Hybrid 2D-3D convolution neural network for hyperspectral image classification

In Research sector for hyperspectral image processing (HSIP), the most challenging and alarming issue is small sample set problem. Deep learning approaches are applied in the classification of hyper spectral image. Using deep learning method, a great accuracy can be achieved for hyperspectral image (HSI) classification. But here using deep learning method, overfitting phenomena can be occurred because of limitation of training sample. Hybrid convolutional neural network is very effective in terms of model size and computational complexity. In our research paper, there are two parts. For dimensionality reduction we have used incremental principal Component Analysis (IPCA). IPCA can fit large dataset into the memory that helpful for this kind of large dataset. Moreover, it processes the data in small batches, updating the principal components iteratively. Here we get the benefit of dealing with large-scale datasets where memory limitations are a concern. Then, for classification, we have used a hybrid architecture with the combination of 2D convolution and 3D CNN. First, 2D convolved is performed with one layer. Then 3D CNN operations perform to extract spatial and spectral features for classification. This results with a great accuracy at a limited number of parameters. We have applied our proposed model on two datasets. The proposed fast hybrid 2D-3D CNN showcases exceptional performance in both accuracy and computational efficiency in compare with the state-of-the-art method.