Soil Taxonomy (Soil Survey Staff, 2014) and finally, the soil samples were taken from their diagnostic horizons. The collected soil samples were transferred to the laboratory, and some physical and chemical analyzes were performed based on routine standard methods. Environmental data include the parameters derived from the digital elevation model, Landsat satellite images (remote sensing indexes), geology map, geomorphic units (geomorphology map) and legacy soil map of the study area. All environmental variables were derived using ENVI and SAGA software. In this research, a multinomial logistic regression model was used to predict soil classes and the modeling was done in R software using nnet package. It is worth noting that leave-one-out cross validation was used for validation. Estimation of predictive accuracy of soil classes was also done using the overall accuracy index and Kappa coefficient.

The results showed that the soils of the study area were mainly classified in the Aridisols and Entisols orders. The modeling results showed that the terrain attributes were recognized as the effective auxiliary variables in the prediction process of soil classes. This confirms topographic importance on soil genesis in the studied area. After that, geomorphology map was an important tool in soil mapping that helps to increase predictive accuracy. Among the soil classes, the prediction of Haplocambids was accompanied with low accuracy, while Haplosalids great groups were predicted with high accuracy. The low estimation accuracy of the great group of Haplocambids is probably due to the low sample size of this class of soil in the study area. A good identification of the relationships between the predictor variables and the target variable depends primarily on the size and distribution of the sample in the layers. There were only two examples of Haplocambids in the area. Therefore, low accuracy is expected because the model has failed to establish a relationship between this class with environmental variables and makes it difficult to identify threshold values for classifying soil classes and, consequently, a poorly trained model. It is also possible that low prediction accuracy is the result of the conceptual model being incomplete, since there is no characteristic feature that can help model training and ultimately prediction. Among the soil great groups, the best predictions were obtained for the great group of Haplosalids, which demonstrates high values of user accuracy and reliability. Accurate prediction of the class of Haplosalids is highly correlated with the spatial distribution of indices such as wetness index and NDVI. Kappa index and purity map were calculated 0.45 and 0.65 for digital soil map derived from multinomial logistic regression. In the predicted map, six major groups of Haplosalids, Haplocambids, Haplocalcids, Haplogypsids, Calcigypsids and Torrifluvents were identified. The great groups of Haplocalcids, Haplosalids, and Calcigypsids cover most of the area and the great groups Haplocambids and Haplogypsids occupy lowest of the area. The great group of Haplosalids is located in the north of the region and in the piedmont plain landform. Haplocalcids great groups were most commonly found in alluvial fan landform, while Calcigypsids are located in pediments, alluvial fans, and piedmont plain landforms. Haplocambids and Haplogypsids great groups are located more in the geomorphic surface of the alluvial fan and the piedmont plain, respectively. The parts of the region with the most variations or diversity of soil classes are exactly where the geomorphological map has the most segmentation. Therefore, the presence of different soil classes in the least-differentiated and most similar regions is resulted to an inefficient conceptual model and poor prediction results.

The results showed that topographic parameters were the most important and powerful variable in modeling, and confirms that topography or relief is the most important soil forming factor in the study area. Predictive results of soil classes in Kuhbanan area of Kerman province showed that geomorphological map in the study area is very useful and necessary and also is effective in understanding and communicating between soil and landscape. Using this map as a qualitative auxiliary variable can explain much of the variability of soils in the study area. Careful field observation, satellite imagery consideration, study and interpretation of data obtained from soil profiles indicate that the study area has been evolved by geological, geomorphological, and hydrological processes that lead to the formation of various landforms including rock outcrops, hills, pediment , alluvial fan and plain. For the multinomial logistic regression model in the study area, terrain attributes have the most influence on the prediction of soil classes and soil properties than the remote sensing indices. The strong relationship between soil data and environmental parameters is one of the factors influencing model accuracy. Logistic regression models will have great potential in predicting soil classes if a complete understanding of the study area and proper selection of auxiliary variables are carried out.