Within-fire estimates of biomass loss across national forests of the US West Coast: inventory-informed inference with uncertainty

arXiv:2610.05614 · stat.AP · Submitted 2026-10-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Within-fire estimates of biomass loss across national forests of the US West Coast".

Jane: Here's the gist of the paper. Wildfire is reshaping forests in the western United States (US) yet estimates of wildfire impacts often rely on burned-area summaries or satellite-derived severity metrics that do not directly quantify changes in forest biomass or their uncertainty <ref:2610.05614#pg2>.

Tom: This paper combines US Forest Inventory and Analysis plot measurements with annual 30 m tree canopy cover data in a Bayesian spatio-temporal modeling framework to estimate within-fire changes in live aboveground biomass (AGB) and forest mortality area across National Forest System lands in Washington, Oregon, and California <ref:2610.05614#pg2>.

Jane: Methodological Framework The study employs a zero-inflated spatio-temporal latent Gaussian modeling framework proposed by May and Finley (2025) to estimate forest structure changes <ref:2610.05614#pg4>.

Tom: The model proceeds in two stages: a Bernoulli data model for live forest presence or absence, followed by a LGM for square-root AGB conditional on live forest presence <ref:2610.05614#pg4>.

Title and authors: Tom: So we're looking at this paper today, "Within-fire estimates of biomass loss across national forests of the US West Coast: inventory-informed inference with uncertainty." It’s about moving away from just using burned area summaries to actually measuring how much forest structure changes.

Jane: That sounds like a big shift for how we understand wildfire impacts on forests. The title itself, "inventory-informed inference with uncertainty," suggests they are trying to bring in the solid ground of forest inventory data while still being honest about how unsure they are.

Lu: Exactly, and what’s interesting is that they’re using plot measurements from the US Forest Inventory and Analysis combined with annual three zero meter tree canopy cover data in a Bayesian spatio-temporal modeling framework to estimate within-fire changes in live aboveground biomass AGB and forest mortality area across Washington, Oregon, and California.

Meng: Okay, so it’s combining ground truth plot data with remote sensing imagery to get a better picture of what’s happening inside the burn zone. But how do we know if this method is actually going to be reliable when you scale it up for a whole region?

Lalam: The paper uses a zero-inflated spatio-temporal latent Gaussian modeling framework proposed by May and Finley in two stages, starting with a Bernoulli data model for live forest presence or absence, followed by a latent Gaussian model for the square root AGB conditional on that live forest presence ><ref:2610.05614#pg4>.

Tom: That sounds like they’re modeling the presence of living trees first before they get into estimating the actual biomass loss numbers. So what’s the core idea behind this two-stage process?

Jane: The first stage models a binary process, which is basically determining whether there's live aboveground biomass or not, using a predictor called TCC ><ref:2610.05614#pg4>. Then, if there is live AGB, they use the second stage to model the square root transformed response using a mean zero spatio-temporal random effect uy s t ><ref:2610.05614#pg4>.

Lu: The spatio-temporal process uy s t is specified as a sum of independent Gaussian process components to capture both short and long range spatio-temporal dependence, which is important for understanding how these changes spread across space and time ><ref:2610.05614#pg8>.

Meng: So they’re trying to get the spatial pattern right by modeling it as a sum of independent processes, which sounds complex when you're trying to deploy this in the field. What kind of data are they feeding into that process?

Title and authors: Jane: They integrate plot-level FIA AGB data with remotely sensed TCC raster data to provide that fine scale inference, using ninety thousand five hundred forty seven plot level FIA observations to fit the Bernoulli model for live forest presence or absence ><ref:2610.05614#pg6>.

Tom: That’s a lot of ground truth data feeding into one part of the model. So what do you get out of combining those two very different types of data, plot measurements and raster images?

Lu: They use the annual TCC data extracted from the raster cell containing each exact FIA plot location in the FIA inventory year to connect those observations ><ref:2610.05614#pg6>. This gives them that fine scale inference they were aiming for.

