SESNet for remote sensing image change detection

It is the implementation of the paper: "SESNet: A Semantically Enhanced Siamese Network for Remote Sensing Change Detection". Here, we provide the pytorch implementation of this paper.

Prerequisites

• windows or Linux
• PyTorch-1.4.0
• Python 3.6
• CPU or NVIDIA GPU

Training

You can run a demo to start training.

python train.py

The network with the highest F1 score in the validation set will be saved in the folder tmp.

testing

You can run a demo to start testing.

python test.py

The F1_score, precision, recall, IoU and OA are displayed in order. Of course, you can slightly modify the code in the test.py file to save the confusion matrix.

Prepare Datasets

download the change detection dataset

SVCD is from the paper CHANGE DETECTION IN REMOTE SENSING IMAGES USING CONDITIONAL ADVERSARIAL NETWORKS, You could download the dataset at https://drive.google.com/file/d/1GX656JqqOyBi_Ef0w65kDGVto-nHrNs9;

LEVIR-CD is from the paper A Spatial-Temporal Attention-Based Method and a New Dataset for Remote Sensing Image Change Detection, You could download the dataset at https://justchenhao.github.io/LEVIR/;

Take SVCD as an example, the path list in the downloaded folder is as follows:

├SVCD:
├  ├─train
├  │  ├─A
├  │  ├─B
├  │  ├─OUT
├  ├─val
├  │  ├─A
├  │  ├─B
├  │  ├─OUT
├  ├─test
├  │  ├─A
├  │  ├─B
├  │  ├─OUT

where A contains images of pre-phase, B contains images of post-phase, and OUT contains label maps.

When using the LEVIR-CD dataset, simply change the folder name from SVCD to LEVIR. The location of the dataset can be set in dataset_dir in the file metadata.json.

cut bitemporal image pairs (LEVIR-CD)

The original image in LEVIR-CD has a size of 1024 * 1024, which will consume too much memory when training. In our paper, we cut the original image into patches of 256 * 256 size without overlapping.

When running our code, please make sure that the file path of the cut image matches ours.

Define hyperparameters

The hyperparameters and dataset paths can be set in the file metadata.json.

"augmentation":  Data Enhancements
"num_gpus":      Number of simultaneous GPUs
"num_workers":   Number of simultaneous processes

"image_chanels": Number of channels of the image (3 for RGB images)
"init_channels": Adjust the overall number of channels in the network, the default is 32
"epochs":        Number of rounds of training
"batch_size":    Number of pictures in the same batch
"learning_rate": Learning Rate
"loss_function": The loss function is specified in the file `./utils/helpers.py`
"bilinear":      Up-sampling method of decoder feature maps, `False` means deconvolution, `True` means bilinear up-sampling

"dataset_dir":   Dataset path, "../SVCD/" means that the dataset `SVCD` is in the same directory as the folder `SESNet`.