WarSpotting: AI-Driven Conflict Tracking
- WarSpotting is an AI-assisted, human-in-the-loop OSINT approach that detects, identifies, and tracks conflict activity from noisy visual and sensor inputs.
- It leverages computer vision models like YOLOv8x and UAV tracking to process diverse data sources, including social-media images and low-FPS videos, for military intelligence.
- The research addresses challenges such as data heterogeneity, adversarial camouflage, and ethical concerns while advancing operational and humanitarian insights.
WarSpotting denotes an AI-assisted, human-in-the-loop form of open-source intelligence in which conflict-generated data are processed to detect, identify, track, and interpret warfare-related activity. In its foundational formulation, the task is to use computer vision on social-media imagery to identify specific weapon systems and the insignias of the armed groups using them, then aggregate those detections into signals about proliferation, usage, and conflict dynamics, initially with the Russia–Ukraine war as testbed (Abedin et al., 2023). Related work uses the same operational logic for low-FPS UAV tracking of military vehicles, visually confirmed loss repositories, RF emitter localization, passive radar surveillance, and war-scene mobility mapping, which suggests a broader technical family rather than a single model class (Kostiv et al., 7 Jul 2025).
1. Definition, scope, and strategic rationale
The core WarSpotting problem arises from the scale and heterogeneity of contemporary conflict documentation. Photos and videos posted to Telegram, Twitter/X, Facebook, Instagram, and YouTube frequently show specific weapon systems, combatants, and unit insignias, but manual analysis does not scale to the volume of available material and exposes analysts to violent content (Abedin et al., 2023). The foundational proposal is therefore an AI-assisted pipeline that detects weapons and insignias, links them to armed groups, and aggregates the resulting evidence to estimate who is using which systems, where, and when (Abedin et al., 2023).
Ukraine is treated as an especially suitable testbed because it is described as “the most documented war in history,” because Ukrainian authorities encouraged social-media use for international support and counter-disinformation, and because the war combines dense online documentation, multiple state and paramilitary actors, and large inflows of foreign weapons (Abedin et al., 2023). The stated motivations are not only operational intelligence but also humanitarian ones: documentation of weapons proliferation, safer de-mining, mapping non-state armed groups’ reach, and improved real-time understanding of where humanitarian and medical aid is needed (Abedin et al., 2023).
The concept subsequently broadens. In low-FPS UAV video, WarSpotting becomes persistent identity maintenance for tanks, trucks, and other military vehicles under moving-camera, compressed, combat-like conditions (Kostiv et al., 7 Jul 2025). In forecasting studies, “WarSpotting” also denotes the visually confirmed OSINT repository used to model Russian equipment losses over time (Teagan, 9 Sep 2025). Together, these variants suggest a layered notion of WarSpotting: first detect militarily salient entities; then associate them with actors, locations, or tracks; then aggregate them into operational or humanitarian intelligence products.
2. Data regimes and conflict-generated corpora
WarSpotting research draws on several distinct data regimes. The original social-media vision study assembled 12,000 images from aggregator accounts focused on foreign weapons in Ukraine, with a labeled subset of 567 images annotated by bounding boxes and classes for weapons and insignias (Abedin et al., 2023). These images are deliberately noisy: low resolution, motion blur, occlusion, poor lighting, varied camera distance, camouflage, and clutter all occur frequently (Abedin et al., 2023).
The low-FPS UAV tracking benchmark contributes a different regime: 119 videos, 18,421 frames, and 58,897 bounding boxes collected from public Telegram, X, and YouTube sources, with resolutions from 480p to 4K and a mix of winter and summer conditions (Kostiv et al., 7 Jul 2025). Its classes are coarser than the social-media detector’s taxonomy—Heavily Armored, Lightly Armored / Unarmored, and Trucks—but the benchmark is tailored to combat-relevant convoy and maneuver monitoring rather than object taxonomic detail (Kostiv et al., 7 Jul 2025).
The WarSpotting repository used for forecasting is an equipment-loss ledger built from publicly available images and videos, with entries vetted to confirm authenticity and avoid duplication (Teagan, 9 Sep 2025). In that study, the consolidated series spans February 2022 through early July 2025, with daily granularity at source and monthly aggregation for several models; June 2025 is treated as partially reported and excluded from training and evaluation, and in one ARIMA analysis June and July 2025 are both excluded (Teagan, 9 Sep 2025). The same paper emphasizes that WarSpotting is structurally a lower bound: it records only visually confirmed losses, not all losses (Teagan, 9 Sep 2025).
Other corpora show how broadly the data model can be generalized. “Reviewing War” analyzes 7,330 Tripadvisor posts, 2,200,368 Google Maps reviews, and 3,963,981 VK posts to study how review platforms and a domestic social network became side channels for conflict information under censorship (Moreno et al., 2023). WarNav contributes 5,554 daylight war-scene images from the DATTALION repository, with 5,354 unannotated training images and only 200 annotated validation and test images, for semantic segmentation of navigable versus non-navigable regions in damaged environments (Graviers et al., 19 Nov 2025).
