No abstract

Abstract. The application of visual and near infrared soil spectroscopy (vis–NIR) is an easy and cost-efficient way to gain a wide variety of soil information to cover high spatial and temporal resolution in large-scale soil surveys and in local field-scale studies. However, unlike for conventional methods, the prediction accuracy of vis–NIR spectral models cannot yet be estimated before the data collection, which hampers its application at the local scale where often a high precision is required (e.g., field experiments). In this study we used soil data from six agricultural fields in Eastern Switzerland and calibrated i) field-specific (local) models and ii) general models (combining all fields) for organic carbon, total carbon, total nitrogen, permanganate oxidizable carbon and pH using partial least squares regression. 24 out of 30 local models showed an accurate or even excellent performance (ratio of performance to deviation (RPD) > 2) and the root mean square errors (RMSE) of prediction were, except for pH, maximum five times higher than the lab measurement error. The variability of a specific soil property and the mean carbonate concentration in the dataset were the two factors influencing the performance of the local models. We found a significant relationship between the coefficient of variation in the dataset and the metrics for model performance (R2, percental RMSE and RPD). Starting from a tolerable prediction error for the spectral measurements, the regressions can be used to develop a sampling design that matches the corresponding target variability. The five inaccurately performing local models with RPD < 2 were on the two fields with highest carbonate content raising the question if local vis–NIR models are suitable for soils with high carbonate concentration. General models combining the datasets from all six fields showed an accurate overall performance but the RMSE on the field level were higher compared to the local models.

No abstract