Prediction of Forest Fire Area Using Machine Learning Algorithms

Abstract

Now-a-days occurrence of forest fires has become a common phenomenon due to multiple reasons which are either man-made or natural. The impact on land due to forest fires has been huge lately, specifically in places full of vegetation around the globe that rapidly affect ecosystems around them. This emphasises the urgent need to develop some models which will be able to predict the forest fire leading to better understanding of the components that influence the total burnt areas in specific conditions. Keeping it in mind, the work in this study aims to predict the total area burnt (in acres) due to forest fires by developing different prediction models using six Machine Learning techniques viz. Neural Network, Support Vector Machine, Multi-Layer Perceptron, K-Nearest Neighbour, Decision Tree and Stacking Regressor. Kaggle forest fire dataset has been used to conduct the empirical analysis. This original Kaggle dataset was enhanced and made enriched with the addition of relevant features which played a key role in making the dataset robust, thus leading to accurate predictions. The performance of the prediction models was evaluated using two evaluation measure viz. Mean Absolute Error and R-Squared Error. Correlation heatmaps were also used to the extract the best features from the original dataset as well as enhanced dataset. It was concluded that the prediction models have performed exceptionally well on the enhanced dataset than on the original dataset with Decision Tree technique and Multi-Layer Perceptron technique showing exceptional performances.