Sub-hourly forecasting of fire potential using machine learning on time series of surface weather variables

The problem

Weather that becomes dangerous within hours to days can lead to sudden fire outbreaks. Accurate short-term forecasts give time for evacuations and fire suppression. Fire danger indices, such as the Forest Fire Danger Index (FFDI) (Griffiths, 1999), estimate the likelihood and potential danger of fires on timescales from daily forecasts to seasonal outlooks. However, many existing systems are not designed for conditions that change quickly (Sharples et al., 2009; Fox-Hughes, 2015).

In this study, we asked whether the record from a single weather station could be used to recognise fire-prone conditions every 30 minutes. We adapted a method first developed to forecast volcanic eruptions (Dempsey et al., 2020). The models estimate what we call fire potential, the likelihood of a fire under the current weather, assuming something ignites it.

What we did

We used weather recorded every 30 minutes at Sunshine Coast Airport, Queensland, over 15.4 years from March 2002 to July 2017. Over that time, 15 significant fires burned 1,304 hectares (ha), about 13 square kilometres. The data streams included air temperature, relative humidity, wind speed and vapour pressure deficit, along with indicators of fuel dryness such as the Fine Fuel Moisture Code (FFMC).

For each data stream, we computed over 700 features in windows of 2, 5 and 10 days, using the Python package tsfresh (Christ et al., 2018). A feature is a statistic of the window, such as its mean, a quantile or the number of peaks. Each feature uses only data up to the time it describes.

Data in the 2, 5 or 10 days before a fire were labelled pre-fire, and all other data non-fire. Non-fire data far outnumber pre-fire data. We therefore randomly under-sampled the non-fire data to an approximately 1:1 ratio. We ranked features with the Mann–Whitney U test and used the highest-ranked to train a random forest of up to 300 decision trees. The forest outputs a consensus between 0 and 1. Values near 1 mean recent weather resembles the build-up to past fires.

To test the models on fires they had not seen, we withheld an 8-week period centred on each fire in turn, trained on the rest, and forecast the withheld period every 30 minutes. This pseudo-prospective test approximates real-time forecasting.

What we found

The area under the receiver operating characteristic curve (AUC) was 0.88, 0.88 and 0.98 for the 2, 5 and 10-day models, against 0.62 for the FFDI. A random guess scores 0.5 and perfect separation scores 1. Averaged over the three models, the AUC was 47% higher than for the FFDI. The 10-day model performs best. This is probably because it accounts for weather that accumulates and gradually raises fire risk. However, before a fire in 2007 the 2-day and 10-day models appear to move in opposite directions. This may be because the 10-day model averages out short-term weather variations that affect the 2-day model.

The decision trees used temperature, relative humidity, wind speed, vapour pressure deficit and FFMC most often. The 0.1 quantile of wind speed and the 0.3 quantile of relative humidity rose before a 2004 fire used in training, and again before a 2009 fire the model had not seen. Nevertheless, how often a tree uses a variable gives limited insight into cause and effect.

We issued a warning when the 10-day consensus crossed a threshold, and ended it once it had remained below the threshold for 48 consecutive hours. With a threshold of 0.7, warnings covered 282.6 days of the record, and 11 of the 15 fires occurred during one. Fires were about 50 times more likely during a warning than outside one.

This corresponds to one fire for every 26 days of warning, against one for every 1,342 days without a warning (our arithmetic). At the same threshold, the 2 and 5-day models anticipated 33% and 20% of the fires. With a lower threshold, more fires occur during a warning, but the time under warning increases.

The widget below applies this rule to an illustrative consensus record. At 0.7, its fire occurs during a warning.

The largest fire, in January 2017, burned about 840 ha, or 65% of the total area. With this fire withheld from training, the 10-day consensus crossed 0.7 about 4 days before the fire and peaked a day before it. The warning stayed active until the fire. The FFDI also rose in the days before this fire. The two could be used together, with early indications from our model reinforced by high FFDI values closer to an event.

Limitations

Our models use only weather data, from one station. They do not include fuel properties, live fuel moisture, topography or soil moisture. The fires were mostly in Noosa, Mount Coolum and Maroochy River National Parks, with eucalypt, heathland and melaleuca fuels respectively. A preliminary analysis suggests fires in heathland and melaleuca were less likely to be anticipated, although there are too few fires to draw firm conclusions.

The 50-fold difference assumes a Poisson model, in which fires are independent and occur at constant rates inside and outside warnings. The rule that keeps a warning active beyond 48 hours is arbitrary at this stage. Our rationale is that a prolonged high consensus indicates sustained fire-prone conditions.

We have not yet tested how a model trained at one station performs in other regions, or whether a 10-day window is the appropriate timescale there. Transfer learning could extend the approach to regions with little historical fire data.

Glossary

Machine learning: methods that learn to make predictions from patterns in past data.

Random forest: a machine learning model that averages the outputs of many decision trees, each a sequence of yes-or-no tests on the features.

Consensus: the output of the random forest, from 0 to 1, where values near 1 indicate weather similar to that before past fires.

Feature: a statistic computed from a window of a data stream, such as its mean or a quantile.

Quantile: the value below which a given fraction of the measurements fall, for example a tenth for the 0.1 quantile.

Forest Fire Danger Index (FFDI): the standard fire danger index for the region, which we used as the benchmark.

Fine Fuel Moisture Code (FFMC): an index of fuel moisture, used here as an indicator of fuel dryness.

Vapour pressure deficit: how far the air is from saturation, given by the difference between the vapour pressure of saturated air and the actual vapour pressure, with higher values for drier air.

Mann–Whitney U test: a statistical test of whether values in two groups differ, used here to rank features.

ROC curve and AUC: a plot of the fraction of pre-fire periods a model correctly flags (true positive rate) against the fraction of non-fire periods it wrongly flags (false positive rate), over all thresholds. The area under it (AUC) summarises how well a model separates the two.

Pseudo-prospective testing: forecasting data that were withheld from training, as an approximation of forecasting in real time.

Poisson model: a model in which events occur independently at a constant average rate.

Transfer learning: applying a model trained on one dataset to another, here from one region to another.

Hectare (ha): an area of 10,000 square metres.