# Tile2Vec: Unsupervised representation learning for remote sensing data

**Authors:** Neal Jean, Sherrie Wang, George Azzari, David Lobell, Stefano Ermon  
**arXiv:** [1805.02855](https://arxiv.org/abs/1805.02855)  
**Date:** 2018-05-08  
**Summary page:** https://www.summarizepaper.com/en/paper/1805.02855v1/

## Key points

- Tile2Vec is an unsupervised representation learning algorithm for remote sensing data
- It extends the distributional hypothesis from natural language to geospatial data
- Tile2Vec effectively learns semantically meaningful representations on three different datasets
- The learned representations improve performance in downstream classification tasks
- Similarities to word vectors allow for visual analogies through simple arithmetic in the latent space
- The research paper titled "Tile2Vec: Unsupervised representation learning for remote sensing data" is authored by Neal Jean, Sherrie Wang, George Azzari, David Lobell, and Stefano Ermon
- Tile2Vec enhances the performance of remote sensing applications by leveraging unsupervised representation learning techniques
- Empirical evidence supports the effectiveness of Tile2Vec in learning meaningful representations from geospatial data
- Tile2Vec fills the existing gap in representation learning methods for remote sensing applications.

## Summary

Tile2Vec is an unsupervised representation learning algorithm designed to address the lack of methods like word vector representations and pre-trained networks in remote sensing. To bridge this gap, Tile2Vec extends the distributional hypothesis from natural language to geospatial data, where words that appear in similar contexts tend to have similar meanings. In their study, the authors demonstrate that Tile2Vec effectively learns semantically meaningful representations on three different datasets. These learned representations not only improve performance in downstream classification tasks but also exhibit similarities to word vectors, allowing for visual analogies through simple arithmetic in the latent space. The research paper titled "Tile2Vec: Unsupervised representation learning for remote sensing data" is authored by Neal Jean, Sherrie Wang, George Azzari, David Lobell, and Stefano Ermon. The paper highlights the significance of Tile2Vec as a means to enhance the performance of remote sensing applications by leveraging unsupervised representation learning techniques. The authors provide empirical evidence supporting the effectiveness of Tile2Vec and its ability to learn meaningful representations from geospatial data. Overall, Tile2Vec offers a valuable contribution to the field of remote sensing by introducing a novel approach that fills the existing gap in representation learning methods. Its potential impact lies in improving classification tasks and enabling visual analogies within remote sensing applications.