A remote sensing framework combining high-resolution multispectral imagery with synthetically generated elevation models reduced tree identification errors by up to 23.9 percent compared to classical segmentation techniques, according to a study conducted in Central Oregon. Evaluating methods for individual tree crown detection and height estimation, researchers found that deep learning object detectors consistently outperformed traditional machine learning and watershed approaches. Among the tested architectures, Mask R-CNN achieved the highest overall detection performance with an F1 score of 0.908, followed by YOLOv8 and Faster R-CNN.
The research addressed the ecological challenges posed by the expansion of invasive Western juniper (Juniperus occidentalis) across eastern Oregon landscapes. As the invasive tree species spreads across rangelands, it alters local hydrological regimes and degrades critical sage-grouse habitat. Land managers require accurate individual-tree inventories to coordinate efficient ecological restoration where resources are sparsely distributed, but conventional ground-based surveys are impractical and cost-prohibitive across broad geographic landscapes.
To overcome logistical barriers, the research team developed a scalable remote sensing workflow utilizing 0.5-meter multispectral aerial imagery obtained from the National Agriculture Imagery Program (NAIP). Because three-dimensional structural data are essential for characterizing tree crowns but are frequently unavailable across large areas, the framework paired the two-dimensional multispectral observations with synthetically generated canopy height information.
To capture structural canopy attributes where airborne LiDAR coverage is absent, the researchers trained a Generative Adversarial Network (GAN) to synthesize Canopy Height Models (CHMs). The generative neural network synthesized elevation values directly from NAIP red, near-infrared, and hillshade bands. A Kolmogorov–Smirnov test revealed no statistically significant difference between the GAN-generated canopy heights and reference LiDAR-derived Canopy Height Models, confirming the synthetic data accurately mirrored physical tree height distributions.
The study compared several analytical methodologies, assessing classical segmentation algorithms and machine learning classifiers against modern deep learning object detectors. The baseline techniques included two-dimensional and three-dimensional multi-scale watershed segmentation (MSWD), Random Forest (RF), and Support Vector Machine (SVM). When comparing the watershed approaches, three-dimensional MSWD achieved superior performance over two-dimensional MSWD, directly demonstrating the indispensable value of incorporating vertical canopy height information into individual-tree delineation.
The findings indicate that pairing accessible two-dimensional multispectral imagery with GAN-derived elevation datasets produces accurate, spatially explicit individual-tree maps across extensive rangelands. By removing the dependency on simultaneous and expensive LiDAR data acquisitions, this remote sensing methodology lowers financial barriers to baseline environmental data collection and provides resource managers with an operational, cost-effective framework for rangeland inventory and restoration planning.
Subject of Research: Remote sensing framework for individual tree detection using multispectral imagery and synthetic canopy height models
Article Title: Fusing multispectral and elevation data for tree-level landscape inventory in Central Oregon, USA
Article References: Schriver, R., Sessions, J., Paun, M., & Strimbu, B. (2026). Fusing Multispectral and Elevation Data for Tree-Level LLandscape Inventory of Individual Western Juniper in Central Oregon, USA, by Fusing Multispectral and Elevation Datain Central Oregon, USA. Frontiers in Forests and Global Change, 9, 1954252. https://doi.org/10.3389/ffgc.2026.1954252
Image Credits: AI Generated
DOI: 10.3389/ffgc.2026.1954252
Keywords: Western juniper, Juniperus occidentalis, remote sensing, multispectral imagery, National Agriculture Imagery Program, canopy height models, Generative Adversarial Network, deep learning, Mask R-CNN, YOLOv8, Faster R-CNN, multi-scale watershed segmentation, Random Forest, Support Vector Machine, Kolmogorov-Smirnov test, sage-grouse habitat, rangeland restoration, individual tree crown detection
News Source: Blake Davidson. (November 16, 2026). Deep Learning and Synthetic Canopy Heights Cut Tree Detection Errors in Oregon. Scienmag.



