AerialWaste dataset

AerialWaste is a dataset for the discovery of illegal landfills. Illegal landfills from aerial images present a visual heterogeneity of the scenes in which waste dumps appear and present a diverse nature of the objects that compose a waste deposit. When observed from above, waste dumps appear as complex arrangements of objects of different shapes, sizes, and orientation.

Visit our site for more details: https://aerialwaste.org/

Repository content

This repository contains a series of utility scripts to handle the dataset:

• Link to dataset: Download images from Zenodo[https://zenodo.org/record/7034381], put the training.json and testing.json containing the images medatadata in the root of this repository and put all images into an images folder.
• Statistics: to plot statistics on the dataset.
• Visualizer: to visualize the images with its correspondent classes and segmentation masks. Download and unzip the image folder, then install the ODIN visualizer tool (https://github.com/rnt-pmi/odin).
• DataLoader: to convert the JSON to tensors containing the image classes and the image itself. Download and unzip the image folder before its use.

JSON

For each image the following information is provided:

{
  "id": --> the id of the image
  "file_name": --> name of the image file in the image folder
  "is_candidate_location":--> if this is a candidate location (1-positive) or not (0-negative)

  "evidence": --> evidence perceived by the analyst at annotation time (from 0 to 3) [only if candidate location]
  "severity": --> severity perceived by the analyst at annotation time (from 0 to 3) [only if candidate location]

  "width": --> image width in pixels
  "height": --> image height in pixels

  "site_type": --> type of site (e.g. production area) [only if candidate location]

  "is_valid_fine_grain": --> if this image was analyzed to observe the Waste Objects or Storage modes present
  "categories": --> which of the different Waste Objects or Storage Modes are present on the images [only if is_valid_fine_grained]
}

Segmentation masks are present in the JSON following the COCO format.

License

Creative Commons CC BY licensing scheme (see LICENSE). The usage of Google Imagery must respect the Google Earth terms and conditions [https://about.google/brand-resource-center/products-and-services/geo-guidelines/].

Cite us

@article{torres2023aerialwaste,
  title={AerialWaste dataset for landfill discovery in aerial and satellite images},
  author={Torres, Rocio Nahime and Fraternali, Piero},
  journal={Scientific Data},
  volume={10},
  number={1},
  pages={63},
  year={2023},
  publisher={Nature Publishing Group UK London}
}

Visit our site for more details: https://aerialwaste.org/