Geospatial intelligence & ISR imagery analysis (defence & security)

Defence, border-security and crisis-monitoring organisations collect far more satellite, aerial and drone imagery than analysts can inspect. AI models pre-screen the flood: they detect and classify objects, flag changes between passes and draft reports, so scarce image analysts spend time on judgement rather than searching. This entry covers policy, architecture and vendor landscape only. Targeting, weapon-engagement and strike-decision functions are explicitly out of scope. Tag: public-sector / defence use only.

Problem & value

• Today: imagery analysts search wide areas manually for objects of interest (vehicles, vessels, aircraft, infrastructure) and compare new against archive images. Collection volume (commercial constellations, SAR, drones) grows faster than analyst headcount; cloud cover and night limit optical sensing.
• AI value: faster triage and cueing, persistent change monitoring across large areas, SAR/EO/IR fusion, structured reporting, and tipping-and-cueing between sensors. The analyst remains the decision-maker.
• Why distinct from VIS-021: classified sensors and data, military object classes, scarce labels, disconnected edge deployment, adversarial/deceptive targets and a separate legal regime (see below).
• Public figures on analyst time savings are largely vendor or programme claims; treat them as unverified and measure locally.

Variants

• Strategic/national intelligence: archive-wide search and change monitoring on satellite imagery.
• EU/NATO situational awareness: treaty-compliance, crisis and security monitoring by bodies such as the EU Satellite Centre (SatCen) and Copernicus Security Service for external action.
• Border and maritime security: wide-area vessel and activity detection with SAR (civil-security framing; see VIS-044, VIS-038).
• Tactical ISR on the edge: on-platform detection on drones/aircraft with intermittent connectivity.
• Humanitarian/crisis: damage and displacement mapping (see VIS-058).
• Not covered: weapons targeting, fire control, autonomous engagement.

How it works

1. tasking / archive query
2. ingest (EO, IR, SAR, FMV; metadata, orthorectification, co-registration)
3. pre-processing (tiling, cloud masking, SAR speckle filtering)
4. detection & classification (object detection, segmentation, open-vocabulary / VLM search)
5. change detection (multi-temporal comparison, activity baselining)
6. cueing & fusion (rank candidates, link with other sources, KNOW-026 style entity resolution)
7. analyst review (confirm / reject, confidence + provenance shown)
8. report drafting (LLM, grounded in confirmed detections)
9. release authority sign-off
10. feedback of corrections to model training set

• Detectors: oriented-box object detectors and segmentation models trained on overhead imagery; foundation models for remote sensing support few-shot and text-queried search.
• SAR: works through cloud and darkness; interpretation needs dedicated models and expertise.
• Human-in-the-loop: every detection is a cue, not a finding. Reports are released by accountable humans; engagement decisions are never produced or automated by this system.
• Security: accredited networks, classification handling, model supply-chain control, logging.

Data preparation

Data are satellite, aerial and drone imagery (EO, IR, SAR, video) with sensor metadata, plus archive imagery and analyst reports. They typically sit in image archives, GIS servers, mission systems and classified file stores, not in the ERP or DMS. Typical gaps: inconsistent coordinate systems and metadata, missing co-registration, cloud cover, SAR speckle, few labelled examples for military-specific classes, scattered analyst knowledge in reports and e-mail, and strict classification and need-to-know rules.

Preparation steps: build a catalogue (STAC) of available imagery, orthorectify and co-register, tile and mask clouds, harmonise metadata, and clean up access rights per classification level. Convert past analyst reports into searchable text (PREP-006, PREP-007, DOC-014) and tag them (PREP-010). Harvest labels from confirmed past detections (PREP-020), define a class scheme before training (PREP-019) and build an expert-labelled evaluation set (PREP-018). Classify data for AI use (PREP-024).

Rough share of effort: data engineering, labelling and evaluation make up well over half of the effort; model integration is the smaller part. Labelling for rare classes takes months and needs analysts' time.

