An Improved 3D-2D Convolutional Neural Network Based on Feature Optimization for Hyperspectral Image Classification

As a new technology in the field of remote sensing, hyperspectral sensing has been widely used land classification, mineral exploration, environmental monitoring, and other areas. In recent years, deep learning achieved outstanding results image classification tasks. However, problems such as low accuracy for small sample classes unbalanced datasets lack robustness models usually lead to unstable performance images. Therefore, from perspective feature optimization, we propose an improved hybrid convolutional neural network extraction classification. Different current simple multi-scale extraction, first optimize features each scale, then perform fusion. To this end, use 3D dilated convolution design multi-level block (MFB), which can be extract with different correlation strengths at fixed scale. Then, construct spatial interactive attention (SMIA) module enhancement phase, refine through weights interaction, further improve quality features. Finally, experiments were performed on datasets, including balanced samples. The show that proposed model is more accurate extracted are robust.