Project Description

Abstract

HYPER-EDGE is a six-month Pre-Phase A mission concept study carried out under the ESA SysNova campaign on disruptive computing paradigms (CfP/5-50166/25/NL/MGu/mmo). Funded by the Preparation Element of the Basic Activities.

The study investigates whether Hyperdimensional Computing (HDC), a brain-inspired paradigm that represents information as very long vectors and replaces the multiply-accumulate workload of deep neural networks with cheap bitwise operations, can deliver reliable on-board wildfire detection and fire-type classification for air-quality impact assessment on a small-satellite constellation, at a fraction of the power budget of conventional on-board AI.

The reference mission is an In-Orbit Demonstration (IOD) constellation of four small satellites (about 90 kg each) in a 500 km Sun-synchronous noon–midnight orbit. Each spacecraft carries a co-aligned dual-head optical payload (a hyperspectral VNIR camera for scene context and a SWIR camera for the active-fire signal), both high-TRL COTS units, and an experimental HDC processing chain that turns raw imagery into a compact, decision-ready alert on board within minutes. A two-channel downlink separates the latency-critical alert, sent via an L-band inter-satellite link through a commercial GEO relay, from the volume-critical evidence products, downlinked in S-band to commercial ground-station networks. Study targets are an end-to-end latency of 10 minutes or less (P95), a probability of detection of at least 90% for medium and large fires, at least 85% fire-type classification accuracy, a tenfold reduction of downlinked data volume, and 98% availability, within the 50 M€ reference cost envelope of the Call (launch excluded).

The consortium combines three Italian SMEs: Betadynamiq (prime; project management, mission and system architecture, mission costing), Grid+ (HDC on-board processing concept, data fusion and benchmarking), and IntelligEarth (user needs, service requirements, technology roadmap). The study delivers three Technical Notes (mission concept and preliminary cost and schedule; critical technologies and technology roadmap; benchmarking and trade-off analysis), a Final Report, an Executive Summary, and a Final Presentation.

Status at September 2026 (T0+5 months): user needs, operational scenarios, 19 service requirements and 9 KPIs are consolidated; the mission analysis with orbital propagator is complete (debris-mitigation compliance, coverage, eclipse, ground-station access, link budgets, system budgets closing with margin and a 9-minute end-to-end latency budget against the 10-minute target); the HDC encoding pipeline and its performance evaluation framework are defined. Benchmarking against a conventional deep-learning baseline, the technology roadmap and the consolidated cost assessment are in progress towards the Final Review in October 2026.

Detailed description

1. Context and motivation

Wildfires and industrial fires are among the fastest-evolving hazards that civil protection and environmental authorities have to manage. Today, satellite-based fire information reaches operators hours after acquisition because imagery is downlinked, processed on the ground, and only then turned into an alert. In the decisive first minutes of an event, the users who must decide whether to evacuate, where to send suppression assets, and which air-quality plume to monitor are working with fragmented and delayed information.

HYPER-EDGE addresses this bottleneck by moving detection and classification on board the spacecraft. The ESA SysNova challenge on disruptive computing asks for complete mission concepts, not subsystems, in which a novel computing paradigm enables a new application or significantly improves mission trade-offs, with quantified benefits on application-level and system-level KPIs, a reference cost of at most 50 M€ (launch excluded), and a pre-Phase A maturity at study closure. HYPER-EDGE responds with a small-satellite IOD constellation whose on-board intelligence is built on Hyperdimensional Computing.

Hyperdimensional Computing (also known as Vector Symbolic Architectures) encodes features, values, and positions as very high-dimensional vectors and composes them with binding and bundling operations. Classification becomes a similarity search against class prototypes. The properties that make it attractive on board are direct consequences of the spacecraft constraints: energy-efficient bitwise arithmetic instead of dense floating-point multiply-accumulate, intrinsic robustness to noise and bit errors, compact models that can be updated with small uplinks, and a transparent decision rule that can be inspected and verified.

Funded by the Preparation Element of the Basic Activities. The study is implemented as an ESA Cooperative Agreement within the SysNova campaign, with technical follow-up by the ESA Technical Officer and dissemination support from ESA Φ-lab.

2. Mission objectives and success criteria

The mission concept pursues four objectives, each traced to the KPI regime of the Call:

Success is measured against five quantitative Mission Success Criteria (study targets):

3. Reference mission concept

The reference architecture was consolidated at the Kick-off and refined during the mission analysis conducted between May and August 2026, using an orbital propagator (HPOP with the NRLMSISE-00 atmosphere) and coverage figures of merit. The table summarises the current baseline.

3.1 Orbit and constellation

The 500 km altitude is the highest orbit that meets the 25-year re-entry requirement for this mass class and drag area, and it is also the most favorable condition for the optical payload. The noon–midnight Sun-synchronous orbit maximizes illumination during imaging. Eclipse and beta-angle analyses over a full year (mean eclipse of about 35 minutes) feed the sizing of the power subsystem. Coverage analyses over the ESA member-state area show the revisit performance of the IOD nucleus and how it scales when additional orbital planes are added.

3.2 Payload

The dual-head configuration pairs two complementary spectral domains: the SWIR head carries the signature of the thermal anomaly and sees through smoke and haze where the visible is blind, while the hyperspectral VNIR head provides the scene context (smoke, burned area, vegetation state) needed to discriminate the type of event. The two ground sampling distances are almost identical, so the two data streams can be fused directly on the common co-registered swath. Both heads are high-TRL COTS units, each weighing around 1.6 kg, with a combined imaging power consumption below 20 W.

