نوع مقاله : مقاله پژوهشی
نویسندگان
1 گروه علوم خاک-دانشکده ی مهندسی آب و خاک-دانشگاه علوم کشاورزی و منابع طبیعی گرگان
2 گروه علوم خاک، دانشکده مهندسی آب و خاک، دانشگاه علوم کشاورزی و منابع طبیعی گرگان، ایران.
3 تحقیقات خاک و آب، مرکز تحقیقات و آموزش کشاورزی و منابع طبیعی گلستان، گرگان
4 گروه علوم خاک، دانشکده مهندسی آب و خاک، دانشگاه علوم کشاورزی و منابع طبیعی گرگان. ایران.
چکیده
کلیدواژه‌ها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
There is an increasing need for continuous spatial and quantitative soil information for environmental modeling and management, especially at the national scale. This study was conducted to predict the soil particle size fraction (PSF) using the combination of quantile regression forest model (QRF) and spline function in a part of Golestan province. An equal area spline equation was fitted to the data of 105 soil profiles from the database of the Gorgan University of Agricultural Sciences and Natural Resources for estimating PSFs at five soil depths (0-25, 25-50, 50-75, 75-100, and 100-125 cm). The primary auxiliary variables in this research included 22 environmental variables derived from DEM, 15 remote sensing indicators obtained from the Landsat 7 ETM+ images, rainfall and piezometric maps. Based on principal component analysis (PCA), 15 variables were selected and entered into the modeling process of soil texture components (clay, sand, and silt). The efficiency of the quantile regression forest model was evaluated using the coefficient of determination (R2), the root mean squared error (RMSE), and the mean absolute error (MAE). The results indicated that the coefficient of determination for clay, silt, and sand at different depths varied from 0/12 to 0/22, 0/07 to 0/30, and 0/07 to 0/28, respectively. Also, the relative importance of environmental variables showed that rainfall (thirty-year average), piezometry (ten-year average), B3/B7, and valley depth were the most important factors in predicting soil texture components. To improve model performance and validation results, some structural uncertainties in this study should be addressed.
کلیدواژه‌ها [English]
Preparation of three-dimensional maps of soil particle size fractions by combining quantile regression forest algorithm and spline depth function in Golestan Province
EXTENDED ABSTRACT
Soil texture, which includes three components of sand, clay, and silt, is one of the important physical characteristics of soil that strongly affects many other soil properties, such as water retention curve, fertility, drainage, organic carbon content, and porosity. Knowledge of the spatial variability of soil properties, including soil texture, is essential in precision agriculture because any change in the spatial distribution of physical and chemical soil properties causes changes in crop yield. Digital soil mapping (DSM) techniques apply the analysis of soil spatial relationships, terrain features, and remote sensing images to environmental modeling. Soil properties in general vary continuously with depth in a soil profile. Spline functions are very efficient in modeling soil attribute depth functions. Combining spline functions and digital mapping is a suitable approach for the 3D modeling soil properties. In recent years, many digital mapping studies have been conducted in Iran. However, to our knowledge, no study has been conducted on depth functions to investigate soil properties in Golestan Province. Therefore, this study aims to estimate soil texture components using the spline function at the target depths, spatial modeling of soil texture components with the quantitative regression forest model (QRF), and evaluate the contribution of environmental variables for modeling in a part of the north of Golestan Province.
The data from 105 profiles of the soil information bank of the University of Agricultural Sciences, which were collected between 2013 and 2015, were used in this research. Equal-area quadratic smoothing splines were used to describe the vertical variation of soil particle size fractions. The next step was to select the most useful of the 39 predictor ancillary variables to reduce the dimensionality and allow the QRF algorithm to operate more effectively. Here, principal component analysis (PCA) was used to rank the relevance of auxiliary variables. 15 auxiliary variables including rainfall, piezometry, carbonate index, clay index, B3/B7, NDVI, MrRTF, MrVBF, morphometric feature, landform, valley depth, flow width, surface wetness index, annual insolation, and slope were selected to predict soil texture components. Twenty-fold cross-validation was used to evaluate the performances of the QRF algorithm. To determine the accuracy of the model, three different criteria were used based on the coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE).
The validation results of the QRF learning algorithm in predicting the spatial changes of soil texture components showed that the coefficient of determination for clay ranged from 0/12 to 0/22, sand from 0/07 to 0/28, and silt from 0/07 to 0/30. According to the findings, auxiliary variables contributed differently to predicting soil texture components at different depths. Based on the relative importance of the variables, the distribution of clay and silt was more affected by rainfall and groundwater depth. The higher rainfall in the southern part of the study area increased silt weathering and clay accumulation. In the northern lowlands of the study area, impermeable clay layers were formed due to the accumulation of fine-textured sediments from the Atrak and Gorganrud Rivers, which affected the piezometric factor. However, piezometric in the southern regions is a measure of the groundwater depth, and its spatial distribution pattern is consistent with the spatial distribution pattern of clay in the south. The maximum amount of silt in the study area was found in the middle regions, mainly due to the floodplain and sedimentary plain of the Gorganrud River. The distribution of sand was more influenced by the two environmental factors of valley depth and B3/B7. The pattern of spatial changes of these two indicators is in line with the spatial distribution of sand.
Based on the findings of this research, the amounts of clay, sand, and silt had a strong relationship with rainfall and groundwater levels in the study area. The quantile regression forest algorithm showed poor performance in predicting soil particle size in the study area. Data augmentation is effective in reducing uncertainty and enhancing model accuracy.
)