ReluformerN：轻量化高低频增强高光谱农业地物分类方法

doi:10.12133/j.smartag.SA202406008

摘要/Abstract

摘要：

【目的/意义】 为了智能监测农业地物种类分布情况，一般采用无人机搭载高光谱相机进行高光谱数据采集，之后对高光谱数据分类，实现农作物分布图自动绘制。但不同农作物外形相似，同一农作物不同生长期差别较大，所以对农业地物分类的网络模型要求较高。分类精度较高的网络模型往往复杂程度较高，无法部署在硬件系统中。针对以上问题，本研究提出一种轻量化高低频增强的ReluformerN网络（Reluformer Network）实现农业地物分类。 【方法】 首先提出自适应八倍频卷积，不仅可以对高光谱图像的空间和光谱频域特征进行提取，同时缓解了内部人工参数设置带来的影响。其次针对低频信息可以捕获全局特征的特点，提出Reluformer进行全局特征提取，Reluformer相比transformer具有线性计算复杂度，有利于网络轻量化的同时保持了提取全局特征的能力。将该网络在三个公开的有关农作物品种精细分类的高光谱数据集上进行实验，并与较为流行的五种分类网络进行对比。 【结果和讨论】 ReluformerN在整体精度（Overall Accuracy, OA）、平均精度（Average Accuracy, AA）等精度评价指标中表现较好。在模型参数量（Parameters）、模型计算量（FLOPs）模型复杂度评价指标中，ReluformerN参数量最小，计算量最低。 【结论】 本研究提出的ReluformerN网络在农作物品种分类精度和模型复杂度之间达到了较好的平衡，有望后续部署在资源有限的硬件系统中，实现地物实时分类功能。

关键词: 高光谱图像, 农业地物分类, 轻量化网络, 线性transformer, 深度学习

Abstract:

[Objective] In order to intelligently monitor the distribution of agricultural land cover types, high-spectral cameras are usually mounted on drones to collect high-spectral data, followed by classification of the high-spectral data to automatically draw crop distribution maps. Different crops have similar shapes, and the same crop has significant differences in different growth stages, so the network model for agricultural land cover classification requires a high degree of accuracy. However, network models with high classification accuracy are often complex and cannot be deployed on hardware systems. In view of this problem, a lightweight high-low frequency enhanced Reluformer network (ReluformerN) was proposed in this research. [Methods] Firstly, an adaptive octave convolution was proposed, which utilized the softmax function to automatically adjust the spectral dimensions of high-frequency features and low-frequency features, effectively alleviating the influence of manually setting the spectral dimensions and benefiting the subsequent extraction of spatial and spectral domain features of hyperspectral images. Secondly, a Reluformer was proposed to extract global features, taking advantage of the fact that low-frequency information could capture global features. Reluformer replaced the softmax function with a function of quadratic computational complexity, and through theoretical and graphical analysised, Relu function, LeakRelu function, and Gelu function were compared, it was found that the ReLU function and the softmax function both had non-negativity, which could be used for feature relevance analysis. Meanwhile, the ReLU function has a linearization feature, which is more suitable for self-relevance analysis. Therefore, the ReLU self-attention mechanism was proposed, which used the ReLU function to perform feature self-attention analysis. In order to extract deep global features, multi-scale feature fusion was used, and the ReLU self-attention mechanism was used as the core to construct the multi-head ReLU self-attention mechanism. Similar to the transformer architecture, the Reluformer structure was built by combining multi-head ReLU self-attention mechanism, feedforward layers, and normalization layers. With Reluformer as the core, the Reluformer network (ReluformerN) was proposed. This network considered frequency from the perspective of high-frequency information, taking into account the local features of image high-frequency information, and used deep separable convolution to design a lightweight network for fine-grained feature extraction of high-frequency information. It proposed Reluformer to extract global features for low-frequency information, which represented the global features of the image. ReluformerN was experimented on three public high-spectral data sets (Indian Pines, WHU-Hi-LongKou and Salinas) for crop variety fine classification, and was compared with five popular classification networks (2D-CNN, HybirdSN, ViT, CTN and LSGA-VIT). [Results and Discussion] ReluformerN performed best in overall accuracy (OA), average accuracy (AA), and other accuracy evaluation indicators. In the evaluation indicators of model parameters, model computation (FLOPs), and model complexity, ReluformerN had the smallest number of parameters and was less than 0.3 M, and the lowest computation. In the visualization comparison, the classification effect diagram of the model using ReluformerN had clearer image edges and more complete morphological structures, with fewer classification errors. The validity of the adaptive octave convolution was verified by comparing it with the traditional eightfold convolution. The classification accuracy of the adaptive octave convolution was 0.1% higher than that of the traditional octave convolution. When the artificial parameters were set to different values, the maximum and minimum classification accuracies of the traditional octave convolution were about 0.3% apart, while those of the adaptive octave convolution were only 0.05%. This showed that the adaptive octave convolution not only had the highest classification accuracy, but was also less sensitive to the artificial parameter setting, effectively overcoming the influence of the artificial parameter setting on the classification result. To validated the Reluformer module, it was compared with transformer, LeakRelufromer, and Linformer in terms of accuracy evaluation metrics such as OA and AA. The Reluformer achieved the highest classification accuracy and the lowest model parameter count among these models. This indicated that Reluformer not only effectively extracted global features but also reduced computational complexity. Finally, the effectiveness of the high-frequency and low-frequency branch networks was verified. The effectiveness of the high-frequency and low-frequency feature extraction branches was verified, and the characteristics of the feature distribution after high-frequency feature extraction, after high-low frequency feature extraction, and after the classifier were displayed using a 2D t-sne, compared with the original feature distribution. It was found that after high-frequency feature extraction, similar features were generally clustered together, but the spacing between different features was small, and there were also some features with overlapping situations. After low-frequency feature extraction, it was obvious that similar features were clustered more tightly. After high-low frequency feature fusion, and after the classifier, it was obvious that similar features were clustered, and different types of features were clearly separated, indicating that high-low frequency feature extraction enhanced the classification effect. [Conclusion] This network achieves a good balance between crop variety classification accuracy and model complexity, and is expected to be deployed on hardware systems with limited resources in the future to achieve real-time classification functions.

