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

1 Introduction

Since the early 20th century, wars have inflicted widespread destruction on both civilian infrastructure and military installations worldwide. Rapid and accurate detection of war-induced damage is critical for humanitarian response and post-war reconstruction, yet the inaccessibility of active conflict zones renders traditional field surveys infeasible (Witmer, 2015). Remote sensing overcomes this constraint by enabling wall-to-wall observation from a distance, and bi-temporal high-resolution optical imagery has become the primary data source for change detection (CD), owing to the rich textural, spectral, and geometric information required to characterize ground-object changes (Shi et al., 2020). Building on this foundation, the rapid progress of deep learning has made automated, high-precision war-damage detection increasingly feasible (Zheng et al., 2021).

However, war damage presents interpretation challenges fundamentally different from those posed by the scenarios for which existing benchmarks were designed. Unlike natural disasters, which inflict large-scale destruction through relatively uniform mechanisms such as earthquakes and floods, armed strikes produce localized, point-like damage whose appearance varies sharply with weapon type and strike angle (Wang Y. et al., 2025). More critically, war damage extends well beyond buildings to a broad spectrum of heterogeneous targets, each of which may follow distinct change trajectories across multiple states. This multi-target, multi-state complexity is what makes CD in war scenarios particularly demanding.

Such complexity exceeds the design scope of existing CD datasets: binary change detection datasets carry no semantic labels, transitional semantic datasets target land-use transitions rather than conflict, and damage-oriented datasets such as xBD (Gupta et al., 2019) address only natural disasters, while the few war-specific datasets cover a single conflict or city and offer no means to evaluate cross-regional generalization. A large-scale benchmark with fine-grained pixel-level annotations spanning multiple target categories and change states, collected from geographically diverse war theaters, is therefore needed. To fill this gap, we introduce WD-CD, which, to the best of our knowledge, is the first large-scale, high-resolution bi-temporal optical remote sensing benchmark for war-damage change detection. WD-CD comprises 9,687 bi-temporal image pairs of 512 × 512 pixels collected from real war zones across Europe, Africa, and Asia, covering a total area of 1,055.83 km 2 and spanning five major wars: the Russia–Ukraine War, the Iraq War, the Syrian Civil War, the Myanmar Civil War, and the Sudanese Civil War (Figure 1). The main contributions of this paper are as follows:

• We introduce WD-CD, a large-scale high-resolution bi-temporal optical benchmark for war-damage change detection, providing pixel-level annotations across 16 target categories and up to six change states over five war theaters on three continents.
• We evaluate seven models on the primary building-damage task and in quantitative multi-target experiments, while using six of these models for cross-dataset comparisons to characterize WD-CD’s performance and task difficulty.
• Through intercontinental transfer experiments, we show that building geometry transfers robustly across regions, whereas damage semantics remain domain-specific, confirming WD-CD’s geographic diversity and yielding design insights for cross-domain generalization.
• We identify five downstream applications and present four potential research directions enabled by WD-CD, extending its utility to a broader range of scenarios.

The structure of this paper is as follows: Section 2 reviews related CD datasets and mainstream deep learning-based methods. Section 3 details the construction of WD-CD. Section 4 reports the benchmark evaluation, cross-dataset comparison, and intercontinental transfer analysis. Section 5 presents downstream applications and potential research directions. Section 6 discusses challenges and outlook. Finally, Section 7 concludes the paper.

2 Related work

With the widespread availability of remote sensing imagery and the rapid rise of deep learning, a large body of research has emerged on benchmark datasets and algorithms for CD. In this section, we review existing optical remote sensing CD datasets (Section 2.1), mainstream deep learning-based CD methods (Section 2.2), and prior work on remote sensing CD in war scenarios (Section 2.3).

2.1 Optical remote sensing change detection datasets

A number of public CD datasets based on optical remote sensing imagery have been released and widely used by the research community. According to task type, these datasets can be grouped into three categories: binary change detection (BCD) datasets, transitional semantic change detection (TSCD) datasets, and damage-oriented semantic change detection (DSCD) datasets. Table 1 summarizes the key statistics of representative datasets across all three categories.

TABLE 1. Overview of representative optical remote sensing change detection benchmarks across three categories: BCD, TSCD, and DSCD. The class count in the Classes column includes the background class.

For WD-CD, the class count consists of all unique category–state label combinations across the 16 target categories, plus one additional background class (see Figure 2).

2.1.1 Binary change detection datasets

BCD datasets represent the most fundamental data form in remote sensing CD. They provide binary change/no-change masks as supervision for identifying changed regions in multi-temporal imagery, without encoding semantic information about target categories or change states. Owing to their relatively low annotation cost and clearly defined task setting, BCD datasets have been widely adopted and have driven the development of a broad range of CD methods.

From the perspective of application targets, existing BCD datasets can be divided into two main categories. One focuses on building change detection, with representative examples including WHU-CD (Ji et al., 2018) and LEVIR-CD (Chen and Shi, 2020), which provide high spatial resolution and clear structural contours for learning stable geometric features. The other focuses on multiclass land-cover change, with representative datasets including SYSU-CD (Shi et al., 2021) and DSIFN-CD (Zhang et al., 2020), which introduce more complex spatial semantics and background interference, placing higher demands on model generalization.

Although BCD datasets have driven algorithmic advances in changed region localization and noise suppression, their binary supervisory signals cannot distinguish the semantic types of change or their evolutionary states. This representational limitation makes BCD datasets inherently insufficient for higher-level semantic understanding tasks, motivating the development of richer annotation schemes.

2.1.2 Transitional semantic change detection datasets

TSCD datasets extend CD beyond binary localization by introducing semantic annotations that characterize category transitions of ground objects before and after change, enabling a higher-level understanding of surface dynamics. They support structured modeling of land-object category evolution across multiple semantic classes.

Existing TSCD datasets span different spatial scales and resolutions. Low-to medium-resolution datasets such as Landsat-SCD (Yuan et al., 2022) are designed for large-area land-cover change analysis, capturing long-term surface evolution trends at the regional scale. High-resolution datasets such as HRSCD (Daudt et al., 2019), SECOND (Yang et al., 2022), Hi-UCD (Tian et al., 2022) and RB-SCD (Shu et al., 2026) focus on urban areas, covering diverse object types including buildings, vegetation, and water bodies, and providing rich multi-class semantic transition patterns for fine-grained scene analysis.

Although TSCD datasets offer substantially greater semantic expressiveness than BCD datasets, their annotation paradigm assumes that changes are fully captured by category-label transitions. This assumption holds well for land-use change and urban expansion scenarios, but breaks down when the task involves the evolution of internal target states. For instance, a building transitioning from intact to damaged does not involve a category change, yet carries critical structural damage information. TSCD datasets are therefore inherently insufficient for fine-grained tasks such as damage state detection.

