Soil science
Hamid Reza Matinfar; M. Jalali; Z. Dibaei
Abstract
Introduction: Understanding the spatial distribution of soil organic carbon (SOC) is one of the practical tools in determining sustainable land management strategies. Over the past two decades, the use of data mining approaches in spatial modeling of soil organic carbon using machine learning techniques ...
Read More
Introduction: Understanding the spatial distribution of soil organic carbon (SOC) is one of the practical tools in determining sustainable land management strategies. Over the past two decades, the use of data mining approaches in spatial modeling of soil organic carbon using machine learning techniques to investigate the amount of carbon to soil using remote sensing data has been widely considered. Accordingly, the aim of this study was to investigate the feasibility of estimating soil organic matter using satellite imagery and to assess the ability of spectral and terrestrial data to model the amount of soil organic matter.Materials and Methods: The study area is located in Lorestan province, and Sarab Changai area. This area has hot and dry summers and cold and wet winters and the wet season starts in November and ends in May. A total of 156 samples of surface soil (0-30 cm) were collected using random sampling pattern. Data were categorized into two categories: 80% (117 points) for training and 20% (29 points) for validation. Three machine learning algorithms including Random Forest (RF), Cubist, and Partial least squares regression (PLSR) were used to prepare the organic soil carbon map. In the present study, auxiliary variables for predicting SOC included bands related to Lands 8 OLI measurement images, and in order to reduce the volume of data, the principle component analysis method (PCA) was used to select the features that have the greatest impact on quality.Results and Discussion: The results of descriptive statistics showed that soil organic carbon from 0.02 to 2.34% with an average of 0.56 and a coefficient of variation of 69.64% according to the Wilding standard was located in a high variability class (0.35). According to the average amount of soil organic carbon, it can be said that the amount of soil organic carbon in the region is low. At the same time, the high value of organic carbon change coefficient confirms its high spatial variability in the study area. These drastic changes can be attributed to land use change, land management, and other environmental elements in the study area. In other words, the low level of soil organic carbon can be attributed to the collection of plant debris and their non-return to the soil. Another factor in reducing the amount of organic carbon is land use change, which mainly has a negative impact on soil quality and yield. In general, land use, tillage operations, intensity and frequency of cultivation, plowing, fertilizing, type of crop, are effective in reducing and increasing the amount of soil organic carbon. Based on the analysis of effective auxiliary variables in predicting soil organic carbon, based on the principle component analysis for remote sensing data, it led to the selection of 4 auxiliary variables TSAVI, RVI, Band10, and Band11 as the most effective environmental factors. Comparison of different estimation approaches showed that the random forest model with the values of coefficient of determination (R2), root mean square error (RMSE) and mean square error (MSE) of 0.74, 0.17, and 0.02, respectively, was the best performance ratio another study used to estimate the organic carbon content of surface soil in the study area.Conclusion: In this study, considering the importance of soil organic carbon, the efficiency of three different digital mapping models to prepare soil organic carbon map in Khorramabad plain soils was evaluated. The results showed that auxiliary variables such as TSAVI, RVI, Band 10, and Band11 are the most important variables in estimating soil organic carbon in this area. The wide range of soil organic carbon changes can be affected by land use and farmers' managerial behaviors. Also, the results indicated that different models had different accuracy in estimating soil organic carbon and the random forest model was superior to the other models. On the other hand, it can be said that the use of remote sensing and satellite imagery can overcome the limitations of traditional methods and be used as a suitable alternative to study carbon to soil changes with the possibility of displaying results at different time and space scales. Due to the determination of soil organic carbon content and their spatial distribution throughout the region, the present results can be a scientific basis as well as a suitable database and data for the implementation of any field operations, management of agricultural inputs, and any study in sustainable agriculture with soil properties in this area. In general, the results of this study indicated the ability of remote sensing techniques and random forest learning model in simultaneous estimation of soil organic carbon location. Therefore, this method can be used as an alternative to conventional laboratory methods in determining some soil characteristics, including organic carbon.
E. Mehrabi Gohari; H.R. Matinfar; Ruhollah Taghizadeh-Mehrjardi; A. Jafari
Abstract
Introduction: Soil texture is the most important environmental variable because it plays a very important role in reducing the quality of land and water transfer processes, soil quality control and fertility. On the one hand, soil texture components are the basis of environmental predictive models and ...
