Domain Knowledge-Guided Self-Supervised Change Detection for Remote Sensing Images

Abstract

As one of the most popular topics in the field of Earth observation using remote sensing images, change detection (CD) provides great practical and valuable significance for many fields. Although the majority of supervised methods have made great progress by introducing deep learning in the CD field, they are still limited by manually labeled data. In comparison, unsupervised methods do not require manually labeled data, but the accuracy of CD is difficult to be improved due to the lack of constraints or guidance during training. To tackle these issues, we propose a novel domain knowledge-guided self-supervised learning approach for unsupervised CD by fusing the domain knowledge of remote sensing indices during training and inference. Furthermore, we calculate cosine similarity to select the high-similarity feature vectors outputted by the mean teacher and student networks to implement the hard negative sampling strategy, which effectively improves the CD performance. Compared with other supervised and unsupervised CD methods, our proposed approach achieves state-of-the-art performance with a Kap of 53.34% and an F1 of 55.69% on the Onera Satellite Change Detection dataset. Fusing domain knowledge to guide model training and inference obtains an improvement of 5.83% in Kap and 5.13% in F1, which further narrows the performance gap between unsupervised and supervised CD.

Land Cover Change Detection Based on Vector Polygons and Deep Learning With High-Resolution Remote Sensing Images

“…The intrinsically data-driven nature of deep learning often constrains model performance when training data are inadequate. However, the incorporation of prior knowledge can serve as an auxiliary mechanism to enhance both model performance and generalizability, particularly in scenarios with limited data or incomplete labels [49]. For instance, Zhu et al [50] developed a Knowledge-Guided Land Pattern Depicting (KGLPD) framework that leverages OpenStreetMap (OSM) data and prior knowledge to enhance land-use classification accuracy significantly.…”

Forest Disturbance Detection via Self-Supervised and Transfer Learning With Sentinel-1&2 Images

“…The threshold operator was used to generate a binary change map by evaluating the cosine similarity of feature vectors in bi-temporal images. In a similar manner, Siamese-style contrastive learning for change detection combines image and domain knowledge contrastive losses during training and uses self-knowledge distillation from the teacher network during inference to enhance change detection accuracy [58]. Other noteworthy studies in the domain of EO imagery-based landcover change detection using self-supervised and unsupervised learning include using Siamese networks with local and global contrastive losses [59], introducing task-specific Siamese-style contrastive learning with hard sampling and smoothing [60], and extracting features from pre-trained CNNs to generate change maps via clustering [61].…”

Detection of Changes in Buildings in Remote Sensing Images via Self-Supervised Contrastive Pre-Training and Historical Geographic Information System Vector Maps

“…In this approach, the teacher network's weights are updated based on their moving average instead of traditional backpropagation. For example, Yan et al [53] introduced a novel domain knowledge-guided self-supervised learning method. This method selects highsimilarity feature vectors outputted by mean teacher and student networks using cosine similarity, implementing a hard negative sampling strategy that effectively improves CD performance.…”