PAMELA: Probabilistic Amplification Module for Lightweight AI on LEO Satellites

Yufan Wu, Qifan Yang, Lang Wei, Ruichen Li, Ting Zhu

Keywords: Machine Learning, LEO Satellites, Onboard AI Inference

Abstract: Onboard AI inference in low Earth orbit (LEO) satellites enables rapid analysis of Earth observation data for time-critical applications such as wildfire monitoring. Yet, two key challenges remain: (i) critical fire signals are extremely sparse, often confined to only a few pixels, and (ii) onboard models must be highly compact due to the stringent hardware constraints of LEO platforms. As a result, small models may report deceptively high overall accuracy by overfitting to background regions, while suffering from poor precision and recall on fire pixels. To overcome these limitations, we introduce PAMELA, a lightweight and modular amplification framework tailored to enhance the performance of onboard wildfire detection models under resource-constrained conditions. PAMELA employs probabilistic modeling to selectively amplify informative channels and pixels, allowing compact models to better capture sparse but mission-critical signals within complex satellite imagery. Experimental results demonstrate that PAMELA consistently improves detection quality, delivering 1.2 ~ 2× higher F1 scores compared to compressed baselines while simultaneously reducing model size by over 90%. To the best of our knowledge, PAMELA is the first framework explicitly designed to enable reliable onboard wildfire detection in LEO satellites.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 12003