Reproduce “Urban Change Detection by Fully Convolutional Siamese Concatenate Network with Attention” on OpenResearch

About this paper

Farnoosh Heidary, Mehran Yazdi, Maryam Dehghani, Peyman Setoodeh

Researchers at Shiraz University developed an attention-enhanced fully convolutional Siamese network for urban change detection in Very High-Resolution (VHR) remote sensing images, achieving state-of-the-art performance with high F1-scores and accuracy on challenging datasets like OSCD and RIVER-CD. This approach effectively integrates both preprocessing and internal attention mechanisms to improve change identification.

Key insights

• Integrating both preprocessing (Gaussian Attention) and internal network attention mechanisms (trainable attention gates) significantly enhances a network's ability to focus on pertinent features in VHR images, leading to more accurate change detection.
• The U-Net inspired skip connections, combined with attention gates, effectively leverage both high-level semantic features and fine-grained spatial details, crucial for precise segmentation of changes.
• Using a weighted loss function is essential for mitigating the impact of class imbalance in change detection datasets, ensuring the model learns to identify minority change classes effectively.

Summary by alphaXiv.