2.1.3 Damage-oriented semantic change detection datasets

To meet the demands of post-disaster damage recognition and assessment, researchers have constructed DSCD datasets, which are designed with annotations centered on the damage states of target objects.

As shown in Table 1, existing DSCD datasets have been developed predominantly for natural-disaster scenarios, including earthquakes, floods, hurricanes, and tsunamis. A review (Wang L. et al., 2024) has shown that earthquakes, floods, and hurricanes are the most common disaster types in current DSCD studies, among which earthquakes account for the largest proportion. In terms of damage-label design, current DSCD tasks focus mainly on binary and four-level building-damage classification. The former emphasizes the rapid distinction between damaged and undamaged targets, such as in ABCD (Fujita et al., 2017) and Building Damage Dataset (Adriano et al., 2021), whereas the latter further refines the labels into no damage, minor damage, major damage, and destroyed, such as in xBD (Gupta et al., 2019) and Ida-BD (Kaur et al., 2023), but poses greater challenges for annotation and recognition.

Moreover, existing DSCD datasets focus almost exclusively on buildings. War scenarios, by contrast, involve multiple high-value target categories, including airport facilities, farmland, aircraft, and vessels, with damage patterns that differ substantially from those induced by natural disasters. Existing DSCD datasets are therefore insufficient to support the multi-target, multi-state change analysis required in complex war scenarios, highlighting the necessity of a dedicated large-scale benchmark for war-damage change detection.

2.2 Remote sensing change detection methods

Early deep learning approaches to CD relied primarily on convolutional neural networks (CNNs). The U-Net-based fully convolutional Siamese networks (Daudt et al., 2018) (FC-EF, FC-Siam-Conc, and FC-Siam-Diff) represented the foundational application of deep learning to this field. DeepSiameseNet (Yang et al., 2021) subsequently formulated CD as an end-to-end semantic-segmentation problem using Siamese feature extraction together with a high-level feature-fusion and decoding pipeline. Building on this line of research, dense skip connectivity was adopted to preserve high-resolution spatial details across scales (Fang et al., 2021). Attention mechanisms were subsequently incorporated to sharpen model sensitivity to changed regions, spanning the hierarchical attention of HANet (Han et al., 2023) and the multiple-attention design of MASNet (Huang et al., 2021). More recent CNN methods further refine local feature extraction and multi-temporal difference representation (Ma et al., 2023; 2024). Despite these advances, the limited receptive field of convolution operations constrains the modeling of long-range spatial dependencies, which is critical for interpreting complex structural changes in large-scale remote sensing scenes.

Architectures based on the Vision Transformer (Dosovitskiy et al., 2021) addressed this limitation through global self-attention. Transformer encoders were first applied to model long-range context over a compact set of high-level bi-temporal semantic tokens (Chen H. et al., 2021), and hierarchical encoder-decoder structures further enabled multi-scale change identification (Bandara and Patel, 2022). Weight-sharing Swin Transformer backbones were also explored for multi-level bi-temporal feature extraction (Zhang et al., 2022; Liu et al., 2021). To balance local and global modeling, hybrid CNN-Transformer methods such as ICIF-Net (Feng et al., 2022) and SARAS-Net (Chen et al., 2023) combined the complementary strengths of both architectures, achieving improved performance at the cost of increased computational overhead. However, the quadratic complexity of self-attention remains a bottleneck when processing high-resolution remote sensing imagery at scale (Wang R. et al., 2024).

More recently, methods based on state space models (SSMs) have attracted increasing attention for their linear computational complexity and effective long-range dependency modeling. ChangeMamba (Chen et al., 2024) pioneered SSM-based CD by proposing a spatiotemporal modeling mechanism with temporal cross-scanning strategies. RS-Mamba (Zhao S. et al., 2024) introduced an omnidirectional selective scan module tailored for the non-directional nature of remote sensing data. OS-Mamba (Han et al., 2025) further optimized the spatial receptive field on the basis of multi-directional scanning to improve damage-level discrimination. CD-Lamba (Wu et al., 2026) proposed a locally adaptive scanning mechanism to address the locality loss inherent in standard image-flattening approaches, substantially improving fine-grained change detection performance. CDMamba (Zhang H. et al., 2025) combined global Mamba modeling with local convolutional cues through a dedicated ScaleResidual ConvMamba module, further compensating for the limitations of SSM alone in dense prediction tasks. These methods based on SSMs represent a promising direction for balancing efficiency and accuracy in complex CD tasks.

2.3 Change detection in war scenarios

Remote sensing CD in war scenarios supports humanitarian assistance, human rights monitoring, and post-war reconstruction, yet ground-truth acquisition is severely constrained by the inaccessibility of conflict zones (Witmer, 2015). Early automated attempts combined hand-crafted features with contextual modeling: self-similarity descriptors were used to detect building destruction in the Gaza Strip (Kahraman et al., 2016), and shadow information together with the gray-level co-occurrence matrix supported quantitative damage estimation in Syrian war regions (Ghandour and Jezzini, 2018). The limited expressive capacity of such shallow features, however, leaves these approaches inadequate for the highly heterogeneous urban war-damage scenes encountered in practice.

With the rise of deep learning, CNNs and Transformers have been increasingly applied to war-related building damage detection, including a hybrid U-Net with multi-scale feature fusion for pixel-level extraction of destroyed buildings on a self-constructed Syrian dataset (Nabiee et al., 2022) and a YOLOv8-based detector for war-damaged buildings over Mariupol and Volnovakha (Wang Y. et al., 2025). Taken together, these studies illustrate that war-damage training data are often assembled for individual conflicts or a small number of locations. This fragmentation is compounded by restricted access to very-high-resolution imagery, which is commonly commercial, selectively released, or fee-based, thereby limiting its systematic use in scientific and humanitarian applications (Witmer, 2015; Bennett et al., 2022). These constraints have shaped a compromise-driven trend at both the data-source and method-design levels.

