研究目的
To propose a stacked autoencoders-based adaptive subspace model (SAEASM) for hyperspectral anomaly detection that improves upon existing methods by leveraging deep learning for feature extraction.
研究成果
The proposed SAEASM algorithm generally outperforms comparison algorithms (LRX, BJSR, TBASD) in hyperspectral anomaly detection across synthetic and real datasets, as evidenced by higher AUC values and better separability, though it has higher computational complexity and weaker background suppression in some cases. Future work should focus on automating parameter selection.
研究不足
The algorithm requires manual parameter tuning based on experience and repeated experiments for optimal performance; automatic parameter setting is not addressed and is noted as an area for future improvement.
1：Experimental Design and Method Selection:
The study uses a novel SAEASM algorithm that integrates stacked autoencoders (SAE) with an adaptive subspace model based on residual ratio theory. It involves defining three windows (inner, outer, dictionary) around test points in hyperspectral images to extract local features and compute detection results using 2-norms of deep features.
2：Sample Selection and Data Sources:
One synthetic hyperspectral dataset (105x100 pixels, 102 bands from ROSIS sensor over Pavia) and two real datasets from AVIRIS sensor (San Diego with 3 planes: 120x120 pixels, 126 bands; and 38 planes: 100x100 pixels, 126 bands) are used for evaluation.
3：List of Experimental Equipment and Materials:
An Intel Core i7 CPU machine with 16 GB RAM running Matlab 2017a software for implementation and analysis.
4：Experimental Procedures and Operational Workflow:
For each test pixel, extract local background and dictionary pixels using windows, compute deep features of differences using SAE architectures (with sigmoid activation, mean square error loss, SGD optimizer), and calculate detection result based on 2-norms. Compare performance with LRX, BJSR, and TBASD algorithms using ROC curves and separability maps.
5：Data Analysis Methods:
Performance is evaluated using area under the ROC curve (AUC), ROC curves, separability maps, and execution time measurements. Parameters (e.g., window sizes, SAE dimensions) are optimized through sensitivity analysis.
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
