Jatan Buch
LightNews LightNews
banner
Jatan Buch
@jatanbuch.bsky.social
490 followers 470 following 56 posts
Scientist studying clouds, wildfires, air quality, with physics and machine learning
Posts Replies Media Videos
3. Focusing on the Middle Rockies, the ecoregion that includes the Blue Mountains which is the site of 4 large Oregon wildfires that have burned ~3000 km^2, we see that the models forecast fire frequency to be significantly lower than the climatological mean, while the burned area appears to be...
July 28, 2024 at 5:33 PM
August 2024 appears to follow a similar pattern, with elevated fire danger in coastal Central/Southern California -- not a good sign for a region that saw several fires in July, notably the Lake Fire in Los Padres National Forest inciweb.wildfire.gov/incident-inf...
July 28, 2024 at 5:26 PM
As the fire season intensifies over parts of the western United States, I wanted to share some results from ongoing work to improve seasonal and subseasonal-to-seasonal scale forecasts of wildfire frequency and burned area with machine learning. The model forecast for July 2024:
July 28, 2024 at 5:24 PM
We find that fire-month vapor pressure deficit (VPD) is the dominant driver of fire frequencies and sizes across the WUS, followed by moisture in large-diameter fuels (FM1000), total monthly precipitation (Prec), 2m maximum temperature (Tmax), and fraction of grassland cover.
SHAP analysis of the fire frequency MDN model outputs for different western United States’ divisions: (a–c) forests, (d–f) deserts, and (g–i) plains. (a, d, g) Input predictors sorted in descending order of their mean SHAP values aggregated over the entire study period. Each colored point along the x axis represents an individual prediction with the color corresponding to high (yellow) or low (indigo) values of the respective input predictor. (b, c, e, f, h, i) Partial dependence plots for two important predictors shown on the x axis, colored corresponding to high (yellow) or low (indigo) values of the mean daily maximum temperature, Tmax (b, c, e, f), and high (indigo) or low (yellow) values of the 1000 h dead fuel moisture, FM1000 (h, i). The color bar (far right) is normalized in terms of standard deviations (σ s) for all relevant values across the three panels. SHAP analysis of the fire size MDN model outputs for different western United States’ divisions: (a–c) forests, (d–f) deserts, and (g–i) plains. (a, d, g) Input predictors sorted in descending order of their mean SHAP values aggregated over the entire study period. Each colored point along the x axis represents an individual prediction with the color corresponding to high (yellow) or low (indigo) values of the respective input predictor. (b, c, e, f, h, i) Partial dependence plots for two important predictors shown on the x axis, colored corresponding to high (yellow) or low (indigo) values of the mean daily maximum temperature, Tmax (all panels). The color bar (far right) is normalized in terms of standard deviations (σ s) for all relevant values across the three panels.
September 12, 2023 at 7:29 PM
Our model has mixed skill in predicting extreme fire sizes during the study period. We find that the observed AAB for 2012 is in the ∼80th percentile of the predicted WUS AAB, while, for 2020, the total WUS AAB is in the 99.5th percentile of our model predictions.
Boxplots of predicted annual area burned (AAB) for 2 extreme fire years, 2012 and 2020, for the entire western United States (WUS) (teal) and three divisions organized by their primary vegetation types: forests (green), deserts (yellow), and plains (gray). The lower and upper whiskers of each boxplot indicate the 0.5th and 99.5th percentile of the predicted AAB distribution, whereas the black line represents its median value. Also shown for reference are the observed AAB for both 2012 (red diamond) and 2020 (indigo triangle).
September 12, 2023 at 7:10 PM
Overall, our AAB predictions for 15 out of the 18 ecoregions exhibit strong correlations (r ≥ 0.7) with the observed AAB. Moreover, the modeled AAB time series emulates the distinct decadal scale increasing trends in observed AAB over both forested and non-forested ecoregions.
Observed (red) and mean (teal) modeled monthly (MAB) and annual area burned (AAB) from 1984 to 2020. The teal shaded regions within each subplot indicate 1σ uncertainty intervals for the mean regional area burned aggregated over the MC simulations from the combined GPD model for all observed fires. Also shown is the Pearson correlation coefficient (r) between the observed and modeled area burned time series for each ecoregion. As in the previous figure but with different WUS ecoregions.
September 12, 2023 at 7:08 PM
With our modeling framework, we can simulate the burned area using MDN predicted fire locations and frequency. This functionality allows us to contrast the different contributions of local fire-related conditions and the long-term trends in natural and anthropogenic fire ignitions.