Meng: How much effort does integrating those two sources take? I mean, processing millions of plot points and matching them to a raster grid sounds like a massive computational lift for production.

Jane: Well, the paper states that the combination is what allows them to fit the Bernoulli model for live forest presence or absence ><ref:2610.05614#pg6>. They then use a subset of forty-three thousand nine hundred fifty-one positive live AGB observations to fit the latent Gaussian model for square root AGB ><ref:2610.05614#pg6>.

Tom: Okay, so they’re using those forty-three thousand as the basis for estimating the actual biomass loss, which is what we really want to know about forest change. What are some of the main findings from this study?

Lu: They estimate ninety five point three million megagrams of wildfire associated AGB loss and two point three nine million hectares of forest mortality area across six thousand two hundred thirty-nine fires from the year two thousand to twenty twenty-two ><ref:2610.05614#pg2>.

Jane: That’s a substantial number, but what makes this result particularly useful for understanding the fire effects themselves? They reveal substantial spatial heterogeneity within fire perimeters, showing large differences between total burned area, forest mortality area, and associated biomass loss ><ref:2610.05614#pg4>.

Tom: So it’s not just one uniform burn pattern across the entire perimeter; there are really distinct zones of impact that need looking at individually. And what about the severity metric?

Lu: They define three primary quantities of interest: total biomass loss in megagrams, forest mortality area in hectares, and wildfire severity in megagrams per hectare ><ref:2610.05614#pg5>. The estimated average wildfire severity is twenty-two point six three megagrams per hectare of forest mortality area ><ref:2610.05614#pg10>.

Meng: twenty-two point six three megagrams per hectare sounds like a high rate of loss; what’s the biggest example they give to show how this looks in practice?

Title and authors: Jane: They point to the two thousand twenty fire, the Biscuit Fire, as having the greatest total AGB loss at four million four hundred sixty-nine thousand nine hundred twenty megagrams and also having the greatest mortality area of nearly eighty thousand hectares ><ref:2610.05614#pg12>.

Tom: And what about regional trends? Are there any specific areas where this method shows it’s particularly good or bad?

Lu: California exhibited the greatest total AGB loss and forest mortality area at the state level according to their estimates ><ref:2610.05614#pg10>. This suggests that when you look at the whole western region, California is a hotspot for these types of impacts.

Meng: So, putting it into perspective for someone who just listens to this show, what does this paper actually change for the way we think about wildfire impact measurement?

Jane: It changes things by providing a method that quantifies uncertainty directly within the forest structure estimates, which is different from just using burn area summaries or satellite derived severity metrics that don't directly quantify changes in forest biomass ><ref:2610.05614#pg2>.

Tom: They’re giving us a way to estimate within-fire changes in live aboveground biomass and forest mortality area across National Forest System lands by combining US Forest Inventory and Analysis plot measurements with annual three zero meter tree canopy cover data in a Bayesian spatio-temporal modeling framework ><ref:2610.05614#pg2>.

Lu: The model provides posterior predictive estimates of pre- and post-fire conditions at the pixel scale, which allows for within-fire spatial detail while retaining uncertainty quantification ><ref:2610.05614#pg4>.

Meng: If I’m trying to run this at scale, what’s the main limitation they flag? Because when you're deploying something in production, you need to know where it breaks ><ref:2610.05614#pg16>.

Jane: They point out that a limitation is the difficulty in validating estimated AGB loss and forest mortality area due to a lack of fine resolution reference data ><ref:2610.05614#pg16>. Also, they note that changes in AGB within wildfire perimeters may reflect compound disturbance effects rather than fire alone ><ref:2610.05614#pg16>.

Tom: So, if we’re talking about what this means for the future of this research area, where do you think the next steps should be?

