Crown-BERT: a crown-morphology-aware deep learning framework for individual tree species classification using UAV LiDAR and hyperspectral data
Date:2026-06-10 clicks:
Impact Factor:6.9
DOI number:10.1080/15481603.2026.2671600
Affiliation of Author(s):东北林业大学
Journal:GIScience & Remote Sensing (Q1)
Place of Publication:英国（牛津，Taylor & Francis 出版社总部）
Key Words:Individual tree speciesclassification; deep learning;BERT; LiDAR; hyperspectral
Abstract:Accurate individual tree species classification using fused unmanned aerial vehicle(UAV) hyperspectral (HSI) and light detection and ranging (LiDAR) data is fundamentalfor forest inventory and biodiversity assessment, yet remains challenging because ofirregular crown morphology, limited species annotations, and the high model complex-ity induced by high-dimensional multimodal features. To address these challenges, wepropose Crown-BERT (Bidirectional Encoder Representations from Transformers), alightweight, crown-morphology-aware deep learning framework for crown-levelHSI–LiDAR classification. Crown-BERT introduces dynamic crown masking (DCM) andcrown positional encoding (CPE) to explicitly encode valid canopy boundaries andintra-crown spatial structure, and employs crown masked pixel modeling (CMPM) as aself-supervised pre-training strategy to learn transferable feature representations fromabundant unlabeled crown samples. These components are integrated into a task-specific lightweight hybrid architecture that combines efficient convolutional opera-tions with transformer-based global modeling, thereby reducing parameter redundancywhile preserving classification performance. Under independent training and testing onthree UAV datasets, Crown-BERT achieved overall accuracies of 83.1%, 85.4%, and 90.8%on MS-2021, MS-2022, and TH-2024 dataset, respectively, with only 0.9 million parame-ters. It outperformed standard CNN and Vision Transformer baselines by 17.9% to 23.5%in overall accuracy, and further exceeded representative HSI–LiDAR fusion-based clas-sificaition models, improving overall accuracy by 5.1%–5.9% over hierarchical CNN andtransformer (HCT) and by 9.7%–11.2% over 3D-CNN across the three datasets. Resultsfrom MS-2021 and MS-2022 indicate that the proposed framework maintained stableperformance under interannual spectral variation, while the strong performance on TH-2024 further demonstrates its robustness under different ecological conditions; inaddition, transfer-based adaptation with AdaBN further improved cross-year applicabil-ity. Therefore, Crown-BERT provides an efficient and morphology-aware solution forindividual tree species classification in UAV-based forest monitoring under complex andvariable stand conditions, with strong potential for improving classification accuracyand reducing manual annotation.
Co-author:Yang Zhao,Yu Han,Yinghui Zhao
First Author:Xinbo Wang
Indexed by:Journal paper
Correspondence Author:Zhen Zhen*
Document Code:2671600
Discipline:Engineering
Document Type:J
Volume:63
Issue:1
Page Number:2671600
ISSN No.:1947-9345（印刷版）/ 1548-1603（电子版）
Translation or Not:no
Date of Publication:2026-01-01
Included Journals:SCI