Estimating environmental vulnerability in the Cerrado with machine learning and Twitter data

Estimating vulnerability is critical to understand human-induced influenceimpacts on the environmental system. The purpose of the current study was to integrate machine learning algorithm and Twitter data to estimate environmental vulnerability in the Brazilian Cerrado for the years 2011 and 2016. We first selected six exposure indicators and five sensitivity indicators to build an environmental vulnerability model and applied an Autoencoder algorithm to find the representation of exposure and sensitivity, respectively. Then the Displaced Ideal method was used to estimate environmental vulnerability. Finally, related historical Twitter data was mined from these two years to validate the results. The findings showed that the percent of land classified as areas of low, medium and high environmental vulnerability were 6.72%, 34.85%, and 58.44% in 2011 and 3.45%, 33.68% and 62.87% in 2016, respectively and most high environmental vulnerability areas were in the Southern Cerrado. Moreover, the Twitter data results showed that more than 85% of tweets occurred in the areas considered as high environmental vulnerability class. The work revealed that the Autoencoder algorithm can be used for environmental assessment, and the social media data has potential to effectively analyze the relationship between human activity and the environment. Although the study provided a novel perspective to estimate environmental vulnerability at a regional scale, it was necessary to develop a more comprehensive indicator system that can improve model performance in the future.