Segmentation C-SAR

About project

This repository provides codes for training and testing algorithms for semantic segmentation of objects on C-band radar images, income from satellite Sentinel-1. The following architectures of convolutional neural networks are presented in the repository:

• U-Net
• DeeplabV3+
• MPResNet
• IncFCN

The following convolutional networks are used as basic networks for segmenters:

• ResNet-34/ResNet-50
• Xception

Used as a machine learning framework PyTorch. Agricultural fields are selected as segmentation objects in the images.

During training, the following variations of network training policies were used:

1. Loss functions

• Dice
• BCE
• Dice + BCE
• Focal

2. Optimizers

• Adam
• AdamW
• RMSProp
• SGD

Installation and running

Installation

NOTE: Installation instructions are provided for Linux.

Install the virtual environment in the project:

pip3 install -r requirements.txt

You can also run the project in Docker container. Pull docker image from repository:

docker pull nikkotov98/kotov_sar:dev

Create a docker container based on the downloaded image:

docker run -it --gpus all -v /local_data_root/:/docker_data_root/ /local_code_root/:/docker_code_root/ --name container_name nikkotov98/kotov_sar:dev

You can also use the official image from Nvidia:

docker pull nvcr.io/nvidia/pytorch:21.06-py3

Running

Start training the corresponding algorithm with the following command (optionally, the output of the terminal is written to the log file):

python3 train_xxx.py &> log_name.log

Start testing the corresponding algorithm on the saved scales with the following command:

python3 test.py

Segmentation of a large satellite image with the chosen architecture and the chosen stored weights is done by the script:

python3 segment_full_image.py

Results

Examples of the operation of the segmentation algorithm are presented (performed using the network DeeplabV3+ with base network ResNet-50).

Test results for networks by Dice and F1 metrics:

	Dice	F1
Deeplabv3+/ResNet34	0.665	0.598
Deeplabv3+/ResNet50	0.686	0.621
Deeplabv3+/Xception	0.668	0.605
IncFCN	0.722	0.669
MPResNet	0.666	0.603
U-Net/ResNet34	0.689	0.624
U-Net/ResNet50	0.5	0.603

The weights of the trained networks can be downloaded from the Yandex.Disk.

References

• Kotov Nikita ([email protected])