Background
Over the past 30 years, the total area destroyed by wildfires has increased roughly by a factor of three to four. Over the same period, the number of wildfires has declined from a record high number of 96,385 in 2006 to 58,985 in 2021. This implies that wildfires now are more likely to spread at a faster rate and, therefore, cause more damage and take more resources and time to be extinguished than before.
To minimize the risk of escalation of a fast-growing wildfire, it is not only essential to quickly detect the fire but also to make an early assessment of how fast and in which direction the fire perimeter is expanding. Infrared (IR) images collected by instruments on NASA Earth-observing satellites are useful in tracking the movement of a wildfire's perimeter for a week or longer. However, these images are not useful for early wildfire detection and growth assessment because of the coarse resolution (>375 m), low acquisition frequency (every 12 h) and processing delay (2.5 h).
To address the current limitations of using IR satellite data for detection and real-time mapping of wildfires and to reduce the cost and increase the efficiency of using this technology on the ground, we plan to propose the following approach to NASA: (1) build an inexpensive imaging system that is optimized for detection and real-time mapping of wildfires, (2) develop and train machine learning data analysis tools for rapid on-board image analysis and prediction of spread, and (3) deploy the system on a constellation of small satellites to establish a facility to rapidly detect and track wildfires. In the current IR&D program we are developing a prototype of the sensor imaging and demonstrate its potential for accelerated detection and real-time mapping of wildfires based on test data and wildfire modeling.
Approach
The IRD work consists of the following seven tasks:
- The program team defined the requirements for the system capable of detecting wildfires at high resolution and for the design of the fire experiments for Task 5.
- The team assembled and integrated the breadboard systems. All cameras were tested and co-aligned to a reference target so that they observe the same field of view.
- The team used an existing monochromatic IR source to calibrate the system for narrow bandwidths centered at various wavelengths. Part of the task is being repeated because the setbacks that were encountered as discussed below.
- The next task involved thermal calibrations of the system using known heat sources. The calibration data are used to map measured incident radiation to source radiant power. Part of this task is also being repeated.
- Task 5 involved full-scale fire experiments. The purpose of these experiments was to determine the fraction within the area covered by one pixel that needs to be burning for the fire to be detected. Due to problems with the instrumentation the full set of experiments could not be conducted. More specifically, a second set of experiments to determine how well the system can track a growing fire shortly after detection.
- The team used models of various levels of complexity to simulate several well-documented wildfires. The purpose was to show the potential benefits of the increased image acquisition frequency and on-board processing of our system in terms of the improving short-term wildfire spread model predictions through data assimilation techniques.
- The team is defining the specifications for a low-cost orbital observation platform, which involves establishing the requirements for the flight data acquisition software, developing important low-level operational modes, and developing a new or repurposing an existing compression algorithm for downlinking flight data.
Accomplishments
The study was completed and the final report submitted in December 2024.
The first set of fire tests required multiple test rounds due to covid-related supply chain procurement issues and a short-term substitution of the near-infrared imaging camera lens that proved inadequate for the task. Finally, after the original lens was received, we were able to detect a relatively small fire and analyze the data. One of the CMOS detectors showed damage, presumably from heat, suggesting a more robust system may be needed for a space-flight application (where fire will not be an issue, but other environmental factors will). Because of the hardware failure, the remaining full-scale fire experiments could not be conducted.
In our fire modeling we found that although the FARSITE package is the most suitable wildfire simulation software package to support operations on the fireground, its inability to accurately predict the location of the fire perimeter versus time for the simulated historical wildfires is disappointing. The relatively poor accuracy of the model calculations can be attributed for a large part to the uncertainties in many of the input parameters, primarily fire behavior fuel model maps and wind speed data. However, the ability to collect and transmit real-time fire radiative power data will not only be helpful to guide operations on the fireground but also will allow us to build a database that can be used to reduce these uncertainties through machine learning techniques and thus greatly improve the predictive capability of FARSITE and other wildfire spread models that are capable of providing real-time predictions (such as cellular automata models), in particular during the initial stages of the fire.
Another important finding in this project was that there is insufficient satellite data at present to serve as training sets for machine learning algorithms we hope to use for rapid, on-board data processing. NASA has been conducting over-flight tests within their FireSense program that may provide more useful data (more rapid cadence, higher spatial resolution) in the future.
Moving ahead, we plan to pursue parallel tracks in hardware and software development, including developing a flight-tested demonstration unit to advance the TRL level of our camera system. For software, we will work to develop algorithms that can rapidly process data and more accurately predict near future spread either on board or in ground processing.