Evaluation and Exploitation of Retrieval Algorithms for Estimating Biophysical Crop Variables Using Sentinel-2, Venus, and PRISMA Satellite Data

Abstract

This paper is devoted to the development and testing of the optimal procedures for retrieving biophysical crop variables by exploiting the spectral information of current multispectral optical satellite Sentinel-2 and Venus and in view of the advent of the new Sino-EU hyperspectral satellite (e.g., PRISMA, EnMAP, and GF-5). Two different methodologies devoted to the estimation of biophysical crop variables Leaf area index (LAI) and Leaf chlorophyll content (Cab) were evaluated: non-kernel-based and kernel-based Machine Learning Regression Algorithms (MLRA); Sentinel-2 and Venus data comparison for the analysis of the durum wheat-growing season. Results show that for Sentinel-2 data, Gaussian Process Regression (GPR) was the best performing algorithm for both LAI (R2=0.89 and RMSE=0.59) and Cab (R2=0.70 and RMSE=8.31). Whereas, for PRISMA simulated data the Kernel Ridge Regression (KRR) was the best performing algorithm among all the other MLRA (R2=0.91 and RMSE=0.51) for LAI and (R2=0.83 and RMSE=6.09) for Cab, respectively. Results of Sentinel-2 and Venus data for durum wheat-growing season were consistent with ground truth data and confirm also that SWIR bands, which are used as tie-points in the PROSAIL inversion, are extremely useful for an accurate retrieving of crop biophysical parameters.

Keywords

• |biophysical crop parameters|prisma|gf-5|sentinel 2|venus