Figure 5.

The architecture diagram for the image-wise classification. The original dimensions of the hyperspectral images were 512 × 512 pixels × 203 spectral bands. The automated feature extraction network used was an architected based on AlexNet [78], but the hyperparameters filters, kernel sizes, and pool sizes were experimented on in parallel and were kept the same for both the 2D- and 3D-CNN for a comparison. The 2D-CNN reduced the data cubes to 16 spatial features × 16 spatial features × 64 convolutional filters, while the 3D-CNN reduced the data cubes to 16 spatial features × 16 spatial features × 5 spectral features × 64 convolutional filters. Reduced features were flattened and supplied to the RF and SVM machine learning classification pipelines.