Assessing the Added Value of Onboard Earth Observation Processing with the IRIDE HEO Service Segment

1. Problem Statement

Current operational Earth Observation (EO) services, such as the Copernicus Emergency Management Service (CEMS) and the European Forest Fire Information System (EFFIS), rely on remote, ground-based processing pipelines. In this traditional architecture:

• Full-scene imagery must be downlinked from the satellite to the ground before any information extraction occurs.
• Latency and Bandwidth: Product timeliness is strictly coupled to downlink availability and ground-segment throughput. During crisis events (e.g., wildfires), this leads to latency bottlenecks and bandwidth saturation.
• Resolution Constraints: Medium-resolution baselines often limit the Minimum Mapping Unit (MMU), making it difficult to detect small, fragmented burn scars that are critical for national and local management.
• Lack of Autonomy: Existing systems lack the capability for autonomous prioritization of observations, often relying on periodic or crisis-triggered ground processing rather than real-time onboard decision-making.

While previous missions (e.g., ΦSat-1, ΦSat-2) have demonstrated the technical feasibility of onboard AI, these remain largely technology demonstrators. There is a gap in embedding onboard AI into institutional service chains with formally defined product specifications, accuracy targets, and operational end-users.

2. Methodology

The paper evaluates the added value of onboard processing using the IRIDE (International Report for an Innovative Defence of Earth) programme as a case study, specifically focusing on the HEO (Hawk for Earth Observation) microsatellite constellation and its application to burnt-area mapping.

• System Architecture Analysis: The authors analyze the IRIDE architecture, which integrates a "constellation of constellations." They focus on the HEO segment, which operates at 500–600 km altitude with sub-3m spatial resolution.
• Hardware Implementation: The study details the onboard computing stack, utilizing the Unibap iX5-106 SpaceCloud® platform equipped with an Intel® Myriad X Vision Processing Unit (VPU). This hardware supports AI inference directly in orbit.
• Comparative Service Analysis: The paper compares the IRIDE HEO service against existing European services (EFFIS, CEMS Emergency Mapping, CLMS) based on key figures of merit: acquisition mode, processing location, spatial resolution, MMU, and end-to-end latency.
• Workflow Modeling: The authors model a hybrid service chain where:
  1. Satellites acquire data.
  2. Onboard AI performs radiometric pre-processing and first-order inference (e.g., fire detection).
  3. Only thematic outputs or prioritized Regions of Interest (ROI) are downlinked, rather than full raw image strips.
  4. Ground segments handle further refinement and distribution via the IRIDE marketplace.
• Model Design Constraints: The methodology considers the constraints of low-power, radiation-hardened hardware, requiring model optimization techniques such as quantization (INT8/FP16), pruning, and Neural Architecture Search (NAS) to ensure real-time inference.

3. Key Contributions

• Operational Service Definition: The paper moves beyond algorithmic validation to define a national-scale operational service (Burnt-Area Mapping) with specific targets: 3m spatial resolution, a 3-hectare Minimum Mapping Unit (MMU), and >95% thematic accuracy.
• Hybrid Architecture Proposal: It proposes a hybrid onboard-ground processing architecture where onboard intelligence acts as a complementary layer to existing Copernicus services. It provides "image-driven pre-classification" to support downstream emergency workflows.
• Quantifiable Service Improvements: The study quantifies the shift from "Days" to "Minutes-Hours" for time-to-first-actionable-information by shifting processing from the ground to the edge (satellite).
• Hardware-Aware Design Framework: It outlines the necessary co-design of AI models and hardware (specifically the Myriad X VPU) to meet strict power, memory, and latency constraints in space.

4. Results

The comparative analysis (Table I in the paper) highlights the following advantages of the IRIDE HEO service over traditional ground-only systems:

• Latency Reduction: By performing inference onboard, the system reduces end-to-end latency from days to hours, enabling rapid situational awareness for civil protection.
• Bandwidth Efficiency: Downlinking only thematic outputs or specific ROIs significantly reduces data volume, alleviating bandwidth bottlenecks during large-scale fire events.
• Detection Capability: The 3m resolution and 3ha MMU allow for the detection of small, fragmented burn scars in wildland-urban interfaces, which are often missed by medium-resolution continental services.

5. Significance

• Paradigm Shift: The paper demonstrates a shift from "downlink-first" to "intelligence-first" architectures. It proves that onboard AI is not just a technical novelty but a critical enabler for low-latency, intelligence-supported EO services.
• Complementarity: It positions IRIDE HEO not as a replacement for Copernicus, but as a complementary layer that fills gaps in spatial detail and responsiveness for national-level emergency management.
• Scalability: The framework provides a blueprint for integrating onboard processing into future national and commercial EO constellations, addressing the growing pressure on ground infrastructures caused by increasing data volumes and revisit frequencies.
• Policy and Operational Impact: By formalizing service-level requirements (latency, accuracy, MMU) and translating them into constellation architectures, the work supports the development of robust, user-driven national EO systems (like Italy's PNRR initiative) capable of supporting real-time decision-making for authorities like ISPRA and Civil Protection.

In conclusion, the paper validates that integrating onboard processing into operational service chains significantly enhances the timeliness, resolution, and efficiency of Earth Observation, particularly for time-critical applications like wildfire monitoring.