Hyperspectral PRISMA images and geophysical proximal sensing data fusion to map topsoil features: vineyard and arable land case studies

Hyperspectral images of new generation satellites, such as the PRISMA system of the Italian Space Agency (ASI), offer an important advantage for monitoring topsoil properties from the field to the regional scale. Although numerous studies about the prediction of soil features by proximal soil spectroscopy have been carried out, the hyperspectral remote sensing of soil features still shows several limits and difficulties. Disturbance of soil surface, namely grass or crop cover, surface stoniness, roughness, and soil moisture, influence the results and need corrections. In addition, the resolution of satellite hyperspectral images is rather low (PRISMA = 30 m) for many objectives.

The aim of this work is to test geostatistical techniques to group proximal electromagnetic induction and remote hyperspectral data to increase the resolution and the accuracy of topsoil texture prediction. Two types of croplands have been used for this study: i) JOL- an arable field in Northern Italy (Jolanda di Savoia, Ferrara) of about 15 ha; ii) BRO- seven vineyards of a winery in central Italy (Brolio castle, Siena), for a total surface of about 30 ha.

In JOL, three dates of PRISMA images were selected: BARE- during the bare soil period (14/2/2021), VEG1- during the summer crop, namely corn (4/6/2021), and VEG2- during the winter crop, namely wheat (30/4/2022), to obtain information about the soil surface and the response of vegetation. In BRO, only two dates of PRISMA images with scarce cloud cover were available, 18/12/2020 and 01/12/2022. In both the dates, the grapevines have no leaves and the interrow was tilled by chisel plow 2-3 weeks earlier. The weeds partially covered the grapevine rows and inter-rows.

To reduce the dimensionality of the hyperspectral data and to preserve as much as possible their information content, a principal component analysis (PCA) of the spectra extrapolated from each image pixel was carried out. The first PCs (PC1) explained most of the variance of the images, therefore, it was selected for analysis. Regarding proximal electromagnetic induction, apparent electrical conductivity of shallower depth (ECa₁, about 0-50 cm) has been used. To predict topsoil spatial variation, two geostatistical methods have been tested: i) Regression Kriging (RK), forward stepwise for p < 0.5, and ii) Multiple Geographically Weighted Regression (GWR) with gaussian weighting function.

In the study field of arable land (JOL), 36 samples were used for model calibration and cross-validation. The prediction of clay and sand was unsuitable because the spatial correlation with ECa₁ or PRISMA images was lacking. On the other hand, SOC showed spatial correlation with BARE-PC1, whereas pH with VEG2-PC1.

In the vineyards, a calibration dataset of 70 points and a validation dataset of 20 points have been used. Clay prediction showed the best results using RK with ECa₁ and PC1_2020, providing R² of 0.68 and RMSEP of 6.43 g·100g^-1. Sand prediction showed slightly better results using GWR (R² of 0.78 and RMSEP of 8.74 g·100g^-1), although PC1 of hyperspectral data did not show clear improvements in prediction. Further analyses will include additional PRISMA images acquired during the growing season.