Hyperspectral Image Classification Using Pyramid Multilayer Perceptron Network and Sample Augmentation

Convolutional neural networks (CNNs) combined with attention mechanism have become the preferred method for hyperspectral image (HSI) classification due to their strong feature extraction and locally contextual modeling ability. However, CNNs with limited receptive fields cannot accurately model long-term dependence between objects in HSI. In this article, we build a new HSI classification framework using multi-layer perceptron (MLP) architecture. To preserve spatial details, our model avoids convolutions or pooling operations that reduce spatial dimensions. Instead, we maintain the spatial dimension and utilize the global receptive field of MLPs to efficiently exploit long-term dependent spatial and spectral features. Specifically, two improved MLPs are proposed for learning long-range spectral and spatial representations, respectively. Additionally, the spatial-spectral residual module is carefully designed to mix spatial and spectral features and capture more discriminative features. Furthermore, to capture rich multiscale hyperspectral features, our MLP-based model is designed as a pyramid structure and includes a pyramid aggregation layer to transmit spatial and spectral feature maps of different scales without losing the fine features, thereby enhancing classification performance. Moreover, to alleviate the problem of small and imbalanced training samples in classification, our method uses sample augmentation techniques to expand and balance the training samples. Experimental results on three benchmark data sets demonstrate that the proposed pyramid MLP achieves better classification performance compared with several recent state-of-the-art models.