چكيده انگليسي :
The current study was conducted in three parts: I: Determining the best soil sampling scheme and exploring the effect of density on certainty of sampling, II: evaluation capability and efficiency of soil spectroscopy in estimating some soil properties in the Vis-NIR-SWIR infrared range (350-2500 nm) and III: Spatial prediction of some soil properties by applying machine learning approaches and environmental covariates. For this purpose, in the study area with an approximate of 110903 hectares in a part of Kurdistan province, 346 soil samples were taken using stratified random sampling from the surface layers (0-20 cm depth), Then, soil properties included soil particle size distribution, calcium carbonate equivalent (CCE), soil organic carbon (SOC), pH, electrical conductivity (EC), some heavy metal contents, and magnetic susceptibility at low and high frequencies were measured in the laboratory. To achieve the objectives of this study, different data were used in different sections, which generally include a set of indicators extracted from remote sensing (RS), digital elevation model (DEM), geological maps, thematic maps, and principal components (PCs) of reflection soil spectra. Studying the pattern and density of soil sampling showed that with increasing the density of samples, the statistical parameters of the subset were closer to the main population and the CLHS method was more capable than other soil sampling methods. Results of spectroscopy (350 to 2500 nm) to predict the percentage of clay, sand, silt, SOC, CCE, EC, and pH using MLR, PLSR, SVM, RF, and GPR models with CR, Detrend, SGD, SNV, and MSC preprocessing showed that for clay (RPIQ = 1.91), silt (RPIQ = 1.87), SOC ( RPIQ = 1.65) performance was appropriate and in predicting sand (RPIQ = 2.21), CCE (RPIQ = 3.41) and pH (RPIQ = 2.29) performance was excellent. The weakest performance was observed in predicting EC (RPIQ = 1.16). The results of the prediction of soil properties using spectroscopy indicated that this method could be used as an indirect method to estimate soil properties. The results of spatial prediction of soil properties showed that a combination of RS indices, DEM derivatives, thematic maps, and soil properties had the best performance in predicting heavy metals and magnetic measures of soil in the study area. In addition, the results showed that the combination of proximal measurement variables (PCs of spectroscopy, magnetic measures) with the other variables in the soil parameters prediction models can play a significant role. Based on the scenarios used, variable importance analysis (VIA) showed that for estimating some soil properties such as clay and silt, CCE, SOC and pH, covariate variables included DEM derivatives such as valley depth, RS indices such as satellite bands 1, 2, 3, TDVI, NDVI, Ferric Iron, Gossan and Laterite and thematic maps such as geology and distance from mines and PCs of soil spectroscopy and magnetic measures were identified important with various priority for each of soil properties. The results of modeling and cross-validation showed that random forest and cubist were more powerful to predict soil properties than the support vector machine model. This study proved the high capability of machine learning methods to use easily available environmental data to predict selected soil properties on a large scale that are essential for decision making in sustainable management of agricultural and environmental concerns.
