Crown-BERT: A crown-morphology-aware deep learning framework for individual tree species classification using UAV LiDAR and hyperspectral Data

The dataset includes the unmanned aerial vehicle (UAV) LiDAR and hyperspectral data for ten sample plots in Maoershan Forest Farm of Northeast Forestry University.
Crown-BERT key code and demo data:Code and reproducibility resources for Crown-BERT for individual tree species classification using UAV hyperspectral and LiDAR data. Crown-BERT is a deep learning framework for individual tree species classification using UAV hyperspectral and LiDAR data. It is designed for crown-level classification by jointly leveraging spectral and structural information. The framework integrates Dynamic Crown Masking (DCM), Crown Positional Encoding (CPE), and Crown Masked Pixel Modeling (CMPM) within a BERT-style architecture. This repository provides the main implementation of Crown-BERT together with the associated resources for training, evaluation, and reproducibility.
Data OrganizationThe input dataset is organized in HDF5 (.h5) format. Each HDF5 file contains four keys corresponding to the crown-level inputs and labels used by Crown-BERT:
inputs: hyperspectral patch data, with shape (N, B, H, W) and data type float32attention_mask: crown-region mask, with shape (N, H, W) and data type float32position_encoding: crown positional encoding map, with shape (N, H, W) and data type float32labels: one-hot encoded class labels, with shape (N, C) and data type float32Here, N denotes the number of crown samples, B denotes the spectral dimension, H × W denotes the spatial patch size, and C denotes the number of classes. The inputs key stores the hyperspectral crown patches, attention_mask indicates valid crown regions, position_encoding provides the crown positional prior, and labels stores the class annotations for supervised training and evaluation.
Code StructureThe repository mainly consists of modules for data loading, model definition, training, and testing.
load_data.py: functions for reading HDF5 files and preparing the crown-level inputs used by Crown-BERTmodel.py: implementation of the Crown-BERT network architecturetrain.py: training script for model optimizationtest.py: evaluation script for model testing and performance assessmentmain.py: main script for organizing the overall workflow, including data loading, model construction, training, and evaluationUsageThe environment configuration, dependency settings, and key hyperparameters used in this project are provided directly in the code. Please refer to train.py, test.py, and the related implementation files for details.