ledemo/zijinshan-vegetation-segmentation

Zijinshan Vegetation Segmentation Dataset

This repository contains a self-built high-resolution remote sensing vegetation segmentation dataset for the Nanjing Zijinshan Scenic Area and its surrounding heterogeneous urban-rural interface in Jiangsu Province, China. The dataset was manually annotated by the authors using the CVAT annotation platform and was constructed for small-sample vegetation segmentation, multi-temporal generalization, and cross-temporal pseudo-label self-training experiments.

This dataset is released to support reproducible research on small-sample vegetation segmentation in high-resolution remote sensing imagery.

Related manuscript:

Hu Le and Fuquan Zhang, "Comparative Performance Analysis of Mainstream Deep Learning and Vision Foundation Models for Small-Sample Vegetation Segmentation in High-Resolution Remote Sensing Imagery", 2026.

Dataset Overview

The dataset focuses on binary semantic segmentation of vegetation from RGB remote sensing imagery. It contains multi-temporal image slices from 2022, 2023, and 2024, with each original image patch resized or cropped to 512 x 512 pixels at 0.75 m/pixel spatial resolution.

Item	Description
Study area	Nanjing Zijinshan Scenic Area and surrounding urban-rural interface, Jiangsu Province, China
Task	Binary semantic segmentation
Classes	background, vegetation
Annotation tool	CVAT
Image modality	RGB remote sensing imagery
Spatial resolution	0.75 m/pixel
Patch size	512 x 512 pixels
Years covered	2022, 2023, 2024
Manually annotated images	201 images
Pseudo-labeled images	67 images

Data Splits

The core manually annotated dataset contains 201 image-mask pairs:

Subset	Year / date	Images	Role in the paper
2022-seg	2022 / 2022-10-10	67	Historical manually annotated samples; used in the small-sample training and validation setting
2023-seg	2023 / 2023-11-17	67	Historical manually annotated samples; used together with 2022 samples to form the initial calibration set
2024-seg	2024 / 2024-5-17	67	Manually annotated same-region temporal test set
2024-pseudo	2024	67	Pseudo-label set generated for cross-temporal pseudo-label self-training

In the manuscript, the 2022-2023 historical samples are used as the small-sample initial calibration set, while the 2024 samples are reserved for same-domain temporal generalization testing. The 2022-seg folder additionally contains an explicit train.txt / val.txt split with 53 training samples and 14 validation samples.

Directory Structure

The dataset follows a Pascal VOC-like directory convention:

datasets/
  2022-seg/
    JPEGImages/              # RGB image patches
    SegmentationClass/       # class masks
    SegmentationClassRGB/    # RGB class masks
    SegmentationClassVis/    # visualization masks
    SegmentationObject/      # object-style masks / exported masks
    ImageSets/Segmentation/
      default.txt
      train.txt
      val.txt
    labelmap.txt

  2023-seg/
    JPEGImages/
    SegmentationClass/
    SegmentationObject/
    ImageSets/Segmentation/default.txt

  2024-seg/
    JPEGImages/
    SegmentationClass/
    SegmentationObject/
    ImageSets/Segmentation/default.txt

  2024-pseudo/
    JPEGImages/
    SegmentationClass/

  processed/
    2022-10-10/
    2023-11-17/
    2024-5-17/

Label Definition

The semantic label definition is:

Class ID	Class name	RGB color
0	background	0, 0, 0
1	vegetation	144, 32, 192

The original label map is stored in:

datasets/2022-seg/labelmap.txt

Note: label encodings are not perfectly identical across all exported folders. In 2022-seg/SegmentationClass, masks are single-channel label images with values 0 and 1. In some 2023/2024 exported masks, SegmentationClass is RGB-encoded. In 2024-pseudo/SegmentationClass, pseudo-label masks are single-channel images with values 0 and 255 and have a size of 1024 x 1024. Users should normalize mask values before training, for example by mapping all non-background vegetation pixels to class ID 1.

Intended Use

This dataset is designed for:

• small-sample vegetation semantic segmentation;
• fine-tuning and evaluating vision foundation models such as SAM2, SAM2.1, MobileSAM, and MobileSAMV2;
• comparing classic supervised segmentation models such as U-Net and DeepLabV3+;
• testing cross-temporal generalization under seasonal, illumination, and texture changes;
• studying pseudo-label or semi-supervised learning strategies for remote sensing vegetation monitoring.

It can support practical applications such as urban green-space assessment, forest resource monitoring, vegetation-cover change analysis, ecological environment evaluation, and high-resolution land-cover mapping.

Annotation Notes

All manually labeled masks were produced by the dataset creator using CVAT. The target category is vegetation, including tree canopy, forested areas, shrub-like vegetation, and visible green vegetation regions in the remote sensing image patches. Non-vegetation surfaces such as buildings, roads, bare soil, water, shadows, and other artificial or natural background regions are treated as background.

Because the dataset is intentionally small and high-resolution, it contains challenging boundary cases such as vegetation-building adjacency, crown shadows, fragmented green areas, seasonal appearance differences, and mixed background textures. These characteristics make it suitable for evaluating model robustness under annotation-scarce conditions.

Pseudo-Label Set

The 2024-pseudo subset is provided for experiments related to Cross-Temporal Pseudo-Label Self-Training (CT-PLST). In the accompanying study, high-confidence pseudo-labels from cross-temporal imagery were used to reduce boundary omissions and improve model stability under small-sample supervision.

When using 2024-pseudo, please clearly distinguish pseudo-labels from manually annotated ground-truth masks in your training protocol and reporting.

Suggested Preprocessing

For semantic segmentation training, a typical preprocessing pipeline is:

1. Load RGB images from JPEGImages/.
2. Load the corresponding mask from SegmentationClass/ or the desired mask folder.
3. Convert RGB or 0/255 masks into binary class IDs:

mask = (mask > 0).astype("uint8")

4. Resize images and masks consistently if a model requires a fixed input size.
5. Keep nearest-neighbor interpolation for masks to avoid creating invalid class values.

Citation

If you use this dataset, please cite the associated paper and this dataset release. Replace the placeholders below after the paper or dataset DOI is available.

@article{hu2026zijinshan_vegetation_segmentation,
  title={Comparative Performance Analysis of Mainstream Deep Learning and Vision Foundation Models for Small-Sample Vegetation Segmentation in High-Resolution Remote Sensing Imagery},
  author={Hu, Le and Zhang, Fuquan},
  journal={Remote Sensing},
  year={2026},
  note={Manuscript under review}
}

@dataset{hu2026zijinshan_dataset,
  title={Zijinshan Vegetation Segmentation Dataset},
  author={Hu, Le and Zhang, Fuquan},
  year={2026},
  publisher={Zenodo/Hugging Face/GitHub},
  doi={TODO: add DOI after release},
  url={TODO: add dataset URL after release}
}

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0) unless otherwise specified in the final dataset release page.

Users may share and adapt the dataset for non-commercial research and educational purposes, provided that appropriate credit is given. Before redistribution or commercial use, please confirm the usage rights of the underlying remote sensing imagery and contact the dataset authors if needed.

Contact

For questions about the dataset, please contact:

• Hu Le, College of Information Science and Technology & Artificial Intelligence, Nanjing Forestry University
• Corresponding author: Fuquan Zhang, zfq@njfu.edu.cn