高光谱图像分类在遥感影像分析中有着广泛的应用。高光谱图像包括不同波段的图像。目前高光谱分类常用的模型有支持向量机，二维卷积网络，但这些方法没有同时考虑到图像光谱和空间特征，导致分类效果不佳。三维卷积由于计算复杂度高并不会大规模使用。本文提出了一种引入了小波分解的混合型神经网络，该模型是一种改良后的用于多分辨率高光谱分类的混合神经网络。这个混合模型是由三维卷积和加入了小波分解的二维卷积构成的。三维卷积可以从一堆光谱波段中联合表示空间-光谱特征。而小波CNN的内部利用多层小波变换来提取抽象的空间特征。此外，小波变换的计算量比三维CNN要少，与单独使用三维卷积相比，使用混合网络降低了模型的复杂性。最终得到了一个新的模型，可以对多分辨率HSI数据进行高精度分类。为了测试这种混合方法的性能，在Indian Pines和HyMap采集的白干湖幅遥感数据集上进行了高光谱图像分类实验。研究提出的高光谱图像分类模型取得了优异的实验结果。

Hyperspectral image (HSI) classification plays a crucial role in the analysis of remote sensing imagery, offering insights across various spectral bands. Among the array of deep learning techniques, Convolutional Neural Networks (CNNs) stand out as potent tools. However, much of the recent advancements have primarily revolved around two-dimensional CNN architectures. Despite their efficacy, prevailing methods such as Support Vector Machines (SVM), 2D CNNs, and 3D CNNs fall short in simultaneously capturing both spectral and spatial features, resulting in suboptimal classification outcomes. Moreover, the widespread adoption of 3D CNNs is hindered by their computationally intensive nature.

To address these challenges, we propose a novel approach—a hybrid neural network integrating wavelet decomposition. Our model represents an enhanced version of the 3D-2DCNN architecture tailored for multi-resolution hyperspectral classification. Leveraging the advantages of both 3D CNNs and 2D CNNs with wavelet decomposition, our hybrid model achieves a balance between computational efficiency and classification performance. By incorporating wavelet transform, which is less computationally demanding compared to 3D CNNs, we mitigate the complexity associated with traditional 3D CNN models.

Our experiments, conducted on the Indian Pines Dataset and Baigan Lake remote sensing dataset acquired by HyMap, validate the efficacy of the proposed hybrid method. The results demonstrate superior classification accuracy, underscoring the potential of our approach in accurately classifying multi-resolution HSI data.