ITER: Image-to-Pixel Representation for Weakly Supervised HSI Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning excels in hyperspectral image (HSI) classification but requires extensive pixel-level annotations.
• Pixel-level labeling of HSI data is challenging due to atmospheric variations, sensor differences, and complex land cover.
• Current methods face limitations due to the laborious and time-consuming nature of acquiring detailed HSI annotations.

Purpose of the Study:

• To propose a novel weakly supervised approach for HSI classification using only image-level annotations.
• To develop a method that bridges the gap between readily available image-level tags and the need for dense, pixel-level predictions.
• To reduce the dependency on extensive and costly pixel-level labeling in HSI analysis.

Main Methods:

• Introduced the Image-To-pixEl Representation (ITER) approach, a two-stage pipeline for HSI classification.
• Developed a pseudo-label generation stage incorporating spectral/spatial activation, alignment loss, and geographic enhancement.
• Implemented a pixel-level prediction stage utilizing a high frequency-aware self-attention mechanism within a transformer architecture.

Main Results:

• ITER successfully predicts pixel-level classification maps for HSI using image-level annotations.
• The proposed method demonstrates competitive performance against state-of-the-art approaches on benchmark HSI datasets.
• The approach effectively refines labels and achieves detailed feature representation for HSI classification.

Conclusions:

• ITER offers a viable solution for weakly supervised HSI classification, significantly reducing annotation requirements.
• The method's ability to leverage image-level tags for dense prediction marks a significant advancement in HSI analysis.
• This work paves the way for more accessible and efficient HSI classification using deep learning.