Developing an ISR Geospatial AI Solution for Defence & Intelligence: Part 4

Agentic ISR Geospatial Mapping and Reasoning

1. Introduction: From ISR Analytics to Geo Agent Mapping

Modern intelligence, surveillance and reconnaissance operations rarely suffer from a complete absence of data. The more common challenge is turning large volumes of disconnected observations into a coherent and timely operational picture.

An analyst may need to review infrastructure change detected in satellite imagery, examine AI-generated asset detections, compare those observations with SAR maritime detections, inspect AIS vessel movements and determine whether the combined activity represents routine behaviour or a meaningful deviation from the established baseline. Each source contributes part of the picture, but none provides sufficient context alone.

The first three parts of this project developed the analytical building blocks required for this process.

• Phase 1 — Change Detection explored deep-learning methods for identifying changes between temporally separated EO images.
• Phase 2 — Asset Detection and Operational Activity Assessment used YOLOv8 and the xView dataset to detect and consolidate operationally relevant asset classes.
• Phase 3 — Multi-Sensor Fusion combined EO, SAR, AIS, change and contextual evidence to calculate AOI-level fusion scores, confidence levels, alert states and evidence conflicts.

Phase 4 connects these previously separate analytical capabilities through an agentic geospatial workflow. The resulting demo asset environment, called Gulf Sentinel, simulates a maritime situational-awareness mission around the Strait of Hormuz.

Five fictional Areas of Interest, or AOIs, represent progressively more complex operational contexts — from routine commercial shipping to multi-domain air, maritime, logistics and infrastructure activity. Rather than presenting the previous analytical outputs as independent model results, Gulf Sentinel organizes them into an operational evidence store that can be queried by a stateful geospatial agent. An ISR analyst can use natural-language requests to retrieve evidence, compare AOIs, generate geospatial buffers, identify intersecting tracks, examine anomalies, prepare map actions and produce structured intelligence assessments with recommended Courses of Action.

The earlier phases generated analytical evidence. Phase 4 demonstrates how an agent can retrieve, correlate, map and explain that evidence within an analyst-directed mission workflow. The objective was not to create an autonomous intelligence system or replace the judgement of an experienced analyst. Instead, the project explores how geospatial AI, deterministic GIS tools and controlled language-model reasoning can reduce the effort required to move from sensor observations to an evidence-grounded operational assessment.

About Gulf Sentinel Demo

Gulf Sentinel is a synthetic demonstration. Its AOIs, sensor observations, asset locations, anomaly events and intelligence products do not represent real-world operational intelligence.

The underlying analytical methods are based on capabilities implemented during the previous phases. However, the Gulf Sentinel detections are positioned synthetically within representative geographical settings. For example, the EO evidence retains realistic fields such as asset class, confidence, observation time, model name and source-scene lineage, but it is not presented as the output of running YOLO against real imagery collected over the Strait of Hormuz.

This distinction is important. The demonstration is intended to show how an operational GeoAgent could work with these evidence types, without making unsupported claims about the provenance or geographic precision of the simulated observations.

The final Phase 4 capability combines:

• a standardized operational evidence store;
• reusable geospatial and analytical tools;
• scenario-aware anomaly assessment;
• stateful GeoAgent orchestration using LangGraph;
• interactive Folium mapping;
• controlled ArcGIS Pro integration;
• intelligence-report and CoA generation;
• and an analyst-facing Gradio operational dashboard.

Together, these components transform the project from a sequence of isolated model experiments into an integrated Agentic ISR Geospatial Copilot demonstration.

2. Why Agentic Geospatial Reasoning Matters in ISR

ISR analysis is fundamentally spatial, temporal and evidential. An observation gains operational meaning through its relationship with a location, a period of interest, an expected baseline and other supporting or conflicting observations. A detected vessel may be routine when considered alone. The same observation may become more significant if it is associated with an AIS transmission gap, a SAR–AIS count disagreement, increased logistics activity or recently detected infrastructure change.

This creates a demanding analytical workflow. Analysts may need to move between imagery viewers, GIS applications, vessel-tracking platforms, detection tables, fusion dashboards and reporting templates. They must repeatedly locate the relevant AOI, filter evidence, compare observation windows, inspect metadata, create spatial selections and translate the findings into a concise assessment.

