研究目的
To address the challenge of expensive labeled data acquisition in hyperspectral image classification by proposing a domain adaptation method that uses existing labeled data from similar scenes to classify new scenes, thereby reducing the need for new labeled data.
研究成果
The proposed domain adaptation method effectively improves hyperspectral image classification by incorporating prior class distribution from the source domain and manifold embedding in the target domain, leading to better generalization performance. Experimental results on Pavia Center and Indian Pines datasets show superior accuracy compared to baseline methods, confirming the method's effectiveness. Future work could explore extensions to more diverse datasets and integration with other deep learning techniques.
研究不足
The method assumes that the source and target domains are similar, which may not hold for highly dissimilar scenes. It relies on the availability of labeled data in the source domain and may not perform well if the class distributions are very different. The neural network architecture and parameters (e.g., number of layers, neighbors) are fixed and may require tuning for different datasets. The experiments are limited to two specific hyperspectral images, and generalization to other types of remote sensing data is not verified.
1：Experimental Design and Method Selection:
The method involves a neural network-based approach for domain adaptation, using maximum mean discrepancy (MMD) with class weighting to minimize distribution shift between source and target domains, and manifold regularization to preserve local structures in the target domain. The neural network is trained with minibatch stochastic gradient descent to optimize an objective function combining classification loss, weighted MMD loss, and manifold loss.
2：Sample Selection and Data Sources:
Two hyperspectral datasets are used: Pavia Center (from ROSIS-3 sensor, 1096x715 pixels, 102 bands) and Indian Pines (from AVIRIS sensor, 145x145 pixels, 220 bands). For each, a subset is selected as the source domain with labeled samples, and the rest as the target domain with unlabeled samples, using a 'biased sampling' strategy to induce distribution bias.
3：List of Experimental Equipment and Materials:
Hyperspectral images from ROSIS-3 and AVIRIS sensors; computational resources for neural network training (not specified in detail).
4：Experimental Procedures and Operational Workflow:
Data preprocessing (selecting source and target domains), neural network setup (two fully connected layers with 64 and 32 units), parameter tuning (α and β from {0.01, 0.1, 1, 2}, k=3 neighbors for manifold), training with minibatch SGD, and evaluation using overall accuracy and kappa coefficient.
5：01, 1, 1, 2}, k=3 neighbors for manifold), training with minibatch SGD, and evaluation using overall accuracy and kappa coefficient.
Data Analysis Methods:
5. Data Analysis Methods: Performance is evaluated by comparing overall accuracy and kappa values with baseline methods (SVM and a state-of-the-art domain adaptation method). Parameters α and β are analyzed for sensitivity.
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