3. Visual detection, actor association, and persistent tracking
The first WarSpotting prototype is an object detector over social-media still images. Its input is a conflict image; its output is a set of bounding boxes and class labels such as Javelin anti-tank missile, landmine, grenade, rocket-propelled grenade, and insignia (Abedin et al., 2023). The principal model is YOLOv8x, whose backbone is a modified CSPDarknet53 with a C2f module and SPPF, whose neck is PAN-FPN, and whose head is a decoupled, anchor-free design with self-attention (Abedin et al., 2023). The training objective combines Varifocal Loss for classification, Distribution Focal Loss for box regression, and Complete IoU loss for localization (Abedin et al., 2023).
Evaluation uses standard detection metrics. With
and per-class average precision defined as area under the precision–recall curve, the overall metric is
at an IoU threshold of $0.5$ (Abedin et al., 2023). On the curated Ukraine dataset, YOLOv8x is reported as the strongest detector among the tested baselines.
| Model | [email protected] | Precision / Recall |
|---|---|---|
| YOLOv5x | 83.90 | 0.87 / 0.75 |
| YOLOv6-L6 | 85.39 | 0.90 / 0.77 |
| EfficientDet-D7x | 86.09 | 0.91 / 0.80 |
| YOLOv7-E6E | 89.59 | 0.90 / 0.86 |
| YOLOv8x | 92.91 | 0.96 / 0.91 |
A distinctive aspect of the system is that insignias are detected jointly with weapons. Post-processing can then associate a weapon detection with an insignia detection in the same frame, yielding the crucial actor–equipment linkage: not merely that a weapon appears, but that an armed group marked by a particular patch or badge appears with it (Abedin et al., 2023). The paper treats this as the step that moves WarSpotting beyond object counting toward attribution.
Persistent tracking in UAV video requires a different architecture. The low-FPS vehicle tracker uses instance association from single-frame annotations rather than densely labeled tracks, with a ResNet-50 backbone, ROI-Align region features, MASA-style self-supervised contrastive learning, and an auxiliary object-prior distillation branch (Kostiv et al., 7 Jul 2025). A central empirical claim is that global scene features are critical: road geometry, formation structure, and surrounding terrain stabilize association when IoU and simple motion models fail because the camera moves aggressively and the frame rate is effectively about $5$ FPS (Kostiv et al., 7 Jul 2025). With ground-truth single-frame labels, the reported benchmark numbers are HOTA 48.29, MOTA 51.38, IDF1 56.93, AssA 51.80, and 471 identity switches, and performance degrades only modestly when inference is reduced to resolution and 32-dimensional embeddings (Kostiv et al., 7 Jul 2025).
4. From detections to time series, forecasts, and side-channel signals
The foundational WarSpotting paper explicitly proposes an aggregation layer in which image-level detections become daily counts of Javelins, landmines, RPGs, insignias, and related entities, to be compared against battle fatalities, territorial changes, or the introduction of new weapon systems (Abedin et al., 2023). It discusses Pearson correlation and lag analysis as natural tools for that stage, but does not implement them; the contribution remains at validated object detection plus the argument that temporal integration should be the next step (Abedin et al., 2023).
That next step is taken, in a different form, by the forecasting study based on WarSpotting’s verified loss repository. There the target series are daily and monthly counts of total equipment losses and tank losses, modeled with ARIMA, Prophet, LSTM, TCN, and XGBoost (Teagan, 9 Sep 2025). The study emphasizes that deep learning models, particularly TCN and LSTM, produce stable and consistent forecasts under high temporal granularity, while Prophet provides interpretable daily trends with weekly seasonality and ARIMA offers a conservative baseline on monthly tank losses (Teagan, 9 Sep 2025). Across models, the reported qualitative result is stabilization or gradual decline in confirmed Russian equipment losses through late 2025, with tanks broadly in a plateau regime and all-equipment losses in a slowly declining daily range in the Prophet forecasts (Teagan, 9 Sep 2025).
The same paper formalizes why such outputs must be interpreted cautiously. WarSpotting counts only visually confirmed losses, suffers temporal lag between event and confirmation, exhibits spatial and platform bias, can misclassify similar systems, and lacks denominator information on inventories, reinforcements, and production (Teagan, 9 Sep 2025). Forecasts therefore project the time series of confirmed losses, not the full latent attrition process (Teagan, 9 Sep 2025).
A different kind of aggregation appears in the study of review platforms and VK as wartime side channels. There, conflict-related signals are extracted from Tripadvisor, Google Maps, and VK by a multilingual keyword-scoring classifier calibrated for high precision (Moreno et al., 2023). The reported overall metrics are accuracy 0.8752, precision 0.9773, recall 0.7588, and F1 0.8543, with weekly spikes in war-related content after February 2022 and strong evidence of platform moderation effects, especially on Google Maps (Moreno et al., 2023). This broadens WarSpotting from visual documentation to the monitoring of conflict communication traces embedded in ostensibly non-political platforms.