Cumulative observed (red) and modeled (teal) annual area burned (AAB) across the western United States from 1984 to 2020 for different combinations of fire frequencies and locations. The upper and lower panels show the AAB derived using modeled frequencies from the ZIPD MDN for each Ecoregion along with fire sizes simulated from the combined GPD model evaluated at observed and model fire locations respectively; the middle panel shows the AAB computed as above except with observed frequencies and model locations. The teal shaded regions indicate 1σ uncertainty intervals for the modeled area burned aggregated over the Monte Carlo (MC) simulations of all constituent fires. The mean total area burned over the study period as well as its 1σ uncertainty interval are indicated at the top of each panel.
September 12, 2023 at 7:06 PM
Similarly, MC simulations of a generalized Pareto distribution (GPD) MDN for all observed fires from 1984 to 2020 are aggregated to compute the mean of the monthly and annual area burned (MAB and AAB respectively) and their 1σ uncertainty intervals.
Observed (red) and mean modeled (teal) area burned across the western United States at monthly (MAB) (a) and annual (AAB) (b) timescales from 1984 to 2020. The teal shaded regions indicate 1σ uncertainty intervals for the mean area burned aggregated over the MC simulations from the combined GPD model for all observed fires. The mean total area burned over the study period and its 1σ uncertainty interval are indicated at the top of (b). Also shown within each subplot is the Pearson correlation coefficient between the observed and modeled burned area time series.
September 12, 2023 at 7:05 PM
Our model also successfully captures the interannual variability and monthly extremes across most ecoregions. However, since our model is trained on data over the whole WUS, its performance, on average, is better over ecoregions with a larger number of fires.
Observed (blue) and modeled (orange) fire frequencies at monthly and annual scales from 1984 to 2020 for ecoregions selected based on the total number of fires and goodness-of-fit metrics. The orange shaded regions within each subplot indicate 2σ uncertainty intervals for the mean regional fire frequency aggregated over the Monte Carlo (MC) simulations for all constituent grid cells. Also shown is the Pearson correlation coefficient (r) between the observed and modeled fire frequency time series for each ecoregion. As in the previous figure but for different WUS ecoregions.
September 12, 2023 at 7:05 PM
At the WUS level, our mean modeled frequencies are in good agreement with the total number of observed fires, exhibiting high correlations at both monthly (r = 0.94) and annual (r = 0.85) timescales.
Observed (blue) and modeled (orange) fire frequencies across the western United States at monthly (a) and annual (b) timescales from 1984 to 2020. Orange shaded regions represent 2σ uncertainty intervals for the mean fire frequency aggregated over the Monte Carlo simulations for all grid cells. The mean number of modeled fires over the study period, as well as the 2σ uncertainty interval, is indicated at the top of panel (b). Also shown is the Pearson correlation (r) between the observed and modeled time series.
September 12, 2023 at 7:04 PM
Our framework consists of a pair of mixture density networks (MDNs) trained on a wide suite of dynamic and static predictors. Think of this ML architecture as a regression model with shared weights for the entire WUS grid rather than individual regions.
Figure 2. Schematic diagram illustrating the input (blue), hidden (green), and output (purple) layers of a mixture density network (MDN) model within the SMLFire1.0 framework. While a fully connected neural network is implemented in practice, only a partial connected one is shown here for clarity; the solid black line on the left denotes the direction from the input to the output layer, whereas the dotted black lines represent additional nodes and layers in the network.
September 12, 2023 at 6:56 PM
The WUS saw a > 300% increase in the total area burned between 1984 & 2020, promoted by increasing fuel flammability due to frequent hot temperature extremes, rising atmospheric aridity, as well as prolonged drought-like conditions, and compounded by a legacy of fire suppression.
Wildfire activity in the western United States (WUS) from 1984 to 2020. (a) Locations of fire centroids (black dots) across the WUS with the spatial extent of three ecological divisions characterized by their primary vegetation type – forests (green), deserts (yellow), and plains (gray). Individual ecoregions are delineated with thick black lines and referenced with abbreviated names in rounded boxes. A full list of ecoregion names is given in Table 2. More details about L3 ecoregions, including high-resolution maps, can be found here: https: //www.epa.gov/eco-research/level-iii-and-iv-ecoregions-continental-united-states (last access: 23 October 2022). (b–d) Observed annual fire frequencies (blue) and annual area burned (AAB) (red) for each division. The black curves indicate the statistically significant (p < 0.05) trends for each AAB (solid) and annual frequency (dashed) time series.
September 12, 2023 at 6:54 PM