Key words: hyperspectral images, agricultural object classification, lightweight network, linear transformer, deep learning

中图分类号:

TP79

刘伊, 张彦军. ReluformerN：轻量化高低频增强高光谱农业地物分类方法[J]. 智慧农业(中英文), 2024, 6(5): 74-87.

LIU Yi, ZHANG Yanjun. ReluformerN: Lightweight High-Low Frequency Enhanced for Hyperspectral Agricultural Lancover Classification[J]. Smart Agriculture, 2024, 6(5): 74-87.

图/表 15

图1

图2

图3

图4

图5

表1

表2

精度及复杂度研究IP数据集上不同网络的分类结果

分类	2D-CNN	HybirdSN	ViT	CTN	LSGA-VIT	ReluformerN
1	95 $. 00 ± 4.30$	96.26 $± 3.45$	98.20 $± 4.17$	98.29 $± 4.17$	92.68 $± 0.15$	97.5 $6 ± 1.1$
2	89.77 $± 1.18$	93.96 $± 1.04$	94.10 $± 1.31$	94.17 $± 1.31$	97.43 $± 0.17$	95.64 $± 0.41$
3	97.87 $± 0.62$	97.85 $± 0.76$	97.60 $± 0.71$	97.69 $± 0.71$	98.79 $± 0.36$	99.59 $± 0.66$
4	88.1 $0 ± 4.19$	89.36 $± 3.93$	99.40 $± 1.61$	99.44 $± 1.61$	98.59 $± 0.44$	100.00 $± 0$ .00
5	90.93 $± 0.67$	99.15 $± 0.39$	99.10 $± 1.36$	99.15 $± 1.36$	95.63 $± 0.19$	97.93 $± 0.18$
6	100.00 $± 0.00$	98.85 $± 0.37$	99.30 $± 1.16$	99.33 $± 1.16$	99.54 $± 0.07$	99.84 $± 0.21$
7	100.00 $± 0.00$	90.77 $± 4.21$	99.10 $± 2.39$	99.11 $± 2.39$	92.00 $± 0.9$	100.00 $± 0.00$
8	100.00 $± 0.00$	99.95 $± 0.08$	100.0 $± 0.00$	100.00 $±$ 0.00	99.76 $± 0.17$	100.00 $± 0.00$
9	50 $. 00 ± 0.00$	70.53 $± 8.61$	81.10 $± 5.73$	81.11 $± 5.73$	83.33 $± 2.12$	94.44 $± 0.35$
10	97.70 $± 0.00$	96.35 $± 0.70$	96.30 $± 3.34$	96.31 $± 3.34$	97.48 $± 0.11$	97.71 $± 0.15$
11	98.64 $± 0.00$	97.94 $± 0.56$	99.16 $± 0.5$	99.16 $± 0.5$	96.87 $± 0.08$	99.40 $± 0.20$
12	87.74 $± 2.82$	84.89 $± 3.25$	95.60 $± 6.26$	95.67 $± 6.26$	98.87 $± 0.21$	95.69 $± 0.48$
13	94.74 $± 0.00$	98.43 $± 0.52$	98.30 $± 2.85$	98.38 $± 2.85$	100.00 $± 0.00$	99.94 $± 0.20$
14	100 $. 00 ± 0.00$	100 $. 00 ± 0.00$	99.73 $± 0 . . 4$	99.73 $± 0 . . 4$	100.00 $± 0.00$	99.91 $± 0.1$
15	99.43 $± 1.63$	96.49 $± 1.26$	98.80 $± 2.85$	98.88 $± 2.85$	90.23 $± 0.35$	99.71 $± 0.58$
16	100.00 $± 0.00$	95.58 $± 2.05$	56.0 $± 18.26$	56.31 $± 18.26$	83.33 $± 1.18$	92.85 $± 2.71$
OA	95.90 $± 0.94$	96.59 $± 0.46$	97.50 $± 1.51$	97.52 $± 1.51$	97.60 $± 0.04$	98.4 $4 ± 0.06$
AA	93.12 $± 1.13$	94.15 $± 1.42$	93.90 $± 6.17$	93.93 $± 6.17$	95.29 $± 0.12$	98.39 $± 0.56$
Kappa	95.67 $± 0.26$	96.11 $± 0.52$	97.10 $± 1.73$	97.16 $± 1.73$	97.27 $± 0.04$	98.21 $± 0.06$
Params/M	1.05	3.70	0.31	0.26	0.54	0.25
FLOPs/G	0.11	0.24	0.88	1.15	2.45	0.10
训练时间/s	13.09	30.74	25.46	11.9	36.12	25.07
测试时间/s	0.37	2.92	2.81	1.34	0.46	2.18

表2

表3

精度及复杂度研究WHU-Hi-Longkou数据集上不同网络分类结果

Class	2D-CNN	HybirdSN	ViT	CTN	LSGA-VIT	ReluformerN
1	99.96 $± 0.01$	99.86 $± 0.09$	99.93 $± 0.03$	99.97 $± 0.02$	99.79 $± 0.25$	99.94 $± 0.25$
2	96.86 $± 1.22$	98.92 $± 0.8$ 0	98.78 $± 0.95$	99.53 $± 0.16$	99.67 $± 0.25$	99.72 $± 0.18$
3	92.81 $± 3.06$	94.43 $± 2.18$	96.02 $± 1.15$	97.81 $± 0.70$	96.53 $± 1.62$	92.33 $± 0.87$
4	98.86 $± 0.26$	99.18 $± 0.48$	99.51 $± 0.16$	99.71 $± 0.07$	99.76 $± 0.06$	99.87 $± 0.85$
5	94.54 $± 2.35$	97.84 $± 1.67$	98.06 $± 1.08$	90.17 $± 2.90$	99.02 $± 0.40$	99.85 $± 0.47$
6	99.54 $± 0.39$	99.40 $± 1.14$	99.05 $± 0.76$	99.94 $± 0.07$	99.04 $± 0.46$	99.27 $± 1.01$
7	99.95 $±$ 0.04	99.83 $± 0.16$	99.91 $± 0.05$	99.93 $± 0.09$	99.79 $± 0.08$	99.95 $± 0.34$
8	96.86 $± 0.63$	97.69 $± 0.8$ 0	96.86 $± 1.11$	94.54 $± 0.45$	96.45 $± 0.70$	97.73 $± 0.37$
9	93.05 $± 2.05$	92.34 $± 1.90$	93.62 $± 1.13$	93.56 $± 1.15$	92.59 $± 0.79$	94.32 $± 0.51$
OA	98.97 $± 0.12$	99.18 $± 0.26$	99.34 $± 0.02$	99.27 $± 0.05$	99.38 $± 0.09$	99.54 $± 0.02$
AA	96.94 $± 0.47$	97.72 $± 0.34$	97.98 $± 0.06$	97.24 $± 0.25$	98.08 $± 0.23$	98.11 $± 0.03$
Kappa	98.64 $± 0.16$	98.94 $± 0.34$	99.13 $± 0.03$	99.05 $± 0.07$	99.19 $± 0.12$	99.39 $± 0.03$
Params/M	1.05	1.69	2.97	0.54	0.17	0.23
FLOPs/G	2.93	12.68	30.18	3.05	1.11	0.17
训练时间/s	11.72	35.47	95.19	38.25	20.48	25.45
测试时间/s	3.03	6.61	11.32	7.71	6.22	6.81

