Machine Learning and Social Media Harvesting for Wildfire Prevention

Social media can be exploited to ease or better understand many problems in science and technology. Social media sensing become one of the methods to get the update information from crowd. Text is the most unstructured information sources available on that media comes from their users. Taking useful information from the unstructured text however, still challenging. The conversion from unstructured data to a vector need to be carried out texts analysis. Machine learning offer an opportunity overcome the challenges. However, it needs massive amount of labelled data to train the machine learning. The research community pose the fact that the availability of the dataset for non - English languages are limited. Forest fire as one of topics discussed in many social media platforms in many languages including Indonesian language. This work fills the gap by providing a labelled dataset in forest fire topics. Moreover, an exercise of the dataset with some available machine learning techniques are compared. Due to the limited number of data rows, researchers perform data augmentation and observe the impact of the augmentation towards the classifier performance. According to our experiments, we found that Random oversampling (ROS), in general, improves the accuracy score for all models we employed. Bidirectional LSTM with ROS overperforms other classifier and achieves 91%, 91% and 90.7% of recall, precision and f1-score respectively.