Authors: Gabriel Tseng, Ruben Cartuyvels, Ivan Zvonkov, Mirali Purohit, David Rolnick, Hannah Kerner

Published on: April 27, 2023

Impact Score: 8.38

Arxiv code: Arxiv:2304.14065

Summary

• What is new: Presto introduces a novel approach by focusing on the unique attributes of satellite data, particularly the temporal aspect and the diverse sensor data, which hasn’t been fully exploited in previous self-supervised models.
• Why this is important: Existing machine learning models for satellite data struggle due to the scarcity or absence of labeled data and fail to leverage the specific characteristics of satellite data effectively.
• What the research proposes: Presto is designed to harness the temporal and multisensor characteristics of satellite data through a pre-trained model, allowing for better performance with less compute power.
• Results: Presto demonstrates superior performance on various global remote sensing tasks, outperforming larger models with significantly lower computational requirements.

Technical Details

Technological frameworks used: Self-supervised learning

Models used: Presto (Pretrained Remote Sensing Transformer)

Data used: Remote sensing pixel-timeseries data

Potential Impact

Agriculture, environmental monitoring, urban planning, and defense sectors, as well as companies specializing in satellite imagery and geospatial analytics.

Want to implement this idea in a business?

We have generated a startup concept here: AgroPresto.