جستجوی مقالات مرتبط با کلیدواژه « مدل سازی مکانی » در نشریات گروه « آب و خاک »

This study aims to evaluate and to compare the efficiency of support vector machine (SVM) and random forest (RF) models using digital soil mapping approach to predict soil texture in part of Khuzestan province. In February 2021, before determining soil texture, 200 soil samples were taken using stratified random sampling from the surface layer )0-10 cm(. Auxiliary variables included primary and secondary derivatives of digital elevation model (DEM), remote sensing spectral indices (RS), from which the appropriate category was selected using principal component analysis (PCA). Based on PCA method, nine topographic variables from DEM and eight vegetation indices and spectra from RS were selected to predict soils texture components (sand, silt, and clay). The efficiency of the models was evaluated using the coefficient of determination (R2) and the root mean squared of the error (RMSE). The results indicated that the random forest model had higher accuracy and less error than the support vector machine model (SVM), so that values of R2 in this model were 0.80 for sand, 0.81 for silt, and 0.78 for clay, and the RMSE in the prediction of these particles were 6.02, 5.89 and 6.02, respectively. While the R2 and RMSE in the support vector machine model for prediction of sand, silt and clay were (0.39, 13.70), (0.45, 10.70), and (0.46, 9.32), respectively. Also, the results of this evaluation showed that salinity index, brightness index, and channel network in addition of the 6-band Landsat 8 satellite or the far infrared band were the most important environmental variables predicting clay, silt, and sand particles. In conclusion, we suggest using Random Forest model as a useful and reliable method in preparing digital maps of soil texture in the study area.

Modeling and mapping of plants nutrient elements in soil has importance in increasing the productivity of agriculture and achieving sustainable development. The aim of this research was to prepare digital maps of two nutrients, namely, available phosphorus (Pav) and exchangeable potassium (Kex) using machine learning models (MLM) i.e., random forest (RF), cubist (CB), support vector regression (SVR) and k-nearest neighborhood (k-NN) at two depths of 15-30 and 0-15 cm in a part of Abyek Plain. In this regard, 278 soil profiles were dug, sampled from the desired horizons, and samples were analyzed. MLM performance was implemented by 10-fold cross-valuation. The modeling results demonstrated that the RF model had the highest accuracy and minimum error compared to the other three models in spatial estimation of available Pav and Kex at the two studied depths. According to the results, for Pav at a depth of 0-15 cm, CCC statistics values of 0.84, 0.74, 0.48 and 0.35 and NRMSE values of 0.38, 0.54, 0.70, and 0.80 belonged to RF, CB, k-NN, and SVR, respectively. For Kex at the same depth, CCC values were 0.82, 0.72, 0. 70, 0.47 and NRMSE 0.25, 0.34, 0.36 and 0.45, by RF, CB, SVR, and k-NN models, respectively. Similar results were obtained for 15-30 cm layer. The relative importance of environmental variables showed that soil covariates had a more effective role in the spatial estimation of Pav and Kex than other environmental variables. According to the estimated maps of the two elements and the predominance of agricultural land uses, major parts of the area are Pav deficient based on standard amounts. Therefore, to increase productivity and improve management of soil fertility, use of phosphate fertilizers is recommended under the supervision of soil experts.

To control soil erosion, several measures can be conducted which information on their effects is very important in managing soil erosion. Therefore, this study was conducted to assess and model soil erodibility in two adjacent sites in Behbahan region (Khuzestan province). At one site afforestation and contour furrowing were conducted to control soil erosion and the other site without any controlling measures was considered as control. Totally 150 soil samples were collected from the surface layer (0-5 cm) and K was estimated using the methods introduced by Wischmeier and Smith (method A) and Vaezi et al. (method B). For spatial modelling of K, based on digital soil mapping (DSM) techniques, several environmental covariates were derived from a Landsat 8 image and a digital elevation model (DEM) and two models including random forests (RF) and artificial neural networks (ANN) were employed. The values of K for methods A and B varied from 0.025 to 0.087 and 0.002 to 0.008 t.ha.h/ha.Mj.mm with means of 0.067 and 0.006 t.ha.h/ha.Mj.mm, respectively. Results revealed good correlation between K and remotely sensed covariates. Although K (method B) had significant correlation with some of the covariates derived from DEM, but there was no significant correlation between K (method A) with all covariates derived from DEM. Results indicated a significant difference between two sites it terms of K estimated by method A, while there was no significant difference in case of K estimated by method B. Model validation showed that both RF and ANN models resulted in good and unbiased estimates of K (methods A and B). In general, the findings indicated, although the performance of DSM techniques in modeling K were high, performances of the models and the results of means compassion may significantly differ in terms of the method through which K is estimated.