Title: Leveraging In-Sensor Computing for Energy-Conscious Earth Observation

Abstract: The burgeoning satellite sector has led to a surge in geospatial data collection, exposing a major challenge: the wide gap between massive data generation and the restricted downlink bandwidth of ground stations. Although On-Board Computing (OBC) has alleviated some pressure through orbital pre-processing, this study pushes the boundaries further by proposing an in-sensor computing architecture. We introduce a refined end-to-end Earth Observation (EO) workflow designed for stringent computational limits, combining TinyML methodologies with the Sony IMX500 Intelligent Vision Sensor. By moving processing tasks directly to the sensor level, we reduce the load on the main embedded hardware and significantly cut down the transmission of superfluous or noisy data. Our evaluation utilized efficient Convolutional Neural Networks (ConvNets)—specifically SqueezeNet, ShuffleNetV2, and MCUNetV1—on the EuroSAT dataset. Despite the necessary optimizations for the IMX500’s 8 MB memory limit, our models achieved a robust 96.68% accuracy. Performance metrics indicate an average throughput of 17.40 FPS and a latency of 27.43 ms. Additionally, the system demonstrates superior energy efficiency, consuming only 14.19 mJ per inference and achieving an efficiency rating of 42.26 GMAC/J, confirming its suitability for in-sensor applications.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC