# HyperspectralViTs: Fast and Accurate methane detection on-board satellites

**Authors:** Vít Růžička, Andrew Markham  
**arXiv:** [2410.17248](https://arxiv.org/abs/2410.17248)  
**Date:** 2024-10-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.17248v1/

## Key points

- Integration of machine learning models in hyperspectral data processing can revolutionize tasks like methane detection and mineral identification
- Automated machine learning models on satellites can efficiently detect methane, providing early warning system and enabling automated scheduling within satellite constellations
- New approach involves fast and accurate machine learning architectures for end-to-end training with high spectral dimension data
- Proposed architectures show significant performance improvements in methane leak detection and mineral identification tasks
- F1 score for methane detection increased by over 27% on synthetic dataset and nearly 13% on large benchmark dataset
- Training models on synthetic data enhances performance when finetuned on real event datasets
- Newly created dataset for mineral identification showcases a 3.5% improvement in F1 score compared to default model versions
- Inference speed of proposed models enhanced by 85.19% compared to classical and deep learning approaches by eliminating reliance on classically computed features

## Summary

, , , , 

In the field of hyperspectral data processing, the integration of machine learning models for on-board processing has the potential to revolutionize various tasks, such as methane detection and mineral identification. Methane, as the second most significant contributor to greenhouse gas emissions and climate change, can be efficiently detected using automated machine learning models on satellites. This capability not only enables an early warning system but also opens up possibilities for automated scheduling within satellite constellations. Traditional methods for methane detection often suffer from high false positive rates, while previous deep learning models have been limited by their computational requirements. To address these challenges, a new approach is proposed that involves fast and accurate machine learning architectures capable of end-to-end training with high spectral dimension data. These models have been evaluated on two key tasks - methane leak detection and mineral identification. The results of the evaluation demonstrate significant improvements in performance compared to existing state-of-the-art models. Specifically, the proposed architectures have increased the F1 score for methane detection by over 27% on a synthetic dataset and nearly 13% on a large benchmark dataset. Additionally, training models on synthetic data has shown to enhance performance when finetuned on real event datasets. Furthermore, a newly created dataset for mineral identification has been utilized to showcase a 3.5% improvement in F1 score compared to default model versions. Notably, the inference speed of the proposed models has been enhanced by 85.19% compared to classical and deep learning approaches by eliminating the reliance on classically computed features. It is worth noting that there is a lack of comprehensive datasets for global mineral identification across diverse regions worldwide. Existing works typically evaluate their methods on limited airborne hyperspectral captures or compare against sparsely sampled scenes like Cuprite or Alunite Hill scenes. Overall, this research presents promising advancements in hyperspectral data processing through innovative machine learning architectures that offer improved accuracy, speed, and autonomy for critical applications like methane detection and mineral identification onboard satellites.