metadata

task_categories:
  - image-classification
tags:
  - biology
  - environment
  - plants
  - animals
  - insects
size_categories:
  - 1K<n<10K

Dataset Card for Dataset Name

This dataset is comprised of 1,079 observations that were posted on the iNaturalist app. iNaturalist is a website and mobile app that 'aims to provide a crowd-sourced identification system' for plants, insects, and animals.

Dataset Details

Dataset Description

For each of the 1,079 observations included in this dataset, there is information about the quality of the associated image (quality_grade), a species label (species_guess), a date of observation (observed_on), the location it was observed (both latitude/longitude coordinates and a place name, location and place_name, respectively), a name that usually matches the species guess but sometimes provides other information such as the genus (name), the number of times that species has been observed before on the app (observations_count), a higher-order taxonimic classification (taxon), a common name for the species/genus/class (common_name), a link to a relevant wikipedia article (wikipedia_link), a url to a photo from that observation (photo), and an embedding (embedding). \

*Note: the embeddings were created using the ResNet50 model with imagenet weight, the top layer removed, and average pooling. They are in the dataset as strings of varying length and should be converted to arrays of shape (2048,) by splitting at ',' before use for modelling.

Dataset Sources

API: Data originally accessed via the iNaturalist API at https://api.inaturalist.org

Uses

Direct Use

Example use: This dataset may be used for training and testing image classification models as in this example (https://colab.research.google.com/drive/19yogkKM7Sxuk92LuC5kkNiZGOWeaH2TH?usp=sharing).

Dataset Structure

Columns:

'quality_grade': string with possible values 'needs_id', 'research', or 'casual'; describes the quality of the photograph associated with the observation
'species_guess': string that describes the species (or sometimes a higher order taxa) of the observation
'observed_on': string that contains the date of the observation (yyyy-mm-dd)
'location': string that contains the geographical coordinates of where the observation was made (latitude, longitude)
'place_name': string that describes the location where the observation was made (not in standard format)
'name': string containing some higher order taxonomic classification of the observation (e.g. family, genus)
'observations_count': integer value describing how many times that species has been recorded on iNaturalist
'taxon': string containg the kingdom or class of the observation
'common_name': string containing a common name for the observation (sometimes the species, sometimes a higher order taxon; only included in 70% of the data)
'wikipedia_link': string url to a relevant wikipedia article
'photo': string url to a photograph associated with the observation
'embedding': strings of varying lengths created from ResNet50 embeddings (should be converted into arrays of shape (2048,) before use)

The dataset has a 80/20 training/testing split (training set contains 863 rows, testing set contains 216 rows) created using the train_test_split from sklearn.

Dataset Creation

This dataset was created from data pulled from the iNatualist API (linked above). 9,990 requests were made to the API, then filtered to only include those observations that had a photograph and a species guess in the Latin alphabet. There were 1,079 remaining observations, all of which are included in this dataset. \

All of the information aside from the embeddings was pulled directly from the API. The process of creating the embedding column is described above (see: Dataset Description).

Curation Rationale

While iNaturalist readily provides this data for public use via an Amazon Web Services (AWS) library, the mentioned API, and for direct download, all of these methods take time to obtain and the data then requires processing before it can be used. This dataset takes a subset of iNaturalist's data and filters/processes it for immediate use, including embeddings of the photos that can be used for training or testing image classification models.

Data Collection and Processing

The data was originally downloaded from the iNaturalist API (linked above) on April 22nd, 2024. The criteria used for filtering the data is described above (has a photo and has a species guess in the Latin alphabet).

Who are the source data producers?

The data was downloaded from iNaturalist. It was uploaded to iNaturalist by individual users of the app from all over the world.

Personal and Sensitive Information

This dataset includes specific geographical coordinates, but no user data. The iNaturalist API allows for download of more specific user data, but this dataset does not include any. Additionally, iNaturalist jitters geographical coordinates of specific species for various reasons (e.g. they are protected/endangered) before making data available to viewers.