Preliminary On-board Image Processing Solution for the H2020 EO-ALERT Project

In this paper, we provide an overview of the H2020 EU project EO-ALERT and specify the preliminary on-board (OB) image generation and processing solution that is being developed within the project, that makes use of a combination of classical and AI solution concepts. The aim of the EO-ALERT project is to propose the definition and development of the next-generation Earth observation (EO) data and processing chain, based on a novel flight segment architecture moving optimised key EO data processing elements from the ground segment to on-board the satellite. In particular, we will focus in this paper on the on-board image generation and processing algorithms, showing the capability of the system of producing EO image products on-board and with very low latency (minutes), in two main scenarios defined for testing the potential of the proposed high-speed data chain: ship detection and extreme weather monitoring. The first step of the processing chain, identified as Image Generation (IG), is responsible for the conversion of the raw data captured by the optical sensor into a radiance calibrated and artefact-free image, which could optionally be sent to the ground as an intermediate EO product. The second step, identified as Image Processing (IP), is responsible for extracting the product information from the images, and producing a scenario related alert that will be transferred to the ground; since the size of the final alert will be considerably smaller than the acquired raw data, the EO product and alert will be available to the final user with very low latency, much lower than in the classical EO data chain. The paper will describe the functional blocks of the proposed IG/IP chain, as part of the overall avionics architecture. It is based on a combination of computer vision algorithms for efficient image generation and extraction of visual features from optical images, tailored to the specific scenario of interest and optimized to the lower computational capabilities of on-board hardware, in conjunction with machine learning algorithms for visual feature classification and discrimination of events of interest. In particular, a pre-trained Support Vector Machine (SVM), a popular classification tool that uses supervised machine learning theory to maximize the predictive accuracy and whose efficient implementation in field programmable gate arrays (FPGAs) has been proven by several studies, will be used as the final classifier. The paper will provide examples of the performance of the on-board IG/IP algorithms using acquired raw data from the DEIMOS-2 payload (sub-meter) and Meteosat Second Generation (MSG) payloads from these satellite missions, showing the performance of on-board solution against the ground-based classical solution. The examples are presented to show the detection capabilities of the system in the two proposed scenarios. The paper will also present an analysis of the avionics needs for the implementation of the proposed solution, showing its feasibility in an FPGA-based on-board avionics architecture. Copyright 2019 as per EO-ALERT Consortium Agreement - All Rights Reserve