Estimating Soil Organic Carbon Using Hyperspectral Data in Visible, Near-infrared and Shortwave-infrared (VIS-NIR-SWIR) Range

Background and objectives: Soil organic carbon is a main soil property and particularly important for development and sustainable management of agricultural systems. Soil Organic Matter content, which is typically measured in the form of soil organic carbon SOC content, is commonly regarded as a key indicator of soil quality and utilization (Liu et al., 2015). The presence of SOM has been proved to be beneficial for soil productivity, water holding capacity, and carbon sequestration (Prescott et al., 2000; Munson and Carey, 2004; Seely et al., 2010; Six and Paustian, 2014). Earlier studies showed that SOM is vulnerable to anthropogenic activities such as farming practices, and other economic development activities (Huang et al., 2007; Kissling et al., 2009; Mao, et al., 2014). Conventional laboratory analyses for measuring soil organic carbon, especially in large scale, are expensive and time consuming. For this reason, fast and accurately assess the amount of organic carbon can be a very valuable measure for long-term management of soil. The objectives of this study were: i) studying of proximal spectral reflectance of soils for estimating soil organic carbon by PLSR and bagging-PLSR methods and ii) investigating the effects of different preprocessing methods on performance of estimated soil organic carbon.
Materials and methods: A total of 200 composite soil samples on watershed scale (calibration data) and 40 soil samples on farm scale (validation data) from two different depth (0-10 and 10-30 cm) collected within a radius of 10 meters and after air drying, they were passed through 2 mm sieve. Some physicochemical characteristics of soils were measured in the laboratory. Consequently, proximal spectral reflectance of the soil samples within the VIS-NIR and SWIR (400-2500 nm) range was measured using a handheld spectroradiometer apparatus and correlation between 2000 bands and soil organic carbon were determined.
Results: Results indicated that the best preprocessing methods to calibrate PLSR model were wavelet deterending (RPD=1.94) and SNV with median filter (RPD=1.92). The best PLSR and bagging-PLSR model for the estimation was obtained with 17 factors. Bagging-PLSR method has high performance (RMSE=0.142-1.03 %) than PLSR method (RMSE=0.167-1.11 %) for estimating soil organic carbon. In both methods, the accuracy was decreased while soil organic carbon was bigger than 1.2 percent.
Conclusion: Using the soil spectral reflectance in the range of VIS, NIR and SWIR can examine the soil organic carbon. The spectral curves of different soils showed three absorbance properties at wavelengths 1414, 1913 and 2207 that was the amount of water in clay network and soil hygroscopic water, so they can be considered as a unique characteristics for each soil. These spectral bands are very important to estimate the amount of soil organic carbon. Soil spectral data pre-processing and selection of the most suitable pre-processing method was one of the most important factors affecting the accuracy of bagging-PLSR and PLSR method to estimate the amount of soil organic carbon. Based on the results, the method of bagging-PLSR showed higher accuracy than the PLSR method to estimate the amount of soil organic carbon.