Evaluation of the potential of the current and forthcoming multispectral and hyperspectral imagers to estimate soil texture and organic carbon

- The capability of seven satellite imagers to estimate soil variables was investigated.
- Satellite data (resampled and simulated) were obtained from soil spectral libraries.
- The effect of spectral resolution and SNR was evaluated comparing PLSR models.
- Future hyperspectral imagers provided better results than current imagers.
- Soil properties estimation from hyperspectral sensors can be limited by the low SNR.

In this study the capabilities of seven multispectral and hyperspectral satellite imagers to estimate soil variables (clay, sand, silt and organic carbon content) were investigated using data from soil spectral libraries. Four current (EO-1 ALI and Hyperion, Landsat 8 OLI, Sentinel-2 MSI) and three forthcoming (EnMAP, PRISMA and HyspIRI) satellite imagers were compared. To this aim, two soil spectra datasets that simulated each imager were obtained: (i) resampled spectra according to the specific spectral response and resolution of each satellite imager and (ii) resampled spectra with declared or actual noise (radiometric and atmospheric) added. Compared with those using full spectral resolution data, the accuracy of Partial Least Square Regression (PLSR) predictive models generally decreased when using resampled spectra. In the absence of noise, the performances of hyperspectral imagers, in terms of Ratio of Performance to Interquartile Range (RPIQ), were generally significantly better than those of multispectral imagers. For instance the best RPIQ for sand estimation was obtained using EnMAP simulated data (2.56), whereas the outcomes gained using multispectral imagers varied from 1.56 and 2.28. The addition of noise to the simulated spectra brought about a decrease of statistical accuracy in all estimation models, especially for Hyperion data. Although the addition of noise reduced the performance differences between multispectral and hyperspectral imagers, the forthcoming hyperspectral imagers nonetheless provided the best RPIQ values for clay (2.16-2.33), sand (2.10-2.17), silt (2.77-2.85) and organic carbon (2.48-2.51) estimation. To better understand the impact of spectral resolution and signal to noise ratio (SNR) on the estimation of soil variables, PLSR models were applied to resampled and simulated spectra, iteratively increasing the bandwidth to: 10, 20, 40, 80 and 160 nm. Results showed that, for a bandwidth of 40 nm, i.e., a spectral resolution lower than that of current and forthcoming imagers, the estimation accuracy was very similar to that obtained with a higher spectral resolution.Forthcoming hyperspectral imagers will therefore improve the accuracy of soil variables estimation from bare soil imagery with respect to the results achievable by current hyperspectral and multispectral imagers, however this improvement is still too limited, to allow an accurate quantitative estimation of soil texture and SOC. This work provides useful indications about what could be expected, for the estimation of the most important soil variables, from the next generation of hyperspectral satellite imagers.