Lu: Future work could incorporate additional disturbance and management information such as USFS Aerial Detection Survey products to distinguish rapid fire effects from other processes ><ref:2610.05614#pg16>. The Bayesian framework itself provides full uncertainty quantification, which is complementary to conventional burn severity products ><ref:2610.05614#pg18>.

Meng: From a production standpoint, if we’re looking at deployment, how do you handle the computational cost of running that spatio-temporal latent Gaussian model? Is it feasible for routine use?

Title and authors: Jane: The framework itself is what they are proposing to improve upon existing approaches ><ref:2610.05614#pg3>. They're demonstrating the value of borrowing information across both space and time to reduce domain level uncertainty and highlighting the importance of accommodating spatial and temporal dependence when estimating wildfire impacts ><ref:2610.05614#pg3>.

Tom: So, to wrap up on this paper, we’ve seen how they use inventory-informed inference with uncertainty to get within-fire estimates of biomass loss across national forests of the US West Coast: inventory-informed inference with uncertainty ><ref:2610.05614#pg1>. It gives us concrete numbers like ninety five point three million megagrams of AGB loss and highlights significant spatial heterogeneity within fire perimeters ><ref:2610.05614#pg4>.

Jane: And it shows that when you look at California, it’s a hotspot for these types of impacts according to their estimates ><ref:2610.05614#pg10>. It moves us past relying solely on burned area summaries by quantifying changes in live aboveground biomass and forest mortality area across National Forest System lands ><ref:2610.05614#pg2>.

Meng: The main caveat is that they admit that validating these estimated AGB loss and forest mortality areas is hard because there isn't really fine resolution reference data available across the whole region ><ref:2610.05614#pg16>.

Lalam: And another thing they mention is that changes in AGB within wildfire perimeters might be reflecting compound disturbance effects instead of just fire itself ><ref:2610.05614#pg16>.

Tom: So, the paper, "Within-fire estimates of biomass loss across national forests of the US West Coast: inventory-informed inference with uncertainty," provides a more direct way to estimate forest structure changes than previous methods ><ref:2610.05614#pg2>.

Jane: It’s a big step because it moves us toward using ground truth data in a way that incorporates the inherent uncertainty, which is something existing remote sensing metrics often miss ><ref:2610.05614#pg2>.

Lu: Future work could incorporate additional disturbance and management information such as USFS Aerial Detection Survey products to distinguish rapid fire effects from other processes ><ref:2610.05614#pg16>.

Meng: From an operational view, the challenge remains how to make this complex Bayesian modeling framework run efficiently enough for regular use across large areas ><ref:2610.05614#pg8>.

Tom: It’s a big piece of work that bridges the gap between sparse inventory data and high resolution remote sensing data to get better estimates of forest structure changes ><ref:2610.05614#pg6>.

Jane: We’ll keep an eye on how this framework is applied in future studies, because getting these within-fire estimates right is crucial for understanding the true impact of wildfire on ecosystems ><ref:2610.05614#pg18>.

The paper's summary: Tom: So we’re looking at how this paper moves away from just using burned area summaries to actually trying to measure how much forest structure changes within a fire zone on National Forest System lands in the West Coast.

Jane: Exactly, it’s about combining ground truth plot measurements from the Forest Inventory and Analysis with three zero meter tree canopy cover data to get a better picture of what’s happening inside the burn area.

Tom: And they use this whole Bayesian spatio-temporal modeling framework to estimate changes in live aboveground biomass and forest mortality area, which is way more detailed than just looking at total burned land.

Jane: The core idea is they use plot measurements from thousands of trees to fit a model that accounts for both where things are located on the map and how time affects those changes.

Lu: They use a zero-inflated spatio-temporal latent Gaussian modeling framework, which means they’re using a math structure designed to figure out the presence of live forest first, and then estimating the biomass loss if there is live forest present.

Meng: So, what does this actually give us? What are the main numbers we should be paying attention to?

