Uncertainty propagation in VNIR reflectance spectroscopy soil organic carbon mapping

Visible and near infrared (VNIR) diffuse reflectance spectroscopy (DRS) offers high potential as a fast and accurate proximal soil sensing technique for soil carbon estimation. The objective of this study is to evaluate the use of VNIR soil spectroscopy for mapping soil organic carbon (SOC) spatial distribution on a 100 ha arable field strongly affected by erosion. The analysis was performed in two main steps: firstly, we focused on the uncertainty in the VNIR spectroscopy regression model (PLSR) under varying number and locations of training samples from which an optimal number of input samples were selected; secondly, we analysed uncertainty propagation in the coupled PLSR and spatial prediction for the selected optimal number of training samples.The PLSR quality parameters are changing exponentially with increasing number of input training samples. The PLSR model constructed using only 37 samples provided a good predictive capability with R2 over 0.7 and RPD over 1.5. The uncertainty of the final map as expressed by mean standard deviation, lowered 3 times when the number of input training samples changed from 37 to 128. The accuracy of the soil map was assessed through the uncertainty propagation analysis for the purpose of evaluating how the uncertainties were partially associated with the PLSR model and partially coming from the spatial prediction propagate to the final output map. We conclude that the PLSR predictions caused lower uncertainty in comparison with uncertainty coming from spatial predictions by kriging algorithm. The study confirms that the SOC prediction as made by VNIR spectral characteristics is a powerful tool, which together with digital soil mapping techniques (DSM) provides the basis for high resolution field scale mapping.

► Study evaluates the use of VNIR soil spectroscopy for mapping soil organic carbon.
► Uncertainty in the VNIR spectroscopy regression model (PLSR) was evaluated.
► PLSR model using only 37 training samples provided a good predictive capability.
► Coupled PLSR and spatial prediction was evaluated by uncertainty propagation.
► PLSR predictions caused lower uncertainty than kriging in the final map.