研究目的
To design a model that uses as few labelled samples as possible and reduces the computational time for the classification of hyperspectral images (HSI) using an active learning (AL) approach based on Extreme Learning Machine (ELM).
研究成果
The proposed ELM-AL approach significantly reduces computation time compared to SVM-AL while maintaining comparable classification accuracy. The results suggest that ELM classifier can be further utilized for other existing AL approaches, and the classification accuracy can be improved by integrating kernel function in ELM classifier.
研究不足
The study focuses on the reduction of computation time while maintaining classification accuracy, but the classification accuracy of ELM-based methods is slightly lower than SVM-based methods. The performance depends on the activation function and the number of hidden nodes in ELM.
1：Experimental Design and Method Selection:
The study proposed an AL approach based on ELM for HSI classification, comparing its performance with SVM-based AL in terms of classification accuracy and computational time. Two query strategies were analyzed: random sampling and multiview with adaptive maximum disagreement (MV-AMD).
2：Sample Selection and Data Sources:
Two HSI datasets, Kennedy Space Centre (KSC) and Botswana (BOT), were used. Pre-processing was performed, and datasets were split into training and testing sets, with initial training samples randomly selected from each class.
3：List of Experimental Equipment and Materials:
The experiment was carried out in Matlab-2016 running on an i3-4130, cpu@3.40GHz. SVM-based AL was implemented using active learning toolbox for remote sensing (ALTB), and ELM-based AL was implemented by replacing SVM with ELM in ALTB.
4：40GHz. SVM-based AL was implemented using active learning toolbox for remote sensing (ALTB), and ELM-based AL was implemented by replacing SVM with ELM in ALTB.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The datasets were segmented into five disjoint feature subsets (views) using correlation-partition-based clustering. ELM models were constructed for each view, and informative samples were selected using maximum disagreement query function. The process was repeated until a predefined stopping criterion was met.
5：Data Analysis Methods:
The performance was evaluated based on overall accuracy vs. samples in training set curve and computation time. The experiment was repeated 20 times with different initial training sets, and the average execution time was recorded.
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
