link: https://infoscience.epfl.ch/record/309368?v=pdf

github: https://github.com/marcCoru/meteor

*Note: There is a longer full version of this paper that I came across right before putting out this article. This version is rich with all kinds of details and should be referred to in the future: https://www.nature.com/articles/s43247-023-01146-0#Sec9*

Goal

To introduce METEOR, a MAML-based deep learning model, to help adapt quickly to new remote sensing tasks where one lacks annotated data. A new few-shot deep learning model.

Motivation

• Remote sensing using satellite imagery is a super useful and novel way to get insights into how our planet is behaving eg: climate change effects, human activities, etc.
• Deep learning is a super useful tool
• Applying deep learning on satellite imagery however isn’t trivial due to the lack of annotated data/images in this space. New tasks intend to solve important problems but often don’t have enough data to train a deep learning model.

Why is this paper interesting?

• Tackling important problem that I have come across multiple times working in the ecology space.
• Releasing a customised few-shot learning model that performs well on tasks with very little annotated data
• Allows anyone working on remote sensing tasks to use the model on data differing from the original data with variations like:
  • Geographical regions being considered
  • Type of sensors used eg: RGB data, SAR data, MS data, etc.
  • GSD of one’s data

The Actual Model

Uses a ResNet-12 deep learning model, which outputs a one-against-all classification score.

It is pre-trained using MAML on a large dataset consisting of prediction of land cover based on multi-spectral satellite images.

Basic standard MAML algorithm is used with the following modifications:

• Replace all batch normalization layers with instance normalization layers since batch normalisation doesn’t give good results with downstream high class imbalance problems.

  Guessing because the statistics of the dominant class will overpower the statistics of any minority class.