Wildfire Progression Prediction and Validation Using Satellite Data and Remote Sensing in Sonoma, California

Publication Date

1-1-2023

Document Type

Conference Proceeding

Publication Title

Proceedings - 17th IEEE International Conference on Service-Oriented System Engineering, SOSE 2023

DOI

10.1109/SOSE58276.2023.00037

First Page

262

Last Page

271

Abstract

Wildfire has been a serious disaster in the United States for a long time with increasing devastation power, and occurrence risk. Compared to wildfire detection in which numerous kinds of research have been conducted, there is limited research on real-time wildfire progression. To bridge this gap, this paper proposes to develop a dynamic daily wildfire progression model using deep reinforcement learning, which can help to track and contain the wildfire. This paper focuses on wildfires that occurred in Sono1ma County for model development training and testing purposes. However, the proposed approach is not limited to Sono1ma County and can be applied to any location if the data is available. This research collected various input data, including weather data, land cover data, land fuel remote sensing data, and fire history data, for the proposed models. The proposed models contain three major components: the wildfire detection model, the land fuel vegetation data classification model, and the wildfire progression prediction model using deep reinforcement learning. The research conducts comparative experiments to evaluate various integrated deep learning models and different combinations of deep learning approaches. Specifically, the wildfire progression prediction model uses DQN with MLP, A2C with MLP, DQN with CNN, and an innovative model incorporated with the Recurrent neural network LSTM (long short-term memory) and deep reinforcement learning. The top-performing model yields an accuracy of 85.87% and an F1 score of 92.17%. Considering the importance and potential devastation caused by wildfires, this paper introduces an intelligent solution consisting of three interconnected models for wildfire detection, land fuel vegetation confirmation, and progression prediction. The approach relies primarily on Deep Learning Models to ensure high accuracy in detecting and predicting the spread of wildfire, making it applicable to real-world scenarios for effective wildfire management.

Keywords

machine learning, reinforcement learning, remote sensing, satellite images, wildfire

Department

Computer Engineering

Link to Full Text

Share

COinS