Sampling Strategies for Soil Property Mapping Using Multispectral Sentinel-2 and Hyperspectral EnMAP Satellite Data

Designing a sampling strategy for soil property mapping from remote sensing imagery entails making decisions about sampling pattern and number of samples. A consistent number of ancillary data strongly related to the target variable allows applying a sampling strategy that optimally covers the feature space. This study aims at evaluating the capability of multispectral (Sentinel-2) and hyperspectral (EnMAP) satellite data to select the sampling locations in order to collect a calibration dataset for multivariate statistical modelling of the Soil Organic Carbon (SOC) content in the topsoil of croplands. We tested different sampling strategies based on the feature space, where the ancillary data are the spectral bands of the Sentinel-2 and of simulated EnMAP satellite data acquired in Demmin (north-east Germany). Some selection algorithms require setting the number of samples in advance (random, Kennard-Stones and conditioned Latin Hypercube algorithms) where others automatically provide the ideal number of samples (Puchwein, SELECT and Puchwein+SELECT algorithm). The SOC content and the spectra extracted at the sampling locations were used to build random forest (RF) models. We evaluated the accuracy of the RF estimation models on an independent dataset. The lowest Sentinel-2 normalized root mean square error (nRMSE) for the validation set was obtained using Puchwein (nRMSE: 8.7%), and Kennard-Stones (9.2%) algorithms. The most efficient sampling strategies, expressed as the ratio between accuracy and number of samples per hectare, were obtained using Puchwein with EnMAP and Puchwein+SELECT algorithm with Sentinel-2 data. Hence, Sentinel-2 and EnMAP data can be exploited to build a reliable calibration dataset for SOC mapping. For EnMAP, the different selection algorithms provided very similar results. On the other hand, using Puchwein and Kennard-Stones algorithms, Sentinel-2 provided a more accurate estimation than the EnMAP. The calibration datasets provided by EnMAP data provided lower SOC variability and lower prediction accuracy compared to Sentinel-2. This was probably due to EnMAP coarser spatial resolution (30 m) less adequate for linkage to the sampling performed at 10 m scale.