表3

表4

精度及复杂度研究SA数据集上不同网络分类结果

Class	2D-CNN	HybirdSN	ViT	CTN	LSGA-VIT	ReluformerN
1	99.95 $± 0.08$	99.81 $± 0.38$	99.82 $± 0.18$	100.00 $± 0.00$	99.99 $± 0.02$	99.94 $± 2.97$
2	100.00 $± 0.00$	100.00 $± 0.00$	100.00 $± 0.00$	100.00 $± 0.00$	99.89 $± 0.10$	100.00 $± 0.00$
3	99.38 $± 1.39$	98.40 $± 4.33$	99.68 $± 0.22$	100.00 $± 0.00$	99.99 $± 0.02$	99.97 $± 0.46$
4	98.31 $± 3.81$	99.00 $± 1.39$	96.85 $± 1.04$	99.53 $± 0.38$	99.55 $± 0.37$	99.79 $± 1.19$
5	98.22 $± 0.55$	98.77 $± 1.73$	99.43 $± 0.47$	98.95 $± 0.63$	98.98 $± 0.87$	96.83 $± 0.88$
6	99.93 $± 0.05$	99.74 $± 0.28$	99.78 $± 0.23$	99.74 $± 0.16$	100 $. 00 ± 0.00$	99.71 $± 0.10$
7	99.83 $± 0.22$	99.96 $± 0.10$	99.93 $± 0.06$	99.78 $± 0.26$	100.00 $± 0.00$	98.42 $± 0.37$
8	96.78 $± 0.49$	97.36 $±$ 0.95	96.9 $0 ± 0.88$	98.03 $± 0.35$	94.76 $± 1.49$	99.65 $± 0.76$
9	99.99 $± 0.02$	100.00 $± 0.00$	100.00 $± 0.00$	100.00 $± 0$ .00	100 $. 00 ± 0.00$	99.59 $± 0.32$
10	99.18 $± 0.53$	99.15 $± 1.41$	99.42 $± 0.32$	99.45 $± 0.38$	98.72 $± 0.60$	99.38 $± 0.10$
11	99.10 $± 0.15$	98.58 $± 0.87$	99.79 $± 0.27$	99.44 $± 1.14$	100.00 $± 0.00$	99.71 $± 0.93$
12	99.57 $± 0.42$	98.69 $± 3.3$ 0	99.10 $± 0.83$	99.86 $± 0.15$	99.61 $± 0.37$	97.57 $± 1.73$
13	97.95 $± 2.45$	94.50 $± 10.8$	97.65 $± 3.35$	92.34 $± 7.22$	98.11 $± 1.68$	94.72 $± 2.87$
14	98.94 $± 1.24$	96.58 $± 3.34$	98.68 $± 0.44$	98.86 $± 0.57$	99.19 $± 0.35$	93.56 $± 0.44$
15	88.81 $± 5.40$	93.34 $± 8.81$	93.76 $± 1.60$	95.19 $± 1.27$	96.45 $± 1.77$	97.98 $± 1.35$
16	99.22 $± 0.19$	99.72 $± 0.42$	99.36 $± 0.26$	99.32 $± 0.47$	99.65 $± 0.23$	99.47 $± 0.62$
OA	97.49 $± 0.91$	98.08 $± 1.74$	98.21 $± 0.14$	98.62 $± 1.27$	98.21 $± 0.28$	98.81 $± 0.19$
AA	98.45 $± 0.77$	98.35 $± 1.99$	98.76 $± 0.21$	98.80 $± 0.46$	99.05 $± 0.12$	97.79 $± 0.19$
Kappa	97.20 $± 1.02$	97.86 $± 1.95$	98.01 $± 0.16$	98.46 $± 0.30$	98.01 $± 0.31$	98.67 $± 0.26$
Params/M	1.05	3.70	2.14	0.27	0.54	0.23
FLOPs/G	0.33	2.25	24.97	1.40	2.45	0.17
训练时间/s	10.77	16.12	19.74	11.67	26.12	19.07
测试时间/s	2.34	1.90	3.84	1.73	2.46	5.01

