Title: Implementing Rapid Onboard Methane Detection via Mag1c-SAS and LinkNet on Satellites

Abstract:

Hyperspectral satellite imagery offers a powerful tool for early methane leak detection, a critical component in mitigating climate change. However, traditional hyperspectral missions often rely on manual targeting by operators, which can cause them to overlook significant events. While onboard detection presents a cost-effective alternative to slow and expensive data downlinks, conventional methane detection algorithms are typically too computationally intensive for resource-constrained satellite hardware. This study addresses these challenges by prioritizing efficient, low-power algorithms to accelerate detection capabilities.

We introduce Mag1c-SAS, a substantially faster iteration of the current state-of-the-art Mag1c algorithm, and evaluate its performance alongside fast target detection methods—ACE and CEM—that have not previously been applied to methane detection. To further enhance detection potential, these methods are integrated with machine learning models based on U-Net and LinkNet architectures.

Our evaluation utilizes the STARCOP dataset and EMIT-MSeg, a new dataset we have introduced and open-sourced, complete with a robust annotation strategy. The results demonstrate that Mag1c-SAS operates approximately 80 times faster than the original Mag1c algorithm, yielding visually comparable results, albeit with increased noise. When combined with the lightweight LinkNet model, this noise is effectively mitigated, leading to an AUPRC score increase of over 30 percentage points on the EMIT-MSeg dataset compared to the baseline Mag1c approach, and an F1 score improvement of roughly 4 percentage points on STARCOP.

Furthermore, we assessed two novel band selection strategies and validated the system’s feasibility for onboard deployment through hardware profiling. The analysis confirms marginal power consumption and efficient utilization of CPU and RAM resources. To support further research and application, we have released the complete system as a lightweight, user-friendly PyPI library at https://pypi.org/project/onboard-methane-detection/, along with all associated experimental code, models, and data at https://github.com/zaitra/methane-filters-benchmark.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC