This work presents an algorithm to perform autonomous navigation in spacecraft using onboard magnetometer data during GPS outages. An Extended Kalman Filter (EKF) exploiting magnetic field measurements is combined with a Single-Hidden-Layer Feedforward Neural Network (SLFN) trained via the Extreme Learning Machine to improve the accuracy of the state estimate. The SLFN is trained using GPS data when available and predicts the state correction to be applied to the EKF estimates. The CHAOS-7 magnetic field model is used to generate the magnetometer measurements, while a 13th-order IGRF model is exploited by the EKF. Tests on simulated data showed that the algorithm improved the state estimate provided by the EKF by a factor of 2.4 for a total of 51 days when trained on 5 days of GPS data.

We present a proof-of-concept onboard cloud detection system for the PRISMA Second Generation (PSG) mission, which combines a secondary forward-looking RGB camera with deep learning (DL) models on a system-on-a-chip (SoC) field-programmable gate array (FPGA). The proposed system enables real-time cloud coverage assessment, optimizing primary hyperspectral payload data collection and supporting adaptive acquisition scheduling, offering an effective solution for enhancing onboard data processing in Earth Observation (EO) missions. To support efficient model development, we derived the design specifications of the secondary camera and constructed a dedicated dataset using astronaut-captured images from the International Space Station – the CloudISS-RGB dataset – employing a self-training approach to generate high-quality pseudo-labels. Through an extensive architectural design exploration and systematic optimization, we tested two fully convolutional networks: U-Net, offering higher segmentation accuracy, and a lightweight convolutional autoencoder (CAE) designed for lower latency inference. The models were deployed on an AMD/Xilinx Zynq UltraScale+ MPSoC using the deep learning processing unit (DPU) IP core for hardware acceleration. The FPGA-deployed U-Net achieved 98.16% with a false positive rate of 0.9%, providing robust segmentation even in challenging conditions, making it the preferred model for reliable inference onboard PSG. The CAE model maintained 97.02% accuracy while achieving over 2× faster inference (31.81 ms vs. 74.29 ms per image). Both models enable accurate cloud segmentation with real-time inference, meeting the operational constraints derived from the secondary camera design, while operating at an average power consumption of 2.6 W (U-Net) and 2.4 W (CAE), well within the mission power constraint for the HW accelerator. Our results validate the feasibility of integrating DL-based cloud coverage assessment onboard PSG, contributing to the broader effort of advancing artificial intelligence-powered computing for EO missions to enable more autonomous data processing and decision-making.