Canopy Height Maps v2

What Canopy Height Maps v2 Does

Canopy Height Maps v2 (CHMv2) is an open-source computer-vision model for estimating vegetation canopy height from high-resolution satellite imagery. In practical terms, it examines an image of a landscape and predicts how tall the vegetation is across the scene. Those predictions can be assembled into a raster map, where each location contains an estimated canopy-height value.

The model was developed by Meta AI in partnership with the World Resources Institute. It belongs to a specialized geospatial and remote-sensing part of Meta’s model catalog rather than its conversational or generative-AI offerings. CHMv2 is intended for researchers, conservation groups, governments, forestry teams, and developers building geographic information or environmental-analysis workflows.

Its primary output is not a caption, report, or chat response. It is a spatial prediction that can be analyzed alongside other geographic data. This distinction matters when evaluating the model: CHMv2 is useful because it converts satellite imagery into a measurable environmental layer, not because it provides general-purpose intelligence.

How the Model Works

CHMv2 uses a DPT-based depth-estimation decoder and a frozen DINOv3 ViT-L/16 backbone. A vision transformer, or ViT, processes an image by learning relationships between visual regions. In this case, the backbone was pretrained on satellite imagery, which gives the model a representation better suited to overhead geographic scenes than a backbone trained only on ordinary photographs.

The decoder converts the visual features into a dense prediction corresponding to canopy height. The model was trained against high-quality lidar examples. Lidar provides reference measurements that can be used to teach the system how visible satellite patterns relate to the three-dimensional structure of vegetation.

The canonical model identifier is facebook/dinov3-vitl16-chmv2-dpt-head. The model is available through Meta AI’s DINOv3 repository and the Hugging Face Hub, with support in the Transformers ecosystem according to the supplied model documentation. Access to the model weights follows Meta’s model-access process.

Accuracy and Training Improvements

Meta reports that CHMv2 improves substantially on the earlier canopy-height mapping system. The reported coefficient of determination, or R², increased from 0.53 for the earlier system to 0.86 for CHMv2. R² is a statistical measure of how much of the variation in reference measurements is explained by model predictions; it is useful for understanding reported evaluation performance, but it should not be treated as a guarantee of accuracy for every location or image.

The updated system uses a more geographically diverse collection of lidar training examples and automated matching methods intended to align satellite observations with real-world canopy-height measurements. Meta also describes a specialized loss function and improvements aimed at reducing bias in predictions for tall trees.

These changes are especially relevant for global mapping. A model trained on limited geographic conditions may perform unevenly when applied to different forest types, climates, image sources, or viewing conditions. More diverse training data can improve consistency, although the supplied research still identifies geographic coverage and viewing geometry as areas where further improvement is needed.

Inputs, Outputs, and Supported Modalities

CHMv2 accepts imagery rather than natural-language prompts. It does not independently explain a map, search the web for supporting information, call external functions, or produce a conversational summary. A larger application could use a separate geographic-processing or language system after CHMv2 has generated the canopy-height layer, but those capabilities are not part of this model itself.

The supplied specifications do not publish a context window, maximum output-token limit, or conventional context-length limit. Those fields are primarily relevant to language models. For CHMv2, practical limits are more likely to depend on the image dimensions, preprocessing pipeline, available memory, and the implementation used for inference; no exact universal image-size or output-size limit is established in the supplied research.

Main Use Cases

• Forest structure mapping: Estimate the height distribution of vegetation over large areas and identify differences between forest stands.
• Restoration monitoring: Compare canopy development in restored or reforested areas and support assessment of whether vegetation is becoming established.
• Forest degradation analysis: Detect areas where canopy structure has been reduced, subject to appropriate validation and temporal data.
• Carbon estimation: Use canopy-height information as one input to models that estimate biomass or carbon storage. CHMv2 does not itself provide a complete carbon estimate.
• Biodiversity and habitat planning: Use vegetation structure as a spatial indicator when analyzing habitat conditions.
• Urban tree planning: Support analysis of urban tree height, shade, and cooling-planning initiatives.
• Forest inventories: Provide a remotely sensed layer for national, regional, or project-level inventories.
• Climate and land management: Add canopy structure to broader geographic analyses of land use, conservation, and environmental change.

These applications generally require additional geographic processing, calibration, and validation. A canopy-height prediction is an input to an analysis rather than a complete answer to questions about forest health, carbon, or biodiversity.

