A hybrid PROSAIL inversion framework for winter wheat LCC using hyperspectral data and transfer learning

Leaf chlorophyll content (LCC) is a key indicator for assessing the photosynthetic capacity and nutritional status of winter wheat. Among traditional LCC estimation methods, empirical models lack a physical basis and have poor generalisability, while physical models are widely applicable but suffer from ill-posed inversion problems. Hybrid inversion methods, which integrate radiation transfer models such as PROSAIL with machine learning, offer both the interpretability of physical models and the efficiency of machine learning; however, they are still affected by the domain shift between simulated and measured data, which limits their generalisation performance. Transfer Component Analysis (TCA), a domain adaption method, can effectively alleviate this problem. In this study, hyperspectral and LCC data were collected in the field, and simulated data were generated using the PROSAIL model; a sensitivity analysis was conducted to identify LCC-sensitive bands. A genetic algorithm was applied to the measured data for band selection and, together with the results of the sensitivity analysis, yielded an optimal set of 30 characteristic bands for subsequent modelling. Three datasets were constructed: measured data only, a direct mixture of measured and simulated data, and a TCA-fused mixture of measured and simulated data. Four models-gradient boosting regression (GBR), random forest (RF), support vector regression (SVR) and deep neural network (DNN)-were developed for each dataset. The results show that: (1) the LCC-sensitive bands are concentrated in the 450-660 nm and 680-720 nm ranges; (2) the model built on the TCA-fused data (R² = 0.722, RMSE = 6.792) outperformed those built on the measured-only data (R² = 0.682, RMSE = 7.259) and the directly mixed data (R² = 0.616, RMSE = 7.976); (3) for the TCA-fused data, the four models differed considerably in accuracy, with SVR performing best (R² = 0.723, RMSE = 5.363), followed by RF (R² = 0.630, RMSE = 6.201) and GBR (R² = 0.575, RMSE = 6.650), whereas the DNN performed worst (R² = 0.388, RMSE = 7.947), probably owing to the limited sample size.