5. Extended sensing across RF, passive radar, and ground mobility
WarSpotting is not limited to optical OSINT. In electronic-warfare settings, a cooperative UAV wireless sensor network can localize an emission source in an urban environment by spectrum sensing and the signal Doppler frequency method (Kelner, 2018). Each UAV measures Doppler shift, computes an SDF-based source estimate from its own trajectory, and shares both the estimate and received power; fusion then uses only LOS/OLOS estimates when available, falling back to arithmetic averaging when all UAVs are in NLOS (Kelner, 2018). In the reported simulations, mean localization errors for five individual UAVs range from 81 m to 207 m, simple averaging gives about 108 m, and LOS/OLOS-aware fusion reduces the mean error to about 38 m (Kelner, 2018).
Passive multistatic radar extends the same logic to wide-area aerial surveillance. SkyWatch uses commercial FM broadcast transmitters as illuminators of opportunity and geographically distributed receivers to estimate the 3D position and velocity of aerial objects (Randall et al., 2023). The system supports direct reference, remote reference, and hybrid modes; delay–Doppler measurements are sent to a central server, where triangulation and an Extended Kalman Filter are used for tracking (Randall et al., 2023). Its stated envelope is altitudes up to 80 km, horizontal distances up to 150 km, and velocities up to km/s (Randall et al., 2023). The paper is framed around UAP detection, but the sensing architecture is directly relevant to passive battlespace awareness.
An actively optimized battlefield variant appears in the aerial RIS-assisted ISAC model. There, a full-duplex access point, a UAV-mounted reconfigurable intelligent surface, multiple users, and a sensing target share a joint communication-and-sensing waveform (Cho et al., 2024). Deep Deterministic Policy Gradient is used to optimize the ARIS trajectory, while transmit beamforming and RIS phase shifts are jointly adjusted under SINR constraints to minimize a target-localization CRB-based objective (Cho et al., 2024). The reported result is that the proposed technique outperforms conventional benchmark schemes by suppressing self-interference and clutter echo signals or optimizing RIS phase shifts (Cho et al., 2024).
On the ground, WarNav reframes WarSpotting as semantic segmentation for autonomous mobility in damaged war scenes. Its label space is navigability-oriented—road, drivable, pedestrian, vehicle, background, with overlay ignored—and evaluation uses a weighted mIoU that emphasizes near-field and region interiors (Graviers et al., 19 Nov 2025). Because the training split is entirely unannotated, the benchmark targets “navigability without annotation” by transfer from Cityscapes, RUGD, and Earthquake (Graviers et al., 19 Nov 2025). The strongest zero-shot baseline trains SegFormer on the combined source datasets and reaches 64.9% weighted mIoU on WarNav test and 63.9% on validation (Graviers et al., 19 Nov 2025).
6. Reliability, adversarial adaptation, and governance
The main technical limitations recur across the literature. The original image detector is trained on only 567 labeled images, covers four weapon categories plus insignias, and is evaluated by internal model comparison rather than external validation against conflict datasets such as ACLED or UCDP (Abedin et al., 2023). The forecasting study reiterates structural undercounting, temporal lag, spatial bias, platform volatility, and the absence of inventory denominators, all of which constrain interpretability (Teagan, 9 Sep 2025). In both cases, the gap between proof-of-concept and operational deployment is acknowledged rather than concealed.
WarSpotting systems are also vulnerable to deliberate deception. In aerial imagery, patch-based adversarial camouflage can sharply reduce object-detector performance against large military assets (Adhikari et al., 2020). On DOTA airplane detection with YOLOv2, a large on-top adversarial patch reduces AP to 5.6% compared with 52.3% for an equally sized random-noise patch; a small on-top patch reduces AP to 37.8% compared with 94.0% for the corresponding noise baseline; and a large side patch still degrades AP to 83.3% versus 99.0% for noise (Adhikari et al., 2020). The paper therefore argues that adversarial patch attacks are a realistic alternative to traditional camouflage and should be considered in automated aerial-surveillance analysis (Adhikari et al., 2020).
Ethical and legal concerns are not peripheral. The original WarSpotting paper emphasizes risks to civilians and combatants, privacy and dignity, dual-use misuse, and psychological harm to analysts exposed to graphic imagery, and argues for transparency, safeguards, governance, and explicitly human-supervised deployment rather than unsupervised black-box use (Abedin et al., 2023). WarNav’s construction process adds GDPR compliance, encrypted storage, and the removal of images with blood, cadavers, or partial blurring to avoid both ethical harms and learning artifacts (Graviers et al., 19 Nov 2025). The review-platform study likewise avoids multimedia collection, hashes usernames, and rejects deanonymization, reflecting the risk that monitoring conflict side channels can itself become a surveillance instrument (Moreno et al., 2023).
Taken together, these constraints define WarSpotting less as a finished system than as a technically diverse research program. Its current state includes strong object detection on noisy conflict imagery, realistic military-vehicle tracking under harsh UAV conditions, verified-loss forecasting, RF emitter localization, passive radar surveillance, and war-scene mobility segmentation. Its unresolved problems are equally clear: larger and more representative datasets, external validation, robust multimodal fusion, protection against adversarial concealment, and governance regimes capable of handling a dual-use technology whose operational value and ethical risk are both unusually high.
