Delta-QA

Delta-QA is a visual question answering dataset for multi-temporal remote sensing change detection and understanding. Built upon SECOND, LandsatSCD, and WUSU-SCD, it contains question-answer annotations associated with bi-temporal and tri-temporal imagery. The annotations cover image-level change understanding and spatially prompted questions, with grounding information for applicable tasks.

Paper: Decoding the Delta: Unifying Remote Sensing Change Detection and Understanding with Multimodal Large Language Models (ACM Multimedia 2026)

Authors: Xiaohe Li, Jiahao Li, Kaixin Zhang, Yuqiang Fang, Leilei Lin, Hong Wang, Haohua Wu, and Zide Fan.

Dataset Composition

Delta-QA comprises three subsets, each distributed in training and test archives:

• Delta-SECOND: Derived from SECOND, with bi-temporal imagery in im1/ and im2/ and corresponding question-answer annotations.
• Delta-Landsat: Derived from LandsatSCD, with bi-temporal imagery in im1/ and im2/ and corresponding question-answer annotations.
• Delta-WUSU: Derived from WUSU-SCD, with tri-temporal imagery in im1/, im2/, and im3/ and corresponding question-answer annotations.

File Organization

Delta-QA/
├── README.md
└── data/
    ├── Delta-SECOND/
    │   ├── train.zip
    │   └── test.zip
    ├── Delta-Landsat/
    │   ├── train.zip
    │   └── test.zip
    └── Delta-WUSU/
        ├── train.zip
        └── test.zip

Each archive contains image directories and JSON annotation files. Image references and annotation structures are preserved in the distributed archives. The train and test archives contain distinct multi-temporal image pairs or triplets. In Delta-Landsat, some individual temporal images appear in both splits as parts of different image pairs.

Annotation Format

Annotation records associate questions with image samples through fields including:

Field	Description
image_id, file_name	Image identifiers and file references.
question, answer	Question text and corresponding answer.
options	Answer choices for applicable question types.
image_size	Spatial dimensions of the referenced image.
visual_prompt	Spatial prompts, including points or bounding boxes where applicable.
groundings	Region-level grounding information, including masks and answer-text spans where available.

Grounding masks may use COCO compressed run-length encoding (RLE). Field availability varies by task and subset.

Usage

Download the dataset with huggingface_hub:

from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="HUG-ABC/Delta-QA",
    repo_type="dataset",
)

Read JSON annotations directly from an archive:

import json
import zipfile
from pathlib import Path

archive_path = Path(dataset_dir) / "data" / "Delta-SECOND" / "train.zip"

with zipfile.ZipFile(archive_path) as archive:
    json_paths = sorted(
        path for path in archive.namelist() if path.lower().endswith(".json")
    )
    with archive.open(json_paths[0]) as stream:
        records = json.load(stream)

example = records[0]
print(example["question"])
print(example["answer"])

Image samples are referenced by image_id and file_name in the annotation records. Identify a sample using its subset and split together with its image reference.

Source Datasets and Licensing

Delta-QA is derived from the following remote sensing datasets:

• SECOND: Official dataset website | Source repository
• LandsatSCD: Dataset (CC BY 4.0) | Source publication
• WUSU-SCD: Official dataset website | Source repository and usage terms | Source publication. The source project specifies academic, non-commercial use.

Source imagery and source-derived annotations remain subject to the applicable terms of their respective datasets. Please cite the original datasets when using their resources.

Citation

If you use Delta-QA in your research, please cite:

@inproceedings{li2026decoding,
  author        = {Li, Xiaohe and Li, Jiahao and Zhang, Kaixin and Fang, Yuqiang and Lin, Leilei and Wang, Hong and Wu, Haohua and Fan, Zide},
  title         = {Decoding the Delta: Unifying Remote Sensing Change Detection and Understanding with Multimodal Large Language Models},
  booktitle     = {Proceedings of the 34th ACM International Conference on Multimedia},
  year          = {2026},
  publisher     = {Association for Computing Machinery},
  address       = {New York, NY, USA},
  doi           = {10.1145/3767308.3835592},
  url           = {https://doi.org/10.1145/3767308.3835592},
  numpages      = {10},
  location      = {Rio de Janeiro, Brazil},
  series        = {MM '26},
  eprint        = {2604.14044},
  archivePrefix = {arXiv}
}