PRISMA SOC, pH and CCE Soil Properties Retrieval Using GPR algorithm with pre-treated datasets: Italy case studies

Abstract

Soil organic carbon (SOC), pH and calcium carbonate content have an important role in the availability of nutrients for plants. Estimation of soil pH, calcium carbonate equivalent (CCE) and SOC (in the case of low variation) using satellite multi and hyperspectral data is still a challenging issue. Hyperspectral data acquired by new-generation spaceborne imagers like PRISMA and EnMAP offer new opportunities to accurately quantify soil properties. In this research the capability of Gaussian Process Regression (GPR) algorithm for SOC, pH and CCE retrieval from different pre-treated PRISMA spectra has been evaluated. To cover a wide topsoil variability, three different study areas in Italy were selected: Jolanda di Savoia (Lat. 44.87°N, Lon. 11.97°E), Maccarese (Lat. 41.87°N, Lon. 12.22°E) and Pignola (Lat. 40.56°N, Lon. 5.76°E). Soil samples were collected according to a 30 m squares elementary sample unit scheme and the SOC (n=635, min =0.19%, max=6.4%, std=1.55), CCE (n=518, min=0%, max=15.1, std=4.614) and the pH (n=460, min=5.035, max=8.075, std=0.769) was measured. The pH values of the samples show a -0.57 and 0.55 correlation with SOC and CCE, respectively. An overall total of 46 clear sky PRISMA images, acquired between 2019 and 2023, were used for this study. The L2D images were co-registered by the AROSICS algorithm which uses the Sentinel-2 image acquired at the closest date, to assure the co-registration (of about 0.5pixel of RMS). Noisy spectral bands and those affected by atmospheric water absorption in PRISMA images were removed, leaving a total of 173 spectral bands. The spectra were smoothed using a Savitzky-Golay filter (SG) with a second-order polynomial and a filter length of 7. To minimize the impact of the soil moisture (SM) effects, the spectra of 198 soil samples, at different SM levels, were acquired in our laboratory using a FieldSpec 4 spectroradiometer and then resampled to the PRISMA bands to be used for developing the external parameter orthogonalization (EPO) of the reflectance. A Principal Component Analysis (PCA) was also applied on the pre-treatments of the reflectance dataset (i.e., reflectance, first derivative reflectance, and EPO-projected reflectance). The first 10 PCs were selected and used for training the GPR Machine Learning (ML) models. A k-fold (k=10) cross-validation method was applied for SOC, pH and CCE modelling. The results indicate that optimal performance is achieved for SOC (R2=0.84, RMSE=0.618%) and CCE (R2=0.70, RMSE=2.527%) by employing the first derivative of EPO-projected reflectance. In the case of pH, the use of reflectance yields the most favorable outcomes (R2=0.72, RMSE=0.411). Improving the accuracy in estimating the SOC, pH and CCE soil properties, which are critical components of productive soils, is very important to allow for an efficient allocation of resources, agricultural management, and the maintenance of fertile soils for an optimal crop growth and many other purposes. Future work will include a much wider range of soil types in different soil moisture conditions.