A 3D-2D Convolutional Neural Network and Transfer Learning for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification faces challenges due to pixel inseparability and the need for efficient spectral-spatial data interpretation.
• Existing methods struggle with precise computational time for complex HSI data.
• Advancements in remote sensing and spectral imagery necessitate improved classification techniques.

Purpose of the Study:

• To propose an efficient and accurate hyperspectral image classification model.
• To address the spectral-spatial separability challenge in HSI data.
• To leverage transfer learning and advanced convolutional neural networks for improved classification.

Main Methods:

• A 3D-2D convolutional neural network (CNN) model integrating ResNeXt-50 blocks was developed.
• Early layers utilized 3D convolutions for spectral-spatial information modeling, followed by 2D convolutions for semantic abstraction.
• Principal Component Analysis (PCA) was employed to enhance class separability and data representation before network input.

Main Results:

• The proposed model achieved high classification accuracies on five public HSI datasets: Indian Pines (99.85%), Pavia University Scene (99.98%), Salinas Scene (100%), Botswana (99.82%), and Kennedy Space Center (99.71%).
• The model demonstrated superior performance compared to several state-of-the-art (SOTA) deep learning and standard classification methods.
• Instantaneous representation of spectral-spatial information was efficiently achieved.

Conclusions:

• The developed 3D-2D CNN model with PCA effectively improves hyperspectral image classification accuracy.
• The model offers a robust solution for analyzing spectral-spatial information in HSI data.
• This research provides valuable insights into advanced HSI classification techniques.