The fraction of absorbed photosynthetically active radiation (FAPAR) is a critical parameter for measuring the vegetation photosynthetic capacity. The Hi-resolution Global LAnd Surface Satellite (Hi-GLASS) FAPAR product (version 1, V1) from Landsat imagery has been successfully applied to the ecosystem productivity modeling; however, this product algorithm still exhibits some limitations, including the poor adaptability to heterogeneous surfaces and limited physical interpretability, due to the absence of real-world knowledge guidance. To address these issues, we integrated deep transfer learning and radiative transfer models to update the Hi-GLASS FAPAR algorithm and generate the corresponding product (i.e., version 2, V2). A long short-term memory (LSTM) model was pre-trained on Soil-Leaf-Canopy (SLC) simulations and then optimized using physical knowledge-guided transfer learning, which was used to generate the new FAPAR product from Landsat image series. Validation results demonstrated that the Hi-GLASS FAPAR V2 (R2 = 0. 95, RMSE = 0.08) significantly outperformed V1 (R2 = 0.94, RMSE = 0.11), with notable improvements in various vegetation categories and sensors. The greatest improvement of FAPAR was found over multiple forest types, where different forest categories showed substantial gains, with R2 increasing by 2 % - 11 % and RMSE decreasing by 15 % - 55 %, confirming the improved adaptability of our proposed method to heterogeneous canopies. Moreover, the Hi-GLASS V2 product preserved better spatial details than MODIS,GLASS,GEOV2 products, and its temporal dynamics were more closely aligned with field measurements than the V1 product. These advancements highlight the potential of Hi-GLASS FAPAR V2 as valuable data for supporting terrestrial ecosystem studies.