Reproduce “Performance Analysis of Various EfficientNet Based U-Net++ Architecture for Automatic Building Extraction from High Resolution Satellite Images” on OpenResearch

About this paper

Tareque Bashar Ovi, Nomaiya Bashree, Protik Mukherjee, Shakil Mosharrof, Masuma Anjum Parthima

Researchers from the Military Institute of Science and Technology systematically evaluated various EfficientNet-based U-Net++ architectures for automatically extracting buildings from high-resolution satellite imagery. Their EfficientNetb4-based U-Net++ model achieved a mean accuracy of 92.23% and a mean IoU of 88.32% on the Massachusetts Buildings Dataset, surpassing the performance of several established deep learning methods.

Key insights

• The synergistic combination of EfficientNet's efficient feature extraction capabilities with U-Net++'s advanced segmentation architecture effectively addresses the challenges of semantic information extraction in high-resolution satellite imagery.
• Deeper EfficientNet backbones generally lead to improved performance, with the EfficientNetb4 variant yielding the highest accuracy and IoU among the configurations tested.
• The proposed models demonstrated stable training behavior without overfitting, indicating good generalization potential for accurate building extraction in complex geospatial scenarios.

Summary by alphaXiv.