Document Type : Research Article
Authors
1 University of Shahrekord
2 Shahid Bahonar University of Kerman
3 Vali-e-Asr University of Rafsanjan
Abstract
Introduction: Soil classification generally aims to establish a taxonomy based on breaking the soil continuum into homogeneous groups that can highlight the essential differences in soil properties and functions between classes.The two most widely used modern soil classification schemes are Soil Taxonomy (ST) and World Reference Base for Soil Resources (WRB).With the development of computers and technology, digital and quantitative approaches have been developed. These new techniques that include the spatial prediction of soil properties or classes, relies on finding the relationships between soil and the auxiliary information that explain the soil forming factors or processes and finally predict soil patterns on the landscape. These approaches are commonly referred to as digital soil mapping (DSM) (14). A key component of any DSM mapping activity is the method used to define the relationship between soil observation and auxiliary information (4). Several types of machine learning approaches have been applied for digital soil mapping of soil classes, such as logistic and multinomial logistic regressions (10,12), random forests (15), neural networks (3,13) and classification trees (22,4). Many decisions about the soil use and management are based on the soil differences that cannot be captured by higher taxonomic levels (i.e., order, suborder and great group) (4). In low relief areas such as plains, it is expected that the soil forming factors are more homogenous and auxiliary information explaining soil forming factors may have low variation and cannot show the soil variability.
Materials and Methods: The study area is located in the Shahrekord plain of Chaharmahal-Va-Bakhtiari province. According tothe semi-detailed soil survey (16), 120 pedons with approximate distance of 750 m were excavated and described according to the “field book for describing and sampling soils” (19). Soil samples were taken from different genetic horizons, air dried and grounded. Soil physicochemical properties were determined. Based on the pedon description and soil analytical data, pedons were classified according to the ST (20) and WRB (11). Terrain attributes, remote sensing indices, geology, soil and geomorphology map were considered as auxiliary information. All of the auxiliary information were projected onto the same reference system (WGS 84 UTM 39N) and resampled to 50×50 m according to the suggested resolution for digital soil maps (14). Four modeling techniques (multinomial logistic regression (MLR), artificial neural networks (ANNs), boosted regression tree (BRT) and random forest (RF)) were used for each taxonomic level to identify the relationship between soil classes and auxiliary information in each classification system. The models were trained with 80 percent of the data (i.e., 96 pedons) and their validation was tested by remaining 20 percent of the dataset (i.e., 24 pedons) that split randomly. The accuracy of the predicted soil classes was determined by using overall accuracy and Brier score.For each classification system, the model with the highest OA and the lowest BS values were considered as the most accurate model for each taxonomic level.
Results and Discussion: The results confirmed that ST showedmore accessory soil properties compared to WRB. The ST described the cation-exchange activity, soil depth classes, temperature and moisture regime. The different models had the same ability for prediction of soil classes across all taxonomic levels based on ST. Among the studied models, MLR had the highest performance to predict soil classes based on WRB. For all the studied models and both classification system, OA values showed a decreasing trend with increasing the taxonomic levels. Predicted soil classes based on the ST had the higher accuracy. Different models selected different auxiliary information to predict soil classes. For most of the models and both classification systems, the terrain attributes were the most important auxiliary information at each taxonomic level.
Conclusion: Results demonstrated that although ST showed more accessory soil properties compared to WRB, the DSM approaches have not enough accuracy for prediction of the soil classes at lower taxonomic levels. More investigations are needed in this issue to make a firm conclusion whether DSM approaches are appropriate for prediction of soil classes at the levels that are important for soil management. Prediction accuracy of soil classes can be influenced by the target taxonomic level and classification system, soil spatial variability in the study area, soil diversity, sampling density and the type of auxiliary information.
Keywords
)
Send comment about this article