Deep-Learning Based Fusion of Satellite Data with Physics-Based Models for Initializing Wildfire Forecast

Motivation

In the context of forecasting large fire events remaining active for prolonged periods of time it is critical to periodically reinitialize the state of the fire model. Reinitializing the model forecast helps reduce the accumulation of the error in the fire spread simulation and increase model accuracy. The fire observations used for the initialization can include airborne fire perimeters, or satellite observations. The limitations surrounding airborne fire perimeters stem from the fact that they are only available for select fires at irregular time intervals, have long latency, and are usually only produced once a day.

Satellite data on the other hand are routinely available at regular time intervals times but tend to be spatially and temporally sparse and noisy. Still, satellite observations provide a critical information that can be used for generating an accurate estimate of the initial fire state, which is required during the initialization phase of wildfire forecasts. The standard approach for accomplishing this involves interpolating for the missing information by imposing some smoothness requirements on the predicted estimates. This is the approach used in data-driven methods like support vector machines, and krigging using Gaussian processes. While this approach yields interesting results, it fails to constrain the resulting prediction to be consistent with the underlying physics and may lead to erroneous predictions. With this as motivation we present a novel deep-learning based method to estimate the historical fire spread of a wildfire for the purpose of assimilating satellite data into a coupled wildfire-weather forecast model.

Approach

The specific problem we tackle is one of obtaining a high-resolution map (30m in space) of the arrival time of an active fire, given sparse (sampling frequency of ~1/12h), noisy and low-resolution (375m) data acquired by the VIIRS sensor on the Suomi-NPP and NOAA-20 satellites. We focus on generating the fire arrival time map as a field of interest, since it can be used to determine (a) the fire perimeter at any time, and (b) the history of the heat and emission fluxes generated by the wildfire. These fluxes in turn provide an input into the fire spread model during a spin-up that assures a consistent state of the fire and the atmosphere at the initialization stage of a wildfire forecast.

We treat this problem as that of probabilistic inference, by recognizing that the desired fire arrival time map and the measurement map are random vectors. Then given a specific measurement map, the goal is to generate samples of the fire arrival time map from the conditional distribution.

Our approach to solving this problem involves two steps, each of which is challenging (see accompanying Figure). The first step requires generating samples of candidate fire arrival times and their measurements from the joint distribution. To accomplish this, we rely on the coupled wildfire-atmosphere model WRF-SFIRE to generate an ensemble of likely wildfire predictions and their measurements. This ensemble is generated by applying WRF-SFIRE to varying terrain maps, fuel maps, ignition points, and wind conditions to generate high-resolution fire arrival time maps. Thereafter, an approximation of the measurement operator is applied to the fire arrival time maps to compute the corresponding measurements. This process yields pairs of the fire arrival times and measurement maps sampled from the joint distribution. We note that by requiring the fire arrival time maps to be WRF-SFIRE solutions, we introduce a physical constraint into our method.

The second step involves using samples from this joint distribution to train a model which when given a new measurement map, can generate fire arrival-time maps from the conditional corresponding distribution. This is a challenging problem because the dimensions of the fire arrival time and measurement vectors are large (~10⁴ - 10⁵). We address this challenge by mapping the problem to the small-dimensional latent space of a conditional Generative Adversarial Network (cGAN) (see [1,2]). The cGAN is trained using samples from the joint distribution and can efficiently sample from the conditional distribution.

Results

We apply the approach described above to several wildfire case studies. We utilize the active fire data measured by the VIIRS system within the first two days of ignition as measurement and determine high-resolution maps of the fire arrival time. We validate the predicted results by computing their concordance with high-resolution IR data for these fires. We also compare the performance of our method with a method that uses machine learning but does not constrain the results to be consistent with the underlying physics. We conclude that our method avoids predictions which are not physically realistic. Finally, since our method generates an ensemble of likely fire arrival time maps, we demonstrate how this ensemble can be used to quantify the uncertainty in accurately predicting a given wildfire.

References

[1] Adler, J. and Öktem, O., 2018. Deep Bayesian inversion. arXiv preprint arXiv:1811.05910.

[2] Ray, D., Ramaswamy, H., Patel, D.V. and Oberai, A.A., 2022. The efficacy and generalizability of conditional GANs for posterior inference in physics-based inverse problems. arXiv preprint arXiv:2202.07773.