Predictive modeling of wildfires: A new dataset and machine learning approach

Wildfires, whether natural or caused by humans, are considered among the most dangerous and devastating disasters around the world. Their complexity comes from the fact that they are hard to predict, hard to extinguish and cause enormous financial losses. To address this issue, many research efforts have been conducted in order to monitor, predict and prevent wildfires using several Artificial Intelligence techniques and strategies such as Big Data, Machine Learning, and Remote Sensing. The latter offers a rich source of satellite images, from which we can retrieve a huge amount of data that can be used to monitor wildfires. The method used in this paper combines Big Data, Remote Sensing and Data Mining algorithms (Artificial Neural Network and SVM) to process data collected from satellite images over large areas and extract insights from them to predict the occurrence of wildfires and avoid such disasters. For this reason, we implemented a methodology that serves this purpose by building a dataset based on Remote Sensing data related to the state of the crops (NDVI), meteorological conditions (LST), as well as the fire indicator “Thermal Anomalies”, these data, were acquired from “MODIS” (Moderate Resolution Imaging Spectroradiometer), a key instrument aboard the Terra and Aqua satellites. This dataset is available on GitHub via this link (https://github.com/ouladsayadyounes/Wildfires). Experiments were made using the big data platform “Databricks”. Experimental results gave high prediction accuracy (98.32%). These results were assessed using several validation strategies (e.g., classification metrics, cross-validation, and regularization) as well as a comparison with some wildfire early warning systems.