Geospatial agentic AI offers a way to coordinate these activities through a unified mission context.

2.1 Moving beyond a conventional dashboard.

A conventional dashboard normally presents predetermined charts, filters and buttons. It is useful for monitoring known indicators, but the ISR analyst must adjust the question to the interface. A Geo Agent reverses part of this relationship. The analyst can express the operational objective in natural language, while the agent interprets the request and coordinates the required tools.

Example requests might include:

• “Assess AOI-01 for unusual AIS behaviour or transmission gaps.”
• “Compare AOI-04 and AOI-05 across the available sensor domains.”
• “Create a five-kilometre buffer and identify intersecting assets and tracks.”
• “Explain why SAR and AIS vessel counts disagree.”
• “Show the highest-priority anomalies and highlight them on the map.”
• “Recommend additional collection to reduce uncertainty.”
• “Generate an evidence-grounded intelligence assessment and alternative CoAs.”

These requests do not map to a single database query or model prediction. They may require several steps: resolving the AOI, retrieving evidence, calling spatial tools, examining anomaly results, comparing fusion assessments and preparing map updates.

2.2 What makes the workflow Agentic?

The Gulf Sentinel workflow is agentic because it can progress through a structured reasoning cycle:

1. Interpret the analyst request.
   The system identifies the requested AOI, mission objective and relevant analytical intents.
2. Create a bounded plan.
   The GeoAgent selects from an approved set of evidence, anomaly, reporting and geospatial tools.
3. Execute deterministic operations.
   Evidence retrieval, buffering, track intersection, AOI comparison and score lookup are performed by code rather than calculated by the language model.
4. Prepare geospatial actions.
   The workflow generates structured actions such as zooming to an AOI, showing evidence, highlighting anomalies or displaying a buffer.
5. Synthesize an assessment.
   Tool results are converted into a readable response that distinguishes observations, calculated results, analytical inference and uncertainty.
6. Maintain conversational state.
   LangGraph preserves the active mission thread and selected AOI, allowing follow-up questions to build on previous interactions.

This separation between reasoning and computation is central to the design. The language model may help interpret intent or explain results, but it does not invent coordinates, calculate spatial intersections or determine evidence counts. These operations remain inside explicit Python and GIS tools.

The orchestration layer also validates requested AOIs and tool intents against approved lists. If an optional language model is unavailable or returns an invalid plan, the demonstration can fall back to deterministic planning.

2.3 Enhancing the ISR analyst Tasks

In defence and intelligence environments, the quality of an answer is not determined only by how fluently it is written. Analysts need to understand:

• What evidence supports the assessment;
• Which calculations were performed;
• Where sources disagree;
• How confident the system is;
• Which assumptions were introduced;
• What requires human validation.

For that reason, the Geo Agent retains a tool-execution trace and prepares map actions as structured records. ArcGIS execution remains a separate, controlled step rather than an invisible autonomous action. The objective is allowing the system to coordinate approved tools and evidence while preserving traceability, analyst control and the distinction between detected activity and inferred intent.

This makes the Geo Agent particularly useful for repetitive but cognitively demanding tasks such as:

• retrieving evidence distributed across sensor domains;
• applying the same analytical procedure consistently across AOIs;
• identifying conflicts between sensors;
• generating initial summaries;
• producing alternative collection or validation options;
• and maintaining the connection between the written assessment and the operational map.

3. Leveraging Evidence from Previous ISR Phases

Phase 4 was designed as the integration point for the complete ISR analytics workflow. The previous phases did not disappear once their model development was complete. Their outputs, analytical concepts and metadata structures became evidence sources for the Geo Agent.

Phase 1 investigated how temporally separated EO images could be compared to identify structural change. The implementation explored Siamese U-Net, Attention Siamese U-Net and a lightweight ChangeFormer architecture using the LEVIR-CD dataset.

Phase 2 used the xView dataset and YOLOv8 to explore military-relevant asset detection. The original xView taxonomy contained 59 classes, which were consolidated into a smaller ISR-oriented taxonomy covering air assets, ground mobility, logistics and maritime activity.

Phase 3 as sensor fusion connected EO detections with synthetic SAR observations, AIS tracks, change evidence and contextual metadata through a late-fusion methodology.

