Abstract
Urban change detection in remote sensing is frequently hindered by environmental noise such as shifting shadows, varying lighting conditions, and parallax errors. Traditional pixel-wise comparison methods often fail to capture the structural and semantic context necessary for accurate detection in dense urban environments. This paper introduces the Local Similarity Siamese Network (LSS-Net), which addresses these challenges through a multi-level attention mechanism and a specialized content loss function. By leveraging cosine similarity across feature maps and perceptual loss, LSS-Net significantly improves detection accuracy, achieving state-of-the-art results on major benchmarks including a newly introduced high-resolution Urban Change Detection (UCD) dataset.
Introduction
Detecting land-use changes in urban landscapes using satellite or aerial imagery is a complex task. Standard convolutional neural network (CNN) models often rely on simple feature subtraction or pixel-level loss functions like Binary Cross-Entropy. These approaches treat pixels as isolated units, making them highly sensitive to non-semantic changes—such as a building's shadow moving throughout the day. To accurately distinguish between real construction or demolition and environmental noise, models must understand the broader structural and semantic relationships within a scene.
Methodology
LSS-Net is built upon three primary architectural innovations designed to enhance structural understanding:
- Siamese Architecture with SE-ResNet-50: The network employs two weight-sharing branches to extract features from temporal image pairs ("before" and "after"). Squeeze-and-Excitation (SE) blocks are integrated to prioritize the most informative feature channels.
- Local Similarity Attention Module (LSAM): Moving beyond simple subtraction, the LSAM calculates the cosine similarity between feature maps at multiple levels. This generates an attention map that highlights structural changes while remaining robust to lighting variations.
- Change Attention map-based Content (CAC) Loss: Inspired by perceptual loss in image synthesis, CAC loss uses a pre-trained VGG-16 network to compare the predicted change areas against the ground truth. This ensures that the detected changes are contextually and structurally consistent with real-world urban features.
Key Insights: Similarity over Subtraction
A core finding of this research is that cosine similarity is a far more robust metric for change detection than traditional feature subtraction or concatenation. While subtraction is sensitive to absolute pixel value changes (like those caused by shadows), cosine similarity focuses on the relationship between feature vectors, effectively capturing changes in the underlying scene structure.
Furthermore, the multi-level fusion approach allows the model to use low-level features (edges and shapes) to refine the boundaries of detected changes, while high-level semantic features confirm the change's location and validity.
Experimental Results
The researchers evaluated LSS-Net on three benchmarks, including the newly introduced UCD (Urban Change Detection) Dataset, which features high-resolution (50 cm/px) imagery of dense, complex urban areas. LSS-Net achieved state-of-the-art (SOTA) performance across all metrics:
- CDD (Change Detection Dataset): F1 score of 0.9630.
- WHU Building Dataset: F1 score of 0.9377.
- UCD Dataset: F1 score of 0.7751.
On the challenging UCD dataset, LSS-Net improved the F1 score by nearly 5% over previous leading models. Qualitative analysis showed that the model produced much cleaner boundaries for new buildings and significantly fewer false positives in shadowed regions compared to baseline methods.
Conclusion and Future Work
LSS-Net demonstrates that incorporating structural similarity and perceptual loss is vital for robust urban change detection. While the model excels at binary change detection, future research will focus on multiclass change detection—identifying the specific nature of the change (e.g., from vegetation to building). Additionally, the authors aim to improve recall in extremely cluttered environments and optimize the computational overhead associated with multi-level attention and auxiliary loss networks.
Summary updated