Hyperspectral remote sensing data for retrieving crop biophysical variables and topsoil properties in agricultural fields

Effective agricultural management hinges on the comprehensive monitoring of both crops and soil. This is because soil health, encompassing nutrient levels, structure, and pH balance, fundamentally dictates crop performance. Consequently, a synergistic approach to monitoring both is essential. This thesis investigates the combined application of the Italian Space Agency's (ASI) PRISMA (PRecursore IperSpettrale della Missione Applicativa) hyperspectral satellite and machine learning (ML) to analyse key agronomic variables and soil properties. A central objective is to leverage multi-temporal PRISMA data to identify the driest bare soil pixels. This focused approach aims to enhance the training of ML algorithms, ultimately boosting prediction accuracy across the broader dataset. Furthermore, this research will examine soil moisture’s influence on retrieving soil variables. Accurate retrieval of vegetation biophysical parameters also depends critically on the parameter distributions within the PROSAIL radiative transfer model. While machine learning inversion methods using look-up tables can lessen the sensitivity to these distributions, they do not entirely resolve the issue. To address this, this study will employ domain adaptation techniques, utilising a hybrid dataset of both simulated and real-world data. This innovative strategy seeks to decouple the performance of machine learning training from inherent assumptions about parameter distributions within the radiative transfer model. Soil Organic Carbon (SOC) is shown to be significantly affected by soil moisture, and Clay content exhibits some sensitivity. In contrast, Sand and Silt appear less responsive to moisture variations. Notably, SOC estimation using Savitzky-Golay first derivative applied to absorbance and Gaussian Process Regression (GPR) showed a performance decline (R² = 0.64, RMSE = 1.22, RPD = 1.67) compared to dry soil conditions using a Standard Normal Variate (SNV) applied to absorbance with GPR (R² = 0.73, RMSE = 0.73, RPD = 1.97), underscoring moisture's impact. Domain adaptation techniques, particularly TrAdaBoostR2, were evaluated for PROSAIL model inversion. While acknowledging TrAdaBoostR2's need for in-situ calibration, the study highlights the effectiveness of baseline detrending with spline interpolation and Partial Least Squares Regression (PLSR) on Leaf Area Index (LAI) retrieval. Without domain adaptation, this combination achieved an R² of 0.65, an RMSE of 0.89, and an RPD of 1.70. Crucially, with domain adaptation, performance significantly improved to R² of 0.95, an RMSE of 0.34, and an RPD of 4.46., demonstrating the substantial benefit of this approach for PRISMA data analysis in agricultural monitoring.