表4

图6

图8

图7

图9

图10

表5

IP、WHU-Hi-LongKou、SA数据Reluformer和相关模块分类精度及复杂度对比

数据	模型	OA	AA	Kappa	Parameters/M	FLOPs/G
IP	ViT	98.17 $± 0.11$	95.40 $± 0.53$	97.92 $± 0.18$	0.25	0.10
	Linformer	96.54 $± 0.24$	95.08 $± 0.71$	95.37 $±$ 0.35	0.22	0.15
	LeakyReluformer	98.21 $± 0.18$	97.86 $± 0.62$	97.75 $±$ 0.17	0.21	0.13
	Reluformer	98.4 $4 ± 0.06$	98.39 $± 0.56$	98.21 $± 0.06$	0.21	0.13
WHU-HI-Longkou	ViT	99.42 $± 0.03$	98.31 $± 0.17$	99.23 $± 0.07$	0.35	0.05
	Linformer	98.25 $± 0.05$	97.67 $± 0.21$	97.54 $± 0.09$	0.26	0.20
	LeakyReluformer	99. 32 $± 0.06$	98.06 $± 0.15$	99.03 $± 0.05$	0.23	0.17
	Reluformer	99.54 $± 0.02$	98.11 $± 0.03$	99.39 $± 0.03$	0.23	0.17
SA	ViT	97.39 $± 0.06$	95.29 $±$ 0.21	97.09 $± 0.18$	0.35	0.12
	Linformer	95.78 $±$ 0.15	93.82 $±$ 0.36	95.36 $± 0.27$	0.26	0.22
	LeakyReluformer	97. 54 $± 0.32$	96.12 $± 0.26$	96.94 $± 0.19$	0.23	0.17
	Reluformer	98.81 $± 0.19$	97.79 $± 0.19$	98.67 $± 0.26$	0.23	0.17

