A Model Based on Dense Connection 3D-2D-CNN for Hyperspectral Image Classification

Hyperspectral image (HSI) classification is the basis of various hyperspectral applications. From classical machine learning theories such as SVMs to 2D-CNN and 3D-CNN, deep learning model frameworks represented by CNN have been widely used in HSI classification and have made remarkable achievements. However, the tagging cost of HSI is high, and the amount of training data is small, which leads to the challenges of HSI classification based on CNN. Aiming at the problem of limited samples, from the point of view of network optimization, we design a CNN model with 15-layer network structure D-HybridSN. On the basis of FA dimension reduction, the HSI are classified using Cross-Entropy function through the organic combination of 3D-2D-CNN, dense connection mechanism, attention mechanism and depth-separable convolution. D-HybridSN can better learn deep spectral spatial characteristics in the case of small samples. The experimental results show that D-HybridSN achieves good classification effect in both Indian Pines and Salinas datasets when the amount of training data is small.