AI-powered terrain analysis for military planning: mobility corridors, observation and cover

Terrain has always determined what forces can do, where they can go, and who can see whom. That has not changed. What has changed is the volume of data available to analyze terrain — digital elevation models at 5-metre posting, multispectral satellite imagery updated daily, SAR passes that see through cloud — and the time available to process it. In a fast-moving operational environment, a terrain analysis that takes 12 hours to produce manually is often worth less than a 90-percent-complete one delivered in 90 minutes. AI does not replace the military geospatial analyst; it eliminates the computational drudgery that consumes most of an analyst's time, freeing that judgment for the questions the algorithm cannot answer. This article walks the full AI-assisted terrain analysis pipeline, from DEM ingestion to finished JIPB software tools output, covering each of the five OAKOC terrain factors.

Terrain analysis in the JIPB process — the four terrain OAKOC factors

The JIPB process structures terrain analysis around five factors remembered by the acronym OAKOC: Observation and fields of fire, Avenues of approach, Key terrain, Obstacles, and Cover and concealment. Together these five factors describe what the terrain permits and denies — where forces can move, what they can see and shoot, where they can hide, and what ground is worth fighting for. The intelligence preparation product that maps these factors forms the foundation of course-of-action development and synchronization planning.

In traditional manual analysis, a trained terrain analyst works through each factor sequentially: overlaying 1:50,000 topographic maps with slope overlays derived from DTED, manually tracing potential avenues of approach along ridgelines and valley floors, annotating vegetation density from imagery for concealment, and computing observation coverage by line-of-sight sketches from a handful of prominent terrain positions. For a 20 km × 20 km area of operations, this process can take a 2–3 person team 8–12 hours to complete to the standard expected for a division or brigade operational order. The product is then largely static — updated only when a new analyst sits down to revise it, not when conditions change.

AI-assisted analysis restructures this workflow. DEM processing, cost-surface computation, viewshed calculation, and image classification are all highly parallelizable operations well-suited to GPU acceleration. The analyst defines the area of operations, specifies the platform types and time window, and the system produces draft OAKOC overlays in minutes. The analyst's role shifts to validation, military judgment, and narrative — assessing whether the AI-identified corridors reflect ground reality, whether the key terrain rankings match the commander's scheme of maneuver, and whether the concealment classification accounts for seasonal vegetation changes not visible in the imagery. This is a more productive use of scarce analytical expertise than manually computing slopes.

DEM processing for mobility analysis

Every quantitative terrain analysis product originates in the digital elevation model. The DEM is the foundational data layer from which slope, aspect, curvature, drainage, and line-of-sight are all derived. DEM selection is therefore the first engineering decision, and it directly controls the spatial resolution of the analysis.

The most widely available sources span a range of resolutions and accuracies. SRTM (Shuttle Radar Topography Mission) provides near-global coverage at 30-metre posting with vertical accuracy of roughly 6–10 metres RMSE — adequate for identifying major ridgelines, broad valley floors, and macro-scale avenues of approach, but insufficient for resolving gullies, ditches, and embankments that are vehicle-scale obstacles. IfSAR products from commercial providers cover large areas at 5-metre posting and 1–2 metre vertical accuracy; at this resolution, field fortifications, anti-vehicle ditches, and terrain microfeatures begin to resolve. Airborne lidar at 0.5-metre point density, where available, resolves individual tree rows and shallow water obstacles. Military DTED (Digital Terrain Elevation Data) at Level 1 through Level 5 parallels these tiers and remains a standard data source for many planning systems.

From the DEM, the pipeline derives several primary products for mobility analysis. Slope in percent grade is the most important single input: slopes below 5 percent are generally trafficable by all vehicle types; slopes of 15–30 percent are trafficable by tracked vehicles but marginal or non-trafficable for wheeled; slopes above 30 percent deny most military vehicles and force dismounted movement. Aspect (slope direction) is secondary but matters for drainage and solar exposure. Curvature identifies terrain shapes — ridgelines, valley floors, saddles, cliff edges — that are militarily significant as channelizing or controlling features. The topographic wetness index, derived from slope and catchment area, serves as a proxy for soil moisture accumulation that modulates trafficability.

Soil trafficability is the DEM-derived analysis factor most sensitive to dynamic conditions. The static trafficability floor is set by soil texture: sandy or gravelly soils maintain bearing capacity even when wet; clay soils lose bearing capacity rapidly with moisture. Standard military trafficability models (based on cone index) use soil texture class from HSWD or national survey data to assign baseline trafficability by vehicle class. Dynamic adjustment requires a moisture layer — typically derived from SAR backscatter intensity (C-band Sentinel-1 or similar), which is sensitive to surface soil moisture — updated after precipitation events. Integrating static soil texture with dynamic SAR-derived moisture produces a trafficability surface that reflects current conditions rather than a static seasonal assumption.

AI-based mobility corridor extraction

With slope, land cover, and soil trafficability assembled into a cost surface, the mobility corridor extraction step finds the paths a specific vehicle type would most plausibly take between defined origin and destination zones. This is a routing problem on a weighted raster graph, and it is the step where AI contributes the most qualitative improvement over traditional analysis.

