Assessing Plant Functional Type Contributions to Carbon Modeling in Mixed Forest Ecosystems Using High-Resolution Satellite Data

Temperate mixed forest ecosystems consist of diverse plant functional types (PFTs) that exhibit variations in phenology and physiological responses to climate change. Consequently, the traditional big-leaf assumptions in carbon modeling have been criticized for oversimplifying these ecosystems, for they overlook the variability in PFT composition and their sensitivity to climate within these ecosystems. However, incorporating PFT composition into carbon and climate sensitivity simulations in heterogeneous mixed forest ecosystems presents two major challenges: (1) accurate fine-scale PFT composition mapping across large forest landscapes remains lacking, which further leads to (2) incomplete assessments of these fine-scale PFT contributions in interpreting ecosystem-scale carbon dynamics and climate sensitivity response. The recent increase in high-resolution satellite and ground observation data offers an unprecedented opportunity to resolve these challenges.

To address the first challenge, we developed a novel approach integrating Fisher-transformation-based unmixing analysis with time-series spectral and radar data. We examined this approach in three representative temperate mixed landscapes in the northeastern United States, using time-series Sentinel-1 and -2 data for calibration and local airborne-derived PFT fraction maps for validation. Our results demonstrate that (1) the synergy of spectral and radar time-series features significantly improves accuracy compared to spectral time-series models; (2) optimized features based on the Fisher-transformation approach minimize within-PFT variability and maximize between-PFT variability, enhancing model generalizability across landscapes. Integration of this approach with Google Earth Engine enables accurate ecoregion-wise PFT fractional mapping. 

To address the second challenge, we integrate different levels of PFT-related characteristics (e.g., PFT fraction map, PFT-specific physiology approximated by satellite vegetation index) with a machine learning-based carbon modeling scheme, examining how these PFT-related characteristics and climate variables separately and jointly determined the net ecosystem carbon exchange (NEE) in real mixed forest ecosystems. Specifically, we used the CHEESEHEAD19 dataset, which includes the world's most densely distributed eddy-covariance (EC) flux towers (13+ towers) within a 10 km × 10 km domain in the mixed forest ecoregion of Wisconsin, US, providing half-hourly flux records. Daily, 3-meter resolution, gap-free maps of vegetation index (NIRv) were calculated using the PlanetScope surface reflectance product. Our results demonstrated that PFT-related characteristics play a significant role (˜50%) in interpreting half-hourly NEE dynamics, with PFT-specific NIRv playing a dominant role (~30%), followed by PFT-fraction (~20%). Furthermore, by partitioning PFT-related effects, our results reveal distinct NEE sensitivity responses to specific environmental variability within and between PFTs in the 10 km × 10 km forest landscapes. 

Collectively this work advances the mapping of PFT composition and highlights the importance of integrating these fine-scale forest compositions into carbon modelling and climate sensitivity assessments, particularly in heterogeneous temperate mixed ecosystems.