【目的】 异源影像变化检测是遥感领域的研究热点，现有研究通常先把异源影像通过域迁移方法得到伪同源影像，再从伪同源影像中提取变化信息，存在异源影像中地物特征的互补信息利用不充分问题。本文围绕域迁移得到的多时相异源影像对（每个时相都有一对异源影像）开展变化检测，为解决多时相异源影像对的时相-空间-频谱联合特征融合提取，并在无标注样本情况下进行无监督变化检测等极具挑战性的问题，提出了一种联合混合卷积网络和相似性评估的异源影像无监督变化检测方法。【方法】 首先利用本文提出的基于斑块相似性评估的伪标签样本数据集生成方法，克服深度学习变化检测模型对人工标注样本数据的依赖。然后，构建了融合异源影像的2维、3维混合卷积神经网络模型，用于提取变化信息。【结果】 该模型不仅可充分利用异源影像之间的特征互补优势，还可有效提取多时相异源影像对的时相-空间-频谱联合特征，可在伪标签样本数据驱动下实现无监督变化检测。为了验证本文提出方法的有效性，在基于循环一致生成对抗网络获取的多时相异源影像对数据上进行了实验，包括曙光和格洛斯特数据集，并与传统方法、深度学习方法和本文方法的消融方法进行了定性和定量比较。【结论】 本文方法通过样本数据的无监督选取，提高了异源影像变化检测的自动化程度；同时，与前述对比方法中最好的变化检测结果相比，本文方法在2个数据集上的平均F1-Score提高了2.54%，有效提高了变化检测结果的可靠性。

[Objectives] Change detection in heterogeneous images is a research hotspot in the field of remote sensing. Existing studies typically obtain pseudo-homogeneous images from heterogeneous images through domain adaptation methods and then extract change information from these pseudo-homogeneous images, which often leads to insufficient utilization of complementary information from features in heterogeneous images. This study focuses on change detection using multi-temporal heterogeneous image pairs (each time having a pair of heterogeneous images) obtained through domain adaptation. To address the challenging issues including the temporal-spatial-spectral joint features fusion extraction from multi-temporal heterogeneous image pairs and perform unsupervised change detection without labeled samples. This study proposes an unsupervised change detection method for heterogeneous images that combines a hybrid convolutional network with similarity assessment. [Methods] To overcome the dependence of deep learning change detection models on manually labeled sample data, this method first utilizes a pseudo-labeled samples generation method based on patch similarity assessment proposed in this study. Then, a 2D-3D hybrid convolutional neural network model that integrates heterogeneous images is constructed to extract change information. [Results] This model can not only fully utilize the complementary advantages of features between heterogeneous images but also effectively extract the temporal-spatial-spectral joint features from multi-temporal heterogeneous image pairs, enabling unsupervised change detection driven by pseudo-labeled sample data. To validate the effectiveness of the proposed method, experiments were conducted on Shuguang and Gloucester datasets. Qualitative and quantitative comparisons were made with traditional methods, deep learning methods, and ablation methods derived from the proposed method. [Conclusions] The experimental results show that the proposed method enhances the automation level of change detection in heterogeneous images through unsupervised selection of sample data. Meanwhile, compared to the best change detection results among the aforementioned comparison methods, the average F1 Score of the proposed method is improved by 2.54% on both datasets, effectively enhancing the reliability of the change detection results.