Method

The method of annual FSS prediction is based on Chen et al. (2011) with some modifications. We developed our empirical model of FSS using fire counts detected by MODIS onboard NASA's Terra satellite along with Oceanic Nino Index (ONI) and Atlantic Multidecadal Oscillation index (AMO) SST anomaly time series. Sea surface temperatures prior to the onset of the fire season have the strongest relationship with the number of satellite observed fires during the fire season in many areas of South America. The lead times enable us to make a prediction for the upcoming fire season.

Active fire counts

We used MODIS collection 5 global monthly fire location product (MCD14ML). We sampled the geographic coordinates of individual fire pixels (at a 1×1 km spatial resolution) that had a confidence level greater than 30%, and calculated the monthly FC within each 0.5° pixel after applying a cloud fraction correction. Persistent hot spots from MODIS observations and gas flare pixels in NOAA Global Gas Flare Estimates were excluded because the burning in these pixels is primarily associated with petroleum production rather than landscape fires. We then calculated the monthly FC for each region (6 states in Brazil (Acre, Amazonas, Maranhao, Mato Grosso, Para, Rondonia), 3 departments in Bolivia (El Beni, Pando, Santa Cruz), and one country (Peru)). The sum of FC during the fire season (defined as the 9-month period centered at the peak fire month) was recorded as the annual FSS for each region.

The Oceanic Niño Index (ONI) is a 3-month mean SST anomaly in the Niño 3.4 region (5°N-5°S, 120°-170°W) of the Pacific. We obtained the ONI time series from the NOAA National Weather Service Climate Prediction Center.

The Atlantic Multidecadal Oscillation index (AMO) represents a similar 3-month mean for the North Atlantic (0°-70°N). We obtained the AMO index time series from the NOAA Earth System Research Laboratory website.

We defined our empirical predictive model as a linear combination of the two climate indices sampled during the months of maximum correlation:

FSS_predicted(x,t,τ_c)=a(x,τ_c)×ONI[t,m(x)-τ_ONI(x,τ_c)]+b(x,τ_c)×AMO[t,m(x)-τ_AMO(x,τ_c)]+c(x,τ_c).

FSS_predicted is the predicted FSS in region x and year t. The parameter τ_c indicates the lead time (number of months before the peak fire month) when the prediction was made. a and b are spatial varying coefficients that represent the sensitivities of FSS in each region to ONI and AMO, individually, and c is a constant. ONI and AMO were sampled each year during months with lead times τ_ONI and τ_AMO relative to the peak fire month (m) in each region. Given a target τ_c, the optimal τ_ONI and τ_AMO values were derived from a series of linear regressions using ONI and AMO values at different months (with a cutoff(minimum) lead time of τ_c).

Based on the data (ONI and AMO) availability and the peak fire month, we derived the τ_c for each region. We then applied the predictive model with corresponding coefficients (a, b, and c) and optimal lead times (τ_ONI and τ_AMO) to derive the FSS in the target fire year. The range of the prediction was calculated using the 1-sigma uncertainty estimates for the parameters of the predictive model. Therefore, we have a set of predictions derived from different months (though with different confidence).

Version 1.0 (method used in Chen et al., 2011)

Model derived for 6 regions in South America: Amazonas, Para, Rondonia, Mato Grosso, El Beni, and Acre.
Optimal lead times for ONI and AMO were calculated using separate regressions between ONI and FSS and AMO and FSS.
The model was derived using a constant cutoff lead time (3 months before the fire peak month).
The prediction model was based on 2001-2009 data.
The map of individual regions had a spatial resolution of 0.5 degrees.

Version 2.0 (used for 2012 fire season forecast)

Four more regions were included: Maranhao, Pando, Peru, and Santa Cruz.
Optimal lead times for ONI and AMO were calculated using an optimization approach that simultaneously considered ONI and AMO lead times.
The model includes a set of regression equations derived using different cutoff lead times (0-11 months before the fire peak month). Depending on the availability of ONI/AMO data, the corresponding regression equation was used to predict FSS in the upcoming fire season.
The prediction model was derived using 2001-2011 fire and sea surface temperature data.
The map of individual regions had a spatial resolution of 0.05 degrees.
A summary table of lead times and model performance statistics based on the comparison with 2001-2011 observations is given here.

Version 2.1 (used for 2013 fire season forecast)

The method of forecast is same to that in version 2.0.
NOAA Climate Prediction Center has changed the strategy to update the base period for ONI trending removal (see here for detailed description). We used this new version of ONI in this prediction.

The model predictions contained on this website are highly experimental. They cannot be used to predict the occurrence of individual fires. Use of this information for planning purposes should also draw upon other independent and reliable climate information sources. The Regents of The University of California will not be liable for any consequences that may occur if you rely on this information.