Joint superpixel and Transformer for high resolution remote sensing image classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks and superpixel segmentation advance high-resolution remote sensing image (HRI) classification.
• Existing methods often stack features across scales, neglecting inter-object contextual dependencies.

Purpose of the Study:

• To introduce a Joint Superpixel and Transformer (JST) framework for HRI classification.
• To address the limitation of not considering contextual dependencies between segmented objects in current methods.

Main Methods:

• HRI data is segmented into superpixel objects.
• A Transformer model is employed to capture long-range dependencies between superpixel objects.
• An encoding and decoding Transformer architecture is designed to model contextual relationships and predict object classes.

Main Results:

• The JST framework achieved high classification accuracy on two HRI datasets, with overall accuracy reaching 0.91 and Kappa coefficients up to 0.89.
• The study explored the impact of semantic range on classification performance.
• Qualitative and quantitative comparisons demonstrated JST's competitive and superior performance against benchmark methods.

Conclusions:

• The proposed JST framework effectively models inter-object contextual dependencies for improved HRI classification.
• JST offers a novel and effective approach for analyzing complex remote sensing imagery.
• The method shows significant potential for advancing HRI classification tasks.