研究目的
To solve the 'massive information but low accuracy' problem in hyperspectral image classification by developing a method that integrates spatial and spectral features using belief propagation and small sample learning.
研究成果
The proposed MFSSL-BPMRF method effectively addresses the low accuracy issue in HSI classification by integrating spatial and spectral features with belief propagation and small sample learning, achieving high accuracy on tested datasets. However, challenges such as long run times and class imbalance remain, indicating a need for further optimization in future research.
研究不足
The method has high time complexity, especially with larger training sets, and may perform poorly on imbalanced classes where small class instances have low probabilities. Future work includes parameter adjustment and optimization for better balance and reduced computation time.
1：Experimental Design and Method Selection:
The method involves extracting spatial features using the extended morphological multi-attributes profiles (EMAPs) algorithm, fusing spatial and spectral features, and applying belief propagation Markov random field (BPMRF) for classification. Small sample learning is used to improve efficiency.
2：Sample Selection and Data Sources:
Two public hyperspectral datasets are used: AVIRIS Indian Pines (145x145 pixels, 220 bands, 16 classes) and ROSIS Pavia University (610x340 pixels, 115 bands, 9 classes). Small sample training sets are selected randomly (e.g., 4% for Indian Pines, 0.5% for Pavia University).
3：5% for Pavia University).
List of Experimental Equipment and Materials:
3. List of Experimental Equipment and Materials: A workstation with an Intel Xeon CPU (@3.3 GHz) and 8 GB RAM is used for computations. Software includes MATLAB
4：3 GHz) and 8 GB RAM is used for computations. Software includes MATLAB Experimental Procedures and Operational Workflow:
2015.
4. Experimental Procedures and Operational Workflow: The workflow includes feature extraction via EMAPs, feature fusion, random sampling for small sample sets, computation of neighborhood matrices, RBF learning, MLR regression using LORSAL, probability computation, and BP iteration for classification.
5：Data Analysis Methods:
Classification accuracy is evaluated using average accuracy (AA), overall accuracy (OA), and kappa coefficient. Comparisons are made with reference methods such as SPE_MLR, SPE_BP, EMAP_MLR, EMAP_BP, and SP_EM_MLR.
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
