Manually Annotated High Resolution Satellite Image Dataset of Mumbai for Semantic Segmentation

Data Description

The dataset is divided into two groups, each including satellite images and their corresponding semantic segmentation masks. The images in each group are of two different sizes. The first group comprises 110 satellite images of size 600×600 pixels and manually labelled semantic segmentation masks. Figure 1 depicts some sample images and their labelled semantic segmentation masks from the first group. The second group contains three sets: training, validation, and testing. Each set consists of images of size 120×120 pixels which are derived and processed from the first group. Researchers can employ it to train and evaluate machine learning models. The first group is provided in the Main_Dataset directory, and the second group is provided in the Prepared_Dataset directory. All the image-mask pairs were named with the same file names, for example, tile_5.37.tif and tile_5.37.png. For the sample images, directory tree and description of the dataset, see Figure 1, Figure 2 and Table 1, respectively.

Figure 1. Original satellite images (a, c, e, g) and their corresponding segmentation masks (b, d, f, h).

Figure 2. The directory structure of the dataset.

Table 1. Overview of the dataset directory structure and files.

File/Directory Path	Description
Dataset/Main_Dataset/images/	Directory of original satellite images of 600×600 px size
Dataset/Main_Dataset/masks/	Directory of labelled masks of 600×600 px size
Dataset/Main_Dataset/class_dict.csv	CSV file containing RGB color codes of classes
Dataset/Main_Dataset/train_files.csv	CSV file containing the file names of the image-mask pairs used in the Prepared_Dataset's training set.
Dataset/Main_Dataset/val_files.csv	CSV file containing the file names of the image-mask pairs used in the Prepared_Dataset's validation set.
Dataset/Main_Dataset/test_files.csv	CSV file containing the file names of the image-mask pairs used in the Prepared_Dataset's testing set.
Dataset/Prepared_Dataset/train/images/	Directory of training set images
Dataset/Prepared_Dataset/train/masks/	Directory of training set masks
Dataset/Prepared_Dataset/val/images/	Directory of validation set images
Dataset/Prepared_Dataset/val/masks/	Directory of validation set masks
Dataset/Prepared_Dataset/test/images/	Directory of test set images
Dataset/Prepared_Dataset/test/masks/	Directory of test set masks

Experimental Design, Material, and Methods

The dataset was created from high-resolution, true-color satellite imagery of Pleiades-1A acquired on March 15, 2017. Pleiades is an Airbus product that provides imagery with a 0.5m resolution at different spectral combinations. A total of 110 patches of size 600×600 pixels were selected by visually eyeballing random locations in the city that contain a wide variety of urban characteristics such as vegetation, slums, built-up, roads, etc. The patches were then manually labelled with polygons using Intel's Computer Vision Annotation Tool (CVAT). Six unique classes were used to categorize the images, namely (1) vegetation; (2) built-up; (3) informal settlements; (4) impervious surfaces (roads/highways, streets, parking lots, road-like areas around buildings, etc.); (5) barren; and (6) water. In addition to these six major classes, the dataset also contains another class termed ‘Unlabelled’, which makes up only 0.08% of the dataset. It primarily consists of airplanes and a few other obscure spots and structures. The ground truth's color legend is described in Table 2 below. Each 600×600 pixels patch was further divided into 120×120 pixels sized tiles with 50% horizontal and vertical overlapping (see Figure 3), making a total of 8910 tiles. This helped in the generation of more training data that would result in better classification. Out of the total 8910 labelled patches, 80% patches (total: 7128) are present in the training set, 10% as the validation set (total: 891), and the remaining 10% for testing (total: 891). Table 3 illustrates the distribution of the classes in the datasets.

Table 2. The semantic classes, their RGB codes, and respective colors in the segmentation masks.

Class	RGB Values	Color
Built-Up	200, 200, 200
Vegetation	80, 140, 50
Barren	200, 160, 40
Water	40, 120, 240
Impervious Surfaces	100, 100, 150
Informal Settlements	250, 235, 185
Unlabelled	0, 0, 0

Figure 3. Tile preparation methodology of Prepared_Dataset: 50% horizontal and vertical overlapping.

Table 3. Distribution of labels in the main dataset and prepared dataset (in percentages).

Dataset		Semantic Classes
		Built-Up	Vegetation	Barren	Water	Impervious Surfaces	Informal Settlements	Unlabelled
Main Dataset		9.49	18.82	15.89	31.86	12.15	11.7	0.08
Prepared Dataset	Training	9.72	19.2	16.04	32.36	11.51	11.06	0.1
	Validation	5.8	13.98	12.74	31.26	15.49	20.73	0
	Testing	7.86	19.93	17.36	28.82	16.84	9.19	0

Access

The dataset is freely available on Mendeley Data.

Citing this Work

This dataset was used in our research paper, which was published in Neural Computing & Applications, Springer. It is available here.

If you use this dataset in a scientific publication, please consider citing:

Dabra, Ayush; Upadhyay, Vaibhav (2023), “Manually Annotated High Resolution Satellite Image Dataset of Mumbai for Semantic Segmentation”, Mendeley Data, V1, doi: 10.17632/xj2v49zt26.1

Dabra, A., Kumar, V. Evaluating green cover and open spaces in informal settlements of Mumbai using deep learning. Neural Comput & Applic (2023). https://doi.org/10.1007/s00521-023-08320-7

@misc{https://doi.org/10.17632/xj2v49zt26.1,
  doi = {10.17632/XJ2V49ZT26.1},
  url = {https://data.mendeley.com/datasets/xj2v49zt26/1},
  author = {{Ayush Dabra}},
  title = {Manually Annotated High Resolution Satellite Image Dataset of Mumbai for Semantic Segmentation},
  publisher = {Mendeley},
  year = {2023},
  copyright = {Creative Commons Attribution 4.0 International}
}

@article{Dabra2023,
  doi = {10.1007/s00521-023-08320-7},
  url = {https://doi.org/10.1007/s00521-023-08320-7},
  year = {2023},
  month = feb,
  publisher = {Springer Science and Business Media {LLC}},
  author = {Ayush Dabra and Vaibhav Kumar},
  title = {Evaluating green cover and open spaces in informal settlements of Mumbai using deep learning},
  journal = {Neural Computing and Applications}
}

License

This dataset is released under the CC BY-NC-SA 4.0 license.

Contact

For comments and feedback, contact the lead researcher Ayush Dabra at ayushd19@iiserb.ac.in or the principal investigator Vaibhav Kumar at vaibhav@iiserb.ac.in.