Few-Shot-Satellite-Image-Classification

Repository for "Few-Shot Learning for Image Classification on Spacecraft" benchmarking scripts

Installation

Python version: 3.12

The following packages are required:

• PyTorch
• numpy
• tqdm
• configargparse
• scikit-learn
• pandas
• matplotlib
• seaborn

Get Datasets

Download datasets into data folder.

Should be in this directory format:

data
│   ├── dataset name
│   │   ├── class 1
│   │   ├── class 2
│   │   ├── class 3
│   │   ├── .json files

Create .json files for classes you want included in dataset. Files for the following datasets are already included:

• UC Merced Land Use (http://weegee.vision.ucmerced.edu/datasets/landuse.html)
• NWPU-RESISC45 (https://www.tensorflow.org/datasets/catalog/resisc45)
• PatternNet (https://sites.google.com/view/zhouwx/dataset)

Adding More Datasets

To add a new dataset

• Download the dataset into the data folder as shown above
• Create .json file(s) for classes you want to use
• Create .config file(s) in the configs folder to specify dataset, path to .json file, and file-types for images

How to Use

Feel free to edit any scripts to meet your needs

Pretrain Backbones

Run from root: ./pretrain.sh

Evaluate Models

To sweep over different shots, run from root: ./run_sweep.sh

To sweep over different shots on CNN baseline, run from root: ./run_baseline_sweep.sh

To evaluate class accuracies, run from root: ./run_confusion.sh

Parameters

The full list of command-line arguments can be found in src/utils/parser_configs.py

--n-tasks: number of few-shot testing episodes
--n-way: number of classes
--n-shot: number of samples per class
--n-query: number of queries per few-shot task


--train-split: path to training dataset .json file
--test-split: path to testing dataset .json file
--algorithm: which classification head to use (PrototypicalNetworks, LaplacianShot, SVM)
--pretrained_dataset: name of dataset backbone was pretrained on