表5

参考文献 32

1	冯凡, 张永生, 张津, 等. 基于混合卷积网络的高光谱图像自监督特征学习方法[J/OL]. 光学学报, 2024: 1-21. (2024-04-22).
	FENG F, ZHANG Y S, ZHANG J, et al. Self-supervised feature learning method for hyperspectral images based on mixed convolutional neural networks[J/OL]. Acta optica sinica, 2024: 1-21. (2024-04-22).
2	李世隆, 许辰一, 王楠, 等. 基于BWO-ELM的水稻氮素无人机高光谱反演研究[J]. 智能化农业装备学报(中英文), 2024, 5(3): 14-21.
	LI S L, XU C Y, WANG N, et al. Research on rice nitrogen unmanned aerial vehicle hyperspectral inversion based on BWO-ELM[J]. Journal of Intelligent Agricultural Mechanization, 2024, 5(3): 14-21.
3	孙梦花. 基于扩展形态学与主动学习的高光谱影像分类方法[D]. 西安: 西安电子科技大学, 2019.
	SUN M H. Hyperspectral image classification based on extended morphology and active learning[D].Xi'an: Xidian University, 2019.
4	杨珺雯, 张锦水, 朱秀芳, 等. 随机森林在高光谱遥感数据中降维与分类的应用[J]. 北京师范大学学报(自然科学版), 2015, 51(S1): 82-88.
	YANG J W, ZHANG J S, ZHU X F, et al. Application of random forest in dimensionality reduction and classification of hyperspectral remote sensing data[J]. Journal of Beijing normal university (natural science), 2015, 51(S1): 82-88.
5	李玉, 宫学亮, 赵泉华. 基于张量径向基核函数支持向量机的高光谱影像分类[J]. 仪器仪表学报, 2020, 41(12): 253-262.
	LI Y, GONG X L, ZHAO Q H. Hyperspectral image classification based on tensor-based radial basis kernel function and support vector machine[J]. Chinese journal of scientific instrument, 2020, 41(12): 253-262
6	CHEN Y S, JIANG H L, LI C Y, et al. Deep feature extraction and classification of hyperspectral images based on convolutional neural networks[J]. IEEE transactions on geoscience and remote sensing, 2016, 54(10): 6232-6251.
7	ZHANG X. Improved three-dimensional inception networks for hyperspectral remote sensing image classification[J]. IEEE Access, 2023, 11: 32648-32658.
8	ROY S K, KRISHNA G, DUBEY S R, et al. HybridSN: Exploring 3-D-2-D CNN feature hierarchy for hyperspectral image classification[J]. IEEE geoscience and remote sensing letters, 2020, 17(2): 277-281.
9	汪乐, 彭江涛, 陈娜, 等. 结合协方差池化与跨尺度特征提取的高光谱分类[J]. 遥感学报, 2024, 28(1): 203-218.
	WANG L, PENG J T, CHEN N, et al. Hyperspectral classification algorithm based on covariance pooling and cross scale feature extraction[J]. National remote sensing bulletin, 2024, 28(1): 203-218.
10	王彩玲, 张静, 王洪伟, 等. 一种波段聚类和多尺度结构特征融合的高光谱图像分类模型[J]. 光谱学与光谱分析, 2024, 44(1): 258-265.
	WANG C L, ZHANG J, WANG H W, et al. A hyperspectral image classification model based on band clustering and multi-scale structure feature fusion[J]. Spectroscopy and spectral analysis, 2024, 44(1): 258-265.
11	王庆超, 付光远, 汪洪桥, 等. 多核融合多尺度特征的高光谱影像地物分类[J]. 光学精密工程, 2018, 26(4): 980-988.
	WANG Q C, FU G Y, WANG H Q, et al. Fusion of multi-scale feature using multiple kernel learning for hyperspectral image land cover classification[J]. Optics and precision engineering, 2018, 26(4): 980-988.
12	高玉鹏, 闫伟红, 潘新. 基于卷积神经网络与注意力机制的高光谱图像分类[J]. 光电子·激光, 2024, 35(5): 483-489.
	GAO Y P, YAN W H, PAN X. Hyperspectral image classification based on convolutional neural network and attention mechanism[J]. Journal of optoelectronics·Laser, 2024, 35(5): 483-489.
13	孙杰, 杨静, 丁书杰, 等. 基于多注意力机制与编译图神经网络的高光谱图像分类[J]. 农业机械学报, 2024, 55(3): 183-192, 212.
	SUN J, YANG J, DING S J, et al. Hyperspectral image classification based on multi-attention mechanism and compiled graph neural networks[J]. Transactions of the Chinese society for agricultural machinery, 2024, 55(3): 183-192, 212.
14	翟希辰, 刘军. 基于注意力机制辅助的空谱联合残差网络的高光谱图像分类[J]. 信息对抗技术, 2024, 3(2): 54-69.
	ZHAI X C, LIU J. Spatial-spectral residual network with attention mechanism for hyperspectral image classification[J]. Information countermeasure technology, 2024, 3(2): 54-69.
