🛎️News

• Feb 6th, 2026: To facilitate future research in the field of multimodal remote sensing foundation models, RGB image resources have been uploaded on HuggingFace.
• Notice☀️☀️: SARLANG-1M has been accepted by [IEEE TGRS 2026] on January 5, 2026!! Related data and benchmark suites has been released!
• Apr 27th, 2025: SARLANG-1M was presented at the 3rd Workshop on Machine Learning for Remote Sensing (ML4RS) at ICLR 2025. See the demo!!
• Apr 7th, 2025: The arXiv paper of SARLANG-1M is now online. If you are interested in details of SARLANG-1M, do not hesitate to take a look!!

🔭Overview

• SARLANG-1M is a large-scale benchmark tailored for multimodal SAR image understanding, with a primary focus on integrating SAR with textual modality. SARLANG-1M comprises more than 1 million high-quality SAR image-text pairs collected from over 59 cities worldwide. It features hierarchical resolutions (ranging from 0.1 to 25 meters), fine-grained semantic descriptions (including both concise and detailed captions), diverse remote sensing categories (1,696 object types and 16 land cover classes), and multi-task question-answering pairs spanning seven applications and 1,012 question types. Fine-tuning with our SARLANG-1M dataset can significantly enhance VLMs' performance in SAR image interpretation, reaching performance comparable to human experts.

• SARLANG-1M dataset supports seven SAR applications:

Application	Application Description	Text Numuber
Image Description	Describe the SAR image	45,650
Object Identification	Determine the presence of specific objects	484,620
Object Classification	Identify the predominant category within the SAR image	132,525
Instance Counting	Quantify instances within the SAR image	117,382
Region Referring	Determine the category present in the specific location	221,450
Object Positioning	Determines the approximate location of a category	106,171
Others	Predict the object shape, direction, reasoning etc	18,479

• The Statistics of Text Annotations in SARLANG-1M dataset:

The image shows the distribution of seven applications provided in the SARLANG-1M benchmark (a), the numbers of each question type in the 'others' application (b), and the distribution of the 30 most frequent object categories (c).

🗝️Let's Get Started with SARLANG-1M!

A. Installation

Note that the code in this repo runs under Linux system. We have not tested whether it works under other OS.

Step 1: Clone the repository:

Clone this repository and navigate to the project directory:

git clone https://github.com/Jimmyxichen/SARLANG-1M.git
cd SARLANG-1M

Step 2: Environment Setup:

It is recommended to set up a conda environment and installing dependencies via pip. Use the following commands to set up your environment:

Create and activate a new conda environment

conda create -n SARLANG1M
conda activate SARLANG1M

Install dependencies

pip install -r requirements.txt

B. Data Preparation

Please download the SAR images from HuggingFace and make them have the following folder/file structure:

${DATASET_ROOT}   # Dataset root directory, for example: /home/username/data/SARLANG_1M
│
├── SARimages
│    ├── France_Alpes-Maritimes_Nice_3.png
│    ├── SV_Darwin_-12.4278_130.8688.png
│    ├── 0009787.jpg  # Add all SAR images in the SARDet_100K dataset to this folder
│         ...   
│   
└── Text
     ├── caption
     │
     └── VQA
         ...

Both the original SAR images [tif format] and preprocessed SAR images [png format] are provided. Notably, SAR image preprocessing is an optional strategy to improve the performance of VLMs by significantly enhancing image clarity and effectively highlighting key objects within the SAR images. You can choose any version according to your needs.

RGB Image Resources

To facilitate future research in the field of multimodal remote sensing foundation models, we provide the RGB image resources used in our paper on HuggingFace.

Text Annotations

Our text data is available in the Text.zip file, which can be accessed via HuggingFace. We provide the text data in both JSON and CSV formats, allowing users to select the appropriate format based on their specific requirements.

Caption Labels

The dataset includes two types of captions: complex captions and concise captions. Mixed captions, which combine both complex and concise descriptions, are used for training and are stored in the Caption-training.csv file. For evaluation, Caption_test.csv contains only complex captions as the test set, while Caption_test1.csv contains only concise captions as the test set.

VQA Labels

We partition the VQA data in our dataset into two subsets: SARVQA1_train/test and SARVQA2_train/test. The files prefixed with SARVQA1_train/test contain VQA text annotations for SAR images sourced from the SARDet_100K dataset. These annotations are generated based on bounding box information, providing highly accurate descriptions of both the location and quantity of objects. However, the SARDet_100K dataset only includes six common categories of remote sensing targets, which limits the diversity of object types represented in these annotations.

In contrast, the files prefixed with SARVQA2_train/test contain VQA textual annotations for SAR images derived from the SpaceNet6, DFC2023 and OpenEarthMap-SAR datasets. These datasets offer high-resolution SAR imagery, with ground sampling distances ranging from 0.1 to 1 meter, and encompass a richer variety of remote sensing target categories. If your research primarily focuses on high-resolution SAR image interpretation, you may opt to use only this subset for model training.

Note: If you intend to train your model on the combined VQA data from both subsets, it is necessary to prepend prompts such as “The classes are: tank, car, aircraft, harbor, bridge, ship.” or "The classes are: top, center, bottom, left, right." to the questions originating from the SARDet_100K dataset to clarify the available object/position categories.

C. Inference & Evaluation

Compute Evaluation Metrics for SAR Image Captioning

Please install pycocoevalcap package and then run the Evaluate_caption.py file.

pip install pycocoevalcap
Or:
pip install git+https://github.com/salaniz/pycocoevalcap

python Evaluate_caption.py

Compute Evaluation Metrics for SAR image VQA

Please install openai package and then run the Evaluate_caption.py file.

pip install openai

python Evaluate_VQA.py

📜Reference

If this dataset or code contributes to your research, please kindly consider citing our paper and give this repo ⭐️:

@article{wei2026sarlang,
  title={{SARLANG-1M}: A benchmark for vision-language modeling in {SAR} image understanding},
  author={Wei, Yimin and Xiao, Aoran and Ren, Yexian and Zhu, Yuting and Chen, Hongruixuan and Xia, Junshi and Yokoya, Naoto},
  journal={{IEEE} {T}ransactions on {G}eoscience and {R}emote {S}ensing},
  year={2026},
  publisher={IEEE}
}

🤝Acknowledgments

The authors would also like to give special thanks to SARDet_100K, SpaceNet6, DFC2023 and OpenEarthMap-SAR for providing the valuable SAR Images.

🙋Q & A

For any questions, please feel free to leave it in the issue section or contact us.

Author’s Note

To construct the dataset and ensure high-quality textual annotations, we engaged several SAR experts who dedicated substantial time and effort to meticulously review and correct the text data. Through this process of manual verification, the quality of the annotations has been significantly improved compared to those originally generated by models such as GPT-4o. Nevertheless, it is inevitable that a small proportion of the text may still contain inaccuracies. We welcome feedback regarding any erroneous annotations and are committed to continuously updating and refining the dataset. We hope that our dataset will be a valuable resource for your research.