Retrieval of Vegetation Biophysical Variables from Top-of-Atmosphere Radiance Data

The retrieval of vegetation biophysical variables from satellite and airborne optical data usually takes place after atmospheric correction (AC). However, it is also possible to retrieve them directly from top-of-atmosphere (TOA) radiance data if algorithms account for variability in the atmosphere. In this context, hybrid methods are of interest for building more efficient and robust retrieval models as they combine the advantages of physically-based radiative transfer models (RTMs) with the flexibility of machine learning (ML) regression algorithms. For this reason, this Thesis aimed to develop Gaussian process regression (GPR)-based hybrid models for the retrieval of essential crop traits applicable to Sentinel-2 (S2) TOA data (L1C product). Another objective was the optimization of the GPR models and the further integration into GEE for large-scale mapping. To achieve this, GPR and variational heteroscedastic GPR (VHGPR) models were trained on a look-up table (LUT) of TOA radiance data and associated input variables simulated using the coupled leaf-canopy-atmosphere RTM PROSAIL-6SV. They were also trained with a bottom-of-atmosphere (BOA) LUT from PROSAIL. The models were then applied to cloud-free S2 L1C (TOA) and L2A (BOA) reflectance products for mapping leaf chlorophyll content, leaf water content, leaf dry matter content, fractional vegetation cover, leaf area index (LAI), and upscaled leaf biochemical compounds, i.e., canopy chlorophyll content, canopy water content, and canopy dry matter content. VHGPR delivered superior accuracies and lower uncertainties than GPR for LAI estimations at BOA and TOA scales when validated against an in situ dataset over Marchfeld agricultural site. Validation using the Munich-North-Isar site showed consistent performance between BOA and TOA, with canopy-level variables outperforming leaf-level variables. Accuracy increased using TOA data instead of BOA data. A successful reduction of the training dataset by 78% was achieved by applying the Active Learning technique Euclidean distance-based diversity (EBD). The optimized EBD-GPR models demonstrated highly accurate validation results for LAI and upscaled leaf variables against the MNI site, with normalized root mean square errors (NRMSE) from 6% to 13%. However, when validated against an independent dataset from the Italian Grosseto site, the models showed moderate-to-good performances for canopy-level variables, with NRMSE ranging from 14% to 50%, but failed for leaf-level estimates. Finally, using GEE, the TOA-based mapping was applied over the entirety of Germany at 20 m spatial resolution including information about prediction uncertainty. In summary, this Thesis demonstrated that essential crop traits can be retrieved from TOA radiance data, thus avoiding the critical AC step. The optimization and integration of ML-based hybrid models in a cloud-based computing platform proved automated fast mapping directly from S2 TOA data from local to large-scale.