For example, an AOI may contain several sensor observations but still have reduced assessment confidence because:

• SAR detects more maritime objects than AIS reports;
• EO confirmation is limited by image quality;
• AIS track contains a transmission gap;
• Observations were collected at different times;
• Spatial association does not prove coordinated behaviour.

Table 1: How the phases contribute to Gulf Sentinel Asset

The different analytical outputs produced during the previous phases needed to be transformed into a consistent evidence foundation before they could support agentic reasoning. Change-detection results, EO asset observations, SAR detections, AIS tracks and fusion assessments originally contained different attributes, geometries and levels of analytical detail. Phase 4 standardized these sources around a common operational context that associates every observation with an AOI, sensor domain, observation time, confidence level, quality indicator, source lineage and geospatial geometry.

This shared evidence foundation allows the GeoAgent to treat information from different sensors as parts of the same mission picture. An EO aircraft detection, an AIS vessel track, a SAR maritime observation and an infrastructure-change footprint can be retrieved through a consistent analytical workflow while retaining their individual origins and limitations. The approach also preserves the distinction between observed evidence, model-generated results and higher-level fusion assessments.

4. The Six-Layer Implementation Model

Gulf Sentinel was implemented as a six-layer geospatial reasoning architecture. Each layer addresses a different stage of the analytical workflow, from organizing sensor observations to presenting evidence-grounded intelligence products to the analyst.

The design deliberately separates data management, deterministic computation, agent orchestration and intelligence interpretation. This separation is important because geospatial calculations, evidence counts and fusion scores should not be generated freely by a language model. They should be produced by controlled analytical functions and then presented to the agent as structured evidence.

4.1 Operational Evidence Store

The Operational Evidence Store provides the shared analytical foundation for Gulf Sentinel. Its purpose is to transform heterogeneous outputs from change detection, EO asset detection, SAR observations, AIS tracks and sensor-fusion assessments into a consistent operational evidence model.

Each observation is associated with an AOI, sensor type, operational domain, timestamp, confidence score, quality indicator, source reference, model lineage and geospatial geometry. This common structure allows evidence from different phases and sensors to be queried through the same analytical workflow while preserving its individual provenance.

For example, an EO aircraft detection and an AIS vessel track have different geometries and operational meanings, but both can be filtered by AOI, observation period, confidence level or proximity to an area of interest. A change-detection polygon can be evaluated alongside asset observations without being reduced to the same type of measurement.

The result is a traceable evidence foundation that supports geospatial reasoning without asking the GeoAgent to invent observations or calculate results from unstructured descriptions.

4.2 Geospatial Analysis and Mission Tools

The second layer exposes the evidence through a reusable set of deterministic geospatial and mission-analysis tools.

These tools perform the calculations requested by the GeoAgent. The agent coordinates them, but the underlying spatial operations remain implemented through Python, GeoPandas and geometric processing functions.

The principal capabilities include:

• Retrieving an AOI and its operational context;
• Filtering evidence by sensor, domain, asset type, time or confidence;
• Retrieving individual records by evidence ID;
• Creating distance-based AOI buffers;
• Identifying assets that intersect an AOI or buffer;
• Selecting AIS tracks that cross a scenario area;
• Comparing evidence between observation periods;
• Retrieving sensor-fusion assessments and evidence conflicts;
• Comparing several AOIs;
• Ranking AOIs by operational priority;
• Generating structured operational summaries.

A particularly important example is AOI buffering. Geographic coordinates are expressed in latitude and longitude rather than metres, so directly applying a five-kilometre buffer in geographic degrees would introduce distance distortion. The Gulf Sentinel tool first estimates an appropriate local projected coordinate system, performs the buffer in metres and then converts the result back into the standard geographic coordinate system for mapping. This approach keeps the geospatial calculation deterministic and reproducible. It also allows the same tool to be used by the Folium dashboard, the LangGraph agent and the ArcGIS integration workflow.

4.3 Anomaly and Multi-Sensor Reasoning

The third layer adds scenario-aware anomaly assessment to the Phase 3 sensor-fusion context.

A universal anomaly rule would be inappropriate across all five scenarios. Missing SAR evidence is not anomalous in AOI-01 because that scenario is intentionally based on AIS monitoring. Similarly, the absence of AIS tracks is not an anomaly for the land-based AOI-03 air-base scenario.

Gulf Sentinel therefore enables only the anomaly checks relevant to each AOI.