15	杜培军, 张伟, 张鹏, 等. 一种联合空谱特征的高光谱影像分类胶囊网络[J]. 测绘学报, 2023, 52(7): 1090-1104.
	DU P J, ZHANG W, ZHANG P, et al. A capsule network for hyperspectral image classification employing spatial-spectral feature[J]. Acta geodaetica et cartographica sinica, 2023, 52(7): 1090-1104.
16	洪娟. 基于金字塔残差结构和注意力机制的高光谱图像分类研究[D]. 武汉: 武汉轻工大学, 2021.
	HONG J. Deep pyramidal residual attention networks for hyperspectral image classification[D].Wuhan: Wuhan Polytechnic University, 2021.
17	汪菲菲, 赵慧洁, 李娜, 等. 基于光谱-空间注意力残差网络的高光谱图像分类[J]. 光子学报, 2023, 52(12): ID 1210002.
	WANG F F, ZHAO H J, LI N, et al. Spectral-spatial attention residual networks for hyperspectral image classification[J]. Acta photonica sinica, 2023, 52(12): ID 1210002.
18	易瑔, 张宇航, 王清, 等. 基于改进残差网络的高光谱图像分类算法研究[J]. 兵工自动化, 2023, 42(10): 15-20.
	YI Q, ZHANG Y H, WANG Q, et al. Research on hyperspectral image classification algorithm based on improved residual network[J]. Ordnance industry automation, 2023, 42(10): 15-20.
19	钱韫竹. 双通道残差网络动态融合的高光谱影像分类方法研究[D]. 阜新: 辽宁工程技术大学, 2023.
	QIAN Y Z. Study on hyperspectral image classification methods for dynamic fusion of dual channel residual networks[D]. Fuxin: Liaoning Technical University, 2023.
20	ZHAO Z G, HU D, WANG H, et al. Convolutional transformer network for hyperspectral image classification[J]. IEEE geoscience and remote sensing letters, 2022, 19: ID 6009005.
21	ROY S K, DERIA A, SHAH C, et al. Spectral-spatial morphological attention transformer for hyperspectral image classification[J]. IEEE transactions on geoscience and remote sensing, 2023, 61: ID 5503615.
22	罗宇轩, 吴高昌, 高明. 自适应卷积和轻量化Transformer的遥感图像超分辨网络[J/OL]. 计算机工程与应用, (2024-04-23)[2024-06-15].
	LUO Y X, WU G C, GAO M. An adaptive convolution and lightweight Transformer-based super-resolution network for remote sensing images[J/OL]. Computer engineering and applications, (2024-04-23)[2024-06-15].
23	MA C, WAN M J, WU J, et al. Light self-gaussian-attention vision transformer for hyperspectral image classification[J]. IEEE transactions on instrumentation and measurement, 2023, 72: ID 5015712.
24	王琪, 常庆瑞, 李铠, 等. 基于主成分分析和随机森林回归的冬小麦冠层叶绿素含量估算[J]. 麦类作物学报, 2024, 44(4): 532-542.
	WANG Q, CHANG Q R, LI K, et al. Estimation of winter wheat canopy chlorophyll content based on principal component analysis and random forest regression[J]. Journal of triticeae crops, 2024, 44(4): 532-542.
25	CHEN Y P, FAN H Q, XU B, et al. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution[C]// 2019 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway, New Jersey, USA: IEEE, 2019: 3434-3443.
26	HONG Q, ZHONG X, CHEN W, et al. Hyperspectral image classification network based on 3D octave convolution and multiscale depthwise separable convolution[J]. Isprs international journal of geo-information, 2023(11): 505-521.
27	LIANG L, ZHANG S, LI J, et al. Multi-scale spectral-spatial attention network for hyperspectral image classification combining 2D octave and 3D convolutional neural networks[J]. Remote sensing, 2023(4): 1758-1784.
28	杨青, 于明, 付强, 等. 基于倍频卷积和注意力机制的图像去雨[J]. 控制与决策, 2023, 38(12): 3372-3380.
	YANG Q, YU M, X, FU Q. Image dewatering based on frequency-doubling convolution and attention mechanism[J]. Control and decision, 38(12): 3372-3380
29	QIN Z, SUN W X, DENG H, et al. cosFormer: Rethinking softmax in attention[EB/OL]. arXiv:2202.08791, 2022.
30	VASWANI A, SHAZEER N M, PARMAR N, et al. Attention is all you need[EB/OL]. arXiv:1706.03762, 2017.
31	Hyperspectral Remote Sensing Scenes[DB/OL]. [2024-05-12].
32	WHU-Hi: UAV-borne hyperspectral and high spatial resolution (H2) benchmark datasets for crop precise classification[DB/OL]. [2024-05-12].