At the data-source level, a trade-off exists between spatial resolution and coverage. One line of work adopts freely available low-to medium-resolution satellite data: Sentinel-1 synthetic aperture radar (SAR) imagery has supported long-term damage monitoring in Raqqa (Braun, 2018) and Mariupol (Bachmann-Gigl and Dabiri, 2024); Sentinel-2 imagery has been used to identify military-induced damage to Ukrainian farmland (Kussul et al., 2023); and Sentinel-1/2 fusion has enabled destruction monitoring in Mosul (Fakhri and Gkanatsios, 2021) and Kyiv (Aimaiti et al., 2022). However, their 10-m spatial resolution limits their ability to capture fine-grained structural damage. Another line of work turns to high-resolution imagery for more detailed damage assessment. For example, the self-constructed dataset used in the cited Ukraine study covers Mariupol and Volnovakha (Wang Y. et al., 2025), illustrating the geographically localized scope of current high-resolution experiments. More generally, cross-geographic evaluations of building-damage models have documented performance degradation under geographic and structural distribution shifts (Kshirsagar et al., 2026). Consequently, geographically narrow datasets provide a limited basis for assessing cross-regional generalization and may encourage models to rely on site-specific architectural, geomorphological, or image-acquisition characteristics.

At the method-design level, the high cost of pixel-level annotation has motivated learning paradigms that reduce data dependence. Cross-domain transfer seeks to leverage prior knowledge from related domains. For example, a building-damage model pretrained on natural-disaster data has been transferred to a real post-war scene in Bucha, Ukraine (Sodeinde, 2024), and CNN models originally developed for natural-disaster building damage assessment have been tested on pre- and post-war imagery of Mariupol (Risso et al., 2025). However, the distribution shift between war-induced and natural-disaster damage patterns induces severe generalization bottlenecks, with models lacking target-domain supervision tending to overfit the “no-damage” category and lose sensitivity to damaged targets. Unsupervised paradigms have likewise been explored, such as using a pretrained denoising diffusion probabilistic model as a feature generator for unlabeled CD tasks (Bandara et al., 2025). Neither cross-domain transfer nor unsupervised feature mining, however, can substitute for high-quality, comprehensive target-domain supervision, the lack of which remains a primary factor limiting war-damage interpretation and directly motivates the large-scale benchmark introduced in this work.

3 The WD-CD dataset

3.1 Data collection and preparation

Data collection follows a conflict-driven geographic sampling strategy, prioritizing regions with verifiable and severe war-induced damage across diverse geomorphological and architectural contexts. We select World Imagery Wayback (Esri, 2018) as the core data source. Drawing on the Uppsala Conflict Data Program (Uppsala Conflict Data Program, 2026) alongside other open-source intelligence sources, we identify severely affected war zones in Europe, Africa, and Asia over the period 2015–2026. We then compile damage-event locations from five major wars: the Russia–Ukraine War, the Iraq War, the Syrian Civil War, the Myanmar Civil War, and the Sudanese Civil War (Figure 1). For each location, we identify a verified, site-specific damage event and select pre- and post-event acquisitions that bracket the relevant event within the 2015–2026 observation window, rather than the initial outbreak of the broader conflict. Across the 66 retained parent scene pairs, the temporal separation between the two acquisitions ranges from approximately 2 months to 3 years. Each pair is manually screened to ensure that the corresponding damage-related ground changes are clearly observable. Using ArcGIS Pro and the GDAL library, we parse the embedded geospatial metadata to automatically co-register each image pair. The spatial resolution of the selected imagery ranges from 0.3 to 1.2 m, depending on the satellite source and acquisition date for each region. After standardized spatial cropping, we obtain a collection of high-quality, large-scale bi-temporal image pairs ready for annotation.

3.2 Annotation procedure

The annotation procedure consists of three interdependent components: labeling schema design, systematic annotation implementation, and data annotation quality control, each addressing a distinct challenge in constructing a reliable large-scale war-damage benchmark.

3.2.1 Labeling schema design

To overcome the limitations of existing datasets in semantic granularity and to provide comprehensive coverage of change types in war scenarios, we construct a hierarchical multi-class labeling schema. We first group ground targets into four macro-categories according to their physical attributes: surface base, fixed targets, mobile targets, and stateless targets. Within this framework, we define 16 fine-grained categories covering the principal ground targets relevant to modern warfare. State labels are assigned according to the physical characteristics and applicable change patterns of each target category in real-world war scenarios. Except for stateless targets, each category uses a category-specific subset of six candidate states: “no-change”, “damaged”, “reduced”, “added”, “extended”, and “replaced” (Figure 2). Operationally, “no-change” denotes the absence of a discernible physical or spatial change in the target between the two acquisitions. “Damaged” denotes visible physical deterioration while the target remains identifiable, whereas “reduced” denotes a category-dependent decrease in the mapped spatial extent or instance count of the target. “Added” denotes the appearance of a new, spatially separate target or target region in the post-event image, whereas “extended” denotes a contiguous outward expansion of a pre-existing target. “Replaced” is used for applicable mobile-target categories when the pre-event target disappears and another target occupies the corresponding location in the post-event image. Borderline cases are independently reviewed by two annotators, and any unresolved disagreement regarding the target category, change state, or spatial boundary is adjudicated by a senior expert, whose decision is used as the final annotation.

3.2.2 Systematic annotation implementation

To construct this large-scale benchmark in a systematic and efficient manner, we develop a standardized three-stage data production pipeline (Figure 3). We first segment the original large-format war-zone imagery by scene type, partitioning it into three categories: Urban-Rural Areas, Airports, and Ports. We crop each scene by extending partition boundaries along linear geographic features such as roads and rivers, preventing individual instances from being split across multiple tiles and reducing the complexity of subsequent annotation and verification. Subsequently, following the approach of Chen D et al. (2021), we use geometric priors from publicly available geospatial data, primarily OpenStreetMap (OSM) (Haklay and Weber, 2008), as a warm-start basis for annotation. This strategy substantially reduces the labor cost of constructing object contours from scratch while preserving geometric regularity and semantic fidelity. Finally, we crop the annotated imagery into patches suitable for deep learning input. For Airports and Ports, images are tiled into non-overlapping 512 × 512-pixel patches. For Urban-Rural Areas, which contain a larger sample volume and richer change patterns, following the cropping strategy of Wang et al. (2025c), we apply overlapping sliding-window cropping. To determine the optimal overlap ratio, we conduct an optimization experiment on the Urban-Rural Areas subset using CD-Lamba (Wu et al., 2026) as a representative model (Table 2). Based on the experimental results, we use an overlap ratio of 25%.

TABLE 2. Patch-overlap ratio optimization results (see Section 4.1 for metric definitions).

The highest scores are in bold, and the second-highest scores are underlined. All scores are represented as percentages.

