A Multi-scale Convolutional Neural Network Based on Multilevel Wavelet Decomposition for Hyperspectral Image Classification

Deep learning methods have outclassed traditional methods in hyperspectral image classification (HIC). However, due to the limited size of input 3D cube, most HIC networks are shallow in depth, resulting in effectiveness constrained. Multi-scale convolution operation has arisen to alleviate this issue by expanding width of neural networks. Though multi-scale CNN is able to enhance the classification capacity in some extent, there are still some inherent disadvantages, such as heavy computation burden and inadequate multi-scale features extraction. To address these, this paper proposes a multi-scale dense network based on multilevel wavelet decomposition for HIC. The proposed approach characterizes multi-scale features by establishing multiple branches with 2D discrete wavelet transform rather than multi-scale convolution and pooling. Various features are joined with others from adjacent branches by level-wise short cut for supporting classification decision. Otherwise, the dense block is modified to fuse multi-frequency features and fetch underlying scale information with dense connections. The experimental results on Pavia University and University of Houston datasets demonstrate that proposed approach acquires competitive performance with the state-of-the-arts and only requires less computation and time consumptions. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022.