TinyML for On-Board Earth Observation: From Ultra-Low-Power MCUs to In-Sensor Computing

Luigi Capogrosso, Pietro Bonazzi, Loris Hoxhaj, Michele Magno

Keywords: Earth Observation, TinyML, CubeSat Satellites

Abstract: The rapid democratization of access to Low Earth Orbit (LEO) has driven an exponential increase in Earth Observation (EO) missions powered by miniaturized CubeSats. However, the traditional paradigm of continuously transmitting raw or minimally processed images to ground stations is severely constrained by the stringent power, memory, and communication bandwidth limitations of these tiny systems. To overcome these limitations, we present our recent advances in shifting computational intelligence directly to the extreme edge of space operations for autonomous image classification. The first is a hardware-aware TinyML pipeline for heterogeneous Microcontroller Units (MCUs) featuring an integrated Neural Processing Unit (NPU), which demonstrates massive model compression without compromising accuracy while achieving millijoule-scale energy consumption. The second is an advanced in-sensor computing strategy that executes Deep Neural Networks (DNNs) directly on the image sensor die, completely bypassing the system bus and offloading the primary On-Board Computer (OBC). This article demonstrates how both paradigms can effectively reduce downlink bandwidth requirements and shift task execution to the edge while maintaining acceptable accuracy, paving the way for sustainable, real-time autonomous space robotics.

Submission Number: 24