改进的残差式3D-CNN和近邻注意力的高光谱遥感图像分类
|
潘增滢(2000-),男,硕士研究生,主要从事高光谱遥感信息处理研究。Email:1157928777@qq.com。 |
Copy editor: 张仙
收稿日期: 2024-05-24
修回日期: 2024-09-21
网络出版日期: 2026-06-03
基金资助
福建省自然科学基金项目“人机协同的自然资源要素提取关键技术研究”(2023J01432)
国家自然科学基金项目“基于深度迁移学习网络的高分影像土地利用分类方法研究”(41801324)
Hyperspectral remote sensing image classification using improved residual 3D-CNN and neighborhood attention
Received date: 2024-05-24
Revised date: 2024-09-21
Online published: 2026-06-03
在引起广泛关注的高光谱遥感图像分类中,同物异谱、同谱异物和少样本都大大限制了分类方法的性能。为了充分挖掘高光谱图像的空间-光谱特征,该文提出了一种改进的残差卷积和近邻注意力网络用于高光谱遥感图像分类。该方法包含3个部分:结合了残差连接和3D卷积神经网络(3D convolutional neural network,3D-CNN)的残差式光谱特征提取模块、使用混合卷积的空间-光谱特征融合模块、用于增强模型对同质区域的关注能力的近邻注意力模块。在3个公开的高光谱数据集Indian pines,Pavia University,Houston2013上的实验结果显示,相比近期先进高光谱分类方法,所提方法有更高的分类精度,且在使用10%以下训练样本的前提下总体精度可分别达到99.39%,99.67%和98.64%,实现了少样本下的高精度分类。
潘增滢 , 吴瑞姣 , 林易丰 , 翁谦 , 林嘉雯 . 改进的残差式3D-CNN和近邻注意力的高光谱遥感图像分类[J]. 自然资源遥感, 2025 , 37(5) : 101 -112 . DOI: 10.6046/zrzyyg.2024191
Hyperspectral remote sensing image classification has attracted widespread attention,yet the performance of classification methods remains greatly limited by challenges such as spectral variability (same object with different spectra),spectral confusion (different objects with similar spectra),and limited availability of training samples. To fully exploit the spatial-spectral features of hyperspectral images,this study proposed an improved network integrating residual convolution and neighborhood attention mechanisms. The proposed method consists of:(1) a residual-based spectral feature extraction module combining residual connections and a 3D convolutional neural network (3D-CNN);(2) a spatial-spectral feature fusion module using mixed convolutions;and (3) a neighborhood attention module designed to enhance the model's ability to focus on homogeneous regions. Experiments were conducted on three public hyperspectral datasets-Indian Pines,Pavia University,and Houston 2013. The results demonstrate that the proposed method achieves higher classification accuracy compared to recent state-of-the-art approaches. Using less than 10% of the samples for training,it attains overall accuracies of 99.39%,99.67%,and 98.64%,respectively,confirming its capability for high-accuracy classification under small-sample conditions.
表1 不同方法在IP数据集上的分类结果Tab.1 Classification Results of Different Methods on IP Datasets |
| 类别 | 1D-CNN | 2D-CNN | SSRN | HybridSN | A2S2K | morphFormer | LSGA | IR3NAN |
|---|---|---|---|---|---|---|---|---|
| 苜蓿 | 22.93 | 55.61 | 72.93 | 79.76 | 79.27 | 89.51 | 88.29 | 97.56 |
| 未翻耕的玉米地 | 59.95 | 78.06 | 84.92 | 95.35 | 98.67 | 97.11 | 98.79 | 98.96 |
| 翻耕过的玉米地 | 51.02 | 82.53 | 89.21 | 93.36 | 98.33 | 96.49 | 95.92 | 98.89 |
| 玉米地 | 38.78 | 82.35 | 91.60 | 88.69 | 97.56 | 95.82 | 97.32 | 98.50 |
| 牧草区 | 73.72 | 95.52 | 96.74 | 92.34 | 98.92 | 96.46 | 98.32 | 98.69 |
| 草地与树木 | 86.51 | 97.84 | 99.47 | 97.69 | 99.27 | 99.74 | 99.89 | 99.82 |
| 已收割的牧草区 | 45.60 | 42.80 | 66.40 | 64.80 | 92.40 | 99.20 | 100.00 | 99.60 |
| 风干的草料 | 94.26 | 99.49 | 98.09 | 97.19 | 99.81 | 99.79 | 99.88 | 99.93 |
| 燕麦 | 18.89 | 65.56 | 9.44 | 94.44 | 99.44 | 84.44 | 97.78 | 98.33 |
| 未翻耕的大豆田 | 48.00 | 76.99 | 86.82 | 96.49 | 98.89 | 98.74 | 99.22 | 99.05 |
| 翻耕过的大豆田 | 67.89 | 84.51 | 97.24 | 95.71 | 99.38 | 98.45 | 99.50 | 99.67 |
| 已清理的大豆田 | 43.73 | 82.73 | 98.45 | 91.46 | 99.29 | 95.13 | 98.65 | 99.16 |
| 小麦 | 93.35 | 99.19 | 99.68 | 98.11 | 100.00 | 99.73 | 99.78 | 100.00 |
| 树林 | 92.99 | 98.13 | 92.68 | 99.08 | 99.93 | 99.87 | 99.99 | 100.00 |
| 建筑物、草地、树木、车道 | 39.39 | 89.57 | 89.34 | 88.1 | 98.67 | 98.30 | 98.79 | 100.00 |
| 石、钢、塔、楼 | 89.76 | 98.81 | 98.57 | 81.79 | 98.10 | 94.76 | 94.88 | 96.79 |
| OA/% | 66.67 | 86.77 | 92.84 | 95.08 | 99.03 | 98.04 | 98.92 | 99.39 |
| AA/% | 60.42 | 83.11 | 85.72 | 90.9 | 97.37 | 96.47 | 97.94 | 99.06 |
| Kappa | 0.619 8 | 0.847 5 | 0.918 8 | 0.943 9 | 0.988 9 | 0.977 6 | 0.987 7 | 0.993 0 |
表2 不同方法在PU数据集上的分类结果Tab.2 Classification results of different methods on the PU dataset |
| 类别 | 1D-CNN | 2D-CNN | SSRN | HybridSN | A2S2K | morphFormer | LSGA | IR3NAN |
|---|---|---|---|---|---|---|---|---|
| 沥青路面 | 87.81 | 95.72 | 99.96 | 96.58 | 99.87 | 98.75 | 99.57 | 99.99 |
| 草地 | 90.85 | 97.50 | 99.60 | 99.77 | 99.97 | 99.91 | 99.88 | 99.98 |
| 碎石 | 57.84 | 69.74 | 92.47 | 86.49 | 94.52 | 89.45 | 89.90 | 97.49 |
| 树木 | 85.84 | 98.36 | 98.66 | 88.23 | 99.07 | 96.19 | 97.81 | 98.61 |
| 涂漆金属板 | 98.75 | 99.94 | 100.00 | 86.66 | 99.98 | 100.00 | 100.00 | 99.98 |
| 裸土 | 72.67 | 84.88 | 99.69 | 99.60 | 100.00 | 99.95 | 99.97 | 99.94 |
| 沥青 | 79.44 | 89.11 | 99.93 | 90.75 | 99.46 | 99.86 | 99.51 | 100.00 |
| 自阻砖 | 84.66 | 90.74 | 98.53 | 89.17 | 98.46 | 93.43 | 96.43 | 99.28 |
| 阴影 | 99.52 | 99.67 | 99.91 | 78.88 | 99.94 | 93.72 | 97.60 | 98.84 |
| OA/% | 85.82 | 93.72 | 99.18 | 95.71 | 99.48 | 98.26 | 98.85 | 99.67 |
| AA/% | 84.15 | 91.74 | 98.75 | 90.68 | 99.03 | 96.81 | 97.85 | 99.35 |
| Kappa | 0.812 3 | 0.916 1 | 0.989 2 | 0.943 0 | 0.993 1 | 0.976 9 | 0.984 8 | 0.995 6 |
表3 不同方法在HO集上的分类结果Tab.3 Classification results of different methods on the HO dataset |
| 类别 | 1D-CNN | 2D-CNN | SSRN | HybridSN | A2S2K | morphFormer | LSGA | IR3NAN |
|---|---|---|---|---|---|---|---|---|
| 健康的草地 | 93.90 | 94.65 | 93.36 | 91.83 | 96.37 | 94.77 | 94.85 | 96.83 |
| 受压力的草地 | 94.81 | 97.63 | 96.54 | 95.50 | 99.13 | 97.95 | 98.85 | 99.21 |
| 人造草皮 | 95.39 | 95.44 | 92.63 | 97.98 | 99.24 | 99.31 | 99.63 | 99.89 |
| 树木 | 94.03 | 98.60 | 99.87 | 95.53 | 99.93 | 97.52 | 97.66 | 99.93 |
| 土壤 | 95.86 | 97.37 | 99.61 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 |
| 水 | 68.01 | 67.21 | 78.16 | 85.37 | 84.04 | 83.75 | 96.25 | 95.44 |
| 住宅区 | 80.29 | 95.12 | 93.66 | 94.52 | 98.24 | 98.68 | 98.94 | 99.05 |
| 商业区 | 78.10 | 84.39 | 91.09 | 91.14 | 94.93 | 94.52 | 94.16 | 94.94 |
| 道路 | 73.29 | 89.06 | 83.47 | 93.73 | 98.55 | 97.80 | 97.79 | 98.11 |
