UAV LiDAR and Hyperspectral Data for Crown-Level Tree Species Classification

Crown-BERT dataset contains UAV LiDAR and hyperspectral data from ten sample plots at Maoershan Forest Farm. The data is structured in HDF5 files with crown-level inputs, masks, positional encodings, and one-hot encoded labels for individual tree species classification. Xinbo Wang published this resource in April 2026.

Use Cases

• Train models like Crown-BERT to classify tree species from `inputs` hyperspectral patches and `labels` using spectral and structural features.
• Develop crown segmentation algorithms using the `attention_mask` data to isolate individual tree regions from background.
• Experiment with positional encoding techniques for remote sensing by utilizing the provided `position_encoding` maps.
• Benchmark multimodal fusion methods by combining the hyperspectral data (`inputs`) with structural information derived from the accompanying LiDAR data.
• Perform few-shot or transfer learning for forestry tasks using the pre-processed crown samples (`N` samples) and their one-hot encoded `labels`.

Strengths

• Data is pre-processed into crown-level samples with four structured keys (inputs, attention_mask, position_encoding, labels) in HDF5 format.
• Includes paired high-resolution data sources: UAV LiDAR for structure and hyperspectral imaging for spectral signatures across ten forest plots.
• Released with full reproducibility resources including model code, training scripts, and evaluation scripts under a permissive CC-BY-4.0 license.

Limitations

• Sample size is limited to ten geographic plots, which may not represent the full diversity of forest conditions.
• The dataset's total number of crown samples (`N`) is not explicitly stated, making it difficult to assess model training scale.
• Data is from a single forest farm (Maoershan), introducing potential geographic and ecological bias.

Provenance

Source

Xinbo Wang, associated with Northeast Forestry University.

Collection Method

Collected using unmanned aerial vehicles (UAVs) equipped with LiDAR and hyperspectral sensors.

Freshness

Last updated in April 2026.

Geography

Maoershan Forest Farm, Northeast Forestry University, China.