Implementation paths

Which path for whom

• Public (armed forces, agencies): Path A, because accreditation, long-term support and integration with existing C2/GIS systems favour accredited defence platforms and government programmes.
• Enterprise (infrastructure, insurance, commodities, security services): Path A, because commercial geospatial analytics and satellite data providers cover standard monitoring and change detection at predictable cost.
• Developer: Path C, because open tooling (GDAL, TorchGeo, open EO foundation models, public benchmarks) allows prototyping and air-gapped deployment without external services.

Path A — Buy (off-the-shelf / government programmes)

• Defence analytics platforms: Palantir (Maven Smart System, Foundry/AIP), Helsing (Germany/EU), Safran.AI (formerly Preligens, France/EU), Anduril (Lattice), Scale AI. EU-based options (Helsing, Safran.AI, Airbus Defence and Space) are preferred for sovereignty.
• Data providers: ICEYE (Finland/EU, SAR), Airbus Defence and Space (Pleiades Neo, SpaceDataHub), Planet, Maxar (Vantor), Capella, Umbra; EU public data via Copernicus.
• Geospatial AI platforms: Esri ArcGIS (deep learning tools), Orbital Insight/Privateer-type analytics, Earth-i, Ursa Space (check EU hosting and export control).
• Choose A when: accreditation, long-term support and integration with existing C2/GIS are required, which is typical for armed forces and agencies.

Path B — Build on cloud APIs

• Sovereign or accredited government cloud only; no unaccredited public API for classified imagery.
• Unclassified/open-source work: Copernicus Data Space Ecosystem, Sentinel Hub, Google Earth Engine, Microsoft Planetary Computer, with Azure/AWS/Google EU regions or EU providers (OVHcloud, STACKIT, Scaleway) for processing.
• Orchestration with STAC catalogues and workflow engines; LLM report drafting through EU-hosted models.

Path C — Self-host (open source)

• Libraries: GDAL, Rasterio, STAC tooling, PyTorch with TorchGeo, MMRotate/Ultralytics for oriented boxes, SAM-family segmentation, ONNX Runtime / TensorRT for edge.
• Models and data: open Earth-observation foundation models (e.g. Prithvi, Clay, SatMAE-type), public benchmarks such as xView, DOTA, fMoW and SpaceNet; SAR sets such as SpaceNet 6 and OpenSARShip.
• Hardware: HW3 for training/inference on large mosaics; HW1 (Jetson Orin class) for airborne edge. Air-gapped deployment expected.

Path D — Custom training

Usually necessary for military-specific object classes and sensors, because public datasets cover civil classes. Needs curated, classified, analyst-labelled imagery, synthetic data/simulation for rare classes, and red-teaming against camouflage, decoys and adversarial patterns. Expect months of labelling and evaluation.

State of the art (2026)

• Programmes: US Project Maven, now run as a programme of record by the National Geospatial-Intelligence Agency, and its Maven Smart System reached NATO use via NCIA; European defence start-ups (Helsing, Safran.AI) and the EU ecosystem are scaling.
• Commercial SAR/EO constellations give frequent revisits, enabling persistent change monitoring and tipping-and-cueing.
• Foundation and vision-language models allow text-queried search ("find ...") and drafting of reports, but still struggle with small objects, oblique views, rare military classes and deceptive targets.
• Limits: domain shift between sensors/regions, label scarcity, false positives at scale, automation bias, adversarial concealment, and weak calibration; open benchmarks do not represent operational data.
• Generative AI can hallucinate in report drafts; restrict it to summarising confirmed detections with references.

Shared resources & combinations

• Shares remote-sensing models and pipelines with VIS-021, VIS-058, VIS-061 (InSAR), VIS-042 and VIS-038.
• Drone and air-defence sensing: VIS-044; video summarisation and NL search: VIS-031.
• Analyst-side fusion and link analysis: KNOW-026; geospatial analytics: DATA-011.
• One GPU cluster (HW3/HW4) can serve training for several earth-observation use cases.

Effort, cost & KPIs

• Path A: quarters, procurement and accreditation dominate; licences plus imagery subscriptions.
• Path B/C: 6-12+ months for an accredited pipeline; labelled data and security accreditation are the main cost drivers.
• KPIs: recall/precision per object class on held-out operational data, analyst time per scene, false-alert rate, latency from collection to report, calibration quality, analyst override rate, and drift per sensor.

