1 of 10

SPATIAL WEB: �MULTI-AGENT FUSION FOR �EARTH OBSERVATION

George Percivall

Percivall@ieee.org

Stephane Fellah

sfellah@geo-knoesis.com

GeoKnoesis, LLC

GeoRoundtable

2 of 10

ROADMAP

Agenda

SWF - Percivall & Fellah · Spatial Web Architecture for Multi-Agent Fusion

The Fusion Challenge

Observation & Feature fusion well established; Decision fusion insufficiently addressed

Spatial Web Architecture

IEEE 2874 Spatial Web: agent-centric, semantically grounded, trust-enabled

Shift from pipelines to knowledge-driven agent-centric fusion

Spatial Web Components: HSML, HSTP, UDG

Formal semantic representation, trusted agent interaction, collective intelligence

Applications & Conclusion

Earth observation systems with collective intelligence and verifiable trust

3 of 10

MOTIVATION

Benefits and Challenges of Information Fusion

Information and Data Fusion in Earth Observation

Data fusion central to EO, ARD, and remote sensing systems

Lunga, et.al., IGARSS, 2023

Information fusion expands to semantically distinct and non-numeric inputs

International Society of Information Fusion

Integration of heterogeneous data into actionable knowledge

e.g., Zaichyk, Semantic Fusion, 2026

Types of Information Fusion

Observation (sensor) fusion: well established theory and methods: GRSS
Feature fusion: benefits from cross-community semantics
Decision fusion: accelerating with AI

The Gap in Decision Fusion for EO

Decision fusion in EO multi-agent environments: insufficiently addressed
Cannot support interoperable reasoning, coordination, adaptation, and trust across heterogeneous agents, platforms, and organizational boundaries

SWF - Percivall & Fellah · Spatial Web Architecture for Multi-Agent Fusion

Schmitt & Zhu IEEE G&RSM (2016).

Fusion has been a topic within IEEE and GRSS for a long time. It has a broader community as well.
Within GRSS, there's been a focus on data fusion.
Several of us over the years have expanded this from data fusion to information fusion, which takes us to the more expanded set of types of fusion, of observation fusion, and geocentric fusion, and by another name. Feature fusion, so the ability to pull together data and look at features that are semantically described and to be able to fuse features.
And then lastly, decision fusion, bring together sensing and observation and features into a decision framework that allows you to support or take action autonomously. And that's accelerating, right? Decision fusion with agents in particular, AI agents in particular.
So it's very timely that we talk about this set of information fusion in these three ways based upon the history that we have in GRSS and other places about information fusion.
The gap is really in decision fusion. This is where the gap and the opportunity, because there is a technology of AI. That allows us to do things we couldn't do before, in particular about reasoning, coordination, and adaptation. And the framework to allow agents to coordinate it is very needed in order to achieve this decision fusion for EO.

4 of 10

ARCHITECTURE

Spatial Web: Fusion based on Space and Time

Foundation: IEEE Spatial Web Standard

IEEE 2874-2025: Spatial Web Protocol, Architecture, and Governance Standard
Developed by the Spatial Web Foundation
Defines semantic, spatial, and governance framework for interoperable entities

Interoperable Spatial Web Entities

Sensors, IoT devices, EO/GIS services, digital twins, software agents, humans

Extension of established fusion concepts

Extends IEEE GRSS and OGC information fusion frameworks

Schmitt & Zhu, IEEE G&RSM, 2016; Percivall et.al., OGC 10-184;

Extends with AI agent-centric, identity/trust, semantically-grounded spatial model
Fusion of observation, feature, and decision across distributed intelligent agents

SWF - Percivall & Fellah · Spatial Web Architecture for Multi-Agent Fusion

So let's talk a little bit about the Spatial Web and how it provides fusion based upon space and time.
The Foundation is an IEEE standard developed within the Computer Society, and it's called 2874. It was published last year.
The Protocol Architecture and Governance Standard of the Spatial Web was developed by the Spatial Web Foundation and then brought into IEEE.
It develops and defines semantic spatial and governance framework for interoperable entities of any kind, very general about entities, but it's all based upon space and time as the organizing framework.
So interoperable entities include things like sensors, IoT devices, GIS data and services come in as a type of, if you will, geo-AI. Digital twins for sure, being able to model both mechanical things as well as environments, digital earth, and software agents and humans.
It's an extension of established concepts.
The application of spatial web to fusion builds on established fusion concepts.
Several of us over the years have provided that in previous IGARSS and GRSS publications.
It extends it with an AI agent-centric, agent-centric identity trust and semantically grounded spatial models for these three types of fusion, observation, feature, and decision across distributed intelligent agents.
So that's the framework.

