研究目的
To address the challenge of classifying hyperspectral images with limited labeled training samples by proposing a deep transfer learning model that constructs and connects higher-level features between source and target domains to overcome cross-domain disparity without requiring prior knowledge of the target domain.
研究成果
The proposed DTLM method effectively handles the semantic gap in hyperspectral image classification by leveraging deep neural networks and canonical correlation analysis, achieving high accuracy without requiring labeled samples from the target domain. It shows promise for transfer learning in remote sensing but is currently limited to binary classification, with future work needed for multi-class extension and broader application.
研究不足
The current framework is only suitable for binary classification and may not extend directly to multi-class scenarios. It requires careful parameter tuning, such as the number of layers and neurons, which can affect performance. The method is tested on specific datasets (Washington DC Mall and Pavia University), and its generalizability to other HSI or non-HSI datasets is not fully established.
1：Experimental Design and Method Selection:
The study employs a Deep mapping based heterogenous Transfer learning Model (DTLM) that uses stacked auto-encoders (SAE) in both source and target domains, with canonical correlation analysis (CCA) applied at each layer to correlate features and fine-tune the network via back-propagation. A support vector machine (SVM) classifier is used for final classification on the common subspace.
2：Sample Selection and Data Sources:
Two hyperspectral image datasets are used: Washington DC Mall Area 2 (target domain) and Pavia University (source domain), with specific numbers of samples, bands, and classes as detailed in the paper. Transfer samples are selected from these datasets.
3：List of Experimental Equipment and Materials:
Not explicitly mentioned in the paper; assumed standard computing equipment for data processing and neural network training.
4：Experimental Procedures and Operational Workflow:
The process involves setting up SAEs for source and target domains, applying CCA to correlate hidden layers, iteratively fine-tuning the networks to minimize reconstruction error and maximize correlation, projecting data to a common subspace, and using SVM for classification. Experiments are repeated 100 times with random training/testing splits to avoid bias.
5：Data Analysis Methods:
Performance is evaluated using overall classification accuracy, with results averaged over 100 runs. Parameter sensitivity is tested for neuron numbers and layers.
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