The implemented checks include:

• AIS transmission gaps;
• AIS route deviation;
• persistent vessel presence;
• EO activity volume above a scenario threshold;
• Change footprint above a defined area threshold;
• and disagreement between represented SAR maritime objects and AIS tracks.

Each generated anomaly event contains a type, severity, score, confidence value, supporting evidence IDs, explanatory text and recommended follow-up action. This makes the anomaly assessment easier to audit than a single unexplained alert.

The overall AOI anomaly score gives greater weight to the strongest event while still accounting for the wider anomaly pattern:

Anomaly score = 55% of the highest event score + 45% of the mean event score

The anomaly score remains separate from the Phase 3 fusion score. Gulf Sentinel also calculates a contextual-priority indicator to help rank analyst attention:

Contextual priority = 65% fusion score + 35% anomaly score

These calculations serve different purposes.

• The fusion score represents the combined level of operational activity across available evidence.
• The anomaly score measures departure from the expected scenario baseline.
• The confidence score reflects the reliability and consistency of the supporting assessment.
• The contextual-priority score provides an analyst-review cue using both fused activity and anomaly context.

A high anomaly score does not establish hostile intent. It indicates that represented activity differs from the configured baseline and warrants further examination or collection.

4.4 Geo Agent Chat Orchestration

The fourth layer transforms the geospatial and analytical tools into a stateful mission workflow.

Gulf Sentinel uses a single LangGraph Geo Agent rather than multiple independent agents. This design keeps the demonstration understandable and allows one agent to coordinate three tool families:

• Base geospatial and evidence tools;
• Anomaly-assessment tools;
• Intelligence-reporting and CoA tools.

The LangGraph workflow follows five workflow stages as defined below.

The agent recognizes intents such as evidence retrieval, anomaly assessment, buffer creation, AIS-track analysis, AOI comparison, map visualization, report generation and CoA development.

For example, the request:

“Assess AOI-04, explain the SAR–AIS disagreement and show intersecting vessel tracks.”

can be resolved into several coordinated operations:

1. Retrieve the AOI-04 operational summary;
2. Examine its evidence conflicts;
3. Run the relevant anomaly checks;
4. Identify AIS tracks intersecting the AOI;
5. Prepare map actions to zoom, display evidence and highlight tracks;
6. Synthesize the findings into a readable assessment.

The workflow operates in a deterministic mode by default, making it possible to demonstrate the complete analytical sequence without external model credentials. An optional OpenAI or Ollama model can assist with planning and narrative synthesis, but its output is validated against explicit AOI and intent allowlists.

Get George Regkas’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Even when an LLM is enabled, it does not calculate coordinates, spatial intersections, sensor counts, fusion scores or anomaly scores. If the model is unavailable or produces an invalid plan, the workflow can return to deterministic planning and synthesis.

LangGraph also maintains the active conversation context. This can support follow-up requests such as:

• “Now compare it with AOI-05.”
• “Increase the buffer to ten kilometres.”
• “Show only AIS and SAR evidence.”
• “Generate a report using the same assessment.”

This stateful behaviour allows the interaction to develop as an analytical mission thread rather than a series of disconnected questions.

4.5 ArcGIS Pro and Map-Action Integration (Optional)

The fifth layer explores how GeoAgent results can be transferred into a professional GIS environment. The workflow prepares AOIs, EO evidence, SAR detections, AIS tracks, change footprints, selected assets and generated buffers for ArcGIS integration.

The ArcGIS Pro section then converts the staged evidence into feature classes inside a geodatabase. The GeoAgent Results group contains derived products such as AOI buffers, selected assets, selected change areas and intersecting AIS tracks. Baseline ISR symbology is applied, and the map can zoom to the AOI selected by the analyst. For the sake of implemenbtation we left this implementation as optional

4.6 Operational Intelligence & CoA Generation

The final layer brings the complete workflow into an analyst-facing operational interface.

The dashboard combines the functionality developed across the previous layers into six workspaces:

• Operational Overview;
• GeoAgent Console;
• Evidence Register;
• Anomaly Assessment;
• Intelligence and CoAs;
• and ArcGIS Integration.

The Operational Overview provides mission-level context through AOI rankings, fusion scores, confidence indicators, alert levels, evidence composition and recommended analytical actions.

