We’re excited to share our new publication in Remote Sensing of Environment, titled “Leveraging transfer learning and leaf spectroscopy for leaf trait prediction across biomes” (link to article), led by Fujiang.
Accurate prediction of leaf traits like chlorophyll content, carotenoids, water thickness, and leaf mass per area is essential for understanding plant function, ecosystem health, and vegetation responses to environmental change. Yet, existing approaches often fall short when applied across different regions, plant functional types, and seasons. In this study, we developed a new generation of models that leverage transfer learning to overcome these limitations. By integrating domain knowledge from radiative transfer models with data-driven learning, our hybrid models significantly outperform traditional statistical models and standalone physical models in both predictive performance and transferability.
Using extensive datasets collected across diverse sites in the U.S. and Europe, and combining pretraining on synthetic data with fine-tuning on field observations, our approach achieved higher R² values (up to 0.79) and lower error rates. These findings highlight how transfer learning can bridge the gap between physical realism and statistical flexibility, offering a powerful path forward for scalable, reliable, and interpretable leaf trait estimation.
This work represents a key step toward building robust and generalizable tools for monitoring plant functional traits at large scales, with implications for ecological research, biodiversity monitoring, and Earth system modeling. Congratulations to Fujiang and the entire team.