Spatial disaggregation semi-detailed soil map using DSMART approach

Background and objectives: Digital soil data with high spatial resolution and enough accuracy and precision are necessary for management of global challenges such as food security, environment problems. Generally, soil data are available in small scale. Nevertheless, in the last decades, with the advent of soil digital mapping and modeling approaches, it is possible to disaggregate soil map units. The spatial disaggregation of soil map units is a method for modeling the spatial distribution of individual soil classes. During this process, the soil map data from a small scale (coarse resolution) is converted to a large scale (fine resolution). The statistical and data mining methods are used to implement it. The purpose of this research was to predict the spatial distribution of soil classes by disaggregating the soil map units of a semi detailed soil map using disaggregating and harmonizing soil map units through resampled classification trees algorithm (DSMART method).
Materials and methods: The study area is located in Kermanshah province. The total area of the study was approximately 14083.9 ha. Soil polygon map include 5 map units and 4 soil subgroups. In this study, elevation, slope, aspect, convexity, direct duration, sediment index, topographic wetness index, valley depth and Vertical distance to channel network as covariates produced using DEM 10 m. Grain size index, clay index and NDVI were also calculated using Landsat 7 ETM+ imagery. Geological map at scale of 1:100,000 were also used as a qualitative covariate. Then dsmart method is run as a novel approach for disaggregation soil maps. In this method, disaggregated soil classes are represented by raster probability surfaces. DSMART samples randomly within the soil map units and uses classification trees (C5.0 algorithm) to produce probability surface maps of soil class distribution. External validation was performed using 82 profiles. The validation dataset was intersected with the corresponding probability surface maps and validation quantified by overall accuracy, producer’s accuracy, user’s accuracy and kappa coefficients. Furthermore, confusion index calculated between the most probable and second-most-probable soil class. The CI expresses concisely degree of confusion about soil class given.
Results: The most important predictive variables in the tree classification model were Vertical distance to channel network, Elevation, lithology, grain size index and MRVBF. The confusion index close to 1 has a large extent in the study area. It shows that occurrence probability of soil subgroups is near equal in each location in both the most probable soil class and second probable soil class maps. Validation of probability surfaces showed that overall accuracy the most probable soil class, second probable soil class and third probable soil class are 44%, 28% and 11%, respectively. These results indicated the relatively good performance of dsmart method for generating digital individual soil class map. However, kappa coefficients for first, second and third probable surfaces soil maps were obtained 0.04, 0.02, -0.08, respectively. Low kappa coefficients can be attributed to the true nature of the data, i.e the dominance of the Typic Calcixerepts subgroup as compared to other subgroups of the soil in the traditional soil map and the dsmart model prediction map and validation data.
Conclusion: Dsmart method is able to predict the occurrence probability of all soil classes which its distribution is unclear in soil map unit. This provides the opportunity to produced digital soil class maps when legacy soil data and covariate information becomes available. Such outputs may help us to recognize better soil- landscape relationship.