1.School of Computer Science and Artificial Intelligence,Northeast Forestry University,Harbin 150040,China
2.School of Information and Intelligence Engineering,University of Sanya,Sanya 572000,China
3.College of Electrical Engineering and Information,Northeast Agricultural University,Harbin 150040,China
Show less
文章历史+
Received
Published
2026-02-01
2026-05-20
Issue Date
2026-06-11
PDF (5458K)
摘要
森林优势树种分类是区域森林资源精细监测与生态评估的重要基础。在中等分辨率遥感场景下,像元混合与边界过渡效应显著,混交区域类别判别不稳定。同时,训练样本中不同树种类别的样本占比存在明显差异,样本占比偏低的类别更易出现漏分、空间破碎与边界漂移等问题。为此,面向光学数据、雷达数据和地形因子构建多模态输入,提出一种少数类增强多模态混合专家网络(minority-class mitigation mixture-of-experts network,M-MoENet)。该方法设计轻量层级融合模块,对不同模态进行专用卷积编码并逐级融合,获得统一特征表示。在编码器前馈子层引入稀疏混合专家多层感知机模块(mixture-of-experts multilayer perceptron,MoE-MLP),并结合置信度调节门控实现自适应专家激活。同时,引入面向低占比类别的训练约束,以缓解主导类偏置并提升少数类的有效学习能力。试验结果表明,M-MoENet总体性能最优,总体精度(overall accuracy,OA)、平均交并比(mean intersection over union,mIoU)和Kappa系数分别达到79.46%、60.08%和0.74,并在常绿松与椴树等少数类上取得更显著提升。该方法能够在多模态和类别样本比例不均衡条件下提升优势树种分类精度,为复杂林区优势树种分类提供有效方法。
Abstract
Dominant tree species classification is an important basis for fine-scale forest resource monitoring and ecological assessment at the regional scale. In medium-resolution remote sensing scenarios, mixed pixels and boundary transition effects are prominent, resulting in unstable class discrimination in mixed forest areas. Meanwhile, the training data exhibit substantial imbalance in the sample proportions of different tree species categories, and categories with lower sample proportions are more prone to omission errors, spatial fragmentation, and boundary drift. To address these issues, this study constructs a multimodal input framework integrating optical data, radar data, and topographic factors, and proposes a multimodal sparse mixture-of-experts network, termed M-MoENet. The proposed method designs a lightweight hierarchical fusion module to perform modality-specific convolutional encoding and progressive fusion, thereby obtaining unified feature representations. A sparse MoE-MLP is introduced into the feed-forward sublayer of the encoder, and confidence-adjusted gating is employed to achieve adaptive expert activation. In addition, a training constraint oriented toward low-proportion categories is incorporated to alleviate dominant-class bias and improve the effective learning of minority classes. Experimental results show that M-MoENet achieves the best overall performance, with OA, mIoU, and Kappa reaching 79.46%, 60.08%, and 0.74, respectively, and yields more pronounced improvements for minority classes such as evergreen pine and linden. These results demonstrate that the proposed method can improve the classification accuracy of dominant tree species under multimodal conditions and class-imbalance settings, providing an effective solution for dominant tree species classification in complex forest regions.
基于时间序列数据,对生长季多时相观测进行像元中位数合成,获取代表性生长季合成影像。在此基础上提取6个光谱波段特征即红、绿、蓝、近红外、短波红外1和短波红外2波段,并进一步计算归一化植被指数(normalized difference vegetation index,NDVI)、增强植被指数(enhanced vegetation index,EVI)、土壤调节植被指数(soil-adjusted vegetation Index,SAVI)。同时,基于逐时相指数序列逐像元提取物候指标生长季开始(start of season,SOS)、生长季结束(end of season,EOS)、生长季长度(length of season,LOS),并引入同期Sentinel-1 SAR散射特征,即垂直发射-垂直接收极化(vertical transmit and vertical receive,VV)和垂直发射-水平接收极化(vertical transmit and horizontal receive,VH)及其极化比值特征(VV/VH)与DEM派生地形因子海拔、坡度和坡向,共同构成模型输入的多通道特征集合。
为定量评估M-MoENet在优势树种分类中的性能,本研究在测试集上构建混淆矩阵,并据此计算各类F1-score(F1)、总体精度(Overall Accuracy,OA)、平均交并比(mean intersection over union,mIoU)以及Kappa系数(Kappa)。F1用于衡量类别表现,OA反映整体正确率,mIoU衡量预测与真实分布的一致性,Kappa在随机一致性基础上修正整体精度。
LIUS Y, WENY B, WUJ W,et al.Classification of dominant tree species in Tieling based on Sentinel-2 multi-dimensional feature fusion[J].Forest Engineering,2025,41(6):1101-1115.
[3]
LIANGX L, KUKKOA, BALENOVIĆI,et al.Close-range remote sensing of forests:The state of the art,challenges,and opportunities for systems and data acquisitions[J].IEEE Geoscience and Remote Sensing Magazine,2022,10(3):32-71.
[4]
LIUX L, GAOL H, JIAC,et al.Classification of broad-leaved forest tree species based on HSFC-DeepLabV3+ in UAV multispectral images[J].IEEE Access,2025,13:174517-174534.
[5]
XUS C, YANGB H, WANGR R,et al.Single tree semantic segmentation from UAV images based on improved U-Net network[J].Drones,2025,9(4):237.
[6]
TANJ, LIJ, MAT Y,et al.Leveraging Sentinel-1/2 time series and deep learning for accurate forest tree species mapping[J].Frontiers in Forests and Global Change,2025,8:1599510.