3.2.3 Data annotation quality control

To ensure annotation accuracy, consistency, and completeness, we establish a quality-control framework combining human–machine collaboration with multi-level verification (Figure 4). During standardized training, candidate annotators independently annotate a calibration set of 100 patch pairs stratified across scene types and target categories. For continuous targets, qualification requires a mean intersection-over-union (IoU) of at least 0.85 between the candidate and expert reference boundaries. For discrete targets, category–state consistency is evaluated using Cohen’s kappa, with a minimum qualification threshold of 0.80. Only candidates satisfying both applicable criteria are admitted to production, after which annotation is performed in two-person teams with continuous mutual cross-checking. Once the initial annotations are complete, a cross-group closed-loop review is conducted, in which each team inspects another team’s output for missing targets, boundary deviations, and category–state inconsistencies. To identify less evident errors, CD-Lamba is trained using five-fold cross-validation to obtain out-of-sample class probabilities. Following the Confident Learning framework (Northcutt et al., 2021), samples whose predicted probability for the assigned class is below 0.50 are flagged for manual re-examination. Confident Learning is used only as a screening mechanism, and no label is modified automatically. As a final quality-control step, an independent team inspects a stratified sample of approximately 20% of WD-CD. The inspected batch is accepted when at least 90% of the sampled patches contain correct category–state labels, no missing target instances, and target boundaries with an IoU of at least 0.85. All identified errors are corrected and independently rechecked before final acceptance.

3.3 Dataset statistics and splits

WD-CD comprises 9,687 annotated bi-temporal patch pairs derived from 66 large-format scene pairs, covering approximately 1,055.83 km 2 across 24 cities in five conflict-affected theaters on three continents. All source imagery was acquired between 2015 and 2026. These theaters are listed in Figure 1. Although hostilities in the Iraqi and Syrian theaters began before the observation window, the imagery records war-related damage associated with documented events during active phases within the 2015–2026 window. This multi-war, multi-continent composition sets WD-CD apart from existing DSCD benchmarks.

The 66 scene pairs in WD-CD are organized into three scene types based on their primary land-use context: Urban-Rural Areas, Airports, and Ports. Because Urban-Rural Areas in WD-CD are tiled using overlapping sliding-window cropping, strict geographic independence between splits is essential to prevent data leakage. Following the principle established in the xView2 challenge (Gupta et al., 2019), we adopt a stratified split strategy based on geographic independence. Specifically, the geographically isolated scene pairs of Urban-Rural Areas produced during the first-stage scene segmentation (Section 3.2) serve as the basic partitioning units, ensuring strict spatial isolation between training and test sets. For Airports and Ports, where tiles do not overlap, we apply random sampling for partitioning. Consistent with the convention of mainstream remote sensing benchmarks, all three scene categories are divided into training, validation, and test sets at an approximate ratio of 3:1:1. The resulting splits are summarized in Table 3.

TABLE 3. Training, validation, and test partitions of WD-CD by scene type.

At the pixel level, WD-CD exhibits a pronounced long-tailed distribution across its 16 fine-grained target categories (Figure 5). The figure summarizes pixel counts over the complete dataset, with “Building” and “Farmland” disaggregated by change state and the remaining categories aggregated across their applicable states for legibility. “Farmland_no-change”, “Building_no-change”, “Farmland_damaged”, and “Building_damaged” dominate the pixel budget, whereas sparse or physically compact targets such as “Vehicle Revetment” and “Tower” each contribute fewer than 1 0 6 annotated pixels, more than two orders of magnitude below the dominant classes. For the quantitative experiments in this study, we construct standalone benchmark tasks only for label groupings with sufficient representation across the training, validation, and test partitions: the grouped building-damage task; the surface-base tasks involving runway, taxiway, apron, and farmland; and the aircraft and vessel tasks, for which the “no-change”, “reduced”, and “added” states are reported separately. Category–state labels not included in a given task, including extremely rare combinations, are consistently remapped to the background class during training, validation, and testing. Their original annotations remain available in WD-CD for semantic completeness, but they are not evaluated as independent classes because their limited pixel or instance counts would make class-wise F 1 or IoU estimates too sensitive to individual samples and dataset partitioning to support stable class-wise benchmarking.

3.4 Data availability and ethical considerations

All imagery in WD-CD is derived from World Imagery Wayback (Esri, 2018), a publicly accessible archive of historical high-resolution satellite imagery, and is used within the scope permitted by its academic and non-commercial research terms. WD-CD is released to support humanitarian response, post-conflict reconstruction, and academic research, and must not be used for operational military targeting or any application that could harm civilian populations. Upon acceptance, the full dataset will be made available by the authors on request for non-commercial academic use.

4 Experiments

This section evaluates WD-CD through three experiments. We first benchmark seven representative deep learning models to establish baseline performance across different target categories (Section 4.3). We then introduce cross-dataset comparisons, quantifying the relative task difficulty that WD-CD presents (Section 4.4). Finally, intercontinental transfer experiments reveal the geographic domain shift inherent to multi-region war-damage data (Section 4.5). Evaluation metrics and experimental settings are defined in Section 4.1 and Section 4.2, respectively.

4.1 Evaluation metrics

We adopt building damage detection (BDD) as the primary evaluation task because buildings are the most abundant target category in WD-CD and represent a central concern in humanitarian relief and post-war reconstruction. For BDD, we consolidate these into two operationally meaningful classes based on structural serviceability. The “no-change”, “added”, and “extended” states are assigned to the no-damage (ND) class because they do not encode visible damage to or loss of a pre-existing building footprint. Specifically, “added” denotes the appearance of a new building, while “extended” denotes the contiguous expansion of an existing building. In contrast, the “damaged” and “reduced” states are assigned to the damage (D) class because they represent visible physical deterioration or a reduction in the mapped extent of a pre-existing building, respectively. This consolidation remains fully compatible with the xBD protocol to support cross-dataset comparisons. Following the xView2 challenge (Gupta et al., 2019), we decompose BDD into building localization and damage classification, scoring both with the F 1 metric. Localization accuracy F 1 loc merges the no-damage and damaged classes into a single building class and reports building-versus-background F 1. Damage classification accuracy F 1 clf is the official xView2 damage score—the harmonic mean of the per-class F 1 of the no-damage and damaged classes, evaluated only over true building pixels. For interpretability, we additionally report the conventional per-class F 1 of the two building classes, F 1 N D and F 1 D. Unlike the foreground-restricted scoring in F 1 clf, the precision of these metrics counts background pixels misclassified as buildings.

Overall performance follows the xView2 weighting, as shown in Equation 1:which assigns greater weight to damage classification (0.7) than localization (0.3), as damage-state discrimination is both the harder and the more consequential subtask for humanitarian assessment. To compare across tasks with different metric systems, we further define F 1 Macro as the arithmetic mean of F 1 N D and F 1 D, allowing binary change detection and BDD to be compared on a more equal footing to some extent.

