Mapping Wildlife Species Distribution With Social Media: Augmenting Text Classification With Species Names (Short Paper)

Shelan S. Jeawak, Christopher B. Jones, Steven SchockaertGIScience 2018 · OpenReview

TL;DR

This paper explores using social media data (specifically Flickr tags) to map wildlife species distributions. They found that a combined approach, using both explicit species name tags and a broader text classifier that considers all tags, significantly outperforms methods relying on either alone, offering a more accurate and comprehensive way to track biodiversity.

What Problem Does This Paper Address?

Mapping the distribution of wildlife species is crucial for conservation efforts, but traditional methods can be time-consuming and expensive. Citizen science initiatives have shown promise, and social media platforms like Flickr, with their vast amounts of geocoded and tagged user-generated content, offer a rich, albeit noisy, source of potential biodiversity data. The challenge lies in effectively extracting reliable species distribution information from this unstructured data, distinguishing genuine observations from incidental mentions and leveraging the full richness of the associated metadata.

Methods

The authors propose and compare three methods for predicting species distribution using Flickr data:

  1. Species Names: This baseline method predicts a species' presence in a region if at least one Flickr photo from that region is tagged with the species' common or scientific name.
  2. All Flickr Tags (SVM1): This approach uses a machine learning classifier (Support Vector Machine - SVM) that considers all tags associated with geocoded Flickr photos. It employs Positive Pointwise Mutual Information (PPMI) to weight the importance of each tag in relation to a specific geographic cell, effectively creating a feature vector for each cell based on its tag content.
  3. Meta Features (SVM2): This is the proposed novel method. It builds upon the "All Flickr Tags" approach by creating a meta-classifier. This classifier combines the confidence score from the SVM1 (based on all tags) with two additional binary features: (a) whether the species' actual name (common or scientific) is present as a tag in the cell, and (b) the percentage of neighboring cells that contain the species' name as a tag. This aims to leverage both the broad environmental context from all tags and the specific signal from explicit species name tags.

Datasets

  • Ground Truth: The National Biodiversity Network (NBN) Atlas, which provides species distribution data for the UK and Ireland.
  • Social Media Data: Approximately 12 million georeferenced Flickr photographs collected within the UK and Ireland in September 2015. The analysis focuses on the tags associated with these photographs.

The study considers a random sample of 50 bird species from the NBN Atlas that have at least 100 observations and whose names appear in at least 100 Flickr photographs.

Key Insights and Ideas

  • Social media is a valuable, but noisy, source: Flickr tags contain a wealth of information that can be used to infer species presence, but it's not a direct one-to-one mapping of observations.
  • Explicit species names are precise but incomplete: Relying solely on photos tagged with a species' name yields high precision (fewer false positives) but suffers from low recall (misses many actual occurrences).
  • All tags provide broader context: Using all tags in a text classifier captures more information and generally leads to better performance than just species names, likely by inferring habitat or associated environmental cues.
  • The synergy of explicit names and broad context is key: The most significant insight is that combining the specific signal of species name tags with the broader environmental context derived from all tags (the meta-classifier) yields the best results. This suggests that a balanced approach is superior.
  • Performance varies by species: For very common species, the "All Flickr Tags" method performs well. However, for rarer species or those with specific habitat requirements, the "Species Names" method can sometimes outperform the broader tag-based approaches, highlighting the need for adaptive strategies.

Key Results

The paper reports quantitative results across different grid cell sizes (10km, 20km, 30km) using metrics like Accuracy, Precision, Recall, F1 Score, and Area Under the ROC Curve (AUC).

  • Meta Features (SVM2) consistently outperformed: The proposed meta-classifier achieved the highest F1 scores and AUC values across all grid cell sizes. For example, at 30km cell size, it reached an F1 score of 0.905 and an AUC of 0.952.
  • "All Flickr Tags" (SVM1) was a strong contender: This method significantly outperformed the "Species Names" baseline, achieving an F1 score of 0.795 at 30km cell size.
  • "Species Names" had high precision but low recall: While it often had high precision (e.g., 0.970 at 30km), its recall was considerably lower (e.g., 0.384 at 30km), leading to lower overall performance metrics like F1 score.
  • Visualizations confirm findings: Figures 2 and 3 show that the meta-classifier's predictions align well with ground truth data, often capturing distributions more accurately than the other methods.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that even with the meta-classifier, there are cases where the "Species Names" method performs better, particularly for less common species or those with niche habitats. This suggests that the model might struggle with sparse data or highly specific environmental contexts.
  • The paper focuses on tags and doesn't delve into the actual image content for validation, relying on the assumption that tags are reasonably indicative of species presence.
  • Future Work:
    • The authors suggest investigating the use of text classifiers to estimate confidence in individual social media postings. This could help filter out postings that mention a species name but in a context unrelated to its actual occurrence (e.g., a discussion about a bird rather than an observation).
  • Exploring how to better handle species that occur in specific geographic contexts or are less common.

Summary updated

Abstract

Social media has considerable potential as a source of passive citizen science observations of the natural environment, including wildlife monitoring. Here we compare and combine two main strategies for using social media postings to predict species distributions: (i) identifying postings that explicitly mention the target species name and (ii) using a text classifier that exploits all tags to construct a model of the locations where the species occurs. We find that the first strategy has high precision but suffers from low recall, with the second strategy achieving a better overall performance. We furthermore show that even better performance is achieved with a meta classifier that combines data on the presence or absence of species name tags with the predictions from the text classifier.
BibTeX
@inproceedings{jeawak2018mapping,
  author = {Shelan S. Jeawak and Christopher B. Jones and Steven Schockaert},
  title = {{Mapping Wildlife Species Distribution With Social Media: Augmenting Text Classification With Species Names (Short Paper)}},
  booktitle = {GIScience},
  year = {2018},
  url = {https://openreview.net/forum?id=bVtCo1vWeL},
}
Figure