Read More
Introduction: Soil texture is the most important environmental variable because it plays a very important role in reducing the quality of land and water transfer processes, soil quality control and fertility. On the one hand, soil texture components are the basis of environmental predictive models and digital mapping of soil and on the other hand, soils are temporally and spatially variable, thus distinguish zoning and their monitoring with traditional sampling methods and laboratory analysis is very costly and time consuming. As a result, the development of methods for analyzing the soil and for required information has become very important. Visible and near infrared spectroscopy (VIS-NIR) is widely used to estimate soil physical properties and estimate soil texture. The present study aims to predict soil texture using spectral measurements and artificial neural network models and partial least squares regression.
Materials and Methods: The study area in southeastern Iran is approximately 70 km from Kerman. In the study area, based on the hypercube technique, 115 profiles were identified and then horizons were sampled. In this way, for each point of study, the necessary information, including the location of the profile on the ground, the type of geomorphic unit and the type of materiel, were recorded and taken from the horizons of each profile. In all soil samples, after drying and passing through 2 mm soil, the soil texture was measured by hypercube. Spectral radiometer was used to measure the spectral reflection of soil samples. The soil samples were air dried and sieved and then placed in a petri dish with an approximate diameter of 10 cm and transferred to the dark room for spectral analysis. Each specimen was tested four times (for each 90 degree sequential rotation) to remove the effects of a change in the radiation geometry. Soil samples were scanned, and absolute reflections at a spectral range of 2500-350 nm yielded 2150 spectral data points (SDPs) per soil sample with a spectral resolution of one nanometer. Finally, to construct a suitable model for forecasting the percentage of clay, sand, and silt, the least squares model was used with the number of factors 1 to 10 by Artificial Neural Network (ANN) modeling using JMP software Work.
Results and Discussion: The reflectance spectrum of the visible range - near infrared - was measured for specimens. Since preprocessing of spectral data has an effective role in improving the calibration, in order to perform spectral preprocessing, two first nodes of the first and the end of the spectra were first removed in the range of 350-400 and 2450-2500 nm. In addition, the interruption due to the change in the detector in the range of 900 to 1000 nm was also eliminated. Types of preprocessing methods were performed on spectral data. Then, using partial least squares regression analysis, the best model was produced when the first derivative was fitted to reflection values. The explanation coefficients for this low and unacceptable model were obtained. Therefore, using partial least squares regression analysis, the best wavelengths were selected to predict the percentage of clay, sand, soil, and extracted from the model. Then it was used as input in the neural network model. To determine the best combination, root error index and error coefficient were used. The results of artificial neural network showed that the number of neurons 9.8 and 10 had the best composition for predicting clay, sand and soil silt. The root-squared error results for clay, sand, and soil silt were 3.42, 6.94, and 4.383 respectively. Also, the results of the explanatory factor were 0.84, 0.83 and 0.81, respectively. After obtaining the optimal structure in the artificial neural network training phase described above, the trained network has been tested on the test data to determine the accuracy of this model to predict clay, sand and silt of surface soil. The root-squared error results for clay, sand and silt components were obtained at 5.54.9.14 and 7.01. Also, the results of the explanatory factor were 0.76.0.70 and 0.73 respectively. The best result of the prediction for partial least squares regression was obtained for the sand sample. The results indicate that the neural network performance is better than partial least squares regression, which is consistent with Mouazenet. al (2010) and also ViscarraRossel R. et. al (2009). Acceptable performance of the artificial-neural network can be attributed to the ability of this model for non-linear behavior of soil texture in visible spectroscopy. In this study, specific wavelengths, which Ben Finder et al. (2003) obtained in the study on the soils of Israel, were used. This conclusion confirms that various types of soil can be modeled using specific wavelengths. The advantage of this study is that, when using the artificial neural network, no pre-processing of reflection data is required before applying the model. Since the relationship between the percentage of soil particles (clay and gravel) and the reflection of the soil is not linear, the neural network method is very useful for analyzing the relationship between soils. Finally, the map of clay, sand and silt and map of soil texture was prepared by artificial neural network method in GIS environment.
Conclusion: The results of this study showed that the neural-dynamic network has a better performance than partial least squares regression. Calibration models designed and used in this study can be transported for use with other soils. When the partial least squares regression model was implemented, it had a very low accuracy (R2 ~ 0.1-0.3); on the contrary, the neural network-based method had high accuracy and less error. Note that although neural-dynamic modeling estimates higher precision results from soil texture, both approaches depend on wavelength selections, and so wavelengths should be selected before using any of the two models. To be finally, a meaningful relationship between the selected wavelengths and the percentage of clay, sand and silt in the present study indicates that soil texture is not only possible but also reliable by reflection spectroscopy.