Tom: They estimate ninety five point three million megagrams of wildfire associated aboveground biomass loss and two point three nine million hectares of forest mortality area across six thousand two hundred thirty-nine fires from two thousand to twenty twenty-two.

Jane: That’s a lot of numbers, but what makes this result interesting is that they show real spatial patterns—not just a uniform burn across the whole perimeter.

Lu: Pixel level estimates reveal substantial spatial heterogeneity, meaning there are really big differences between the total burned area and the actual biomass loss in different spots within a fire zone.

Tom: They also defined three key things to measure: total biomass loss in megagrams, forest mortality area in hectares, and wildfire severity which they calculated as megagrams per hectare of mortality area.

Jane: The average wildfire severity they calculated is twenty-two point six three megagrams per hectare of forest mortality area. That tells us the intensity of the impact on a ground level scale.

Meng: So, what’s the biggest example they give to show how this looks in practice? Which fire stands out?

Tom: They point to the two thousand twenty fire, called the Biscuit Fire, as having both the greatest total biomass loss and nearly eighty thousand hectares of mortality area.

Jane: And when you look regionally, California showed the largest total biomass loss and mortality area at a state level compared to other places in their study.

Lu: The paper provides full uncertainty quantification through this Bayesian approach, which means they aren't just giving you a single number; they are telling you how sure they are about those estimates.

Tom: That’s the big difference, right? It moves us past relying only on burn area summaries or satellite severity metrics that don't directly measure changes in biomass.

Jane: It gives us a way to estimate within-fire changes in live aboveground biomass and forest mortality area across National Forest System lands by combining those ground truth measurements with remote sensing data.

Tom: This method helps us get better estimates of forest structure changes than what we had before because they use real inventory data as their starting point.

Jane: The limitation they admit is that validating these estimated biomass loss and mortality areas is hard because there isn't really fine resolution reference data available across the whole region.

Tom: And they also flag that changes in biomass within a fire perimeter might be caused by compound disturbance effects, not just the fire itself.

Jane: So, while this framework is a step forward by incorporating ground truth and uncertainty, it stops short of being a perfect replacement for everything else we use right now.

Lu: Future work could incorporate more information about other disturbances and management activities to help distinguish rapid fire effects from other processes.

Tom: It’s a big piece of work that bridges the gap between sparse inventory data and high resolution remote sensing data to get better estimates of forest structure changes across the West Coast.

The paper's improvements: Tom: We just talked about how this paper uses ground truth plot data to estimate biomass loss and mortality area across West Coast National Forests, and now we’re looking at what they propose to improve the method itself.

Jane: So, what are they suggesting we should change or add to make this whole system better? Are they just saying the current setup has some holes?

Lu: They are focusing on how to make the model more robust by incorporating more data streams and improving how that spatial and temporal uncertainty is handled in the modeling framework.

Tom: I mean, what’s the actual improvement? Is it just adding another dataset, or a tweak to the math? I need to know if this makes it faster or just more accurate.

Lu: The main thing they push is using additional disturbance and management information. They suggest incorporating products like the US Forest Service Aerial Detection Survey data into the model.

Jane: That makes sense, because they think that extra information could help distinguish rapid fire effects from other types of land change happening around a burn site.

Meng: I’m interested in what that means for deployment at scale. If we add more data to the input, does it make the computational cost explode?

Lu: Well, they are trying to use this new information to reduce domain level uncertainty and show how accommodating spatial and temporal dependence matters when you are estimating these wildfire impacts.

Tom: So it’s about using that extra data not just for a slightly better answer, but to tighten up the confidence in the whole prediction?

Jane: Exactly, they argue that the Bayesian framework itself is good because it gives you full uncertainty quantification, which is different from just taking a single number.

Lu: They emphasize that this uncertainty quantification is complementary to conventional burn severity products we already use.

Meng: If I’m running this in production, how much work does incorporating these new data sources require compared to the original setup? Is it a simple pipeline change or a complete re-train?

