Three-Dimensional Chlorophyll Content Retrieval in Forest Canopies via Physics Guided Transfer Learning and UAV Multi-Source Remote Sensing

Accurately retrieving the vertical heterogeneity of chlorophyll content (Chl) is crucial for under-standing the physiological functions of forest canopies. In this study, we proposed a physical-guided transfer learning (PGTL) framework to achieve three-dimensional (3D) Chl retrieval. The approach integrates multispectral and LiDAR observations from unmanned aerial vehicles (UAVs) with a radiative transfer model (RTM) simulation. The PGTL is pre-trained on RTM simulated data to incorporate the source domain knowledge from the RTM simulations, and then fine-tuned using labeled field measurement data to adapt to real canopy conditions. The results of comparative experiments showed that the model performance of PGTL outperforms that of the pure data-driven method, LUT-based RTM inversion, and the hybrid model. We also investigated the three-dimensional Chl distribution of a Ginkgo plantation with three different tree ages. The results indicated that the shifts of Chl from a bell-shaped vertical pattern in young stands to a descending vertical pattern in older stands, and highlighted the consistent differences between sunlit and shaded leaves in vertical Chl distribution. These findings demonstrated the close coupling between lighting conditions, canopy structure, and biochemical characteristics, under-scoring the significance of considering canopy vertical heterogeneity when evaluating tree physiological status and ecosystem function.