A Data Transfer Fusion Method for Discriminating Similar Spectral Classes.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral data offers advanced discrimination of spectrally similar classes.
• Analyzing hyperspectral signatures can be challenging due to high dimensionality and differing data distributions.
• Applying discriminative information across different feature spaces and distributions is a key challenge.

Purpose of the Study:

• To propose a novel data fusion method for hyperspectral image classification.
• To address the challenge of applying discriminative information from training to testing data with varying distributions.
• To improve classification accuracy by incorporating knowledge from existing datasets.

Main Methods:

• A data fusion method based on transfer learning is proposed.
• Transfer learning is integrated into a boosting algorithm.
• Outdated hyperspectral datasets are utilized to guide the classification of new data.

Main Results:

• The proposed method demonstrates significant improvements in classification accuracy.
• Experimental validation was conducted on EO-1 Hyperion and ROSIS hyperspectral datasets.
• The approach outperforms conventional hyperspectral classification methods.

Conclusions:

• The transfer learning-based data fusion method effectively enhances hyperspectral image classification.
• This approach successfully overcomes challenges related to differing data distributions and high dimensionality.
• The method provides a robust solution for accurate hyperspectral data analysis.