andyvhuynh/NatureMultiView

Nature Multi-View (NMV) Dataset Datacard

To encourage development of better machine learning methods for operating with diverse, unlabeled natural world imagery, we introduce Nature Multi-View (NMV), a multi-view dataset of over 3 million ground-level and aerial image pairs from over 1.75 million citizen science observations for over 6,000 native and introduced plant species across California.

Characteristics and Challenges

• Long-Tail Distribution: The dataset exhibits a long-tail distribution common in natural world settings, making it a realistic benchmark for machine learning applications.
• Geographic Bias: The dataset reflects the geographic bias of citizen science data, with more observations from densely populated and visited regions like urban areas and National Parks.
• Many-to-One Pairing: There are instances in the dataset where multiple ground-level images are paired to the same aerial image.

Splits

• Training Set:
  • Full Training Set: 1,755,602 observations, 3,307,025 images
  • Labeled Training Set:
    • 20%: 334,383 observations, 390,908 images
    • 5%: 93,708 observations, 97,727 images
    • 1%: 19,371 observations, 19,545 images
    • 0.25%: 4,878 observations, 4,886 images
• Validation Set: 150,555 observations, 279,114 images
• Test Set: 182,618 observations, 334,887 images

Acquisition

• Ground-Level Images:
  • Sourced from iNaturalist open data on AWS.
  • Filters applied:
  • Vascular plants
  • Within California state boundaries
  • Observations dated from January 1, 2011, to September 27, 2023
  • Geographic uncertainty < 120 meters
  • Research-grade or in need of ID (excluding casual observations)
  • Availability of corresponding remote sensing imagery
  • Overlap with bio-climatic variables
• Aerial Images:
  • Sourced from the 2018 National Agriculture Imagery Program (NAIP).
  • RGB-Infrared images, 256x256 pixels, 60 cm-per-pixel resolution.
  • Centered on the latitude and longitude of the iNaturalist observation.

Features

• observation_uuid (string): Unique identifier for each observation in the dataset.
• latitude (float32): Latitude coordinate of the observation.
• longitude (float32): Longitude coordinate of the observation.
• positional_accuracy (int64): Accuracy of the geographical position.
• taxon_id (int64): Identifier for the taxonomic classification of the observed species.
• quality_grade (string): Quality grade of the observation, indicating its verification status (e.g., research-grade, needs ID).
• gl_image_date (string): Date when the ground-level image was taken.
• ancestry (string): Taxonomic ancestry of the observed species.
• rank (string): Taxonomic rank of the observed species (e.g., species, genus).
• name (string): Scientific name of the observed species.
• gl_inat_id (string): iNaturalist identifier for the ground-level observation.
• gl_photo_id (int64): Identifier for the ground-level photo.
• license (string): License type under which the image is shared (e.g., CC-BY).
• observer_id (string): Identifier for the observer who recorded the observation.
• rs_classification (bool): Indicates if remote sensing classification data is available.
• ecoregion (string): Ecoregion where the observation was made.
• supervised (bool): Indicates if the observation is part of the supervised dataset.
• rs_image_date (string): Date when the remote sensing (aerial) image was taken.
• finetune_0.25percent (bool): Indicates if the observation is included in the 0.25% finetuning subset.
• finetune_0.5percent (bool): Indicates if the observation is included in the 0.5% finetuning subset.
• finetune_1.0percent (bool): Indicates if the observation is included in the 1.0% finetuning subset.
• finetune_2.5percent (bool): Indicates if the observation is included in the 2.5% finetuning subset.
• finetune_5.0percent (bool): Indicates if the observation is included in the 5.0% finetuning subset.
• finetune_10.0percent (bool): Indicates if the observation is included in the 10.0% finetuning subset.
• finetune_20.0percent (bool): Indicates if the observation is included in the 20.0% finetuning subset.
• finetune_100.0percent (bool): Indicates if the observation is included in the 100.0% finetuning subset.
• gl_image (image): Ground-level image associated with the observation.
• rs_image (sequence of sequences of int64): Aerial image data associated with the observation, represented as a sequence of pixel values.

References

• iNaturalist: www.inaturalist.org
• United States Department of Agriculture: NAIP Imagery. www.naip-usdaonline.hub.arcgis.com.