Abstract
Change detection in heterogeneous remote sensing images is a challenging problem because it is hard to make a direct comparison in the original observation spaces, and most methods rely on a set of manually labeled samples. In this article, a spatially self-paced convolutional network (SSPCN) is constructed for change detection in an unsupervised way. Self-paced learning (SPL) is incorporated into convolutional networks to dynamically select reliable samples and learn the representation of the relations between the two heterogeneous images. In the proposed method, the pseudo labels are initialized by a classification-based method, and each sample is assigned to a weight to reflect the easiness of the sample. Then, SPL is used to learn the easy samples at first and then gradually take more complex samples into account. In the training process, the sample weights are dynamically updated based on the network parameters. Finally, a binary change map is acquired based on the trained convolutional network.
Introduction
Detecting changes in the same geographical area over time is a cornerstone of remote sensing, used for everything from disaster assessment to urban planning. However, two major hurdles persist in this field:
- Heterogeneity: We often need to compare images from different sensors (like Synthetic Aperture Radar and Optical). Because these sensors capture the world through different physical processes, their pixel values cannot be compared directly.
- Label Scarcity: Most high-performing deep learning models require massive amounts of manually labeled data. In remote sensing, labels are expensive and often unavailable.
While unsupervised methods exist, they typically generate "pseudo-labels" that are riddled with noise. If a standard Convolutional Neural Network (CNN) is trained on these noisy labels, it simply learns the errors, leading to poor performance.
Methodology
The authors propose a three-stage approach to bridge the gap between heterogeneous images:
- Pseudo-Label Generation: A classification-based clustering method creates an initial, "coarse" change map. This serves as the starting point for training.
- Convolutional Mapping: A three-layer CNN is constructed to learn the representation of relations between the two images. Notably, the architecture omits pooling layers to maintain pixel-wise precision, which is critical for accurate change maps.
- Spatially Self-Paced Learning (SSPCN): This is the core innovation. Inspired by human learning (starting with easy concepts before moving to hard ones), the model assigns a weight to each sample based on its "easiness" (loss value).
- The "Self-Paced" Part: In early iterations, the model only trains on "easy" samples where the pseudo-labels are likely correct. As training progresses, it gradually incorporates more "complex" samples.
- The "Spatial" Part: Unlike standard self-paced learning, SSPCN considers the neighborhood. If a pixel's neighbors also have low loss, the model gains confidence that the sample is reliable, effectively filtering out isolated noise (like SAR speckle).
Experimental Evaluation
The model was evaluated on five distinct datasets, covering both homogeneous and heterogeneous scenarios, including the Farmland (SAR-to-SAR), Sardinia (Near-infrared TM to Optical), and Yellow River (SAR to Optical) datasets.
Key Results
The SSPCN was compared against traditional unsupervised methods (like GGKI and EMLS) and supervised benchmarks (like RKSVM and standard CNNs).
- Superior Accuracy: On the Sardinia dataset, SSPCN achieved a Kappa Coefficient (KC) of 0.8416, significantly outperforming traditional unsupervised methods which hovered around 0.55–0.69.
- Better Trade-offs: In the Shuguang Village dataset—a particularly difficult case due to complex landforms—SSPCN achieved the highest Intersection over Union (IoU) of 69.06%, whereas the standard CNN trained on the same pseudo-labels only reached 63.15%.
- Visual Clarity: The resulting change maps were qualitatively "cleaner," with fewer isolated false positives compared to methods that don't use spatial self-pacing.
Discussion and Conclusion
The proposed SSPCN has three main advantages compared to traditional methods. First, it is robust to noisy samples because it involves only reliable samples in training. Second, samples have different learning rates for converging to better values, which are dynamically changed based on current sample weights. Finally, spatial information among samples is used to further enhance robustness.
By integrating spatial consistency as a denoising filter, the model assumes that changes occur in clusters, helping it ignore the "salt-and-pepper" noise common in unsupervised clustering. Experiments confirm that even if initial pseudo-labels are poor, SSPCN can "self-correct" by focusing on the most consistent samples first. Future work may explore deeper architectures or attention mechanisms to further enhance feature extraction.
Summary updated