Retrieval of vegetation leaf equivalent water thickness from GF-5 hyperspectral imagery via physics-guided transfer learning

Vegetation leaf equivalent water thickness (EWT) is an important indicator of plant water status and is relevant to ecological monitoring and vegetation stress assessment. However, hyperspectral retrieval of leaf EWT remains challenging because canopy structural effects and discrepancies between simulated and measured observations can limit model transferability. To address this issue, we developed a physics-guided transfer-learning framework for EWT retrieval using Gaofen-5 Advanced Hyperspectral Imager (GF-5 AHSI) data. A diverse simulation domain was constructed by coupling PROSPECT-D with the 4SAIL and 4-Scale canopy models to represent both relatively homogeneous and structurally complex vegetation. Complementary PCA-derived and physically interpretable spectral features were selected, and four base learners were integrated using out-of-fold stacking. Models were first pretrained on simulated data and subsequently fine-tuned using in-situ observations. Simulation-only pretraining achieved RMSEs of 0.083 kg/m2 on the held-out measured test set and 0.091 kg/m2 on the independent external test set. After fine-tuning, the corresponding RMSEs decreased to 0.056 and 0.071 kg/m2, respectively. Moreover, on the independent external test set, the fine-tuned ensemble reduced RMSE from 0.101 to 0.071 kg/m2 relative to measured-only modeling, further supporting improved transferability. Regional applications over Saihanba and an area within and adjacent to the Northeast China Tiger and Leopard National Park produced spatially coherent EWT patterns, while ensemble model disagreement varied among regions and vegetation types. These results demonstrate that physics-guided simulation pretraining combined with measured-data fine-tuning provides a practical approach for improving the transferability of hyperspectral leaf-EWT retrieval under heterogeneous vegetation conditions.