Comparison of different approaches used in digital soil mapping and researching the relationship between digital variables and soil properties: A case study of Atabey plain

This study includes testing the predictability of soil classes with digital soil mapping methodology and the usability of different methods used in creating spatial maps in a sample area with available data. Mapping of categorical soil classes is a decision-making process that falls within the field of general classification. For this reason, a multinomial logistic regression algorithm, which is used for classification in machine learning, has been used. Using this algorithm, which is defined as a supervised learning algorithm, spatial maps were produced in the field with soil class models learned from the existing training set and their accuracy was evaluated. The K-means algorithm, one of the unsupervised learning algorithms, has been applied to define the areas where the formation of soil classes may occur as a result of the combination of environmental variables. Finally, by using multiple linear regression models between the soil data obtained from the field and the environmental variables obtained from the current satellite image and digital elevation model, the relationships between soil properties and environmental variables were tried to be revealed. In order to understand and interpret the existing relationships between soil properties and environmental variables, principal component analysis was conducted. As a result of the study, it was determined that soil class maps produced as a result of supervised classification have higher accuracy. The digital soil maps obtained can be qualified data that can be used as input to digital agriculture systems, different from the way the existing soil maps are used traditionally.