Comparative Analysis of the Multiscale Convolutional Neural Network and Advanced Deep Learning Models for Land Cover Classification Using Hyperspectral Imagery

Hyperspectral images (HSI) provide detailed spectral information, making them invaluable for numerous applications. The current study proposes a multiscale 3D Convolutional Neural Network (CNN) for extracting land cover from Chikusei hyperspectral data. The proposed architecture consists of three parallel convolutional channels, each employing different kernel sizes to extract features across various spatial resolutions. The performance of this model was compared to other deep learning models, including Deep Neural Networks (DNNs), 1D-CNNs, and 3D-CNNs. Results showed that the multiscale 3D-CNN achieved an overall accuracy of 97.73% an average accuracy of 97.54%, and Kappa coefficient of 97.00, indicating excellent agreement between predicted and actual land cover classes. The multiscale 3D-CNN model demonstrated superior computational efficiency and accuracy compared to the evaluated deep learning models.