Crown-BERT: Individual Tree Species Classification from UAV LiDAR and Hyperspectral Data

Crown-BERT achieved overall accuracies of 83.1% to 90.8% on three UAV datasets for classifying individual tree species. The lightweight deep learning framework, proposed by Xin-Bo Wang, uses fused hyperspectral and LiDAR data and was last updated in May 2026. It introduces crown-morphology-aware techniques like dynamic crown masking and self-supervised pre-training to improve accuracy and reduce manual annotation.

Use Cases

• Forest inventory and biodiversity assessment based on individual tree species classification.
• Developing transferable deep learning models for remote sensing based on self-supervised pre-training with unlabeled crown samples.
• Analyzing interannual spectral variation in forests based on cross-year applicability demonstrated on the MS-2021 and MS-2022 datasets.
• Benchmarking lightweight hybrid architectures for multimodal data fusion based on the comparison with CNN and Vision Transformer baselines.

Strengths

• Achieved overall accuracies of 83.1%, 85.4%, and 90.8% on three distinct UAV datasets.
• The model has only 0.9 million parameters, indicating a lightweight design.
• Outperformed standard CNN and Vision Transformer baselines by 17.9% to 23.5% in overall accuracy.
• Demonstrated robustness under different ecological conditions and interannual spectral variation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset size is 85.1 MB, which is relatively small and may limit the scope of analysis.

Provenance

Source

figshare

Collection Method

Data likely collected via unmanned aerial vehicle (UAV) equipped with hyperspectral and LiDAR sensors.

Freshness

Last updated 2026-05-15 04:55:02; freshness should be verified.