Gaza-Change Dataset: Damage & Change Analysis

Updated 1 December 2025

Gaza-Change Dataset is a comprehensive collection of high-resolution optical and SAR imagery providing detailed semantic change and building-level damage mapping in the Gaza Strip.
The dataset leverages advanced methods including a DINOv3-based Siamese network for optical change detection and SAR-based t-test and coherence analytics for robust damage assessment.
It supports humanitarian response and urban reconstruction by delivering precise, quantitative metrics and validated classifications against UNOSAT references.

The Gaza-Change Dataset is a collective designation for several open-access and request-based datasets providing high spatiotemporal resolution damage and semantic change information for the Gaza Strip, 2023–2024. These resources offer varying granularity—pixel-level semantic masks, building-level damage classifications, and raster coherence-change time series—derived from high-resolution optical and synthetic aperture radar (SAR) imagery. The datasets facilitate fine-grained inference and systematic monitoring of conflict-induced damage, supporting research, humanitarian response, and reconstruction planning.

1. Geographic and Temporal Scope

The Gaza-Change datasets comprehensively cover the Gaza Strip, centered at latitudes 31.4°–31.6° N and longitudes 34.4°–34.6° E. They encompass nine major urban localities, including Khan Yunis and Rafah. Temporal coverage varies by modality and task:

Optical/semantic:
- Bi-temporal Beijing-2 imagery (3.2 m GSD), patch-based, with two dates per scene during 2023–2024.
SAR-based:
- Sentinel-1 C-band GRD and SLC data, with reference/inference periods spanning October 2023 – October 2024.
- InSAR monitoring at up to weekly cadence using 321 SLC scenes (Scher et al., 17 Jun 2025).

These datasets are designed for damage assessment, change detection, humanitarian logistics, and longitudinal urban impact studies (Zhenga et al., 24 Nov 2025, Ballinger, 10 May 2024).

2. Data Sources, Structure, and Annotation Protocols

Gaza-Change (High-Resolution Optical, Pixelwise Semantics)

Imagery: Beijing-2 satellite, 3.2 m GSD, 512×512 px patches.
Change annotation: Only pixels exhibiting semantic change between $T_1$ and $T_2$ are labeled; background remains unlabeled (implied “no-change”).
Semantic classes (6 total):
- Building Damage
- New Building
- New Camp
- Farmland Damage
- Greenhouse Damage
- New Greenhouse

Annotations are exclusively change masks, omitting full-scene semantic segmentation. This approach reduces annotation cost and focuses on change detection over static land cover labeling, mitigating error propagation from multi-stage annotation (Zhenga et al., 24 Nov 2025).

SAR-Based Damage Datasets

PWTT (Pixel-Wise T-Test): Area-weighted damage classification at the building footprint level (201,629 footprints) using Sentinel-1 GRD (10 m resolution, both VV and VH polarizations, ascending/descending).
- Binary label per footprint: “damaged” (UNOSAT-intersecting) or “undamaged”, with area and mean $T$ -statistic (Ballinger, 10 May 2024).
- Derived from Microsoft GlobalML building footprints and UNOSAT optical manual damage points as reference.
InSAR/LT-CCD: Time-resolved, building-level coherence-change monitoring (330,079 OSM-based polygons).
- Per-pixel and building-aggregated coherence anomalies at 40 m resolution, with time of first, last, and confirmed damage flag per building.
- Raster stack: 56 weekly GeoTIFF grids.
- Aggregated attributes: pre-war mean and std. coherence, damage flag (binary), fraction of building flagged, confirmation persistence (Scher et al., 17 Jun 2025).

3. Processing Pipelines and Methodological Frameworks

Semantic Change Detection (CSD Task)

The CSD framework, defined in (Zhenga et al., 24 Nov 2025), extends conventional binary change detection (BCD) to multi-class semantic change tasks using only change pixels for annotation. Methodological innovations include:

Backbone: Pre-trained DINOv3 for robust representation from bi-temporal inputs.
Network: Multi-scale cross-attention difference Siamese architecture (MC-DiSNet).
Masking protocol: Only changed semantic regions labeled; background pixels avoided in mask annotation.

SAR-Based Damage Algorithms

PWTT: Pixel-wise disjoint t-tests performed for each orbit/polarization/time period; per-pixel $T$ -statistic aggregated per building. Decision threshold for $T$ empirically set ( $\tau=2.7$ typical for $\alpha=0.01$ ), damage label assigned per building depending on threshold exceedance (Ballinger, 10 May 2024).
LT-CCD: Long temporal-arc InSAR coherence change detection:
- For each week, form $N$ interferogram pairs between conflict and pre-war epochs.
- Compute mean coherence $\overline{\gamma}_{\text{conflict}}(t)$ , $\overline{\gamma}_{\text{pre}}$ , std $\sigma_{\text{pre}}$ .
- Calculate coherence anomaly and z-score: $\Delta\gamma(t) = \overline{\gamma}_{\text{conflict}}(t) - \overline{\gamma}_{\text{pre}}$ , $z(t) = \Delta\gamma(t)/\sigma_{\text{pre}}$ .
- Damage criterion: $\Delta\gamma(t) < k$ and $z(t) < -2$ ( $k = -0.1$ empirically optimal).
- Persistence filter: change must persist for at least one repeat acquisition to avoid transient errors (Scher et al., 17 Jun 2025).

