摘要:
遥感图像域适应分类通常以光谱特征或简单的空间特征作为数据特征，通过对齐域间的特征分布来标注缺乏标签样本的域。域适应方法忽略了遥感图像更深层的空间特征， 导致对空间局部信息的发掘不够；同时，域适应方法很少考虑类别非均衡的问题，导致小类样本的重要性被忽略。针对上述问题，提出了一种多级空间特征的平衡测度学习遥感图像域适应分类方法。首先，通过迭代均值滤波器提取遥感图像更深层的空间特征；然后，构建测度学习来对齐边缘分布，最小化类内距离和类间散度；最后，根据类的先验概率构建自适应权重，最大化平衡类间距离，以缓解类别不平衡的问题。所提方法在Worldview-2 数据集和 Pavia 数据集上设计了 4 个任务来评估有效性，使用遥感图像分类任务中常用的总体精度、 Kappa 系数、可视化结果和时间成本作为评价指标，通过实验分析不同参数对分类结果的影响以及各个组成部分对分类结果的影响。相较于次优的算法，所提方法的总体准确率提高了 1.64%~9.18%， Kappa 系数提高了 0.0231~0.1717。实验结果表明，所提方法不仅通过多级空间特征加强了局部特征的提取和有效利用，还有效缓解了类不平衡问题，使分类器取得了更好的预测效果。
Abstract:
Objective: Domain adaptation classification of remote sensing images usually takes spectral features or simple spatial features as data features, and labels domains lacking labeled samples by aligning feature distributions among domains. The domain adaptation method ignores the deeper spatial features of remote sensing images, resulting in insufficient exploitation of spatially localized information. At the same time, the domain adaptation method seldom takes into account the problem of class imbalance, which leads to the importance of the small class of samples being ignored. Methods: To address the above problems, balanced metric learning in multilevel spatial features for domain adaptation in remote sensing image classification is proposed . First, deeper spatial features of remote sensing images are extracted by an iterative mean filter. Then, metric learning is constructed to align the marginal distributions and minimize the intraclass distance and interclass scatter. Finally, adaptive weights are constructed based on the prior probability of the classes to maximize the balanced interclass distances to alleviate the problem of class imbalance. Results: The proposed method first designed four tasks on the Worldview-2 dataset and Pavia dataset to evaluate the effectiveness. Then, the overall accuracy, kappa coefficient, visualization results, and time cost, which are commonly used in remote sensing image classification tasks, are used as the evaluation indicators. Secondly, a series of experiments are carried out to analyze the effects of different parameters on the classification results. Lastly, the individual components on the classification results. Compared with the suboptimal algorithm, the proposed method improves the overall accuracy by 1.64% to 9.18%, and the kappa coefficient by 0.0231 to 0.1717. Conclusions: The experimental results show that the proposed method not only enhances the extraction and effective utilization of local features through multilevel spatial features but also alleviates the problem of class imbalance effectively so that the classifier achieves a better prediction result.
下载:
