What makes WD-CD genuinely novel is not just its scale but its annotation depth. Existing change detection datasets fall into three broad camps, and each falls short of what war scenarios demand. Binary change detection datasets such as LEVIR-CD and SYSU-CD provide only change or no-change masks, with no semantic information about what changed. Transitional semantic change detection datasets such as SECOND and Hi-UCD label category transitions, for instance farmland becoming urban, but they cannot encode a building shifting from intact to damaged, because the category itself never changes. Damage-oriented datasets such as xBD, the best-known benchmark in this space, do label damage states, but they were built for natural disasters like earthquakes, hurricanes, and tsunamis, and they focus almost exclusively on buildings. War damage is a fundamentally different problem: strikes produce localized, point-like destruction whose appearance varies sharply with weapon type and impact angle, and the targets extend far beyond buildings to airports, ports, farmland, aircraft, and vessels.

To capture that complexity, the team designed a hierarchical labeling schema with 16 fine-grained target categories grouped into four macro-categories: surface base, fixed targets, mobile targets, and stateless targets. Each category uses a subset of six candidate change states: no-change, damaged, reduced, added, extended, and replaced. Operationally, damaged means visible physical deterioration while the target remains identifiable, reduced means a decrease in spatial extent or instance count, added means a new spatially separate target appears, extended means a contiguous outward expansion, and replaced applies to mobile targets when one object disappears and another occupies its location. Across all unique category and state combinations plus background, WD-CD carries 69 distinct label classes, far richer than any prior damage benchmark.

The imagery itself comes from World Imagery Wayback, a publicly accessible archive of historical high-resolution satellite imagery. The researchers identified severely affected war zones using the Uppsala Conflict Data Program and other open-source intelligence, then compiled verified damage-event locations and selected pre- and post-event acquisitions bracketing each event. Across the 66 retained parent scene pairs, temporal separation ranges from roughly two months to three years, and each pair was manually screened to ensure the damage-related ground changes were clearly observable. Image pairs were automatically co-registered using embedded geospatial metadata parsed with ArcGIS Pro and the GDAL library.

Annotation quality received equally rigorous treatment. Candidate annotators first qualified on a calibration set of 100 patch pairs, needing a mean intersection-over-union of at least 0.85 against expert boundaries for continuous targets and a Cohen’s kappa of at least 0.80 for category-state consistency on discrete targets. Production annotation ran in two-person teams with continuous mutual cross-checking, followed by a closed-loop review in which each team inspected another team’s output. The researchers also trained the CD-Lamba model with five-fold cross-validation to flag samples whose predicted probability for the assigned class fell below 0.50, using the Confident Learning framework purely as a screening mechanism with no automatic label modification. As a final step, an independent team inspected a stratified sample of roughly 20 percent of the dataset, accepting batches only when at least 90 percent of sampled patches contained correct category-state labels with no missing instances.

The splits follow a geographic independence principle borrowed from the xView2 challenge to prevent data leakage. Urban-rural scene pairs, tiled with overlapping sliding windows at a 25 percent overlap ratio determined experimentally, are partitioned at the scene level so training and test tiles never share geography. Airports and ports, tiled without overlap, use random sampling. All three scene categories divide into training, validation, and test sets at roughly 3:1:1, yielding 5,955 training, 1,573 validation, and 2,159 test patch pairs.

On the primary building-damage detection task, the researchers benchmarked seven representative deep learning models spanning convolutional, Transformer, and state space model architectures: DeepSiameseNet, DDLNet, STNet, BIT, ChangeFormer, ChangeMamba, and CD-Lamba. CD-Lamba led all metrics with an overall score of 79.53 percent, while the six advanced models achieved building localization F1 scores between 81.54 and 83.04 percent, well above the classical DeepSiameseNet baseline at 72.99 percent. Yet damage classification consistently lagged localization for every model, revealing that distinguishing damage states from intact structures remains the core bottleneck. The gap stems partly from severe class imbalance: in the Aleppo case cited by the authors, buildings assessed as destroyed by the United Nations Satellite Centre accounted for only 2.8 percent of the total, with damaged buildings below 15 percent, so war-damage samples form a small fraction easily overwhelmed by the majority class during feature learning.

Cross-dataset comparisons against xBD and three binary change detection benchmarks confirmed the difficulty stratification. Models scored 80 to 90 percent on binary change detection tasks but dropped well below those levels on building damage assessment for both xBD and WD-CD. WD-CD proved harder than xBD across all six advanced models, a difference the authors attribute to damage morphology: high-intensity natural disasters like earthquakes typically collapse buildings wholesale, whereas kinetic strikes leave facades relatively intact on large buildings while reducing small buildings to irregular fragments that are visually harder to separate from their surroundings.

Intercontinental transfer experiments exposed a striking asymmetry in what generalizes. Models trained on one continent and tested on another showed that building geometry, meaning edges and footprint shapes, transferred far more robustly than damage semantics. The European-trained model dropped only 14.39 and 9.66 percentage points on African and Asian test sets respectively, while African- and Asian-trained models fell by 64.45 and 46.05 percentage points on the European set. Localization declines were far smaller than damage-classification declines in most transfer directions, suggesting that rubble texture, color, and collapse patterns are tightly coupled to local environments while geometric cues are comparatively universal. The authors propose that future models decouple the geometric and semantic pathways, using transferable geometric priors for localization while learning damage representations that suppress region-varying background style.

Beyond benchmarking, WD-CD opens five downstream applications and four research directions. The multi-state building annotations jointly capture destruction and post-war development, enabling recovery monitoring alongside damage assessment. Airport pavement annotations support strike damage detection and emergency repair prioritization. Farmland state labels paired with crater instance masks enable unexploded ordnance contamination risk mapping grounded in the empirically established spatial correlation between crater distributions and subsurface ordnance. Aircraft and vessel state transitions support battlefield situational awareness and civilian infrastructure management alike. The dataset is released for non-commercial academic research upon request, with explicit ethical restrictions against operational military targeting or any application that could harm civilian populations, positioning WD-CD as a foundation for humanitarian response, post-conflict reconstruction, and the computer vision research needed to support both.

Subject of Research: A new large-scale satellite imagery benchmark called WD-CD enables artificial intelligence to detect and classify fine-grained war damage across five conflicts on three continents.

Article Title: WD-CD: a large-scale high-resolution optical remote sensing benchmark for fine-grained change detection in war scenarios

Article References: WD-CD: a large-scale high-resolution optical remote sensing benchmark for fine-grained change detection in war scenarios. (n.d.). https://doi.org/10.3389/frsen.2026.1930744

Image Credits: AI Generated

DOI: 10.3389/frsen.2026.1930744

Keywords: WD-CD, remote sensing, change detection, deep learning, satellite imagery, war damage, benchmark dataset, artificial intelligence, humanitarian response, conflict zones, computer vision, geospatial analysis