研究目的
To propose a new method for feature extraction of hyperspectral images that models spatial information based on spectral features, reduces computational cost and dimensionality, and improves classification performance.
研究成果
The proposed MQWLDH method effectively extracts spatial features from hyperspectral images by integrating quaternion representation and Weber local descriptors, reducing dimensionality and computational cost while achieving superior classification performance across multiple data sets and classifiers, demonstrating robustness and efficiency.
研究不足
The method relies on the first three principal components, which may not capture all spectral information; computational complexity increases with more PCs; parameter tuning (ω, R, M) is data-dependent and may require optimization; applicability to very high-dimensional HSIs or other data types not explored.
1：Experimental Design and Method Selection:
The method involves dimensionality reduction via PCA to extract the first three principal components, representation using quaternions to unify processing, construction of a quaternion Weber local descriptor (QWLD) to characterize spatial variations, multiscale feature histogram construction to capture spatial information at different scales, and fusion of spectral and spatial features for classification using SVM and SRC classifiers.
2：Sample Selection and Data Sources:
Three hyperspectral image data sets are used: Indian Pines (AVIRIS sensor, 145x145 pixels, 200 bands), University of Pavia (ROSIS sensor, 610x340 pixels, 103 bands), and Salinas (AVIRIS sensor, 512x217 pixels, 204 bands). Training samples are randomly selected (e.g., 10% for Indian Pines, 1% for others), with the rest for testing.
3：List of Experimental Equipment and Materials:
Hyperspectral images from AVIRIS and ROSIS sensors; MATLAB software for implementation; computer with
4：60 GHz CPU and 0 GB RAM. Experimental Procedures and Operational Workflow:
Normalize HSI, apply PCA for dimensionality reduction, represent first three PCs as quaternions, compute QWLD (differential excitation and orientation features), construct multiscale feature histograms, fuse with spectral features, and classify using SVM or SRC.
5：Data Analysis Methods:
Performance evaluated using overall accuracy (OA), average accuracy (AA), kappa coefficient (κ), and standard deviations from ten independent runs; parameter tuning for ω, R, and M; comparison with other methods (SOMP, SADL, ELM-CK, KELM-CK, EMPS).
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