The GeoAgent Console places the conversation directly beside the interactive Folium map. This allows the analyst to observe how a request changes the selected AOI, visible evidence layers, asset detections, anomaly highlights, buffers and AIS tracks. EO asset evidence is displayed with class-specific colours, probability labels and bounding-box-style visualization proxies. These boxes improve map readability but are explicitly identified as representative display geometries rather than original image-space YOLO footprints. The Evidence Register exposes the observations supporting the assessment, while the Anomaly Assessment presents scenario-relevant events with severity, confidence, supporting evidence and recommended follow-up.

The Intelligence and CoA workspace converts the analytical context into structured products containing:

• Executive assessment;
• Principal indicators;
• Evidence conflicts;
• Confidence and limitations;
• Intelligence gaps;
• Recommended collection;
• Alternative Courses of Action

The resulting interface does not operate as an autonomous command system. It functions as an analyst-support environment in which evidence retrieval, computation, interpretation and controlled action remain distinguishable.

Layered Geospatial ISR Reasoning Architecture (LLM Generated)

5. Simulating Land and Maritime situational-awareness

To evaluate the Gulf Sentinel workflows, five fictional AOIs were designed around progressively more complex situational-awareness requirements.

The scenarios are inspired by representative geographical settings in and around the Strait of Hormuz, and their boundaries, observations, identities and assessments are synthetically generated. Below you can see an overview of the five scenarios that were generated together with their evidence domains and analytical challenges.

AOI-01: Establishing a maritime baseline

Alpha Corridor represents the least complex scenario. Its evidence consists only of AIS tracks associated with routine commercial shipping.

AOI-02: Correlating logistics and vessel movement

Bravo Logistics Port combines EO asset evidence with AIS port-arrival and departure track

AOI-03: Assessing air-base activity and change detection

Q Air Base is a predominantly land-based scenario containing fixed-wing aircraft, helicopters, ground-support vehicles and an infrastructure-change footprint.

AOI-04: Resolving conflicting maritime evidence

Delta Island Zone contains EO, SAR, AIS and coastal-change evidence, but the sensors do not provide complete agreement.

AOI-05: Full multi-domain complexity

Echo Strait Zone represents the highest-volume and highest-priority scenario.

Operational Scenarios Sentinel Gulf

The purpose was to simulate a real intelligence picture and to create a controlled environment in which the Geo Agent could demonstrate increasingly sophisticated forms of evidence retrieval and reasoning.

Below you can see indicatively AOI scenarios of Geo Agent console with the buffer zones of 5 km, the AIS maritime tracking and the military asset detection.

AOI-02: Bravo Logistics Port including AOI buffer area, AIS tracking, EO asset classification (trucks, container stacks, utility trucks, maritime vessels)

AOI-03: Q Air Base including AOI buffer area, AIS tracking, EO asset classification (trucks, ground support vehicles, helicopters, fixed wing aircrafts)

AOI-04: Delta Island Zone including AOI buffer area, AIS tracking, EO asset classification (trucks, container stacks, utility trucks, maritime vessels)

AOI-05: Echo Strait Zone including AOI buffer area, AIS tracking, EO asset classification (patrol vessels, logistics support vessels, maritime patrol aircrafts, maritime vessels)

6. Evaluating Multi-Sensor Fusion Across AOIs

The Operational Overview uses the Phase 3 fusion assessments to compare activity across the five Gulf Sentinel scenarios.

The fusion score represents the combined level of operational activity derived from the available evidence. The confidence score reflects the reliability, quality and consistency of that assessment. Alert levels translate the calculated context into a simple operational category.

These results demonstrate why fusion score and confidence should be displayed separately.

AOI-01 has the lowest fusion score but the highest confidence. Its represented activity is largely routine and supported by a clear AIS baseline.

AOI-04 has a higher fusion score but the lowest confidence. Multiple sensors indicate activity, yet they disagree about the number of maritime objects and EO quality is reduced. The assessment is operationally important precisely because uncertainty remains unresolved.

AOI-03 has a high fusion score and relatively strong confidence because its EO asset activity is supported by infrastructure-change evidence.

AOI-05 produces the highest fusion score and a critical alert because it combines the largest evidence volume with multiple sensor domains, AIS transmission gaps, persistent maritime activity and change context.

