[An improved classification approach based on spatial and spectral features for hyperspectral data].

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Context:

• Hyperspectral data classification often struggles to synchronously integrate spatial and spectral information.
• Traditional Markov Random Field (MRF) models face challenges with high computational complexity and time consumption during optimization.

Purpose:

• To propose an improved classification approach for hyperspectral data that effectively combines spectral and spatial information.
• To enhance the estimation accuracy of class conditional probability (CCP) using a probabilistic support vector machine (PSVM).
• To develop an efficient belief propagation (EBP) algorithm with multi-accelerated strategies to overcome computational limitations.

Summary:

• The proposed method utilizes an MRF model with maximum a posteriori to integrate spectral and spatial data.
• A PSVM algorithm refines CCP estimation, while an EBP algorithm with strategies like ordinal propagated message, linearized message-updating, and coarse-to-fine approach optimizes the MRF model.
• The approach was validated using real hyperspectral data from the AVIRIS sensor in Indiana, USA.

Impact:

• Achieved an average classification accuracy of 95.78% and a Kappa coefficient of 93.34%, outperforming traditional MRF algorithms.
• Demonstrated over 3 times improvement in computational efficiency compared to standard belief propagation algorithms.
• Provides a more accurate and efficient solution for hyperspectral image classification, crucial for applications in agriculture and environmental monitoring.