| 高速公路 | 71.85 | 87.05 | 88.26 | 98.77 | 99.35 | 99.47 | 99.86 | 99.97 |
| 铁路 | 63.66 | 92.69 | 90.84 | 99.85 | 99.78 | 98.32 | 99.95 | 100.00 |
| 停车场1 | 75.47 | 92.18 | 88.17 | 96.51 | 97.84 | 96.82 | 96.77 | 99.12 |
| 停车场2 | 23.19 | 88.01 | 80.87 | 89.78 | 91.36 | 89.21 | 92.62 | 92.29 |
| 网球场 | 89.44 | 98.91 | 94.84 | 73.33 | 99.96 | 99.89 | 100.00 | 100.00 |
| 跑道 | 94.79 | 100.00 | 99.34 | 97.99 | 100.00 | 99.89 | 100.00 | 100.00 |
| OA/% | 81.11 | 92.91 | 92.14 | 95.19 | 98.21 | 97.45 | 97.96 | 98.64 |
| AA/% | 79.47 | 91.89 | 91.38 | 93.46 | 97.25 | 96.53 | 97.82 | 98.32 |
| Kappa | 0.794 7 | 0.923 0 | 0.914 5 | 0.947 7 | 0.980 6 | 0.972 3 | 0.977 8 | 0.985 2 |
图4 各个方法在IP数据集分类图对比Fig.4 Comparison of classification maps of various methods on the IP dataset |
图5 各个方法在PU数据集分类图对比Fig.5 Comparison of classification maps of various methods on the PU dataset |
图6 各个方法在HO数据集分类图对比Fig.6 Comparison of classification maps of various methods on the HO dataset |
表4 各个模块的消融对比Tab.4 Ablation comparison of each module |
| 数据集 | 精度 | Baseline | RSFEM | NAM | RSFEM+NAM |
|---|---|---|---|---|---|
| IP | OA/% | 99.04 | 99.38 | 99.11 | 99.39 |
| AA/% | 98.31 | 99.06 | 98.38 | 99.06 | |
| Kappa | 0.989 1 | 0.993 0 | 0.989 9 | 0.993 0 | |
| PU | OA/% | 99.26 | 99.45 | 99.30 | 99.67 |
| AA/% | 98.19 | 98.63 | 98.34 | 99.35 | |
| Kappa | 0.990 2 | 0.992 7 | 0.990 7 | 0.995 6 | |
| HO | OA/% | 98.39 | 98.53 | 98.50 | 98.64 |
| AA/% | 98.17 | 98.13 | 98.24 | 98.32 | |
| Kappa | 0.982 5 | 0.984 0 | 0.983 7 | 0.985 2 |
表5 不同大小空间块分类精度对比Tab.5 Comparison of classification accuracy for different patch sizes |
| 数据集 | 精度 | 7 | 9 | 11 | 13 | 15 |
|---|---|---|---|---|---|---|
| IP | OA/% | 99.27 | 99.39 | 99.22 | 98.97 | 98.65 |
| AA/% | 98.60 | 99.06 | 98.85 | 95.83 | 91.74 | |
| Kappa | 0.991 7 | 0.993 0 | 0.991 1 | 0.988 3 | 0.984 6 | |
| PU | OA/% | 99.19 | 99.42 | 99.58 | 99.67 | 99.64 |
| AA/% | 98.56 | 99.04 | 99.18 | 99.35 | 99.33 | |
| Kappa | 0.989 3 | 0.992 3 | 0.994 4 | 0.995 6 | 0.995 2 | |
| HO | OA/% | 98.42 | 98.64 | 98.52 | 98.42 | 98.34 |
| AA/% | 97.98 | 98.32 | 97.95 | 97.78 | 97.68 | |
| Kappa | 0.982 8 | 0.985 2 | 0.983 9 | 0.982 8 | 0.981 9 |
| [1] |
张涛, 王彬沣, 付莹, 等. 基于深度学习的光谱图像超分辨率综述[J]. 中国图象图形学报, 2024, 29(8):2113-2136.
|
| [2] |
马晓瑞, 哈林, 谌敦斌, 等. 融合特征优化的跨数据集高光谱图像分类[J]. 中国图象图形学报, 2024, 29(8):2175-2187.
|
| [3] |
袁静文, 武辰, 杜博, 等. 高分五号高光谱遥感影像的城市土地利用景观格局分析[J]. 遥感学报, 2020, 24(4):465-478.
|
| [4] |
王书伟, 席磊, 邱霜, 等. 高光谱遥感森林资源监测原理与应用[J]. 安徽农业科学, 2023, 51(15):111-114,197.
|
| [5] |
杨燕杰. 基于高光谱的高海拔地区矿物信息提取与分析[J]. 世界核地质科学, 2023, 40(4):1020-1028.
|
| [6] |
吕欢欢, 张峻通, 张辉. 多尺度特征与双注意力机制的高光谱影像分类[J]. 光电子·激光, 2024, 35(2):143-154.
|
| [7] |
|
| [8] |
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
| [16] |
|
| [17] |
|
| [18] |
|
| [19] |
|
| [20] |
|
| [21] |
|
| [22] |
|
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
|
| [29] |
|
| [30] |
|
/
| 〈 |
|
〉 |