A comparison of sensor resolution and calibration strategies for soil texture estimation from hyperspectral remote sensing

Imaging spectroscopy of bare soils has been shown to have considerable potential for the estimation of properties such as soil texture. However, in order to be able to fully exploit data from forthcoming hyperspectral satellites, information on several issues related to sensor spatial and spectral resolution and range, as well as on calibration and validation issues, is still required.In the present study, images acquired over bare soil in Central Italy by airborne MIVIS (430–1270nm; spatial resolution: 4.8m) and space-borne CHRIS-PROBA (415–1050nm; spatial resolution: 17m), were used to explore methods for the quantitative estimation of soil texture. Extensive soil sampling was carried out for the determination of soil particle size fractions. Soil texture was related to the spectral signature of corresponding CHRIS or MIVIS pixels. The spectral behavior of the soil samples was also examined in the laboratory, by using a spectroradiometer in the 400–2500nm range. Spectra were used to calibrate prediction models for the estimation of clay, silt and sand, through partial least-square regression (PLSR). The impact of several factors on the accuracy of estimation of soil texture was studied, such as spectral range and resolution, the effect of varying soil moisture and the geolocation error. The modality of setting up the calibration and validation data sets was also investigated, by employing either randomly selected or spatially separated datasets. The validity of the models was assessed from several statistics, such as bias, root mean square error of prediction (RMSEP) and ratio of performance to deviation (RPD).Results from laboratory data show the importance of SWIR bands to estimate clay, silt and sand fractions. The tests with remote sensing data show a sufficient accuracy of prediction (RPD>1.4) for clay and sand using both MIVIS and CHRIS-PROBA data, but results vary in response to the modality of setting up the calibration and validation sets. Results were found to be sensitive to the difference in support between point and pixel data and to the geometric registration error, especially for MIVIS data. When using a 3×3 window instead of a single pixel, RPD values as high as 1.85 for MIVIS and 1.75 for CHRIS were found.Despite the lack of SWIR bands and a lower spatial resolution, CHRIS did show a comparable potential to MIVIS in terms of accuracy and of prediction ability. This was probably a consequence of the conditions in which the images were acquired, in which the pattern of soil moisture in the field might have played a role in the discrimination of soil texture.