Lu: It suggests the framework can be improved by better handling that spatial and temporal dependence across different scales.

Tom: So, we’re talking about refining the underlying mathematical structure of the modeling process, not just feeding it more raw data points.

Jane: The implication is that this kind of inventory-informed approach, even with its limitations on validation, provides a solid foundation for understanding forest structure changes better than relying only on burned area summaries.

Lu: It moves the research toward using ground truth data in a way that correctly incorporates the inherent uncertainty about those measurements.

Tom: So what’s next? Where does this lead us in terms of actually applying this to real-world fire management decisions?

Jane: We need to see how these improved estimates translate into actionable insights for people managing forests across California, Oregon, and Washington.

Conclusion: Tom: So we’re wrapping up this look at the paper "Within-fire estimates of biomass loss across national forests of the US West Coast: inventory-informed inference with uncertainty." Basically, they used ground truth plot data and remote sensing to estimate forest structure changes inside fires, moving beyond just burned area summaries.

Jane: That’s right, it gives us concrete numbers like ninety five point three million megagrams of biomass loss and shows how much variability there is across different fire zones.

Lu: The core idea is using a Bayesian spatio-temporal model to capture those spatial and temporal patterns, which means they aren't just giving you one answer, but a measure of how sure they are about it.

Meng: From an engineering standpoint, the biggest thing for us is that this method provides posterior predictive estimates of pre- and post-fire conditions at the pixel scale while keeping track of that uncertainty.

Tom: It means we can get a lot more spatial detail inside a fire perimeter than before, which is crucial for understanding how ecosystems actually respond to heat.

Jane: It’s about moving past just seeing what burned and starting to measure the actual loss of living biomass in a way that's informed by real forest inventory.

Lu: And they’re pointing out that while the validation is hard because we don't have fine resolution reference data, it sets a better baseline for what future ground truth efforts should look like.

Meng: I think the biggest change here is the framework itself; it shows how to structure a model that can handle this kind of messy, real-world data combination efficiently enough to be useful.

Tom: So we’ve seen how they use inventory-informed inference with uncertainty to get within-fire estimates of biomass loss across national forests of the US West Coast: inventory-informed inference with uncertainty. It gives us concrete numbers and shows a lot of spatial heterogeneity.

Jane: It really does give us a way to measure the true impact on forest structure, not just the area that burned.

Lu: Future work needs to focus on integrating more disturbance data, like those aerial survey products, to help distinguish rapid fire effects from other things happening in the landscape.

Meng: If we can nail that integration part at scale, this could become a pretty useful tool for real-time impact assessment after a wildfire event.

Tom: It’s definitely something worth watching as we try to build models that are more robust when they hit real field data and complex environments.

Department of Forestry, Michigan State University Department of Statistics and Probability Department of Electrical Engineering and Computer Science South Dakota School of Mines & Technology USDA Forest Service Pacific Northwest Research Station Northern Research Station United States Department of Agriculture Forest Service Rocky Mountain Research Station

stat.AP

Submitted: 2026-10-04

Updated: 2026-11-16

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Wildfire is reshaping forests in the western United States (US) yet estimates of wildfire impacts often rely on burned-area summaries or satellite-derived severity metrics that do not directly…

Key concepts

Bayesian spatio-temporal modeling framework
This is a statistical approach that uses Bayesian methods to model changes over both space (location) and time (year). It allows researchers to combine different types of data—like ground plot measurements and satellite images—to create estimates of forest change while also providing a measure of uncertainty for those estimates.
Live Aboveground Biomass (AGB)
This refers to the total weight of living trees and vegetation above the ground in a forest. The study tracked how much this biomass was lost or changed inside fire zones, which is more detailed than just measuring how much land burned.
Zero-inflated latent Gaussian modeling
This is a complex statistical model used to predict forest structure changes. It first models whether live trees are present or absent (a Bernoulli process) and then models the amount of biomass loss conditional on those trees being present, capturing both presence/absence and magnitude.

