Wildfire risk assessment using deep learning in Guangdong Province, China
•A wildfire dataset applicable to Guangdong Province was constructed, containing over 11,000 historical wildfire incidents. A risk indicator system was implemented to categorize these incidents with four wildfire driven factors and 14 quantitative variables.•A deep learning-based CNN model was desig...
Saved in:
Published in | International journal of applied earth observation and geoinformation Vol. 128; p. 103750 |
---|---|
Main Authors | , , , , , , , , |
Format | Journal Article |
Language | English |
Published |
Elsevier B.V
01.04.2024
Elsevier |
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | •A wildfire dataset applicable to Guangdong Province was constructed, containing over 11,000 historical wildfire incidents. A risk indicator system was implemented to categorize these incidents with four wildfire driven factors and 14 quantitative variables.•A deep learning-based CNN model was designed to extract high-order features under the coupling of various types of wildfire-driven factors, enabling accurate prediction of the probability of wildfire occurrence.•The wildfire risk map in Guangdong Province was constructed using the proposed CNN model. The spatiotemporal characteristics of wildfire occurrence and the contribution of wildfire driven factors in different seasons were analyzed, respectively. These experimental results serve to assist emergency agencies in efficiently managing potential wildfire risks.
The severe wildfires that have ravaged Guangdong province, China, present a significant threat to the local ecosystem, socio-economics, and public health. Effective risk assessment is essential for early warning and timely prevention in wildfire management, thereby mitigating disaster losses. In this study, we compiled a dataset comprising 11,507 historical wildfire incidents in Guangdong Province spanning a decade (2011–2021) and developed a deep learning-based model to predict the likelihood of wildfire occurrence in the region. In addition to analyzing risk characteristics throughout the year, we also trained separate models for different seasons and analyzed the discrepancies in the contribution of driven factors to wildfire occurrence across seasons. Furthermore, the performance of our deep learning-based model was compared with that of traditional machine learning algorithms. The experimental results revealed that: (1) Factors such as relative humidity, temperature, NDVI, and precipitation exerted significant influence on wildfire occurrence. (2) The impact of wildfire driving factors varied across different seasons. (3) Our deep learning model outperformed traditional machine learning models, achieving a superior performance with an AUC of 0.962. |
---|---|
ISSN: | 1569-8432 1872-826X |
DOI: | 10.1016/j.jag.2024.103750 |