Zhuohong's talk at the American Geophysical Union fall meeting 2025

Zhuohong’s abstract: Forests are critical ecosystems that regulate climate, store carbon, support biodiversity, and sustain economies. In the United States, forests cover roughly one-third of the land area and sequester nearly 15% of the annual greenhouse gas emissions. The USDA Forest Inventory and Analysis (FIA) program reports more than 400 tree species across the country, each playing vital ecological roles. Accurate, high-resolution mapping of tree species distribution is essential to understanding forest dynamics and informing conservation and management efforts. Recent advances in remote sensing and deep learning have enhanced the speed, scale, and cost-effectiveness of forest monitoring. However, challenges persist in integrating heterogeneous data and overcoming the scarcity of labeled data needed for species-level classification. Most existing approaches are constrained to detecting basic forest attributes, such as tree height, crown extent, and functional type. In this study, we present a semi-supervised deep learning framework that integrates high-resolution hyperspectral imagery, orthorectified aerial photography, LiDAR point clouds, and ground data to perform large-scale, individual tree-level species classification. Unlike traditional bounding-box methods, our framework leverages pixel-wise segmentation within tree crowns, enabling more precise identification of species and better handling of complex, overlapping crown structures. We demonstrate the application of this framework across different terrestrial sites within the National Ecological Observatory Network (NEON), spanning a wide range of biogeographical regions of the United States. This study delivers the most fine-grained, crown-level, and up-to-date assessments of tree species composition at NEON sites. Our approach offers a scalable, cost-effective solution for nationwide biodiversity monitoring and forest management.