This repository contains code implementations of contrastive learning architecture, called Multimodal SuperCon, for classifying drivers of deforestation in Indonesia using satellite images obtained from Landsat 8. Multimodal SuperCon is an architecture which combines contrastive learning and multimodal fusion to handle the available deforestation dataset.
This project is using several papers as the main references:
This project implements two-stage learning, representation and classification stage for training the models. Training process takes 2 step:
- Representation Stage using Supervised Contrastive Learning.
- Classification Stage using Supervised Learning with Multimodal Fusion.
Main libraries and dependencies:
PyTorch
Shapely
: python package for set-theoretic analysis and manipulation of planar features, beneficial for spatial data analysisAlbumentation
: python package for image augmentations
- To run the program, simply by re-running the available notebook:
Training - Effnet + Resnet.ipynb
andTraining - UNet.ipynb
- If you want to add any available auxiliaries/predictors from ForestNet dataset, you can modify the backbone model where the code implementation can be found under
model
folder (will update the other examples, especially with four auxiliaries/predictors, soon)
@article{10.1117/1.JRS.17.036502,
author = {Bella Septina Ika Hartanti and Valentino Vito and Aniati Murni Arymurthy and Adila Alfa Krisnadhi and Andie Setiyoko},
title = {{Multimodal SuperCon: classifier for drivers of deforestation in Indonesia}},
volume = {17},
journal = {Journal of Applied Remote Sensing},
number = {3},
publisher = {SPIE},
pages = {036502},
keywords = {deforestation driver classification, contrastive learning, class imbalance, multimodal fusion, Machine learning, Education and training, Data modeling, Image fusion, Performance modeling, Atmospheric modeling, Data fusion, Deep learning, Landsat, RGB color model},
year = {2023},
doi = {10.1117/1.JRS.17.036502},
URL = {https://doi.org/10.1117/1.JRS.17.036502}
}