F 1 overall = 0.3 × F 1 loc + 0.7 × F 1 clf , (1)

4.2 Experimental setup

For the WD-CD benchmark, we evaluate seven models: DeepSiameseNet (Yang et al., 2021), which serves as a classical Siamese CNN baseline, and six more advanced models spanning the CNN, Transformer, and Mamba families (Table 4)—DDLNet (Ma et al., 2024), STNet (Ma et al., 2023), BIT (Chen H. et al., 2021), ChangeFormer (Bandara and Patel, 2022), ChangeMamba (Chen et al., 2024), and CD-Lamba (Wu et al., 2026). All seven models are used for the primary BDD and multi-target quantitative evaluations, whereas the six more advanced models are used for the cross-dataset comparison. The cross-dataset comparison includes the established natural-disaster building-damage benchmark xBD (Gupta et al., 2019) and three widely used binary change-detection benchmarks: WHU-CD (Ji et al., 2018), LEVIR-CD (Chen and Shi, 2020), and SYSU-CD (Shi et al., 2021).

TABLE 4. The classification and computational overhead of the benchmark suite. FLOPs are counted for an input image size of 256 × 256 × 3.

All experiments are run on four NVIDIA A100 GPUs (40 GB each) with PyTorch 2.0.0, Python 3.8, and CUDA 11.8. Inputs are resized to 256 × 256 × 3. A batch size of 8 is used for WD-CD, xBD, WHU-CD, LEVIR-CD, and SYSU-CD. All models are trained for 200 epochs using SGD (initial learning rate 0.01, momentum 0.9, weight decay 0.0005) under a poly decay schedule (power 0.9), with image flipping and blurring for augmentation and four data-loading workers; the checkpoint with the best validation F 1 overall is used for testing.

4.3 Benchmark evaluation

We adopt BDD as the principal evaluation task and benchmark DeepSiameseNet (Yang et al., 2021) as a classical Siamese CNN baseline together with six more advanced models using the metrics defined in Section 4.1. As reported in Table 5, across all seven models, F 1 overall ranges from 72.46% to 79.53%, with CD-Lamba leading all reported metrics and attaining the highest F 1 overall of 79.53%. Localization performance remains higher than damage-classification performance for all seven models. The six more advanced models achieve F 1 loc scores between 81.54% and 83.04%, whereas the classical DeepSiameseNet baseline obtains 72.99%, indicating the progress achieved by more advanced architectures in building-footprint extraction. Damage classification F 1 clf nevertheless remains consistently below F 1 loc, primarily because the extreme irregularity of damage morphology and severe inter-class confusion in real war scenes (Zheng et al., 2025) hinder reliable damage-state assignment. The lower F 1 D relative to F 1 N D across all seven models further reflects the class imbalance observed in the WD-CD building subset: with undamaged buildings dominating the distribution, gradient optimization favors the majority class and suppresses feature activation for the scarce damaged samples. Damage-state discrimination under severe imbalance is thus the primary challenge WD-CD poses.

TABLE 5. BDD performance of the benchmark suite on the Urban-Rural Areas subset of WD-CD.

The highest scores are in bold, and the second-highest scores are underlined. All scores are represented as percentages.

Real war scenarios also demand monitoring of high-value strategic targets beyond buildings. We therefore extend the seven-model quantitative benchmark to four surface-base and two mobile-target categories (Table 6). On surface-base fine-grained classification and multi-class damage detection (MCDD), CD-Lamba reaches at least 83.41% across all reported metrics, with runway recognition and MCDD both exceeding 96%. DeepSiameseNet records an MCDD score of 64.30%, while its mobile-target F 1 scores range from 16.09% to 42.98%, providing a lower classical reference against which the stronger performance of the more advanced models can be assessed. Overall, the results demonstrate strong benchmark performance on large continuous targets and indicate that WD-CD can support macro-scale ground-damage monitoring. Mobile targets are substantially harder: unlike bounding-box detection, pixel-level change detection on extremely small targets is a recognized remote-sensing bottleneck, because minor boundary offsets under background clutter can sharply degrade Intersection-over-Union (IoU) and F 1 (Wen et al., 2025; Shim and Lee, 2025). CD-Lamba nonetheless exceeds 60% on most mobile-target metrics, quantifying this challenge while underscoring WD-CD’s value for interdisciplinary research such as battlefield situational awareness and strategic-equipment mobility monitoring.

TABLE 6. Quantitative results on surface base and mobile target categories.

The highest scores are in bold, and the second-highest scores are underlined. All scores are represented as percentages. MCDD denotes the F 1 overall for multi-class damage detection on surface base; the remaining columns report category-wise F 1. Symbols: “=” denotes “no-change”, “−” denotes “reduced”, and “+” denotes “added”.

The qualitative results in Figure 6 visually synthesize these findings across the three target types: most models recover the contours of large continuous surface-base targets reliably, yet exhibit semantic confusion and boundary errors in fine-grained building-damage classification and for small mobile targets against complex backgrounds. This recurring contrast shows that WD-CD supports consistent benchmarking across different target types and that, within the seven models evaluated in this study, spatial localization is more mature than damage-semantic discrimination. Notably, the two strongest models, CD-Lamba and DDLNet, share a core design principle: both combine fine-grained local features with global context (Ma et al., 2024; Wu et al., 2026). This indicates that accurate war-damage assessment hinges on models that couple global perception with effective local detail recovery.

4.4 Cross-dataset comparison

To compare the task difficulty and benchmark performance of WD-CD on its core building-damage task, we evaluate the six more advanced models on three mainstream binary change-detection datasets and the natural-disaster benchmark xBD. For the aligned visualization in Figure 7a, we report the conventional F 1 score for the three BCD datasets and F 1 Macro for the two BDD datasets, as defined in Section 4.1. In this analysis, DeepSiameseNet serves only as a classical in-dataset reference and is therefore not included in the cross-dataset comparison.

Figure 7a reveals a clear stratification across task types. The model’s F 1 scores on three BCD benchmarks generally cluster between 80% and 90%. It achieves over 90% on WHU-CD and LEVIR-CD and maintains scores above 80% even on the more challenging SYSU-CD, indicating that the evaluated models attain higher scores on these BCD benchmarks under the adopted protocol. Once the task shifts to BDD, however, F 1 Macro on both xBD and WD-CD drops well below these BCD scores, reflecting the difficulty of the BDD task.

To examine the differences between xBD and WD-CD, the two datasets supporting building-damage assessment, we apply the BDD-specific metrics defined in Section 4.1, with the per-model results shown in Figure 7b. The F 1 overall scores obtained on WD-CD are lower than those obtained on xBD for all six models. Nevertheless, F 1 loc on WD-CD remains within 81.5%–83.0% and consistently exceeds the corresponding xBD localization score, indicating consistent and competitive building-localization performance under the evaluated protocol. In contrast, F 1 clf on WD-CD is lower than both the corresponding xBD classification score and the WD-CD localization score for every model, thereby reducing overall performance on WD-CD. This consistent “high-localization, low-classification” pattern indicates that damage-state discrimination presents a greater challenge than building localization under the adopted protocol.

