HGINet

Official Pytorch Code base for "Semantic change detection using a hierarchical semantic graph interaction network from high-resolution remote sensing images"

Project

Introduction

Our study aims to develop an effective multi-task network for SCD from high-resolution remote sensing images. We constructed a hierarchical semantic graph interaction network (i.e., HGINet) to simultaneously improve the identification of changed areas and categories. First, we designed an effective multi-level feature extractor to enhance the extraction of bi-temporal semantic features. Second, we qualified the interactions between different feature layers of the extracted features, the extracted difference features, and both, by graph learning. This improved the detection ability for complex SCD scenarios. Last, we concentrated on boosting the correlations between bi-temporal features for the unchanged areas.

Using the code:

The code is stable while using Python 3.7, torch 1.10.0, CUDA 11.3

• Clone this repository:

git clone https://github.com/long123524/HGINet-torch
cd HGINet-torch

To install all the dependencies using conda or pip:

Torch+
timm
TensorboardX
OpenCV
numpy
tqdm
skimage
timm
....

Data Format

Make sure to put the files as the following structure:

inputs
└── <train>
    ├── im1
    |   ├── 001.tif
    │   ├── 002.tif
    │   ├── 003.tif
    │   ├── ...
    |
    └── im2
    |   ├── 001.tif
    |   ├── 002.tif
    |   ├── 003.tif
    |   ├── ...
    └── label1
    |   ├── 001.tif
    |   ├── 002.tif
    |   ├── 003.tif
    |   ├── ...
    └── label2
    |   ├── 001.tif
    |   ├── 002.tif
    |   ├── 003.tif
    └── ├── ...

For validation and testing datasets, the same structure as the above.

Training and testing

1. Train the model.

run train.SCD.py

2. Predict the SCD results.

run pred_SCD.py

3. Evaluate the SCD results.

run Eval_SCD.py

Semantic change detection datasets:

SECOND dataset: https://drive.google.com/file/d/1QlAdzrHpfBIOZ6SK78yHF2i1u6tikmBc/view HRSCD dataset: https://ieee-dataport.org/open-access/hrscd-high-resolution-semantic-change-detection-dataset#files We also release a Fuzhou non-cropland dataset provided by our partner Haihan Lin: https://drive.google.com/file/d/1SlTw3jKr3cE6d3i5XYQhzylG0geMzNZW/view?usp=sharing

Pretrained weight:

The weight of PVT-V2-B2 pretrained on ImageNet dataset can be downloaded from: https://drive.google.com/file/d/1uzeVfA4gEQ772vzLntnkqvWePSw84F6y/view?usp=sharing

Acknowledgements:

We are very grateful for these excellent works CLCFormer, PVT, and Bi-SRNet, etc., which have provided the basis for our framework.

Citation:

If you find this work useful or interesting, please consider citing the following references.

[1] Long J, Li M,  Wang X, et.al. Semantic change detection using a hierarchical semantic graph interaction network from high-resolution remote sensing images. ISPRS Journal of Photogrammetry and Remote Sensing, 2024, 211:318-335.
[2] Lin H, Wang X, Li M, et al. A Multi-Task Consistency Enhancement Network for Semantic Change Detection in HR Remote Sensing Images and Application of Non-Agriculturalization. Remote Sensing, 2023, 15(21): 5106.