Hyperspectral Retrieval of Winter Wheat Leaf Area Index and Canopy Chlorophyll Content: Evaluating Simulation Design and Measured-Data Adaptation

How retrieval performance varies with simulation settings, neural architecture, and measured-data adaptation remains unclear for hyperspectral crop trait retrieval. Here, we evaluated winter wheat leaf area index (LAI) and canopy chlorophyll content (CCC) retrieval using two simulated datasets and four dual-branch neural architectures. Models were pretrained on two PROSAIL-D datasets with default and previously optimized leaf-angle settings and then adapted separately to ground and unmanned aerial vehicle (UAV) observations. Evaluation used five repetitions with physical-plot grouping, with approximately 30% of plot groups allocated to measured-data adaptation. Simulation-trained random forest (RF) and partial least squares regression served as baselines without measured-data adaptation. Adaptation reduced mean test root mean square error (RMSE) in all 32 neural model–dataset–target–platform combinations relative to direct simulation-based prediction. Dataset II, which used the previously optimized leaf-angle setting, reduced mean RMSE in 13 of 16 neural configurations before adaptation and 11 of 16 afterwards. No neural architecture was uniformly best. RF achieved the highest mean R2 in seven of eight simulation-only tasks and outperformed the adapted Dataset II CNN–Transformer on both UAV targets. Despite adaptation gains, all adapted neural configurations retained negative mean R2 for UAV CCC. These findings support measured-data adaptation within the tested protocol, while showing that the effects of simulation design and architecture are task dependent and that relative improvements do not necessarily imply adequate absolute predictive accuracy.