Deep Learning for Satellite Building Change Detection
A deep-learning project for detecting building changes in bitemporal satellite imagery using the LEVIR-CD dataset. It compares a shallow convolutional baseline, an early-fusion U-Net, and a Siamese Temporal Attention U-Net under a shared training and evaluation pipeline. The temporal-attention model achieves the strongest test F1-score of 0.9019, demonstrating the value of shared feature extraction and explicit cross-time comparison for robust change segmentation.
See Project