This shift originates in two intertwined properties of the data. In damage morphology, the high-intensity disasters in xBD (such as strong earthquakes) typically strip buildings of vertical support and collapse them wholesale, whereas the kinetic strikes that dominate WD-CD produce localized or complete roof collapse in large buildings while leaving facades relatively intact, and reduce small buildings to highly irregular fragments (Wang Y. et al., 2025). Such patterns are visually harder to separate from their surroundings than earthquake collapse, inflating inter-class confusion. At a deeper level, the conflict-affected scenes sampled for WD-CD exhibit severe class imbalance: unlike the homogeneous destruction of tsunamis or wildfires, precise strikes against high-value targets leave damage highly localized. In the Aleppo case (Mueller et al., 2021), buildings assessed as destroyed by the United Nations Satellite Centre accounted for only 2.8% of the total, with damaged buildings below 15%. The war-damage samples bearing high-frequency texture mutations thus form only a small fraction of WD-CD and are easily overwhelmed by the majority class during feature activation, producing severe bias in damage-state classification.

In summary, the cross-dataset comparison shows that WD-CD achieves consistent and competitive benchmark performance under the evaluated protocol while presenting additional challenges associated with heterogeneous damage morphology and severe class imbalance. The comparison characterizes the relative difficulty of building localization and damage-state discrimination across the evaluated datasets and models. In particular, the consistent gap between localization and damage-classification performance identifies three important challenges for future research: pseudo-change suppression, fine-grained damage-feature disentanglement, and learning under severe class imbalance.

4.5 Intercontinental transfer analysis

A core challenge for war-scenario CD is the cross-regional heterogeneity of geographic environments, where differences in building style, vegetation, and topography introduce substantial distribution shift. The theaters in WD-CD differ markedly in visual character (Figure 8): the Asian theater is dominated by arid, desert-like landscapes with earth-toned low-rise buildings, the African theater by distinctive red soil, and the European theater by modern architecture set against lush vegetation. This spread provides a natural testbed for examining how well war-damage CD models generalize across domains. We split the Urban-Rural Areas samples into continent-wise training and test subsets. Taking Asia, the smallest, as the reference, we randomly downsample the other two continents to match its size for a fair data volume. We then train CD-Lamba, the best-performing model on WD-CD, separately on each continent and evaluate every trained model on all three test subsets, using only the top-ranked model to isolate geographic domain shift from architecture-specific variance.

The intercontinental transfer matrix (Figure 9) shows, as expected, that models trained and tested within the same continent outperform those in every cross-continental setting. Among cross-continental transfers, the European-trained model generalizes best: its F 1 overall drops by only 14.39 and 9.66 percentage points on the African and Asian test sets relative to in-domain performance. The African- and Asian-trained models instead show much larger declines on the European set, with F 1 overall falling by 64.45 and 46.05 percentage points, respectively. One possible explanation is the greater visual diversity of the European subset, although this hypothesis is not directly tested here. Comparing Figures 9a,b, the cross-domain decline in localization F 1 loc is, in most cases, far smaller than that in damage classification F 1 clf: geometric cues such as building edges and footprint shapes are generally more transferable, although localization still degrades substantially in several transfer directions. By comparison, damage semantics—including rubble texture, color, and collapse patterns—are more tightly coupled to the background environment and are therefore more susceptible to distribution shift.

In summary, the transfer experiments reveal a clear relative gap between geometric and semantic transferability: target geometry generally transfers better than damage semantics, although localization also degrades substantially in several cross-regional settings. Damage morphology and its visual features are more tightly coupled to local environments and show larger declines in most transfer directions. This asymmetry underscores both the geographic diversity of WD-CD and the distinctive generalization challenge it poses, and points to a concrete design principle: rather than treating cross-region interpretation as one homogeneous task, future models should decouple the geometric and semantic pathways—using comparatively transferable geometric priors for localization while learning damage representations that suppress region-varying background style yet preserve the intrinsic appearance of damage for accurate state recognition.

5 Downstream applications and potential directions

The annotation schema of WD-CD, encompassing 16 fine-grained target categories and multi-state change labels, supports a broad spectrum of research tasks, as summarized in Figure 10. Section 5.1 presents applications that can be pursued immediately using the dataset, supported by either our benchmark results or established methods in the literature. Section 5.2 highlights underexplored directions enabled by WD-CD, with methodological pathways outlined for each.

5.1 Current downstream applications

5.1.1 Building change analysis

As direct imprints of human activity, buildings record both urban development and conflict-driven destruction across time. Existing datasets reduce building change to binary detection (Chen and Shi, 2020; Ji et al., 2018) or single-axis damage assessment (Gupta et al., 2019), missing the coexistence of destruction and reconstruction in war scenarios. WD-CD resolves this through a five-state annotation scheme (“no-change”, “damaged”, “reduced”, “added”, “extended”), enabling joint modeling of destruction and development.

For destruction, all six more advanced benchmark models achieve F 1 overall > 75 %, while the classical DeepSiameseNet baseline obtains 72.46% (Table 5). These results show that the evaluated architectures can recover building-damage signals from WD-CD’s annotations under the adopted protocol, while also documenting the performance gap between the classical reference and more advanced models. Such capability provides valuable support for humanitarian response planning (Gholami et al., 2022), strike impact evaluation (Mueller et al., 2021), and asset loss estimation (Lu and Guo, 2021). For development, WD-CD enables methods originally designed for non-war scenarios, such as MTL-CD (Sun et al., 2020) and S 2Net (Wang et al., 2025b), to be applied to war contexts for the first time. Identifying the “added” and “extended” states further enables post-war urban recovery monitoring (Feng et al., 2025; Zhang et al., 2020).

WD-CD provides: five-state pixel-level building annotations which jointly capture destruction and development.

5.1.2 Airport pavement condition monitoring

Airports function simultaneously as civilian transportation hubs and critical military projection platforms, making automated pavement condition assessment a high-priority need in both contexts. The field currently lacks standardized large-scale benchmarks that couple functional area annotations with real strike damage records, severely constraining algorithm development and comparability (Feitosa et al., 2025). WD-CD fills this gap with annotations of runways, taxiways, and aprons alongside weapon-strike crater instances.

