Abstract: In recent years, artificial neural networks (ANNs) trained on data simulated by radiative transfer models (RTMs) have demonstrated substantial potential for estimating vegetation biophysical parameters. However, substantial biases persist between ANN-based estimations and actual values because of model simplifications, discrepancies between simulated and realworld scenarios, and uncertainties in input parameters. In this study, a transfer learning-enhanced inversion method is proposed for retrieving leaf area index (LAI) from Landsat surface reflectance data by integrating a physical model with LAI ground measurements. Using a deep belief network (DBN) as the base model, the optimal hyperparameters are first determined through a Bayesian optimization algorithm (BOA); then, the DBN is pretrained using a dataset simulated by the PRO4SAILT model to construct a source model. All the parameters of the source model are subsequently fine-tuned to construct a fine-tuned model using carefully quality-controlled sparse LAI ground measurements provided by the implementing multiscale agricultural indicators exploiting sentinel (IMAGINES) and validation of European remote sensing (RS) instruments (VALERI) projects. During the fine-tuning process, regularization constraints are introduced to mitigate parameter overfitting, and a robust 10-fold crossvalidation (CV) method is used to verify the performance of the fine-tuned model. The fine-tuned model is then applied to retrieve LAI values from Landsat surface reflectance data. The results demonstrate excellent agreement between the retrieved LAI values and those from the LAI reference maps from the IMAGINES and VALERI projects, with performance significantly outperforming the source model and SNAP software. This method eectively combines the advantages of RTM simulations and LAI ground measurements, oering a viable physical-modelbased solution to mitigate the domain shift that occurs between simulated data and satellite observations.

Citation: Kexin Lv, Zhiqiang Xiao, Juan Li, Yajie Zheng, Jinling Song, Hua Yang, and Hanyu Shi. Transfer Learning-Enhanced Leaf Area Index Retrieval From Landsat Surface Reflectance Data, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 64, 2026, 4413718.    

论文链接: https://ieeexplore.ieee.org/document/11614862/