Nowadays, the use of Machine Learning (ML) onboard Earth Observation (EO) satellites has been investigated for a plethora of applications relying on multispectral and hyperspectral imaging. Traditionally, these studies have heavily relied on high-end data products, subjected to extensive pre-processing chains natively designed to be executed on the ground. However, replicating such algorithms onboard EO satellites poses significant challenges due to their computational intensity and need for additional metadata, which are typically unavailable on board. Because of that, current missions exploring onboard ML models implement simplified but still complex processing chains that imitate their on-ground counterparts. Despite these advancements, the potential of ML models to process raw satellite data directly remains largely unexplored. To fill this gap, this paper investigates the feasibility of applying ML models directly to Sentinel-2 raw data to perform thermal hotspot classification. This approach significantly limits the processing steps to simple and lightweight algorithms to achieve real-time processing of data with low power consumption. To this aim, we present an end-to-end (E2E) pipeline to create a binary classification map of Sentinel-2 raw granules, where each point suggests the absence/presence of a thermal anomaly in a square area of 2.5 km. To this aim, lightweight coarse spatial registration is applied to register three different bands, and an EfficientNet-lite0 model is used to perform the classification of the various bands. The trained models achieve an average Matthew’s correlation coefficient (MCC) score of 0.854 (on 5 seeds) and a maximum MCC of 0.90 on a geographically tripartite dataset of cropped images from the THRawS dataset. The proposed E2E pipeline is capable of processing a Sentinel-2 granule in 1.8 s and within 6.4 W peak power on a combination of Raspberry PI 4 and CogniSat-XE2 board, demonstrating real-time performance.

Automated Band-to-Band alignment is a crucial prerequisite for satellite mission involving image operations. Many techniques are now developed for image registration: correlation-based methods, feature-based methods, hybrid methods and Deep Learning-based methods. Due to perturbations in the spacecraft orientation and dynamic disturbances, the a priori knowledge concerning the misalignment of bands is often insufficient, therefore an appropriate sub-pixel level registration scheme must be implemented on-board the satellite. In this work three registration approaches are tested and compared using raw image data from the European Space Agency Sentinel-2 mission: the Coarse Coregistration, SuperGlue Coregistration and LightGlue Coregistration. The Coarse Coregistration is a simple approach to image registration based on the deterministic shift of the sensed image, but its performance is affected by non-systematic disturbances that are not corrected, often limiting it is accuracy to a pixel level. On the other hand, SuperGlue is a powerful but computationally expensive deep network designed for computer vision tasks. Despite its excellent accuracy, the inference time does not suit the on-board requirements of the mission. For this reason, a lighter version of SuperGlue based on an adaptive network, named LightGlue, has been tested to study its performance and inference times. The maximum number of keypoints, the adaptive depth and the number of layers of the network have been modified in order to test the architecture's response. With this analysis, it has been found that LightGlue brings the precision on the correcting shift up to a sub-pixel level, up to an order of magnitude better than the Coarse Coregistration, while running up to 1.5 times faster than SuperGlue. Sentinel-2 raw data is used here for the first time.