2.3.5. Soil Organic Carbon

Numerous studies have examined the prediction of soil organic carbon (SOC) via the spectral signal [17–20,60–63]. Most recent publications tend to stress the better accuracy of machine learning algorithms [13], especially when predicting more complex soil properties. For the prediction of SOC, however, linear regression models have proven to deliver robust predictions with acceptable accuracy. Since our study solely aims for the rough delineation of strata and our number of prediction variables is quite low, ordinary least squares (OLS) regression was chosen as a rather straightforward prediction model. As already pointed out by Bumgardner et al. [60], the overall reflectance in the visible part of the spectrum decreases when the organic matter content increases. For this reason, only the first four image bands of the multispectral image were used as input data for the prediction of SOC (450, 550, 680, and 720 nm). SOC measurement results were assigned to a set of training pixels, which were extracted from the image bands based on the respective sampling locations. Training pixels comprise all pixels of the digitised sample location. As the number of pixels is quite small due to the limited spatial resolution of the images, further processing of the pixel values via, for example, mean values of random subsets (as suggested for bigger data sets by, e.g., [63]) was not conducted. The fitted model was used for prediction of SOC based on all image pixels. Model fitting was performed with the stats package for R [45] and prediction using the raster package for R [55].
