Wavelet Based Machine Learning Algorithms for Wildfire Prediction

It is estimated that there are 70,000 wildfires every year in the United States alone. These wildfires pose a significant danger to humans, both from property damage and health implications; air quality can drastically decrease from the resulting smoke, leading to over 33000 premature deaths each year [1]. As a result, it is imperative that improvements to wildfire prediction models be made. Previous research has demonstrated the effectiveness of Machine Learning (ML) methods such as XGBoost and Convolutional Neural Network (CNN). In this research, we utilize wavelet analysis to preprocess our data by applying a discrete wavelet transform (DWT) in order to increase the accuracy of our models in predicting the next day's wildfire danger in certain counties in California. We input this DWT transformed data into an XGBoost model and compare the model performance to a traditional model (one trained on non-DWT data). We also apply DWT on aerial images data set from Canada and input the results into a convolutional neural network (CNN) to investigate the effects of DWT transformed image data and to determine if such wavelet based CNN can accurately distinguish between days with and days without a wildfire. We find that the use of DWT significantly improves model performance by denoising the data as well as adding more information in the form of hidden features for the model to train on.