LIUY T, LIUZ J, CAOC H,et al.The study of point cloud tree species classification based on local geometric priors and Transformer[J].Science of Surveying and Mapping,2025,50(8):92-100.
[9]
苗帆.基于多源遥感的城市行道树树种识别与应用[D].杨凌:西北农林科技大学,2024.
[10]
MIAOF.Urban street tree species identification and application based on multi-source remote sensing[D].Yangling:Northwest A&F University,2024.
CAIY L, WANGX L, GAOH Z,et al.Tree species classification method based on multi-modal remote sensing data combined with 3DCNN and vision transformer[J].Laser & Optoelectronics Progress,2025,62(20):250-260.
[13]
NEZAMIS, KHORAMSHAHIE, NEVALAINENO,et al.Tree species classification of drone hyperspectral and RGB imagery with deep learning convolutional neural networks[J].Remote Sensing,2020,12(7):1070.
[14]
CHENS D, WANGX F, SHIM M,et al.Developing interpretable deep learning model for subtropical forest type classification using Beijing-2,Sentinel-1,and time-series NDVI data of Sentinel-2[J].Forests,2025,16(11):1709.
[15]
GUOH Y, BOONPRONGS, WANGS H,et al.Dominant tree species mapping using machine learning based on multi-temporal and multi-source data[J].Remote Sensing,2024,16(24):4674.
[16]
吴梦可.基于多源遥感影像的优势树种分类方法研究[D].哈尔滨:东北林业大学,2024.
[17]
WUM K.Research on dominant tree species classification method based on multi-source remote sensing images[D].Harbin:Northeast Forestry University,2024.
[18]
HUANGY M, OUB T, MENGK X,et al.Tree species classification from UAV canopy images with deep learning models[J].Remote Sensing,2024,16(20):3836.
[19]
ILLARIONOVAS, TREKINA, IGNATIEVV,et al.Tree species mapping on sentinel-2 satellite imagery with weakly supervised classification and object-wise sampling[J].Forests,2021,12(10):1413.
[20]
SADAIYANDIJ, ARUMUGAMP, SANGAIAHA K,et al.Stratified sampling-based deep learning approach to increase prediction accuracy of unbalanced dataset[J].Electronics,2023,12(21):4423.
[21]
QIT, ZHUH, ZHANGJ,et al.Patch-U-Net:Tree species classification method based on U-Net with class-balanced jigsaw resampling[J].International Journal of Remote Sensing,2022,43(2):532-548.
[22]
SHAHEENL, RASHEEDB, MAZZARAM.Tree species detection using hyperspectral and Lidar data:A novel self-supervised learning approach[J].Computer Research and Modeling,2024,16(7):1747-1763.
[23]
MOURETF, MORIND, PLANELLSM,et al.Tree species classification at the pixel level using deep learning and multispectral time series in an imbalanced context[J].Remote Sensing,2025,17(7):1190.
[24]
LEEH, LEEC,WOO H,et al.Optimal training sample sizes for U-Net-Based tree species classification with Sentinel-2 imagery[J].Forests,2025,16(11):1718.
[25]
WANGG Y, CHENJ H, MOL F,et al.Lightweight land cover classification via semantic segmentation of remote sensing imagery and analysis of influencing factors[J].Frontiers in Environmental Science,2024,12:1329517.
[26]
BELOIUM, HEINZMANNL, REHUSHN,et al.Individual tree-crown detection and species identification in heterogeneous forests using aerial RGB imagery and deep learning[J].Remote Sensing,2023,15(5):1463.
[27]
MARTINSG B, LA ROSAL E C, HAPPP N,et al.Deep learning-based tree species mapping in a highly diverse tropical urban setting[J].Urban Forestry & Urban Greening,2021,64:127241.
[28]
JINGW P, LOUT T, WANGZ Y,et al.A rigorously-incremental spatiotemporal data fusion method for fusing remote sensing images[J].IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,2023,16:6723-6738.
[29]
DIAKOGIANNISF I, WALDNERF, CACCETTAP,et al.ResUNet-a:A deep learning framework for semantic segmentation of remotely sensed data[J].ISPRS Journal of Photogrammetry and Remote Sensing,2020,162:94-114.
[30]
CHENL C, ZHUY, PAPANDREOUG,et al.Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of the European Conference on Computer Vision (ECCV).2018:801-818.
[31]
LIUJ R, YANGH, ZHOUH Y,et al.Swin-UMamba:Mamba-based U-Net with ImageNet-based pretraining[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention.Cham:Springer Nature Switzerland,2024:615-625.
[32]
CAOH, WANGY, CHEN JY,et al.Swin-Unet:Unet-Like pure transformer for medical image segmentation[C]//Computer Vision - ECCV 2022 Workshops.Cham:Springer Nature Switzerland,2022:205-218.
[33]
XIEE Z, WANGW H, YUZ D,et al.SegFormer:Simple and efficient design for semantic segmentation with transformers[J].Advances in Neural Information Processing Systems,2021,34:12077-12090.
[34]
HEMMERLINGJ, PFLUGMACHERD, HOSTERTP.Mapping temperate forest tree species using dense Sentinel-2 time series[J].Remote Sensing of Environment,2021,267:112743.
[35]
WANGX L, WANGJ, LIANZ Z,et al.Semi-supervised tree species classification for multi-source remote sensing images based on a graph convolutional neural network[J].Forests,2023,14(6):1211.
[36]
CHENP D, LIUY, RENY R,et al.A deep learning-based solution to the class imbalance problem in high-resolution land cover classification[J].Remote Sensing,2025,17(11):1845.