EGU General Assembly 2023
© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

A hybrid deep learning framework for predicting point-level Alaskan fires

Hocheol Seo and Yeonjoo Kim
Hocheol Seo and Yeonjoo Kim
  • Department of Civil and Environmental Engineering, Yonsei University, Seoul 03722, South Korea (

Fires in high latitudes are becoming more critical in terrestrial ecosystem modeling. With climate warming and dry weather condition, the fires have spread more, and widespread burning has severely damaged the ecosystem. As the fire dynamics cannot be described with the mass or energy balance equations, the fire models have been developed with different input variables, linked with different vegetation models, and widely coupled with the earth system models (ESMs) or land surface models (LSMs) with different complexities of parameterization. Here, we designed a new approach using hybrid deep learning [Long Short-Term Memory (LSTM) - Artificial Neural Network (ANN)] for predicting Alaskan natural fires and aimed to understand the impacts of fires with from the NCAR community land model 5 – biogeochemistry (CLM5-BGC). This study was conducted based on fire information provided by Alaska Interagency Coordination Center (AICC), which provides the data for each fire point, start date, end date, and total burned area from 2016-2020. As the fire duration was identified as the most important in predicting the burned area, we first trained the LSTM for predicting fire duration (i.e., fire ignition and fire persistence period) with ERA5 atmospheric forcings. Also, we trained ANN to predict the burned area with both ERA5 atmospheric forcings and fire duration. Then, we combined two models (LSTM and ANN) to simultaneously predict the fire days and burned area with climate and vegetation datasets. This hybrid model has the strength to capture large fires (>10000ha), comparing the burned area from CLM5-BGC (Correlation: 0.79). When this hybrid model is coupled with CLM5-BGC, we found that the carbon fluxes changed over Alaska. In particular, total net ecosystem exchange (NEE) increased by more than two times that of only CLM5-BGC, which could primarily affect terrestrial carbon exchanges.


This work was supported by the Korea Polar Research Institute (KOPRI, PE22900) funded by the Ministry of Oceans and Fisheries and the Basic Science Research Program through the National Research Foundation of Korea, which was funded by the Ministry of Science, ICT & Future Planning (grant no. 2020R1A2C2007670) and by the Ministry of Education (2022R1A6A3A13073233).

How to cite: Seo, H. and Kim, Y.: A hybrid deep learning framework for predicting point-level Alaskan fires, EGU General Assembly 2023, Vienna, Austria, 24–28 Apr 2023, EGU23-10336,, 2023.