People & change

Start with management and the department lead, who define the mission, the tasking process and where the human decision stays. IT and information security are needed early because accredited networks, classification handling and model supply-chain control determine the architecture. Legal/compliance clarifies export control, sector rules and the AI Act scope: military use is excluded from the Act, dual-use and civil deployments are not, and high-risk duties would apply from 2 Dec 2027. Data owners of imagery archives must release access.

Works council: depends. Imagery of terrain rarely monitors staff, but analyst work logs and performance dashboards would trigger §87(1) no. 6 BetrVG; §90(1) no. 3 requires early information about AI-supported work processes. In authorities the Personalrat applies. A DPIA is only needed if persons are identifiable (for example high-resolution drone imagery); involve the DPO to check.

Support AI literacy (AI Act Art. 4) with analyst training on confidence scores, decoys and automation bias. Pilot with a small analyst cell on archive scenarios. Typical concern: loss of analytic skill or being sidelined; answer with a clear rule that detections are cues, analysts release reports, and engagement decisions are never automated. See GUIDE-009, GUIDE-013, GUIDE-008.

Risks, regulation & compliance

• EU AI Act: Art. 2(3) excludes AI systems placed on the market or used exclusively for military, defence or national-security purposes. Dual-use systems also used for civil purposes fall under the Act: e.g. border control or law-enforcement analytics can be high-risk (Annex III). Art. 5 prohibitions (for example biometric categorisation of sensitive traits, real-time remote biometric identification for law enforcement outside narrow exceptions) apply to in-scope uses; do not repurpose overhead-imagery tooling for person identification. Classification is context-dependent, hence depends.
• International humanitarian law: Art. 36 of Additional Protocol I requires legal review of new weapons, means and methods of warfare; ISR decision support feeding into targeting must be reviewable, with meaningful human control. NATO and EU states publish responsible-AI principles for defence.
• GDPR: generally applies to civil authorities (border, police, crisis); military processing is largely outside the GDPR but national law applies. High-resolution imagery can reveal individuals' activity; apply minimisation, retention limits and DPIAs for civil deployments.
• Export control and sovereignty: dual-use Regulation (EU) 2021/821, imagery resolution licensing, vendor lock-in and foreign-jurisdiction access (US CLOUD Act) risks.
• Operational risks: automation bias, misclassification with grave consequences, spoofing/deception, data poisoning, model theft. Mitigate with human review, adversarial testing, provenance logging and independent evaluation.

Getting started

1. Define the exact decision supported (cueing, monitoring, reporting) and the legal regime (AI Act scope, IHL, GDPR for civil bodies).
2. Start with unclassified data (Copernicus Sentinel, commercial archive) and open benchmarks to baseline detection and change-detection quality.
3. Build an analyst review workflow and labelling loop before any automation; log every model output and human decision.
4. Run a controlled evaluation on representative, held-out imagery with metrics per class and sensor; red-team for deception.
5. Plan accreditation, export-control checks and a legal review before moving to classified or operational data.

Sources

1. Regulation (EU) 2024/1689 (AI Act), Art. 2(3) and Art. 5 - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
2. Protocol Additional to the Geneva Conventions (1977), Art. 36, ICRC - https://ihl-databases.icrc.org/en/ihl-treaties/api-1977/article-36
3. EU Satellite Centre (SatCen) - https://www.satcen.europa.eu/
4. Copernicus Security Service / Copernicus Data Space Ecosystem - https://dataspace.copernicus.eu/
5. NATO - Summary of NATO's revised AI strategy / responsible use principles - https://www.nato.int/cps/en/natohq/official_texts_227237.htm
6. ICEYE - SAR satellite data - https://www.iceye.com/
7. Helsing - https://helsing.ai/
8. Safran.AI (formerly Preligens) - https://www.safran-group.com/
9. TorchGeo - https://github.com/microsoft/torchgeo
10. xView dataset - http://xviewdataset.org/