The cost surface assigns each grid cell a traversal penalty based on the platform-specific inputs: slope (penalties escalate sharply at platform-specific threshold grades), vegetation density (forests penalize wheeled vehicles and reduce tracked vehicle speed), land cover (roads and tracks reduce cost; water bodies, urban rubble, and dense vegetation increase it), and soil trafficability (clay soils after rain approach impassable). The resulting surface is a float-valued raster where low values represent easy movement and high values represent difficult or denied terrain. A graph is constructed by connecting each cell to its 8 neighbors with edge weights derived from the cost values of the two cells, and Dijkstra's algorithm (or A* with a Euclidean heuristic) finds the minimum-cost path between origin and destination zones.

Running the router once produces a single optimal path — the mathematical ideal, which may not correspond to realistic tactical movement. The AI enhancement is to generate an ensemble of paths by perturbing the cost surface stochastically, applying random cost multipliers to different terrain zones, and re-running the router each time. This produces hundreds of candidate paths that cluster into a small number of distinct corridors separated by natural terrain features. Machine learning — in practice, a clustering algorithm applied to path geometries — groups the candidate paths and selects representative centerlines, producing 3–6 distinct avenues of approach that bracket the realistic movement space rather than a single idealized route.

Complementing the routing stage, ML-based obstacle classification from satellite imagery identifies man-made obstacles — anti-vehicle ditches, berms, wire obstacles, minefields indicated by surface disturbance patterns — that the DEM does not capture. A convolutional classifier trained on labeled aerial and satellite imagery flags linear features and regular excavations that are inconsistent with natural terrain morphology. These detections are added to the cost surface as high-penalty zones, ensuring that AI-identified corridors route around known prepared obstacles. The same computer vision defense systems pipeline used for obstacle detection also feeds the cover and concealment classification step.

Cover and concealment classification

Cover and concealment are distinct military concepts that require separate analytical treatments. Cover provides physical protection against weapons effects — masonry walls, earthen berms, terrain folds, reverse slopes. Concealment hides a force from observation without necessarily protecting it from fire — forest canopy, shadows, urban alleys, low-light conditions. An AI terrain analysis system maps both, producing separate layers that planners can overlay independently.

Concealment classification is primarily an image analysis problem. Multispectral satellite imagery — at minimum 4-band (blue, green, red, near-infrared) at 3-metre or better resolution — provides the spectral signatures needed to distinguish bare earth from sparse shrub from dense forest canopy. A semantic segmentation model (typically a U-Net or DeepLabV3+ architecture fine-tuned on labeled military terrain imagery) produces a land-cover map that distinguishes open ground, low scrub (knee-high to waist-high), tall grass, sparse tree cover, dense forest, orchard, built-up area, and water. Each class receives a concealment probability score based on its ability to hide personnel, vehicles, and equipment from direct visual observation and from multispectral sensors.

Shadow analysis adds a time-sensitive concealment layer. Using the DEM and known solar geometry for the planned date and time of operations, the pipeline computes which terrain cells and structure footprints are in shadow at the critical movement time. Shadows cast by ridgelines, forest edges, and buildings create corridors of reduced visual and infrared observability that are significant for dismounted and light vehicle movement. The shadow layer is generated for several time windows — pre-dawn, dawn, midday, dusk — so planners can select movement timing to maximize concealment from overhead observation.

Structure extraction from high-resolution imagery provides the built-up-area component of both cover and concealment classification. A building-footprint detection model identifies rooftops, walls, and compound boundaries. Wall material inference — distinguishing masonry and concrete from sheet metal and wood based on spectral and textural signatures — produces a ballistic cover estimate for urban positions. Dense urban terrain with masonry construction rates as high cover and high concealment; open industrial areas with metal-clad warehouses rate as high concealment but low cover. The change detection satellite imagery pipeline monitors for new construction or destruction that changes the standing cover classification.

Observation and fields of fire analysis

Observation analysis answers the question: from this position, what can be seen? Fields of fire analysis answers the related but distinct question: from this position, what weapon systems can engage, and to what range? Both depend on viewshed computation — the determination of which terrain cells are visible from a given observer location — but they differ in the observer height, target height, and effective range used in the calculation.

Viewshed computation is straightforward in principle: for each candidate observer position, cast lines of sight outward in all directions across the DEM and mark each cell as visible if no intervening terrain block the direct line, or masked if terrain interposes between observer and target. The computational challenge is that a single viewshed over a 20 km × 20 km area at 5-metre posting involves millions of line-of-sight tests, and military terrain analysis requires viewsheds from hundreds or thousands of candidate positions. GPU-accelerated viewshed computation — implemented as a parallel ray-casting kernel — reduces what would take hours on a single CPU core to seconds on a modest GPU, making comprehensive multi-position analysis practical.