IP				WHU-Hi-LongKou				SA
标号	颜色	种类	样本量/个	标号	颜色	种类	样本量/个	标号	颜色	种类	样本量/个
0		背景	10 776	0		背景	15 458	0		背景	56 975
1		苜蓿	46	1		玉米	34 511	1		绿花椰菜_1	2 009
2		免耕玉米地	1 428	2		棉花	8 374	2		绿花椰菜-杂草_2	3 726
3		玉米幼苗	830	3		芝麻	3 031	3		休耕	1 976
4		玉米	237	4		阔叶-大豆	63 212	4		休耕-粗耕	1 394
5		草地-牧场	483	5		窄叶-大豆	4 151	5		休耕-细耕	2 678
6		草地-树木	730	6		大米	11 854	6		秸秆	3 959
7		修剪后草地	28	7		水	67 056	7		芹菜	3 579
8		干草-料堆	478	8		道路和房屋	7 124	8		未修剪的葡萄	11 271
9		燕麦	20	9		混合杂草	5 229	9		葡萄园	6 203
10		免耕大豆地	972					10		玉米-未修剪的杂草	3 278
11		大豆幼苗	2 455					11		罗曼生菜_4周	1 068
12		修剪后大豆地	593					12		罗曼生菜_5周	1 927
13		小麦	205					13		罗曼生菜_6周	916
14		树林	1 265					14		罗曼生菜_7周	1 070
15		建筑-草-数目-机器	386					15		未修剪的葡萄园	7 268
16		石-钢-塔	93					16		垂直支架的葡萄园	1 807


