Deep Learning-based Change Detection in Remote Sensing Images

Remote sensing imagery provides multi-temporal observations of the Earth’s surface and is widely used in change detection tasks. However, practical applications are challenged by limited temporal modelling capability, scarce labelled data, and severe class imbalance. To address these issues, this thesis proposes three complementary change detection methods. First, to better exploit temporal information, a high-resolution triplet network with dynamic multi-scale feature modelling is developed. A triplet input structure is designed, and a dual-branch convolutional architecture with temporal attention is used to capture intertemporal dependencies. A high-resolution backbone with parallel multi-scale convolutions preserves spatial details, while a dynamic fusion module adaptively integrates local and global features. Change maps are generated using cosine similarity and thresholding. Experimental results on multiple high-resolution datasets demonstrate superior performance compared to existing methods. Second, to mitigate the reliance on labelled data, a language-guided change detection framework is proposed. Pre-trained vision-language models are employed to encode textual prompts, which guide visual feature learning via cross-modal attention. A semi-supervised learning strategy is introduced to generate high-quality pseudo-labels. In addition, a hybrid CNN-Transformer architecture is designed to jointly capture local texture and global contextual information. Experiments on public datasets show improved performance over competing approaches. Third, to address sample imbalance and annotation scarcity in SAR imagery, a semisupervised method based on category rebalancing and collaborative learning is proposed. A twin encoder network with multi-scale feature pyramids is used for feature extraction, and a difference enhancement module highlights changed regions. An approximate rank clustering algorithm generates auxiliary pseudo-labels, while a class rebalancing strategy dynamically adjusts sample weights. The collaborative use of network and clustering pseudo-labels effectively expands training data. Extensive experiments on multiple SAR datasets demonstrate significant improvements in robustness and accuracy under imbalanced conditions. Overall, the proposed methods enhance temporal modelling, reduce dependence on labelled data, and improve robustness to data imbalance, advancing the performance of remote sensing change detection.