Trained a ResNet50 model on the EuroSAT satellite imagery dataset w/ PyTorch. Analyzed the model's encoder by visualizing linear interpolations within the embedding space to illustrate the semantic separation in the learned feature representations.
This project utilizes the EuroSAT dataset, which offers satellite images categorized into 10 distinct classes. The dataset is available for public access and can be found on the EuroSAT GitHub repository. I split the data into a training and testing set with an 80%-20% ratio.
The dataset is organized in the following directory hierarchy:
Train_Test_Splits
│
├── train
│ ├── AnnualCrop
│ ├── [Other Classes]
│
└── test
├── AnnualCrop
├── [Other Classes]
Make sure that the data path in the config.yaml is correct.
use pip to install all the requirements.
pip install -r requirements.txt
Showing nearest neighbors at each interpolation step from left (Industrial) to right (Forest). Inspired by plot in page 7 of Tile2Vec paper.