6.1 Ranking operational priority

The base AOI-ranking tool combines the Phase 3 fusion and confidence scores:

Priority score = 70% fusion score + 30% confidence score

The anomaly layer introduces a separate contextual-priority calculation:

Contextual priority = 65% fusion score + 35% anomaly score

6.2 Interpreting the operational overview dashboard

The dashboard’s green, yellow, orange and red bands provide rapid visual prioritization, but they are not automated determinations of threat or intent.

A critical alert means that the represented scenario contains a high combination of fused activity and supporting context. It does not mean that the system has independently identified hostile behaviour.

Similarly, a sensor disagreement is not necessarily a model failure. A SAR–AIS count difference could result from:

• vessels not broadcasting AIS;
• temporal misalignment between observations;
• SAR false positives;
• occlusion or image-quality limitations;
• or differences in sensor coverage.

The correct analytical response is therefore not always escalation. It may be additional collection, improved temporal synchronization or human validation.

Operational Overview for AOI 5- Echo Strait Zone

The Operational Overview supports this process by presenting the fusion score, confidence, alert level, anomaly context, contributing indicators and evidence conflicts together. This gives the analyst a more balanced view than any individual score could provide.

Ranking of AOIs in terms of fusion score and alert level

7. Evidence Register and Anomaly Assessment

Fusion scores provide a useful operational summary, but analysts still need access to the evidence behind them. Gulf Sentinel addresses this through two connected dashboard workspaces: the Evidence Register and the Anomaly Assessment.

7.1 Maintaining traceability through the Evidence Register

The Evidence Register presents the observations associated with the selected AOI across EO, SAR, AIS and change-detection sources. Each record retains its sensor type, operational indicator, timestamp, confidence, quality and source or model lineage.

This allows the analyst to move from a high-level assessment back to the supporting observations. A critical alert can therefore be examined as a collection of individual detections, tracks and change events rather than accepted as an unexplained system judgement.

The register also makes evidence limitations visible. A detection with lower confidence, reduced imagery quality or temporal misalignment remains available for analysis, but it can be distinguished from stronger corroborating evidence.

Evidence Register for AOI-04 Delta Island Zone

7.2 Identifying Anomalies from the operational baseline

The Anomaly Assessment adds a second analytical perspective. Instead of asking only how much activity is present, it asks whether the represented activity differs from what is expected for that scenario.

The dashboard displays anomaly type, severity, score, confidence, supporting evidence and recommended follow-up. Examples include AIS transmission gaps, route deviation, persistent vessel presence, elevated EO activity and disagreement between SAR detections and AIS tracks.

Anomaly Assessment Operational View for AOI-04 Delta Island Zone

8. ISR Mission-GeoAgent Tasking Workflows

The GeoAgent Console turns the Gulf Sentinel evidence and analytical tools into an interactive mission-tasking workflow. Instead of navigating independently through multiple tables and map controls, the analyst can describe the required task in natural language.

Representative prompts include:

• “Assess AOI-01 for AIS transmission gaps or route deviation.”
• “Compare AOI-04 and AOI-05 across the available sensors.”
• “Create a five-kilometre buffer and select intersecting evidence.”
• “Explain the SAR–AIS disagreement in AOI-04.”
• “Recommend additional collection to reduce uncertainty.”

The GeoAgent resolves the relevant AOI and analytical intent, selects approved tools, executes the required evidence or GIS operations and then prepares a grounded response. At the same time, it can update the adjacent map by zooming to the AOI, displaying selected evidence, drawing buffers or highlighting intersecting AIS tracks.

Because the workflow maintains mission context, follow-up requests can build on the previous assessment. An analyst can ask the agent to compare the current AOI with another scenario, increase a buffer distance or generate a report without restating the complete analytical context.

GeoAgent chat interface attached to the Map: Tasking GeoAgent for for asset detection located with 5-km buffer (blue line) of AOI-04 Delta Island Zone

9. Intelligence Assessment & CoA Generation

The final analytical step is converting sensor observations and GeoAgent findings into a structured intelligence product.

Gulf Sentinel generates assessments containing:

• An executive operational summary;
• Principal supporting indicators;
• Evidence conflicts;
• Confidence and analytical limitations;
• Intelligence gaps;
• Recommended collection priorities;
• Alternative Courses of Action.

