Introduction
Accurately measuring the Canopy Height Model (CHM)—the vertical distance between the top of the vegetation and the ground—is a cornerstone of modern forestry. It is essential for calculating timber stocks, monitoring forest growth, and managing carbon biomass. However, foresters have traditionally faced a difficult trade-off between the high precision of local surveys and the broad coverage of satellite data.
Researchers have recently developed a deep learning workflow that bridges this gap, estimating forest canopy height at high resolution (<5m) using multispectral satellite imagery. By combining RGB and Near-Infrared (NIR) bands with digital elevation models in a U-Net architecture, this approach offers a cost-effective alternative to expensive airplane-based LiDAR or drone surveys.
The Resolution Gap in Forest Monitoring
Historically, forest monitoring has relied on three primary methods, each with significant drawbacks. Field measurements are highly accurate but prohibitively slow and expensive for large-scale applications. LiDAR and Unmanned Aerial Vehicles (UAVs) provide excellent precision but are limited by high operational costs and logistical difficulties in remote regions. Conversely, standard satellite imagery offers global coverage but typically produces coarse resolutions (often >30m per pixel), which are too blurry to capture the structural nuances of a forest canopy.
This study aims to fill this "resolution gap" by leveraging very high-resolution (VHR) satellite data and modern computer vision to produce 1m-to-5m resolution height maps, balancing the need for broad coverage with the requirement for actionable detail.
Methodology: Deep Learning for Height Estimation
The core of the study is a deep learning approach utilizing a U-Net architecture equipped with an Inception-ResNet-v2 encoder. The technical workflow involves several key innovations:
- Input Fusion: The model incorporates standard RGB colors, a Near-Infrared (NIR) band, and the ArcticDEM (a digital elevation model). This provides the network with both spectral signatures and topographical context.
- Pansharpening: To maximize spatial detail, the authors used pansharpening techniques to increase the resolution of multispectral bands to 1 meter.
- Weighted Loss Function: To account for the uneven distribution of tree heights (where very tall trees are rarer), the researchers implemented a Weighted RMSE loss. This prevents the model from simply "averaging out" predictions and ensures accuracy across all height classes.
- Regression vs. Classification: The team evaluated both continuous height prediction (regression) and categorical sorting (classification) to determine the most robust output format.
Data and Study Area
The research focused on the Russian boreal forests in the Arkhangelsk region, a vast area characterized by complex vegetation structures.
The dataset integrated multiple sources:
- Satellite Imagery: WorldView-2 (VHR) and Sentinel-2 data.
- Topographic Data: 2m-resolution ArcticDEM.
- Ground Truth: High-density airplane-based LiDAR measurements (1m resolution) and official forest inventory documents covering approximately 50,000 hectares.
- Public Basemaps: The model was also tested on free, cloud-free RGB orthophotomaps from Mapbox to evaluate its feasibility as an open-access solution.
Key Results and Insights
The best-performing configuration (U-Net + Inception-ResNet-v2 + RGB + NIR + ArcticDEM) achieved a Mean Absolute Error (MAE) of 2.47 meters. Several critical insights emerged from the analysis:
- Texture over Color: At 1m resolution, the CNN extracts "spatial features"—texture and patterns—that are more informative for height estimation than raw spectral values.
- The NIR Advantage: Adding the NIR band significantly reduced error, as it is highly sensitive to vegetation health and density.
- Scalability: When applied to free Mapbox RGB data, the model still achieved a respectable MAE of 3.5 meters, suggesting that the workflow is viable even for budget-constrained projects.
- Downstream Benefits: Using the generated height maps as an additional feature for tree species classification (conifer vs. deciduous) improved accuracy to an F1-score of 0.94.
Limitations and Future Work
While promising, the model faces geographic constraints; the ArcticDEM used here does not cover southern latitudes, requiring alternative elevation models for global deployment. Additionally, shadows on steep slopes and slight co-registration shifts between satellite and LiDAR data remain challenges. Future research will focus on adapting the model to tropical and temperate rainforests with more diverse canopy structures.
Summary updated