Multiscale Dual-Branch Residual Spectral–Spatial Network With Attention for Hyperspectral Image Classification

The development of remote sensing images in recent years has made it possible to identify materials inaccessible environments and study natural on a large scale. But hyperspectral (HSIs) are rich source information with their unique features various applications. However, several problems reduce the accuracy HSI classification; for example, extracted not effective, noise, correlation bands, most importantly, limited labeled samples. To improve case training samples, we propose multiscale dual-branch residual spectral–spatial network attention classification model named MDBRSSN this article. First, due redundancy between principal component analysis operation is applied preprocess raw data. Then, MDBRSSN, structure designed extract useful HSI. advanced feature, abstract by convolution neural network, image processing, which can complex data accuracy. In addition, mechanisms separately each branch enable optimize refine feature maps. Such an framework learn fuse deeper hierarchical fewer purpose designing have high compared state-of-the-art methods when samples limited, proved results experiments article four datasets. Salinas, Pavia University, Indian Pines, Houston 2013, proposed obtained 99.64%, 98.93%, 98.17%, 96.57% overall using only 1%, 5%, 5% training, respectively, much better methods.