Evaluation Extrapolation of Random Forest Model to Predict Soil Classes at Subgroup Level

Many soil maps that produced in Iran are in medium scale related to the soil survey projects that have done over the past six decades. In many cases, soil maps have not updated due to the high cost of soil survey activities in conventional methods. A proposed solution to overcome limitations of the conventional soil survey is digital soil mapping (DSM) that extensively used for producing soil maps in many countries recently. The extrapolation method in which soil pattern rules in reference area is used for soil class prediction in other areas as a cost-effective method have been mentioned by some soil surveyors. To achieve the main advantages of extrapolation in DSM, in this research we evaluated the use of random forest model in a reference area (donor area) for producing soil taxonomic classes at subgroup level in a site out of the reference area (recipient area). In this study two neighboring areas in Fars Province in southern Iran were selected: 1) Saadat Shahr plain as donor site and, 2) Seidan plain as recipient area. Two agricultural plain have a moderately similar environmental condition such as elevation, geology, physiography, and climate and agriculture behavior. In donor area, 82 soil profiles were excavated, described and analyzed. Latin hypercube sampling (LHS) was used as a statistical method in donor area. In recipient area, 27 locations were determined on some parallel transects across the plain. All soils were classified according to USDA soil taxonomy System (2014). Random forest (RF) in R statistical software was used to predict soil classes in donor area. Then the constructed model in donor area saved and applied to the recipient area. 25 variables related to soil forming factors consist of 1) primary and secondary train attributes and 2) remote sensing indices obtained from Landsat 8 satellite, OLI sensor imagery were used in this study. All auxiliary environmental covariate layers were resampled to a 30 resolution. Producer's, users and overall accuracy and kappa index calculated according to the agreement of the field surveyed with predicted soil classes. Using RF algorithm from the 25 variables related to soil forming factors, five primary and secondary train attributes consist of slop, multiresolution index of valley bottom flatness (MRVBF), terrain ruggedness index, topographic wetness index and modified catchment area were selected as influential covariates. An overall accuracy of 72%, and a Kappa index of 0.59 in the donor area, illustrating the relatively desirable agreement between observed and predicted soil classes. For extrapolating evaluation, the result of RF model with 70% of soil samples in the donor area was compared with the output of the transported RF model using 27 observations of the validation dataset. The overall accuracy of the external validation was 45%, and the Kappa index was 0.28. Transferring the RF model constructed by all soil samples of the donor area (100%) showed a better result of soil prediction in the recipient area. The overall accuracy and the Kappa index of the external validation was 52% and 0.38, respectively. From the six soil subgroup classes, the best predicted classes were Typic Calcixerepts and Typic Xerorthents. Some classes were too sparse and the model was unable to predict them correctly. The results showed that the model extrapolation in the framework of DSM could be a powerful tool for producing soil map in the area of Iran that soil maps are not available or updating the present soil maps are time and cost consuming. The low-cost and time saving method reported here, encourages soil surveyors to select model extrapolation for their survey activities.