AFDE-Net: Building Change Detection using Attention-Based Feature Differential Enhancement for Satellite Imagery

Shimaa Holail ResearchgateShimaa Holail Linkedin, Tamer Saleh, Xiongwu Xiao, and Deren Li

The train and test code will be released soon. The EGY-BCD dataset is available.

Updates

⚑ AFDE-Net has been Published in IEEE Geoscience and Remote Sensing Letters (GRSL). DOI: 10.1109/LGRS.2023.3283505

πŸ’¬ EGY-BCD Dataset Description

Bi-temporal images in the EGY-BCD dataset are taken from 4 different regions located in Egypt, including New Mansoura, El Galala City, New Cairo, and New Thebes. The figure below shows the building changes in New Mansoura City and New Thebes. Our image data capture time varies from 2017 to 2022. The images feature seasonal changes and different lighting changes in our new dataset, which can help develop effective methods that can mitigate the impact of unrelated changes on real changes.

image-20230201153142126

πŸ’¬ Network Architecture

An overview of the proposed architecture. (a) is the backbone of ADFE-Net. The node $Xi$ features denotes a convolutional block as shown in (b). The feature differential enhancement FDE is described in (c). (d) is the ensemble spatial-channel attention fusion ESCAF. image-20230201153142126

πŸ’¬ Quantitative and Qualitative Results on the EGY-BCD Dataset

πŸ‘‰ Quantitative Results

image-QuantitativeResult

πŸ‘‰ Qualitative Results

image-QualitativeResult

πŸ’¬ Requirements

  • Python 3.7
  • Pytorch 1.7

Please see requirements.txt for all the other requirements.

πŸ”­ Baselines

Encoders

Below is a list of encoders used in this work and their pre-trained weights.

ResNet
Encoder Weights Params, M
resnet50 imagenet / ssl / swsl 23M
resnet101 imagenet 42M
RegNet(x/y)
Encoder Weights Params, M
timm-regnety_120 imagenet 49M
timm-regnety_160 imagenet 80M
timm-regnety_320 imagenet 141M

πŸ’¬ Dataset Preparation

πŸ‘‰ Data Structure

"""
EGY-BCD dataset with pixel-level binary labelsοΌ›
β”œβ€”β€”β€”β€”train
|      β”œβ€”β€”β€”A  
|      β”œβ€”β€”β€”B
|      β”œβ€”β€”β€”label
|
β”œβ€”β€”β€”β€”val
|      β”œβ€”β€”β€”A  
|      β”œβ€”β€”β€”B
|      β”œβ€”β€”β€”label
|
β”œβ€”β€”β€”β€”test
|      β”œβ€”β€”β€”A  
|      β”œβ€”β€”β€”B
|      β”œβ€”β€”β€”label
"""

A: Images of Time1; B:Images of Time2; label: Ground Truth; each file contains the image names (XXXX.png)

🚚 Datasets

You can download our novel public EGY-BCD dataset through the following link:

πŸ“ƒ Citing

@ARTICLE{10145434,
  author={Shimaa Holail, Tamer Saleh, Xiongwu Xiao, and Deren Li},
  journal={in IEEE Geoscience and Remote Sensing Letters}, 
  title={AFDE-Net: Building Change Detection Using Attention-Based Feature Differential Enhancement for Satellite Imagery}, 
  year={2023},
  volume={20},
  number={6006405},
  pages={1-5},
  doi={10.1109/LGRS.2023.3283505}}

Contact Information

If you have any question about EGY-BCD dataset, please contact shimaaholail@whu.edu.cn

πŸ’¬ References

Appreciate the work from the following repositories: