Introduction

Around the world, the risk and prevalence of wildfires is increasing, yet the development of effective models to predict the occurrence and spread of wildfires has remained elusive. Recently, a group of researchers attempted to address this challenge by combining Ecosystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) and wildfire event data into machine learning models, and then applying them to the historic Australian wildfires of 2019-2020. With these inputs, the researchers were able to predict over 90% of the wildfire occurrences one week in advance of this season's events. 

These results indicate that ECOSTRESS data can significantly improve wildfire prediction by capturing drought signals that precede wildland fire events. Further, the approaches the researchers documented in their paper may help policymakers, fire managers, and city planners assess, manage, prepare, and mitigate future wildfire events.

Science Objectives

The increase in the severity and intensity of wildfires around the globe has spurred concern about their impacts on terrestrial ecosystems, air quality, human health and well-being, and local and regional economies. In response, scientists and natural resource managers have sought the development of effective and reliable models to identify areas with increased fire risk to help authorities prioritize monitoring efforts and resource allocation in vulnerable regions. The effectiveness and utility of such models require the availability of high-quality terrestrial and atmospheric data.

One source of this data is the ECOSTRESS instrument, which was installed aboard the International Space Station (ISS) in 2018. Data from ECOSTRESS allow scientists to evaluate the impact of water availability in terrestrial biomes, estimate drought, gauge the health of agricultural crops, and provides these data with a spatial resolution of approximately 70 meters and a temporal resolution of 1 to 5 days. ECOSTRESS data highlight the link between the water cycle and plant health, thereby providing information that researchers, fire management agencies, and others can use to monitor land surface conditions and respond more effectively to the conditions that increase fire risk.

The challenge lies in how to integrate ECOSTRESS data into predictive models to capture the complex interactions between vegetation stress, environmental conditions, and wildfire dynamics. By combining ECOSTRESS data and machine learning for wildfire prediction, the researchers sought to develop wildfire susceptibility models that can inform prediction of the likelihood of wildfire occurrence and spread by assessing pre-fire vegetation and drought conditions.

Data and Techniques Used

The Australian bushfire season of 2019-2020 was unprecedented in its intensity and devastation. Widely known as “Black Summer,” multiple fires scorched large tracts of land in the states of Victoria and New South Wales, resulting in 34 fatalities. New South Wales lost more than 2,400 homes while Victoria lost more than 390, and both experienced huge losses of land and wildlife. The researchers chose to examine the wildfire dynamics in this one-year period due to the extreme nature of the fire season and the availability of high-resolution ECOSTRESS data, both of which complemented their goal of developing robust models capable of depicting severe wildfire events. 

The researchers designed two types of models: “general” and “monthly.” The general models provided an overarching view of wildfire susceptibility throughout the season, while the monthly models were focused on specific time frames. The monthly models built on the insights gained from the general models to provide a finer temporal resolution for understanding and predicting wildfire risk. They then incorporated ECOSTRESS and other datasets (as independent variables) and fire event data (as dependent variables) into machine learning algorithms applied to the aforementioned 2019-2020 Australian fire season. 

In general, the researcher’s machine learning approach was based on algorithms that have the capacity to “learn” from data and make accurate fire predictions based on environmental conditions. This learning process involved modeling the relationships between sets of input (i.e., independent) variables and fire occurrences (i.e., the dependent variable).

Included among those inputs were six NASA datasets, each of which is archived at NASA’s Land Processes Distributed Active Archive Center (LP DAAC). (See these under "referenced datasets" below.)

In addition, the researchers also used a GEODATA 9 second Digital Elevation Model and D8-digital elevation model version 3 and flow direction grid from Geoscience Australia and rainfall data from the Australia’s Bureau of Meteorology.

Major Findings

By combining machine learning algorithms with data on a variety of biophysical factors, the researchers’ general and monthly models were able to predict wildfire occurrence in different application scenarios. The general models provided overall insight into the critical biophysical factors contributing to wildfire occurrence, making them valuable tools for pinpointing areas that merit intensive monitoring. The monthly models were able to predict the likelihood of wildfire spreading in the near future (i.e., one week ahead) and identify the key pre-fire factors during that period. Such information can help firefighters and forest managers plan and implement firefighting measures before a fire occurs.

In addition, the researchers highlighted the significance of using ECOSTRESS data to assess the impact of water availability on key climate biomes worldwide. Both models consistently identified ECOSTRESS Water Use Efficiency (WUE) as the most influential pre-fire factor. WUE, the ratio of carbon uptake to water use, averages 1.88 grams of carbon per kilogram of water (g C kg⁻¹ H₂O) over the study area during the fire season. As the researchers demonstrate, areas with WUE exceeding 2 g C kg⁻¹ H₂O have a 95% probability of experiencing vegetation burning during wildfire events. 

ECOSTRESS Evaporative Stress Index (ESI) was also identified as a significant contributor to wildfire predictions as it captures early signs of “flash drought,” which occurs during extended periods of hot, dry, and windy conditions, leading to rapid depletion of moisture in vegetation. Taken together, these results underscore the value of ECOSTRESS data in current and future wildfire prediction analyses and point to a method of developing effective strategies for predicting, managing, and mitigating wildfires.

Reference

Zhu, Y., Murugesan, S.B., Masara, I.K., Myint, S.W. and Fisher, J.B. (2025), Examining wildfire dynamics using ECOSTRESS data with machine learning approaches: the case of South-Eastern Australia's black summer. Remote Sensing and Ecological Conservation, 11: 266-281. doi:10.1002/rse2.422