These annotations enable a two-stage analysis pipeline. In the first stage, functional area segmentation must overcome the high visual similarity between pavement classes (Ding and Wu, 2020). Architectures that jointly model local geometry and global layout, such as CTMANet (Wu et al., 2024) and topology-aware line detection approaches (Ganeriwala et al., 2025), are well suited to this challenge and directly applicable to WD-CD. Our benchmark results further confirm that the highest-scoring models achieve F 1 > 94 % on runways (Table 6), serving as a reference for future specialist methods. In the second stage, models can further incorporate crater annotations to achieve precise detection of pavement damage. Such a paradigm provides valuable insights for the development of next-generation airport pavement management systems (Feitosa et al., 2025), which are expected to rapidly generate post-war pavement condition maps and estimate the minimum operating strip, supporting emergency repair prioritization and air traffic network restoration.

WD-CD provides: pixel-level airport pavement and crater instance annotations, jointly supporting functional area segmentation and strike damage detection.

5.1.3 Farmland damage detection

War-induced agricultural damage threatens regional food security and imposes long-term ecological costs through soil disruption and unexploded ordnance contamination (Lin et al., 2020). WD-CD provides semantic state labels for agricultural parcels alongside per-instance crater masks, enabling a coherent analysis pipeline that spans farmland boundary delineation and ordnance contamination risk mapping.

Accurate delineation of farmland regions remains challenging due to pseudo-changes and intra-class variability caused by phenological transitions, motivating the development of advanced networks such as MSCANet (Liu et al., 2022) and STMTNet (Lv et al., 2025). These methods can leverage WD-CD for robust feature learning. Building upon reliable farmland delineation, crater density estimation within agricultural areas enables unexploded ordnance contamination risk mapping. This application is grounded in the empirically established spatial correlation between crater distributions and subsurface ordnance (Lin et al., 2020) and can support humanitarian demining prioritization and post-war land-reclamation planning.

WD-CD provides: pixel-level farmland state annotations and crater instance masks, jointly supporting farmland delineation and ordnance contamination risk mapping.

5.1.4 Mobile target dynamic monitoring

Aircraft and vessels are high-priority strategic assets whose state transitions are central to battlefield situational awareness (Li et al., 2024). WD-CD provides multi-state pixel-level annotations for both asset types (“no-change”, “reduced”, “added”, etc.), thereby establishing a data foundation for modeling the spatiotemporal evolution of such dynamic targets.

As reported in Table 6, CD-Lamba (Wu et al., 2026) achieves F 1 > 60 % across most state categories without task-specific design. This result is comparable to results reported on specialized small-target benchmarks (Wang et al., 2023; Shim and Lee, 2025) and establishes concrete baselines for future architectures. Beyond per-pair detection, state transitions between paired acquisitions support higher-level inference of redeployment patterns and asset attrition. In civilian contexts, the same framework transfers directly to airport occupancy analysis (Bingol et al., 2026) and port throughput monitoring (Zhang B. et al., 2025).

WD-CD provides: multi-state pixel-level annotations of aircraft and vessels, jointly supporting military asset monitoring and civilian infrastructure management.

5.1.5 Single-temporal scene understanding

Beyond change detection, WD-CD can be adapted to support single-temporal vision tasks, including semantic segmentation and object detection after task-specific label conversion and validation.

For semantic segmentation, the released class–state annotations must first be mapped to the target categories present in the selected temporal image to produce a single-temporal semantic mask. The resulting masks can support continuous-scene parsing for categories such as runways, taxiways, aprons, and farmland (Tang Y. et al., 2025). For object detection, the masks of discrete targets, including buildings, aircraft, and vessels, must first be separated into verified object instances and then converted into bounding boxes (Cao et al., 2024; Zhao T. et al., 2024; Benchabana et al., 2023). When explicit instance identifiers are unavailable, connected-component analysis can be used to generate initial instance candidates. However, because adjacent or touching objects may form a single connected component, the resulting candidates require manual verification before they can be treated as reliable object-level annotations. Accordingly, task-specific validation is required when the converted labels are used to train or evaluate single-temporal models. Such versatility highlights that WD-CD can also serve as a general-purpose dataset for training and evaluating vision models in remote sensing. In particular, its multi-category, multi-state annotations provide valuable resources for large-scale vision pre-training and multi-task learning, facilitating the development of unified models across different perception tasks.

WD-CD provides: pixel-level class–state annotations that can support single-temporal semantic segmentation and discrete-target detection after task-specific label conversion and, where necessary, instance-level verification.

5.2 Potential directions

5.2.1 Post-war passability analysis

Post-war emergency response in urban areas heavily relies on the availability of functional ground transportation networks. However, collapsed or severely damaged buildings along road corridors often lead to physical obstructions, significantly reducing network connectivity and accessibility.

The joint annotations of building damage and road networks in WD-CD enable modeling the spatial interaction between structural collapse and road accessibility. By identifying the overlap between damaged building regions and adjacent road segments, potential blockage points within the road network can be automatically detected, providing a basis for large-scale passability analysis. Building upon this capability, WD-CD enables the development of dynamic passability assessment models and obstacle-aware routing systems. Such approaches can support post-war rescue operations, evacuation planning, and infrastructure recovery by providing data-driven guidance on accessible routes under complex obstruction conditions.

WD-CD provides: joint annotations of damaged buildings and road networks for obstruction-aware passability analysis and dynamic routing.

5.2.2 Surface disruption correlation analysis

Modern war not only damages infrastructure but also induces surface disruption through dense crater formation and military shelter construction. Here, military shelters refer to two categories annotated in WD-CD: aircraft shelters (the “Shelter” category) and vehicle revetments. Existing studies have largely focused on crater detection, while the role of military shelters remains underexplored. In fact, both craters and military shelters represent typical forms of war-induced surface disruption, with the latter providing additional insights into force deployment, defensive configurations, and engagement locations.

Based on WD-CD, craters and military shelters can be jointly modeled within a unified visual analysis framework. In wartime scenarios, the combined analysis of crater-based strike density and the spatial distribution of military shelters can support the inference of fire-coverage patterns and frontline configurations, contributing to battlefield situational awareness. In post-war contexts, high-precision spatial localization can assist in unexploded ordnance detection, debris removal, and the estimation of earthwork required for land leveling and reclamation. This cross-object spatial analysis paradigm provides a promising direction for integrating remote sensing interpretation with environmental and geotechnical engineering, supporting comprehensive post-war terrain assessment and restoration.

WD-CD provides: joint annotations of craters, aircraft shelters (the “Shelter” category), and vehicle revetments for cross-target spatial correlation analysis and terrain recovery modeling.

5.2.3 Multimodal data fusion

In war scenarios, single-sensor observations often suffer from limitations. Optical imagery, for instance, is sensitive to illumination and weather conditions and lacks structural depth, while other modalities such as SAR, infrared, and multispectral data offer complementary information (Zhu et al., 2017). WD-CD is constructed from high-resolution optical imagery that provides rich textural and semantic information for multimodal data fusion. Each sample of WD-CD retains geolocation metadata and acquisition timestamps, facilitating cross-modal alignment.