Strengths and Limitations

Key Strengths

• Specialized satellite-image design: The DINOv3 backbone was pretrained on satellite imagery, matching the model’s intended data domain.
• Dense spatial output: The model produces location-specific canopy-height predictions rather than only classifying an entire image.
• Global mapping orientation: Its design and training are intended to support canopy mapping across geographic regions.
• Open-source availability: Researchers and developers can access the implementation and model weights through Meta’s repository and the Hugging Face Hub, subject to the applicable access process and DINOv3 License.
• Reported improvement over the earlier system: Meta reports an increase in R² from 0.53 to 0.86, along with better handling of bias for tall trees.

Important Limitations

• Not a general-purpose AI model: It cannot replace a language model for question answering, writing, coding, or document analysis.
• No hosted pricing or standard API is documented: The supplied research identifies local or research use, but no official hosted inference price is published for this exact model.
• Geographic variation remains important: Performance may differ in areas that are underrepresented in the training data.
• Viewing geometry can affect results: Satellite acquisition conditions and image geometry may introduce errors that require local evaluation.
• Limited temporal coverage for change detection: CHMv2 can contribute to monitoring workflows, but the supplied research does not establish a complete time-series change-detection system.
• Predictions are not automatically field truth: Results should be checked against local measurements or other reliable reference data before being used for high-consequence decisions.

The reported R² value is a provider-reported research result, not an unconditional accuracy guarantee. Local image quality, forest type, season, geography, and preprocessing can all affect the usefulness of the output.

Pricing and Access

No official hosted API pricing has been published for CHMv2. The model is positioned for local or research use, with weights distributed through Meta’s access process and the Hugging Face Hub. This means the direct model price is not a recurring per-image or per-token service price in the supplied documentation.

Using CHMv2 locally can still create infrastructure costs. Users may need suitable hardware, storage for satellite imagery, preprocessing tools, and geographic-analysis software. Those costs depend on the size of the imagery and the scale of the mapping project, but the research does not provide a standardized amount, so no numeric estimate should be assumed.

The model is released under the DINOv3 License. Teams should review the license and access requirements before incorporating the weights into a commercial or operational system.

Reasoning, Coding, and Tool Capabilities

CHMv2 does not perform reasoning in the language-model sense. It applies a learned image-to-raster prediction process, but it does not explain its conclusions, follow multi-step natural-language instructions, or deliberate over competing hypotheses.

It also does not provide coding assistance, function calling, web search, streaming responses, or agent actions. Developers can integrate it into a Python or Transformers-based workflow, but the surrounding software is responsible for loading imagery, preparing inputs, storing outputs, validating predictions, and performing subsequent analysis.

When to Choose Canopy Height Maps v2

Choose CHMv2 when the central task is estimating vegetation or tree-canopy height from satellite imagery and you need a model that can be run as part of a geospatial workflow. It is a particularly good fit for research teams that need a reproducible model, organizations creating regional or global forest layers, and developers who can manage local inference and geographic validation.

Its reported performance improvement and satellite-specific backbone make it more appropriate for this task than a generic image classifier or a general multimodal chat model. A general-purpose vision model may describe an image, but that does not mean it can produce a calibrated, dense canopy-height raster.

Another option may be more appropriate when the requirement is conversational analysis, text generation, coding, web research, or a managed hosted API with published usage pricing. A conventional remote-sensing platform may also be preferable when the priority is an end-to-end production service, operational monitoring, or integrated access to imagery and geographic tools rather than direct control of model inference.

For carbon accounting, regulatory reporting, or decisions affecting land management, CHMv2 should generally be combined with local reference data and domain-specific validation. Its output can be valuable evidence, but it is not presented as a complete forestry, biodiversity, or carbon-accounting solution.

Bottom Line

Canopy Height Maps v2 is a focused Meta AI research model for turning high-resolution satellite imagery into canopy-height maps. Its strongest distinction is specialization: it uses a satellite-pretrained DINOv3 ViT-L/16 backbone and a DPT decoder for a geospatial prediction task, rather than attempting to serve as a general-purpose AI assistant.

For users who can run an open-source model locally and validate results, CHMv2 offers a practical foundation for forest mapping, restoration assessment, and ecological analysis. Its lack of hosted pricing, conversational features, published context limits, and guaranteed performance across every geography means it should be evaluated as a research and geospatial component, not as a turnkey general AI service.