[An unsupervised classification of hyperspectral images based on pixels reduction with spatial coherence property]

Area of Science:

• Remote Sensing
• Image Processing
• Machine Learning

Context:

• Improving classification accuracy and edge definition in remote sensing data is crucial for accurate analysis.
• Existing unsupervised classification algorithms like K-MEANS and ISODATA have limitations in handling complex spectral variations and spatial relationships.

Purpose:

• To introduce a new unsupervised classification algorithm that enhances classification accuracy and edge definition.
• To leverage Principal Component Space (PRSCP) and linear regression analysis for improved pixel and block-level analysis.

Summary:

• The proposed algorithm initiates classification by assessing pixel spectral similarity, subsequently forming blocks of adjacent, similar pixels using a minimum related window.
• Linear regression models the spectral vectors within each block, with significance validated by F-statistic. Optimal Decision Level Regression (ODLR) estimates block basic vectors, merging blocks with similar vectors into classes.
• The algorithm was evaluated using AVIRIS data and compared against K-MEANS and ISODATA, demonstrating superior performance.

Impact:

• The developed algorithm offers enhanced classification accuracy, improved edge representation, and greater robustness compared to traditional methods.
• This advancement has the potential to improve the analysis and interpretation of hyperspectral and other remote sensing datasets.