Multiscale superpixel depth feature extraction for hyperspectral image classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Superpixel segmentation is common in hyperspectral image (HSI) classification.
• Single-scale superpixels struggle with diverse land-cover structures and irregular shapes, hindering feature extraction.
• Existing methods face challenges in accurately classifying HSI due to these limitations.

Purpose of the Study:

• To propose a novel multiscale superpixel depth feature extraction (MSDFE) method for improved HSI classification.
• To effectively integrate spatial-spectral information by addressing limitations of single-scale superpixel segmentation.
• To enhance the accuracy and robustness of land-cover classification in remote sensing applications.

Main Methods:

• Applied multiscale superpixel segmentation to hyperspectral images (HSI) to capture diverse spatial information.
• Constructed unified two-dimensional statistical features for superpixels of varying shapes and scales.
• Utilized a convolutional neural network for depth feature extraction and classification based on statistical features.
• Implemented an adaptive strategy for fusing multiscale classification results.

Main Results:

• The proposed MSDFE method demonstrated superior performance compared to state-of-the-art methods.
• Experiments on three real hyperspectral datasets validated the effectiveness of the MSDFE approach.
• The method successfully integrated spatial-spectral information for more accurate land-cover classification.

Conclusions:

• The MSDFE method effectively overcomes the limitations of single-scale superpixel segmentation in HSI classification.
• Multiscale superpixel segmentation and depth feature extraction significantly enhance classification accuracy.
• The proposed approach offers a robust solution for remote sensing HSI classification tasks.