The pipeline runs viewsheds from a systematic grid of candidate observation points at observer heights typical for dismounted soldiers (1.5 m), vehicle-mounted optics (3–4 m), and elevated structures (10–30 m). Each viewshed is computed to the maximum relevant range for the observation task — 3–5 km for visual observation, 10–15 km for thermal or electro-optical sensors, the weapon's maximum effective range for fields of fire. Aggregating all viewsheds produces a cumulative observation coverage map: a raster where each cell value represents the fraction of candidate positions from which it is visible. Cells with high coverage are well-observed; cells with low coverage are in defilade.

Defilade position identification — automatically finding locations that provide protection from observation and direct fire — inverts the coverage map. An automated search identifies positions that are in defilade from adversary observation zones (the set of cells where an adversary observer could plausibly be), maximally concealed from overhead sensors, but with their own good observation coverage of the avenue of approach they are tasked to cover. This automated search produces a shortlist of candidate fighting positions that a human analyst evaluates for tactical and logistic feasibility. The line-of-sight matrix — a pairwise visibility table for a defined set of positions on both sides — supports fire planning by identifying which friendly positions can mutually support one another and which adversary positions can be engaged by direct fire from which friendly locations.

Key terrain identification

Key terrain is defined operationally, not geometrically. A hilltop that dominates the surrounding ground and overlooks three avenues of approach is key terrain in one scheme of maneuver and an irrelevant sideshow in another if the commander is bypassing the area entirely. This context-dependence is why AI key terrain identification produces candidate lists for analyst review rather than definitive designations — the algorithm can evaluate terrain attributes, but only the human planner understands the scheme of maneuver that determines what matters.

The terrain dominance scoring algorithm evaluates each candidate position across four quantitative dimensions. Observation dominance is the cumulative viewshed score over the area of operations — what fraction of the operational area is visible from this position, weighted by the military significance of the visible ground. Avenue control counts how many of the AI-extracted mobility corridors pass within the direct-fire range of this position, treating choke points and corridor intersections as multipliers. Accessibility scores how difficult this position is to approach under fire — positions that are only reachable through terrain already dominated by the position itself score higher. Bypass difficulty estimates how easily an adversary could avoid or neutralize the position by flanking or masking.

Positions that score highly across all four dimensions — visible to a large fraction of the area of operations, controlling multiple avenues of approach, difficult to approach and bypass — are flagged as tier-1 key terrain candidates. Positions scoring highly on subset of dimensions are flagged as tier-2, with the controlling dimension noted (observation-dominant, avenue-controlling, etc.). The analyst receives a ranked list with the supporting score breakdown and an overlay showing each candidate's viewshed and the corridors it controls. Comparison of control points — identifying which positions cover the most corridor intersection or choke points — allows direct comparison between candidates that might score similarly overall but differ in which threats they address.

Integration with C2 planning tools

Terrain analysis products have no value sitting in a GIS workstation. They must reach the planners and commanders who act on them, in the formats and at the update cadence that operational tempo demands. Integration with C2 planning tools is therefore not an afterthought but a core engineering requirement of the AI terrain analysis system.

The output layer types span several geospatial formats. Mobility corridor products are published as vector polygons — corridor centerlines with width buffers, choke point markers, and obstacle annotations — in GeoJSON or SHP format compatible with any GIS-capable C2 system. Cover and concealment rasters are exported as GeoTIFF with defined classification schemes and symbology so they render correctly in military GIS displays. Observation coverage maps export similarly as classified GeoTIFFs. Key terrain designations export as point or polygon features with attribute tables containing the scoring data. For ATAK-compatible C2 environments, all products can be packaged as CoT events or KML/KMZ files that load into an ATAK EUD without any additional conversion step.

Direct integration with JIPB product templates goes further. An API connection between the terrain analysis pipeline and the JIPB documentation system enables automated population of terrain analysis product sections: mobility corridor tables, key terrain nominations, observation coverage summaries, and cover and concealment classification maps are inserted into the JIPB product template with the analyst's selection and annotations, reducing the product generation step to review and approval rather than manual transcription from analysis tools to document.

Update cadence is governed by imagery ingestion events. The terrain analysis system monitors the imagery archive for new satellite or UAV passes over the registered area of operations. When a new pass arrives, the pipeline automatically re-runs the imagery-derived layers — cover and concealment classification, obstacle detection, structure extraction — compares the results against the standing terrain picture, and flags cells where the classification has changed above a defined threshold. The change flag surfaces in the C2 display as a terrain update alert, prompting the terrain analyst to review the affected areas and decide whether to update the published overlays. DEM-derived products (slope, viewshed, corridor routing) are stable unless new elevation data is ingested and do not require re-computation with each imagery pass. This update architecture keeps the operational terrain picture current without requiring an analyst to manually re-run the full analysis each time new imagery arrives.

The full AI terrain analysis workflow — from DEM ingestion through OAKOC factor derivation to C2 overlay publication — represents one of the most mature applications of machine learning in military geospatial intelligence. DEM processing and viewshed computation are established algorithms; the AI contributions are in cost-surface optimization, image-based obstacle and land-cover classification, and the automated scoring and ranking that prioritize the analyst's attention. The boundary between what the algorithm does well and what requires human military judgment is clearly defined: the algorithm handles computation at scale, and the analyst handles context, scheme-of-maneuver alignment, and final authority over what reaches the commander's planning product.