5 of 10

ARCHITECTURE

Pipeline-Centric vs. Spatial Web Approach

SWF - Percivall & Fellah · Spatial Web Architecture for Multi-Agent Fusion

Here's an architecture diagram.
The current kind of traditional pipeline is to be able to take heterogeneous sources and do some kind of a local process-focused processing on observation and future fusion, some kind of a centralized processing, and then make decisions about that.
The Spatial Web agentic approach allows for same access to all those heterogeneous data sources, but allows for agents to work on a shared knowledge graph, which we'll call the universal domain graph here.
And that provides for a modeling of observation features and decisions in a knowledge graph format that's accessible to agents, and then to allow agents in a space to be able to coordinate and adapt and provide collective intelligence and adaptive action based upon that shared knowledge graph and the ability, the various abilities of different agents to be able to do that.
So that's where we're at and where we're headed with respect to a more decentralized agent focused fusion environment that allows for fusion on heterogeneous data sources.

6 of 10

SEMANTIC CORE

Spatial Web Component: HSML

SWF - Percivall & Fellah · Spatial Web Architecture for Multi-Agent Fusion

Conceptual Model: IEEE 2874

Hyperspace Modeling Language: HSML

Knowledge Graph

W3C stack: RDF, OWL, SPARQL

Hyperspace: context for every entity

Space: elements related by paths
Topological, vector, cell, graph spaces

Holonic Model: RDF-H

Every entity is whole and part
Entities may be in multiple domains

Identity and Trust: SWIDs

W3C DIDs and Verifiable Credentials
Identity, policy travel with data

Agents form contracts, conduct activities

Fusion: Activity performed by Agents

So let's go into the components. I mentioned there are three, HSML, HSTP, and UDG.
The first one is HSML. It's a hyperspace modeling language.
The conceptual model for HSML is defined in IEEE 2874 last year, and it's based upon the set of concepts at the base of the framework is space and time and what we call hyperspace.
Hyperspace includes space and time. It also expands the idea of space that I'll show you momentarily.
Concepts that persist over time in that framework of space and time are called domains.

And domains can be a variety of things, including things as well as to the Earth.
And domains have identity and that persist over time.

We then move towards types of domains that describe and enable change.
So change over time is achieved by agents and activities.

Agents enacting activities, coordinated agents working together through defined activities in order to produce change on domains.

We want that to be controlled. We want that to operate within a context in which norms are defined.

It's not a top-down control, but it's a contextually defined norms control of what agents can do based upon the credentials of what agents have, what the norms are, and the ability to form contracts between agents.
Norms are defined within the Domain in which an agent operates.

So that's a conceptual model that's laid out.

What Stefan in particular did was take that and make it into a modeling language, HSML,which is now its own draft standalone spec using the W3C stack of RDF, RDF star, OWL, and SPARQL.
And the hyperspace that goes with that then is a context for every entity.
But here we have space. Elements that are related by paths and that allows for coordinate spaces the paths being between the coordinates as you go through a vector but it also allows for more discrete topological cell graph spaces all within the context of what we call hyperspace and being able to move between these definitions of space using some foundations of category theory
The entities within this space are allowed to have a holonic model, which is defined in an emerging extension to RDF called RDF-H,

where every entity is both a whole in itself and a part, perhaps, of another domain, a higher domain, another entity.
And these entities may be in multiple domains simultaneously. That's a holonic approach of a mereological formation of organization.

And then key to this notion about governance and coordination is identity and trust.

Spatial Web IDs are based upon W3C DIDs, which are distributed identifiers and verifiable credentials.
And that allows for identity policy and policy to travel with the data.

And then, as I mentioned, agents can form contracts, which are defined in HSML and allow for agents to negotiate and determine whether or not the activity has been achieved.
So that's HSML. It's an RDF based language of this concept that you see on the left and is key to the communication between agents.

7 of 10

INNOVATION

Spatial Web Component: HSTP

SWF - Percivall & Fellah · Spatial Web Architecture for Multi-Agent Fusion

Hyperspace Transaction Protocol (HSTP)

Secure, verifiable protocol for HSML
Ensuring seamless interoperability between diverse AI and Digital Twin systems.
HSTP Operations based on Actor Paradigm

Incorporates W3C DID for identity

SW Domain Registration Authority

8 of 10

KNOWLEDGE SUBSTRATE

Universal Domain Graph (UDG)

Knowledge Graph (KG)

HSML hypergraph of all entities
Observations, features, and decisions
Reason over shared knowledge

Hyperspace Features

Independently-trained world models converge on common features.

Anthropic J-Space, Neuronpedia

AI analog of geo-objects, e.g. Goodchild

Need Hyperspace Reference Systems (HRS)

Spatial Web achieves World2Vec, Y. LeCun

SWF - Percivall & Fellah · Spatial Web Architecture for Multi-Agent Fusion

Distributed Registries of Domains

Domain membership and SWID registry
Trustworthiness of data and decisions
Credentials, trust, DIDs

Social Network of Agents

Agent coordination: activities, contracts
Coordinate sensing and analysis strategies
Collective intelligence

Architectural Shift From data pipelines toward knowledge-driven, agent-centric, trust-aware fusion

Shared, distributed knowledge substrate for trusted multi-agent fusion

A key concept, though, to this ability to do diffusion by agent coordination is the universal domain graph.
Think of this as an analogy of the World Wide Web, if you will. It's a shared distributed knowledge substrate for trusted multi-agent fusion.