e. ViT	f. CTN	g. LSGA-VIT	h. ReluformerN

[1]	罗友璐, 潘勇浩, 夏顺兴, 陶友志. 基于改进YOLOv8的苹果叶病害轻量化检测算法[J]. 智慧农业(中英文), 2024, 6(5): 128-138.
[2]	年悦, 赵凯旋, 姬江涛. 基于改进DeepLabCut模型的奶牛滑蹄检测方法[J]. 智慧农业(中英文), 2024, 6(5): 153-163.
[3]	张岩琪, 周硕, 张凝, 柴秀娟, 孙坦. 基于改进实例分割算法的区域养殖生猪计数系统[J]. 智慧农业(中英文), 2024, 6(4): 53-63.
[4]	翁智, 范琦, 郑志强. 基于多模态图像信息及改进实例分割网络的肉牛体尺自动测量方法[J]. 智慧农业(中英文), 2024, 6(4): 64-75.
[5]	侯依廷, 饶元, 宋贺, 聂振君, 王坦, 何豪旭. 复杂大田场景下基于改进YOLOv8的小麦幼苗期叶片数快速检测方法[J]. 智慧农业(中英文), 2024, 6(4): 128-137.
[6]	李豪, 杜雨秋, 肖星竹, 陈彦羲. 基于深度学习的四川盆地丘陵区县域耕地遥感识别研究[J]. 智慧农业(中英文), 2024, 6(3): 34-45.
[7]	聂刚刚, 饶洪辉, 李泽锋, 刘木华. 基于改进YOLACT的油茶叶片炭疽病感染严重程度分级模型[J]. 智慧农业(中英文), 2024, 6(3): 138-147.
[8]	张京, 赵泽瑄, 赵艳茹, 卜泓超, 吴星宇. 基于Bi-GRU和空-谱信息融合的油菜菌核病侵染区域高光谱图像分割方法[J]. 智慧农业(中英文), 2024, 6(2): 40-48.
[9]	庞春晖, 陈鹏, 夏懿, 章军, 王兵, 邹岩, 陈天娇, 康辰瑞, 梁栋. 用于小麦多生长阶段倒伏边界精准检测的分层交互特征金字塔网络[J]. 智慧农业(中英文), 2024, 6(2): 128-139.
[10]	张玉玉, 邴树营, 纪元浩, 严蓓蓓, 许金普. 基于改进YOLOv8s的玫瑰鲜切花分级方法[J]. 智慧农业(中英文), 2024, 6(2): 118-127.
[11]	张建华, 姚琼, 周国民, 吴雯迪, 修晓杰, 王健. 作物农艺性状与形态结构表型智能识别技术综述[J]. 智慧农业(中英文), 2024, 6(2): 14-27.
[12]	郭旺, 杨雨森, 吴华瑞, 朱华吉, 缪祎晟, 顾静秋. 农业大模型：关键技术、应用分析与发展方向[J]. 智慧农业(中英文), 2024, 6(2): 1-13.
[13]	王鹤榕, 陈英义, 柴莹倩, 徐玲, 于辉辉. 融合VoVNetv2和置换注意力机制的鱼群摄食图像分割方法[J]. 智慧农业(中英文), 2023, 5(4): 137-149.
[14]	李政凯, 于嘉辉, 潘时佳, 贾泽丰, 牛子杰. 冬季猕猴桃树单木骨架提取与冠层生长预测方法[J]. 智慧农业(中英文), 2023, 5(4): 92-104.
[15]	唐辉, 王铭, 于秋实, 张佳茜, 刘连涛, 王楠. 融合改进UNet和迁移学习的棉花根系图像分割方法[J]. 智慧农业(中英文), 2023, 5(3): 96-109.