研究目的
To investigate how insufficient ground truths affect the performance evaluation of hyperspectral image classification methods using accuracy indexes such as OA, AA, and Kappa coefficient.
研究成果
Insufficient ground truths can lead to overoptimistic accuracy assessments and unreliable performance evaluations in hyperspectral image classification. Future work should focus on designing more objective accuracy indexes for cases with insufficient ground truth.
研究不足
The study is limited to the Indian Pines dataset and specific classification methods; results may not generalize to other datasets or methods. The ground truth downsampling strategy may not cover all real-world scenarios.
1：Experimental Design and Method Selection:
The study uses a framework with two stages: constructing experimental environments and performing experiments. It involves subjective and objective comparisons using various classification methods and evaluation metrics.
2：Sample Selection and Data Sources:
The Indian Pines dataset is used, acquired by the Airborne Visible Infrared Imaging Spectrometer, with a spatial size of 145x145 pixels. Ground truths are downsampled from the original.
3：List of Experimental Equipment and Materials:
Hyperspectral image data, classification methods (SVM, KBTC, EPF, LBP), and evaluation metrics (PLCC, RMSE, SRCC, KRCC).
4：Experimental Procedures and Operational Workflow:
Classification is performed multiple times with different training sample sizes. Ground truths are constructed by downsampling, and accuracies are measured and compared.
5：Data Analysis Methods:
Statistical analysis using correlation coefficients and error metrics to assess robustness and reliability.
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
