To address the issues of low accuracy and high computational cost in hyperspectral image classification tasks of grassland, this study proposed the FRCNet network model. Hyperspectral images of grasslands were captured using an unmanned aerial vehicle equipped with a hyperspectral imager, and a hyperspectral dataset for four grassland categories was established. Gaussian filtering and principal component analysis (PCA) were applied to noise reduction and dimensionality reduction of the hyperspectral images. The FRCNet network model, primarily consisting of the FRC module and the FAC module, was established for classification tasks. Average accuracy (AA), overall accuracy (OA), F1 score, and runtime were taken as performance indicators, and comparative experiments were conducted between FRCNet and eight methods. The results showed that the FRCNet network performed best with AA of 93.36%, OA of 93.49%, and F1 score of 96.64, improving accuracy by 10%-20% compared to the other methods. In addition, the comparative experiments on three public datasets demonstrated that FRCNet performed best, with accuracy improvements ranging from 2%-20%. The research results proved the effectiveness of the FRCNet network model in hyperspectral image classification tasks of grassland. It can serve as an efficient solution to the current issues of low hyperspectral accuracy and high computational cost.
草地是一种复合型的植物群落,占世界总陆地面积1/2,在保持水土、防风固沙、调节气候方面发挥重要作用[1⁃2],草地健康状况和类型识别对生态环境监测、土地利用规划和农业生产具有重要意义[3⁃4]。高光谱成像技术基于窄波段中获取物体的光谱信息,提供丰富的光谱信息和空间信息[5],在地物识别、食品安全和医疗诊断得到了广泛的应用。高光谱图像具有地物的丰富光谱信息可以作为识别草地种类的方法,如何高效识别草地物种类别成为其关键任务。传统深度学习地物分类方法,不能高效解决高光谱数据波段多、数据量大的问题[6]。皮伟强[7]创建轻量化DGC-3D-CNN网络,并利用无人机高光谱数据的空间信息来解决草地退化问题,总体精度达到98.11%;乌尼乐等[8]通过图像预处理得到三个波段,使用支持向量机(SVM)和随机森林方法进行物种识别,精度分别达到96.92%、97.32%;张杰等[9]基于残差网络结构,提出了一种密集残差卷积网络进行分类研究,李浩等[10]利用卷积神经网络[11](convolutional neural network,CNN),结合CNN与LSTM(long⁃short term memory)有效地对大规模光谱数据进行分类,准确度达到97.35%,并且在恒星光谱数据分类中表现出色。卷积神经网络作为一种深度学习方法,已被广泛应用于各种图像分类任务中[12]。
FAC block模块主要由初始化、特征融合、映射输出三大部分组成。模块数据初始化部分,使用普通卷积块对通过残差连接传入的数据进行整理计算;特征融合部分,使用特征融合的方法将Deformable与Coordinate相结合;映射输出部分,将数据形状重新映射到初始形状并输出。FAC block模块首先使用Conv2D卷积块将残差连接后的数据进行计算,计算结果分别输出到两个不同的注意力机制当中(DA与CA),后将两种不同类型的参数使用特征融合方法结合,输出到Fusion Conv中将128通道映射到64通道,最后输出到平均池化层,将结果重塑为(res.size(0),-1)形状,展平成一个向量后,输入到全连接层中使用。
ALZUBAIDIL, ZHANGJ L, HUMAIDIA J, et al. Review of deep learning: Concepts, CNN architectures, challenges, applications, future directions[J]. Journal of Big Data, 2021, 8(1): 53.
[12]
BELLOI, FEDUSW, DUX Z, et al. Revisiting ResNets: Improved training and scaling strategies[D]. New York: Cornell University, 2021.
[13]
CAIZ, DINGX, SHENQ, et al. Refconv: Re-parameterized refocusing convolution for powerful convnets[J]. arXiv preprint arXiv, 2023:2310.10563.
[14]
LIANGJ, FANY, XIANGX, et al. Recurrent video restoration transformer with guided deformable attention[J]. Advances in Neural Information Processing Systems, 2022, 35: 378-393.
[15]
ZHUX, SUW, LUL, et al. Deformable detr: Deformable transformers for end-to-end object detection[J]. arXiv preprint arXiv, 2020:2010.04159.
[16]
WANGJ, LIF, ANY, et al. Towards robust lidar-camera fusion in bev space via mutual deformable attention and temporal aggregation[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2024, 34(7):5753–5764.
[17]
DAIL, ZHANGG, ZHANGR. RADANet: Road augmented deformable attention network for road extraction from complex high-resolution remote-sensing images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 1-13.
[18]
ZHANGX, YUW, PUNM O. Multilevel deformable attention-aggregated networks for change detection in bitemporal remote sensing imagery[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-18.
[19]
ZHOUZ, ZHUX, CAOQ. AAGDN: Attention-augmented grasp detection network based on coordinate attention and effective feature fusion method[J]. IEEE Robotics and Automation Letters, 2023, 8(6): 3462-3469.
[20]
XIEC, ZHUH Y, FEIY Q. Deep coordinate attention network for single image super-resolution[J]. IET Image Processing, 2022, 16(1): 273-284.
[21]
CHENR, LIG H, DAIC L. Feature fusion via deep residual graph convolutional network for hyperspectral image classification[J]. IEEE Geoscience and Remote Sensing Letters, 2022, 19: 1-5.
[22]
MOHAMMADIZADEHM, MOZHDEHIA, IOANNOUY, et al. Meta-GCN: A dynamically weighted loss minimization method for dealing with the data imbalance in graph neural networks[J]. arXiv preprint arXiv, 2023: 2406.17073.
[23]
BOC J, LUH C, WANGD. Spectral-spatial K-Nearest Neighbor approach for hyperspectral image classification[J]. Multimedia Tools and Applications, 2018, 77(9): 10419-10436.
[24]
THECKEDATHD, SEDAMKARR R. Detecting affect states using VGG16, ResNet50 and SE-ResNet50 networks[J]. SN Computer Science, 2020, 1(2): 79.
[25]
WUQ F, CHENY P, MENGJ. DCGAN-based data augmentation for tomato leaf disease identification[J]. IEEE Access, 2020, 8: 98716-98728.