Researchers at Imperial College London have developed a new machine learning model that uses social media data to predict and monitor wildfires more accurately in real time

As the use of social media has increased over the past decade, people have developed a tendency to write about events around them on social media. Thus, there are large amounts of data containing a lot of information, for example, reporting disasters, spreading diseases, etc. In recent years, scientists and researchers have also harnessed social media to create systems to detect and predict natural disasters. A group of scientists from Imperial College, UK, have created the first-ever real-time wildfire detection system using data from physical sensors and social media data.

Researchers used data from Twitter to create a wildfire dataset using advanced query filtering and optimization methods. Next, they performed a sentiment analysis of this Twitter data to extract human sentiments associated with the wildfires. The result is a robust dataset containing not only regional wildfire data, but also local wildfire data. Multiple ML algorithms then process this high-quality dataset to best predict wildfires, creating a robust disaster response system.


The researchers used historical satellite wildfire data containing various physical attributes in this work. They used some of these attributes, namely latitude, longitude, and start and end times of wildfires, to generate queries for Twitter data and to select relevant tweets. For maximum tweet relevance, researchers used multiple criteria to select tweets, such as whether the tweet contained relevant keywords and hashtags. However, it could also contain loud tweets; for example, “fire”, “burn”, etc. can express several meanings, from political opinions to several other analogies. Therefore, the tweets are again filtered to remove these noises.

The filtered tweets go through NLP (Natural Language Processing) models to generate social sentiment variables, namely the sentiment score (it describes the emotional attribute of the text, a value between -1 and 1) and the magnitude of the sentiment (it describes the magnitude of the emotional attribute, the value is between 0.0 and +inf. For an individual message, the score and magnitude are calculated by averaging all the scores and the magnitude of all words.Then, ML models are used to learn mapping to predict social feelings towards a wildfire from the physical attributes of the fire and vice-versa.Researchers experimented with gradient-boosted random forest, the neural network regressor, support vector regressor, etc. in the prediction module.

Thus, the researchers demonstrated a relationship between online social sentiments and wildfire activity in this work. This would help to reliably predict forest fires, locate dangerous areas as well as emergency services. In the future, this approach can be applied to other natural events and disasters.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'Sentimental wildfire: a social-physics machine learning model for wildfire nowcasting'. All Credit For This Research Goes To Researchers on This Project. Checkout the paper and reference article.

Please Don't Forget To Join Our ML Subreddit

I am Arkaprava from Kolkata, India. I finished my B.Tech. in Electronics and Communications Engineering in 2020 from Kalyani Government Engineering College, India. During my B.Tech. I developed a keen interest in signal processing and its applications. Currently, I am pursuing MSc studies in IIT Kanpur in Signal Processing, researching Audio Analysis using Deep Learning. Currently I am working on unsupervised or semi-supervised learning frameworks for several audio tasks.

Sherry J. Basler