4. Dataset Scale, Class Balance, and Organization

Dataset/Modality	Granularity	N (primary unit)	Temporal Coverage	Label Type
Gaza-Change (optical)	Patch/pixel	922 patch pairs (512×512)	2023–2024 (bi-temporal)	6-class pixel-level change masks
PWTT (SAR, S1 GRD)	Building	201,629 footprints	2023.10–2024.03 (aggregate)	Binary (damaged/undamaged, per bldg)
LT-CCD (SAR, S1 SLC)	Building/pixel	330,079 buildings	2023.10–2024.10 (weekly)	Damage flag + time series

Class distribution is highly imbalanced: change pixels constitute a small minority in optical patches; building damage dominates among classes, with “new camp” and “farmland damage” much rarer. In SAR-derived tables, damaged buildings are ~33.9% (PWTT, 2024.03) and ~57.9% (LT-CCD, cumulative over war) (Zhenga et al., 24 Nov 2025, Ballinger, 10 May 2024, Scher et al., 17 Jun 2025).

5. Quantitative Benchmarks and Evaluation Metrics

All Gaza-Change models and datasets utilize standard pixel-wise or instance-wise classification, segmentation, and detection metrics—expressed as:

OA (Overall Accuracy): $\frac{\sum q_{ii}}{\sum_{i,j} q_{ij}}$
Precision (per class): $TP_c/(TP_c + FP_c)$
Recall (per class): $TP_c/(TP_c + FN_c)$
IoU (per class): $TP_c/(TP_c + FP_c + FN_c)$
Mean IoU (mIoU): $(1/C) \sum_{c=1}^C IoU_c$
F1 (per class): $2 \cdot \frac{\text{Precision}_c \cdot \text{Recall}_c}{\text{Precision}_c + \text{Recall}_c}$

MC-DiSNet (Gaza-Change, 6-way CSD):

Precision: 86.10%
Recall: 60.98%
mIoU: 55.16%
F1: 69.25%
Per-class F1 ranges: 44.79% (farmland damage) to 86.91% (new greenhouse) (Zhenga et al., 24 Nov 2025).

PWTT (SAR, building-level, area-weighted):

AUC: 0.81
F1: 0.64
Precision: 0.53
Recall: 0.82
N: 201,629 footprints

Balanced sample (N=273,224): AUC 0.82, F1 0.76 (Ballinger, 10 May 2024).

LT-CCD (InSAR, building-level, validation vs UNOSAT):

Overall agreement: 92.5%
True positive rate: 86.2%
False positive rate: 1.2%
F1: 91.8%
CSI: 85.2%
N: 928,397 UNOSAT point labels (Scher et al., 17 Jun 2025).

A plausible implication is that multi-temporal coherence approaches are highly effective for rapid cumulative damage tracking, while semantic optical change datasets offer richer fine-grained category information but with considerably more challenging annotation and class imbalance characteristics.

6. Access Modalities and Data Formats

Gaza-Change (optical/CSD): Access to full data and file-format documentation is provided to qualified researchers upon request; formats likely follow de facto patch-wise GeoTIFF for imagery and PNG/TIFF for masks, with associated metadata for date, scene, coordinates, and class legend (Zhenga et al., 24 Nov 2025).
PWTT and LT-CCD SAR datasets: Open access via GitHub and Zenodo (static DOI), providing:
- Building-level attribute tables (CSV/GeoJSON)
- Weekly coherence-change GeoTIFF raster stacks (LT-CCD)
- Metadata JSONs.
API/Earth Engine: RESTful JSON endpoints delivering per-building or per-time step SAR change status; persistent Earth Engine public assets for direct geospatial analysis (Ballinger, 10 May 2024, Scher et al., 17 Jun 2025).

7. Applications, Limitations, and Context

Gaza-Change datasets are essential for near-real-time and retrospective assessment of conflict-induced damage:

Primary applications: Damage mapping, humanitarian logistics, post-conflict planning, trend analysis, and resource allocation.
Optical semantic change detection is challenged by the limited spatial extent of semantic regions, high intra-class similarity, and boundary ambiguity—yielding low percentage of change pixels and significant class imbalance (Zhenga et al., 24 Nov 2025).
SAR-based approaches (PWTT, LT-CCD) enable building-level and area-wide monitoring with timely updates and robust reference-controlled change detection. Their limitations include reduced discriminatory capacity in areas with low pre-conflict coherence and inability to attribute specific semantic class to detected change (damage vs. new construction, etc.) (Ballinger, 10 May 2024, Scher et al., 17 Jun 2025). Both validation and false-positive/negative analyses rely on UNOSAT manual optical damage points, considered high-quality reference but potentially missing subtle or partial damages.

Collectively, these datasets anchor open, reproducible, and multi-modal methodologies for damage assessment in conflict zones, providing granular as well as high-frequency monitoring resources intimately tied to operational and humanitarian needs.