But now we're not talking about documents, we're talking about entities that make up this domain graph.
It is a knowledge graph. So there is a hypergraph of all the entities that includes observation, fusion and decisions and allows for reasoning over the shared knowledge structure.
And it'll be very heterogeneous in terms of lumpiness, if you will.

let's go to the distributed registry of domains.

Domains are where the authorities are vested and domain membership and registry is key to that have authority and the trustworthiness of data decisions.
That's where the SWIDs come from. That's where the credentials come from. That's where the authority and the domains define the norms and the rules for their domain.

Hyperspace features.

That is novel in the sense based on what we see with world model, that have been independently trained in different ways, they converge on common features.
You can look at the individually trained world models and see structures that are isomorphic between them, not only in terms of structure, but also in terms of meaning.
I think that's astounding to see that this analysis task of LLMs and the like are producing identifiable features ended up how they were trained.
We see this in particular in anthropics, J-space, and Neuronpedia is a great way to visualize this.
This allows for us to discuss the AI analog of geo-objects, if you will, features. Goodchild and the whole community of geospatial science has developed for us.
It allows for this notion of features to be those things that have meaning that are available in space and perhaps multiple representations.
But the key thing is the feature itself has the meaning.
We need for hyperspace reference systems to be able to talk about the relationship between these features.
But this allows us to achieve what LeCun talked about as world-to-vec, the ability to represent the world as vectors, and the spatial web allows that to happen.

That then allows for the social network of agents where agents coordinate.

They have activities and contracts.
They coordinate, and for this discussion, sensing and analysis strategies, the ability to collect, analyze, make decisions, and have essentially a collective intelligence.
This is the basis for the architectural shift from pipelines to an agent-centric distributed environment.

9 of 10

APPLICATIONS

New Capabilities & Application Domains

New Capabilities for Earth Observation

Scalable collective intelligence across heterogeneous platforms and organizations
Interoperable decision-making across agent boundaries
Verifiable trust across heterogeneous platforms and organizations

Application Domains: Spatial Web applied to specific domains

Coordinated satellite and airborne observation
Disaster response: near real-time integration of observations, features, and decisions
Climate and environmental monitoring
Smart infrastructure management with autonomous agents

SWF - Percivall & Fellah · Spatial Web Architecture for Multi-Agent Fusion

What does that allow us to do then? New capabilities and application domains,
new capabilities,

scalable collective intelligence across heterogeneous platforms and organizations,
interoperable decision making across agent boundaries, verifiable trust across heterogeneous platforms and organizations.

What we need to do now is build this out beyond the individual places that allowed us to form kind of this general model, but use it in more specific domains.

Certainly the notion of coordinated satellite and airborne observation, tasking and tipping and queuing and all that stuff.
Disaster response, amazing advances in GeoAI and disaster response are benefiting from the ability to do agentic kind of coordination.
Spatial web provides even stronger abilities there.
Climate environmental monitoring, smart infrastructure with autonomous agents.

So a lot of opportunity for domains as we move forward for broader

10 of 10

CONCLUSION

Summary & Conclusion

Unified Fusion Framework

Observation, feature, and decision fusion unified within a semantic, spatial, and trust-enabled framework
Shifts from pipeline-centric architectures to knowledge-driven, agent-centric models

Grounded in Open Standards

IEEE 2874-2025 Spatial Web, IEEE GRSS fusion frameworks, OGC standards
W3C DIDs and Verifiable Credentials for decentralized trust
Analysis Ready Data principles for multi-modal feature integration

Foundation for Next-Generation Earth Observation

Interoperable systems spanning heterogeneous agents, platforms, and organizations
Autonomous agents capable of collective and trustworthy reasoning
Complex spatial environments with real-time adaptation

SWF - Percivall & Fellah · Spatial Web Architecture for Multi-Agent Fusion

spatialwebfoundation.org

georoundtable.xyz

Well, thanks for your time.
This provides perhaps what you see as a unified framework for what we've called fusion, including observation, feature, and decisions, but now with a semantic, spatial, and trust-enabled framework.

It shifts us from this pipeline-centric to a knowledge-driven agent models.

It's already grounded in open standards, more are needed.

The IEEE 2874 was approved last year.
It's being built out through the components and the testing and development that is continuing through an iterative process within the Spatial Web Foundation, GRSS Fusion, OGC standards, W3C.
And really relates to an GRSS article that we several of a few years ago on ARD and multimodal feature integration.

The Spatial Web allows for a foundation for the next generation Earth observation,

where you've got interoperable heterogeneous agents, platforms and organizations worried about trust, worrying about complex spatial environments, real-time adaptation

Thank you for your time.