Identifying Occurrences of the Cnidarian Physalia physalis in Social Media Data

Abstract

The Portuguese man-of-war (Physalia physalis), though beautiful, poses a risk to the population due to its potential to cause severe burns. Tracking their occurrences can prevent accidents through alerts to the population and predictive simulation models. However, traditional sources do not always provide records of their sightings. On the other hand, previous studies indicate that social media can be an effective source of information for conservation science. This work uses natural language processing and computer vision to obtain machine learning models to classify data extracted from Instagram. Such models can be used as part of an automated Extract-Transform-Load process to build a database on occurrences of Physalia physalis on the Brazilian coast. In preparation for training the models, we collected and manually labeled Instagram posts in order to distinguish the ones about the animal from other subjects, such as ships and tattoos. Given the nature of the problem, the spatial and temporal information associated with the sightings are essential for biologists. Thus, the absence or nonvalidity of such data is often used as a rationale to reject the post. However, the same criteria may not be suitable for training machine learning models to classify new posts automatically. The main goal of this article is to highlight the importance of choosing appropriate labels to train both text and image models, as well as to take into consideration the rejection criteria of the biologist before using a classification model. An experimental study is presented to show the effect of unquestioning adoption of labels given by a specialist, compared to labels adapted for machine learning training.