OpenSARShip Classification

Based on https://doi.org/10.1016/j.asr.2021.08.042

Methods

1. Parsed OpenSARShip dataset and retrieve all PATCH_CAL images.
2. Used CFAR to create bounding boxes and calculate 14 scale-variant features used in training.
3. Resized all images to 64 x 64 x 2 (Stacked VH and VV polarisations).
4. Split dataset to 70 - 20 - 10 (Train - Val - Test) using stratified split to maintain same distribution of classes.
5. Augment dataset using oversampling for fishing classes and undersampling for cargo classes.
6. Standardise images.

How to Run

Install

pip install -r requirements.txt

Preprocess Data

python data_preparation.py

Train

Set learning parameters on the config object in main.py then run below command.

python main.py