Densely Connected Multiscale Attention Network for Hyperspectral Image Classification

Hyperspectral images (HSIs) are characterized by high spatial resolution and rich in spectral information. In the process of HSI classification, extraction spectral-spatial features directly influences classification results. recent years, hyperspectral method based on convolutional neural networks has demonstrated excellent performance. However, as network structure deepens, degradation occurs, learned from fixed-scale kernels usually specific, which is not conducive to feature learning thus impairs accuracy. To solve problem difficult underutilization information data, a densely connected multiscale attention 3-D convolution proposed for classification. First, reduce redundancy HSIs, principal component analysis algorithm performed raw data; then, several blocks comprised parallel factorized spatial-spectral modules different sizes adopted extract enriched HSIs; furthermore, dense connections introduced further fuse obtained depths, thereby enhancing reuse propagation helping alleviate vanishing gradients. Besides, channel-spectral-spatial block put forward spontaneously reweight fused emphasize that more relevant results while weakening less ones. The experimental show effective extracting discriminative target outperforms other state-of-the-art methods.