3.3 Two-channel communications architecture

The architectural choice the team focused on most is separating the latency-critical path from the volume-critical path. The alert packet produced on board is transmitted via an L-band inter-satellite link to a commercial GEO relay service. It reaches the ground within seconds, regardless of a ground station's visibility. RF access analysis shows contact is available about 69% of the time over a year, with a 50th-percentile contact duration of over 27 minutes; the gaps concentrate over the polar regions, where fire events are rare. Evidence products and payload data are downlinked in S-band to commercial ground-station networks, for which access analysis across the constellation provides ample contact time and allows the mission to select a compact set of stations.

3.4 System budgets

The preliminary system budgets close with margin: about 86 kg per satellite against the 80–110 kg target, 191 W peak power against a 320 W end-of-life solar array, and an end-to-end latency budget of 9 minutes against the 10-minute target of MSC-1.

4. Users, use cases, and service requirements

The user side of the study, led by IntelligEarth, established in a traceable manner who the service must serve and what they need. A central finding drives the whole concept: in the Italian regulatory framework, a vegetation fire and an industrial or contaminated fire activate two legally distinct response chains, with different lead authorities, timelines, and objectives (wildfire law and regional forest-fire services on one side, Seveso III and the fire brigade with environmental agencies on the other). This is the institutional reason why the service must classify the type of event at the moment of detection, not after human interpretation on the ground.

• Stakeholder taxonomy and need register covering national and regional civil protection, fire brigades, forestry services, environmental agencies, and the EU emergency coordination layer.
• Three operational use cases (reference vegetation fire, industrial and contaminated fire, wildland–industrial interface) anchored to documented Italian events: Aspromonte 2021, Bellolampo 2023, and Vicopisano 2026.
• Nineteen service requirements, each with an acceptance criterion and a verification method, and nine KPIs traced to the Mission Success Criteria.
• A dataset strategy built on Sentinel-2 and Landsat optical imagery, atmospheric composition and meteorological reanalysis, static context layers and independent reference sources, with acquisition-by-acquisition labeling, bias-controlled splits and a literature benchmark as reference baseline.

5. Hyperdimensional on-board processing

Grid+ defined the on-board processing chain from raw data to a trained classifier. The hyperdimensional representation was selected after comparing three alternative constructions from the literature, motivated by the platform's real constraints. Feature identities are represented by an item memory, feature values by a continuous item memory with graded level vectors, and the pixel position by an axis dictionary; binding and bundling compose them into a single pixel hypervector. Class prototypes are built during training and queried by similarity at inference, with two candidate decision structures under evaluation: a flat three-way classification and a two-stage cascade.

The performance evaluation framework is built around the asymmetry of operational costs: missing an industrial fire, the error with the most severe consequences for population safety, weighs more than any other error, and the metrics are chosen accordingly. A three-level fallback strategy to a conventional classifier, always activated by ground command, is documented as a risk-reduction measure for a technology at this stage of maturity. Benchmarking against a deep-learning baseline on the reference dataset is the core of the study's final phase.

6. Work plan and consortium

Betadynamiq brings mission design and system architecture expertise and the BetaEdge edge-computing heritage; Grid+ brings the Hyperdimensional Computing technology, on-board processing, and FPGA acceleration; IntelligEarth brings Earth Observation application expertise, user requirements engineering, and technology assessment heritage from ESA InCubed and national programs.

7. Collaboration with ESA and expected outcomes

The study is conducted under an ESA Cooperative Agreement within the SysNova campaign, with monthly progress reports to the ESA Technical Officer and formal reviews. At the Final Review in October 2026, the consortium delivers the three Technical Notes, the Final Report, the Executive Summary, the Final Presentation, and public highlight material. The SysNova board evaluates the Challenge Analyses on scientific merit, technical feasibility, and quality of the analyses and roadmaps; the best analysis is rewarded with an ESA Concurrent Design Facility assessment.

Beyond the study, HYPER-EDGE aims at a Phase A and an IOD flight of the four-satellite constellation, at a technology roadmap that brings the HDC on-board chain to flight readiness within the acceptance horizon of the Call, and at an operational service that gives civil protection and environmental authorities a decision-ready fire and air-quality alert within minutes of acquisition.

Project Outcomes

• Technical Note 1 – Mission Concept and Preliminary Cost and Schedule — Final Review, October 2026 (draft in preparation)
• Technical Note 2 – Critical Technologies and Technology Roadmap — Final Review, October 2026
• Technical Note 3 – Benchmarking and Trade-off Analysis — Final Review, October 2026
• Final Report, Executive Summary Report, Final Presentation, and public highlight images — Final Review, October 2026; to be published on the ESA activities.esa.int project page [link TBC]
• User Needs, Use Cases and Service Requirements package (WP2010, IntelligEarth) — Completed August 2026, consortium-internal; consolidated into TN1
• Mission Analysis Note for the IOD constellation, rev. 2 (WP2020, Betadynamiq) — Completed August 2026, consortium-internal; consolidated into TN1
• On-board HDC Encoding and Evaluation Note (WP3010, Grid+) — Completed August 2026, consortium-internal; consolidated into TN2/TN3
• Reference dataset for fire-type classification (multi-source labeled acquisitions over three Italian events) — In construction; public release to be decided at Final Review

The collaboration with ESA under the SysNova campaign gave the consortium a structured, review-driven framework to turn a research idea into a mission-level concept: a user-anchored requirement set, a propagator-verified mission architecture, an on-board HDC processing chain with a fallback strategy, and a benchmarking plan against the state of the art. The two-channel alert and evidence architecture and the fire-type classification at the moment of detection are the elements the team considers most valuable for future ESA and national programs.