Verifying the predictive performance for soil organic carbon when employing field Vis-NIR spectroscopy and satellite imagery obtained using two different sampling methods

Abstract

• Failure to select an appropriate sampling method may affect the accuracy of SOC prediction. • Different sampling design methods provide different prediction accuracy. • Combined dataset precision could be affected by vulnerabilities in individual datasets. • SOC's predictive performance could be improved by combining different pre-treatments. In soil research, the most employed sampling design techniques can be categorized as random sampling (stratified or simple random (SR)) or systematic techniques (transects or grid). Many other sampling approaches have also been developed by researchers based on these sampling principles. The purpose of this study is to compare the differences in SOC prediction when using field spectra (FS) and Sentinel-2 (S2) data collected separately through SR and grid design (GD) on the same agricultural field. Additionally, the impact of spectral indices on S2 data in a merged data approach under the two-sampling strategies will also be tested. The data for each sampling method were obtained based on a previous study in which 130 soil samples were collected from a full grid design (with 40 m spacing) covering the entire area. Although the full GD method was used for this current study, the distance between the samples was increased (80 m apart). The schemes were therefore structured for the collection of 65 samples in the field for each sampling technique. However, 63 samples were collected with the GD because two of the sampling points fell on rocky areas and were eliminated accordingly. For SR sampling, the study field was not stratified, and no requirements were used for minimum sample spacing. Sixty-five samples and spectral data were collected at various locations. To achieve the mentioned objective, this study builds a five-fold cross-validation model based on support vector machines (SVMs). Different pretreatment combinations were also implemented. The results showed that the GD was better than the SR approach using the merged dataset (R 2 CV = 0.45, RMSE CV = 0.26, RPD = 1.41, bias = −0.0073); however, SOC prediction under SR sampling using FS yielded the highest accuracy and lowest error margin (R 2 CV = 0.60, RMSE CV = 0.21, RPD = 1.66, and bias = 0.0045). Despite the above-mentioned disparity between the single and merged data, this study shows that using different sampling design methods on the same field separately is a very promising approach for SOC estimation, particularly in fields with low SOC. Based on these results, the robustness of this approach should be investigated next in future studies using larger sample sizes as well as other modeling techniques. Based on these results, the robustness of this approach should be investigated next in future studies using larger sample sizes as well as other modeling techniques.