The CoAs are derived from the selected AOI’s evidence, fusion assessment, anomaly events and unresolved intelligence gaps. They are designed to present alternative analytical responses rather than produce a single prescriptive decision.

Example: CoA generation for AOI-05

Consider an ISR analyst tasking:

“Generate an intelligence assessment and recommended CoAs for AOI-05.”

The Geo Agent retrieves the critical fusion assessment, high-confidence multi-sensor activity, AIS transmission gaps, persistent vessel presence, SAR maritime observations and relevant change evidence. It also retains the central limitation: spatial and temporal association between these observations does not independently establish coordinated or hostile activity. This creates a more natural interaction with the operational picture while keeping the underlying computations deterministic and traceable.

COA generation for AOI-05 based on sensor fusion overall assessment

The Geo Agent may recommend CoA 2 as the preferred analytical option because synchronized collection directly addresses the principal intelligence gap: whether the SAR, AIS and EO observations represent the same activity within a comparable time window.

However, that recommendation remains conditional. If new collection is unavailable, the analyst may select CoA 1 or combine enhanced monitoring with the targeted validation described in CoA 3. This example illustrates that CoA generation is not simply a list of generic actions. Each option is connected to the available supporting evidence, The uncertainty it attempts to reduce, its expected analytical benefit and its operational or collection limitation.

The CoAs remain analytical and collection-oriented recommendations — not autonomous operational decisions. Gulf Sentinel helps structure the alternatives, while the analyst evaluates feasibility, risk, mission priorities and the appropriate course to approve.

Intelligence Assessment for the significance of elevated multi-sensor activity for AOI-05 and COA generation

10. Human Oversight and Agentic Explainability

Explainability in Gulf Sentinel asset is not treated as a single text explanation generated after the result. It is built into the workflow through evidence traceability, deterministic tools and visible execution records.

The analyst can inspect:

• the observations supporting an assessment;
• the confidence and quality of those observations;
• the tools selected by the GeoAgent;
• the calculated fusion and anomaly results;
• the conflicts and limitations retained in the final narrative;
• and the map actions prepared for execution.

This creates several levels of AI explanability. Evidence lineage explains where the information originated. Tool traces explain how it was processed. Confidence and conflict indicators explain why the conclusion may remain uncertain. The final narrative explains the operational significance without hiding those limitations.

Human control is also preserved at the execution layer. The Folium map can provide an immediate visual representation, but structured ArcGIS actions remain available for analyst review before being applied in the professional GIS environment. Intelligence reports and CoAs are similarly presented as decision-support products rather than authoritative commands.

The resulting model is one of bounded analytical agency: the Geo Agent can coordinate approved tools, maintain mission context and prepare assessments, but consequential interpretation and action remain under human control.

11. Summary and Conclusions

Gulf Sentinel asset demonstrates how separate geospatial AI capabilities can be transformed into a coherent analyst-support workflow.

Change detection, EO asset detection, SAR observations, AIS tracks and multi-sensor fusion are consolidated into a shared operational evidence foundation. Deterministic geospatial tools expose that evidence to a stateful Geo Agent, while anomaly assessment, interactive mapping and intelligence-report generation extend the workflow from detection toward operational interpretation.

The central achievement is not simply the addition of a chat interface. It is the connection between natural-language mission tasking and controlled analytical operations. The GeoAgent can retrieve evidence, compare AOIs, execute spatial analysis, identify anomalies, prepare map actions and generate structured assessments while maintaining traceability to the underlying observations.

The demonstration also highlights several principles that are essential for responsible agentic AI in ISR:

• Sensor activity and inferred intent must remain distinct;
• Fusion, anomaly and confidence scores serve different purposes;
• Geospatial calculations should be performed by deterministic tools;
• Conclusions should remain linked to evidence and uncertainty;
• Operational decisions must remain under human authority.

Gulf Sentinel remains a synthetic asset rather than a deployed intelligence platform. Nevertheless, it shows how geospatial AI, GIS, agent orchestration and human-centred explainability can be combined into a practical foundation for next-gen ISR decision-support systems.

Disclaimer: Part of the views expressed here are those of the article’s author and do not represent the views of IBM Corporation. Part of the content (including images) on the blog is copyright from the author or 3rd party and all rights are reserved — but, unless otherwise noted- under IBM Corporation