[Paper Review] Animal Wildlife Population Estimation Using Social Media Images Collections

Matteo Foglio, Lorenzo Semeria+3 more|arXiv (Cornell University)|Aug 5, 2019

Species Distribution and Climate Change4 citations

TL;DR

This paper proposes a novel framework to estimate wildlife population sizes using social media image collections, accounting for user sharing bias by modeling image collection structure and shareability. Using Grevy’s zebra as a case study, it applies XGBoost to predict unshared animal counts from shared images, achieving an R² of 0.417 and a population estimate close to official figures, demonstrating that social media bias is learnable and correctable for conservation monitoring.

ABSTRACT

We are losing biodiversity at an unprecedented scale and in many cases, we do not even know the basic data for the species. Traditional methods for wildlife monitoring are inadequate. Development of new computer vision tools enables the use of images as the source of information about wildlife. Social media is the rich source of wildlife images, which come with a huge bias, thus thwarting traditional population size estimate approaches. Here, we present a new framework to take into account the social media bias when using this data source to provide wildlife population size estimates. We show that, surprisingly, this is a learnable and potentially solvable problem.

Motivation & Objective

• To address the critical lack of reliable wildlife population data due to the limitations of traditional monitoring methods.
• To develop a method that corrects for social media sharing bias in wildlife image collections, which distorts population estimates.
• To model both the number of animals in a photographer's original image collection and the likelihood of images being shared.
• To demonstrate that image shareability and collection structure are learnable signals for improving population estimation.
• To provide a scalable, species-agnostic framework for population estimation using publicly available social media data.

Proposed method

• The framework treats the population estimation problem as a regression task, using features from shared images and collection structure to predict the total number of animals in a photographer's original SD card collection.
• It models image shareability as a binary classification problem, using features such as animal individuality, image beauty, and collection context to predict whether a photo would be shared.
• A survey was conducted on Grevy’s zebra image sets from 2016 and 2018 to label shareability, creating ground truth for training models.
• The method uses XGBoost for both regression and classification tasks, with cross-validation to evaluate performance on R², MSE, F1-score, and accuracy.
• The final population estimate is derived using the Jolly-Seber capture-mark-recapture method, adjusted by an average sharing coefficient to account for undetected recaptures.
• The framework computes a correction factor $ k_i $ per user based on predicted shareability, scaling the number of shared animals to estimate the total photographed.

Experimental results

Research questions

• RQ1 Can social media image collections be reliably used to estimate wildlife population sizes despite significant user sharing bias?
• RQ2 To what extent does the structure of an image collection (e.g., number of animals, diversity) influence image shareability?
• RQ3 Can machine learning models accurately predict the number of animals in a photographer’s original image set from only the shared subset?
• RQ4 How effective is the combination of image features and collection-level features in modeling user sharing behavior?
• RQ5 Can a learnable bias correction improve population estimates compared to naive counts of shared images?

Key findings

• XGBoost achieved an R² of 0.417 (±0.155) and MSE of 0.070 (±0.020) on the regression task, significantly outperforming baseline models that scored negative R².
• For the shareability classification, XGBoost achieved 82% accuracy and 0.62 F1-score on the combined dataset, and 92% accuracy with 0.77 F1-score on the more regular GGR1 dataset.
• The framework estimated the Grevy’s zebra population at 2,341 in 2016, close to the official count of 2,250, with a root mean square error (RMSE) of 1,194.
• The model's performance was higher on GGR1 (more regular data) than on the combined dataset, indicating that data structure affects learnability.
• The study demonstrates that social media bias in wildlife image sharing is learnable and correctable using collection-level features and machine learning.
• The final population estimate using the Jolly-Seber method and a correction coefficient yielded a plausible result, validating the framework’s potential for conservation applications.