23.02.24 - METEOR: Adaptable Meta-Learning for Earth Observation Problems

Who has never struggled with different dataset resolutions, class imbalances or few labelled images?

Let's have a look at the recent paper Meta-learning to address diverse Earth observation problems across resolutions by Russwurm, M., Wang, S., Kellenberger, B. et al. (https://lnkd.in/epVzQ2yC). This paper presents Hashtag#METEOR, an intelligent approach to help computers better understand the Earth from satellite imagery.

METEOR is a meta-learning methodology tailored for Earth observation challenges, utilizing optimization-based meta-learning with a small deep learning model. It is pre-trained using the model-agnostic meta-learning (MAML) algorithm.

It introduces key features such as replacing batch normalization with instance normalization, adapting dynamically convolutional kernels for different spectral bands, and handling a varying number of classes through a binary meta-model, fine-tuned for each class and by ensembling with the one-vs-all approach.

In practice, METEOR demonstrates its adaptability by allowing domain experts to fine-tune a single deep neural network for different downstream problems with minimal labelled examples. These downstream tasks can vary significantly from the land cover source tasks, covering diverse domains such as land cover classification, urban scene classification, and marine debris segmentation.

The effectiveness of METEOR lies in its ability to handle variations in spatial and spectral resolution, class imbalance, and limited annotated data. The applicability of the methodology is showcased across a spectrum of Earth science use cases, making it a valuable tool for addressing real-world challenges encountered by experts in the field.

Image and content source: https://lnkd.in/epVzQ2yC