WD-CD’s fine-grained annotations and semantic coverage provide a reference for aligning complementary sensing modalities. For example, integrating SAR data can introduce structural and scattering information to enhance scene understanding (Ling et al., 2024; Wang et al., 2025a); combining infrared data enables robust perception under low-light or nighttime conditions (Yang et al., 2025); and incorporating multispectral data allows long-term monitoring of vegetation dynamics and land surface changes (Pettorelli et al., 2005). Such multimodal integration enables models to learn more robust and discriminative representations across diverse observation conditions. Taken together, these integration pathways allow WD-CD to serve as an optical reference for cross-modal learning, facilitating all-weather, multidimensional situational awareness and post-war assessment.

WD-CD provides: a high-quality optical reference for cross-modal alignment and multimodal data fusion.

5.2.4 Foundation model pre-training

In recent years, vision foundation models have demonstrated robust generalization capabilities in remote sensing applications (Lu et al., 2025). However, existing models are mainly pre-trained on urban, agricultural, and natural imagery, which limits their robustness when transferred to war scenarios. The highly unstructured debris patterns, dense crater distributions, and deliberately camouflaged targets introduce significant domain shift, leading to degraded performance (Mouradi and Kshirsagar, 2026).

WD-CD fills a critical gap in the current data ecosystem. WD-CD not only provides large-scale coverage but, more importantly, also captures distinctive visual characteristics of the war-damage domain, making it a valuable resource for domain-adaptive pre-training and self-supervised learning. By leveraging paradigms such as masked image modeling and contrastive learning (Pang et al., 2026), models can internalize the underlying physical semantics of terrain disruption and structural damage, thereby achieving improved generalization under complex conflict scenarios with limited or no labeled data.

WD-CD provides: a large-scale war-damage domain corpus for domain-adaptive pre-training, with the potential to support robust feature learning in war scenarios.

6 Challenges and outlook

Although WD-CD provides a solid data foundation for fine-grained war-damage interpretation, it also presents several challenges due to the extreme physical conditions of real war scenarios and the inherent difficulty of data acquisition.

6.1 Class imbalance

Within the conflict-affected scenes selected for WD-CD, the annotated categories and change states exhibit substantial class imbalance (Tian et al., 2022). This challenge manifests in two aspects. From the perspective of semantic categories, large-scale land-cover types such as farmland and buildings account for the majority of annotated pixels, making sparse high-value targets difficult to detect. From the perspective of change states, key categories such as “damaged” regions, “extended” fixed facilities, and “replaced” mobile targets are substantially underrepresented compared to the “no-change” class. To mitigate this dual imbalance, future research should incorporate specialized long-tailed learning paradigms for WD-CD. For instance, decoupled training frameworks or adaptive re-weighting loss functions can be explored to balance gradient contributions (Liu et al., 2023; Chen et al., 2025). Alternatively, data augmentation strategies based on feature synthesis or diffusion models may be adopted to enhance model sensitivity to minority classes (Wan et al., 2025).

6.2 Image-domain nuisance variation

WD-CD is also affected by substantial image-domain nuisance variation, as illustrated in Figure 11. One source arises from varying imaging conditions. Differences in solar elevation and sensor viewing angle produce global radiometric shifts, occlusions, shadows, and viewpoint-related displacement, particularly in dense urban environments containing buildings and vegetation. Moreover, different parts of the same object may be observed from different perspectives (Bruzzone and Bovolo, 2012). A second source comprises pseudo-changes caused by seasonal variation, changes in object clutter, and spectral differences resulting from roof refurbishment. Although high-resolution imagery provides rich spatial detail, it can amplify false positives induced by these observation effects. Robust models must therefore learn features that are insensitive to such nuisance variation while preserving genuine change signals.

6.3 Scarcity of specific categories and change states

Although WD-CD includes 16 target categories and their category-specific change states, some categories and category–state combinations remain too sparsely represented for reliable standalone benchmarking. At the category level, special structures such as towers and bridges have limited annotated support in the conflict-affected scenes sampled for WD-CD. At the category–state level, particular “damaged”, “extended”, and “replaced” combinations within sparsely represented categories can contain only small numbers of pixels or target instances. As described in Section 3.3, these labels are retained in the complete dataset annotations but are mapped to background when they fall outside the label space of a reported task, and no independent performance values are reported for them in the present study. Their limited support would make class-wise F 1 or IoU values highly dependent on individual samples and partition composition. Future work will continuously expand the instance pool of rare categories by tracking conflict-related open-source information and incorporating newly released or declassified imagery. In parallel, we will investigate physics-informed data-generation techniques to enrich extremely scarce change states (Tang D. et al., 2025), thereby progressively extending the category–state coverage that can support statistically meaningful evaluation.

7 Conclusion

In this work, we addressed the absence of a large-scale, geographically diverse benchmark for fine-grained change detection in war scenarios by constructing WD-CD, which couples pixel-level, multi-state annotations with imagery spanning five conflicts across three continents. Beyond providing this resource, our evaluation clarifies where the core difficulty of war-damage interpretation lies: across all seven benchmark models, building-localization accuracy exceeds damage-classification accuracy, although the classical DeepSiameseNet baseline performs below the six more advanced architectures. This localization–classification gap is further widened by the long-tailed distribution observed in the conflict-affected scenes sampled for WD-CD, in which damage-class pixels and high-value mobile-target pixels are substantially less numerous than unchanged-background pixels. Robustness under such severe imbalance, rather than average accuracy, therefore emerges as an important criterion for future progress. Our intercontinental transfer study sharpens this picture by revealing that geometry and semantics generalize in fundamentally different ways: building geometry features transfer more consistently across the evaluated continental splits, whereas damage morphology and background landform remain strongly domain-bound. We regard this dissociation as a key finding of WD-CD, as it not only justifies the deliberate three-continent design but also points toward a concrete algorithmic principle—decoupling a transferable geometric pathway for localization from a damage-semantic pathway that is explicitly regularized against background style. Building on these findings, as well as the five downstream applications and four research directions detailed in Section 5 and the open challenges outlined in Section 6, we envision WD-CD not as a static dataset but as an extensible foundation that we will continue to enrich through long-term hotspot tracking and physics-informed sample synthesis, ultimately bridging algorithmic research with the humanitarian and reconstruction efforts that motivate it.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Xinghua Li, Wuhan University, China

Reviewed by: Wen Yang, Wuhan University, China

Zhengxia Zou, Beihang University, China