Terminology

Summary

Wildfire is reshaping forests in the western United States (US) yet estimates of wildfire impacts often rely on burned-area summaries or satellite-derived severity metrics that do not directly quantify changes in forest biomass or their uncertainty <ref:2610. p. 005614#pg2>. This paper combines US Forest Inventory and Analysis plot measurements with annual 30 m tree canopy cover data in a Bayesian spatio-temporal modeling framework to estimate within-fire changes in live aboveground biomass (AGB) and forest mortality area across National Forest System lands in Washington, Oregon, and California <ref:2610. p. 005614#pg2>.

Methodological Framework

The study employs a zero-inflated spatio-temporal latent Gaussian modeling framework proposed by May and Finley (2025) to estimate forest structure changes <ref:2610.05614#pg4>. The model proceeds in two stages: a Bernoulli data model for live forest presence or absence, followed by a LGM for square-root AGB conditional on live forest presence <ref:2610. p. 505614#pg4>. The first stage models the binary process p(s, t) representing positive or zero live AGB using TCC as a predictor <ref:2610. p. 505614#pg4>. Conditional on positive AGB, the second stage models the square-root transformed response ˜y(s, t) using a mean-zero spatiotemporal random effect uy(s, t) <ref:2610. p. 505614#pg4>. The spatio-temporal process uy(s, t) is specified as a sum of independent Gaussian process components to capture short- and long-range spatio-temporal dependence <ref:2610. p. 505614#pg8>.

Data Integration and Modeling

The research integrates plot-level FIA AGB data with remotely sensed TCC raster data to provide fine-scale inference <ref:2610.05614#pg6>. The model uses 90,547 plot-level FIA observations to fit the Bernoulli model for live forest presence or absence <ref:2610. p. 305614#pg6>. The subset of 43,951 positive live AGB observations is used to fit the LGM for square-root AGB <ref:2610. p. 305614#pg6>. Annual TCC data is extracted from the raster cell containing each exact FIA plot location in the FIA inventory year <ref:2610. p. 405614#pg6>. The spatio-temporal process uy(s, t) is specified with two components to capture local- and landscape-scale patterns in AGB effectively <ref:2610. p. 505614#pg8>.

Key Findings on Forest Change

The analysis estimates 95.3 million Mg of wildfire-associated AGB loss and 2. p. 039 million ha of forest mortality area across 6,239 fires from 2000 to 2022 <ref:2610. p. 205614#pg2>. Pixel-level estimates reveal substantial spatial heterogeneity within fire perimeters, including large differences between total burned area, forest mortality area, and associated biomass loss <ref:2610. p. 005614#pg4>. Total AGB loss is estimated at 95,260,702 Mg (94,700,728 Mg) across the study period <ref:2610. p. 805614#pg10>. Forest mortality areas were estimated at 2,391,984 ha (2,388,519 ha) transitioning from positive live AGB before fire to zero live AGB after fire <ref:2610. p. 805614#pg4>.

Derived Metrics and Regional Trends

The study defines three primary quantities of interest: Total biomass loss (Mg), Forest mortality area (ha), and Wildfire severity (Mg/ha) <ref:2610.05614#pg5>. The estimated average wildfire severity is 22. p. 963 Mg/ha of forest mortality area <ref:2610. p. 005614#pg10>. The 2002 Biscuit Fire resulted in the greatest total AGB loss, estimated at 4,469,920 Mg (4,423,774 Mg) and the greatest mortality area of 79,230 ha (78,881 ha) <ref:2610. p. 905614#pg12>. California exhibited the greatest total AGB loss and forest mortality area at the state level <ref:2610. p. 905614#pg10>.

Limitations and Future Directions

Limitations include the difficulty in validating estimated AGB loss and forest mortality area due to a lack of fine-resolution reference data <ref:2610.05614#pg16>. Furthermore, changes in AGB within wildfire perimeters may reflect compound disturbance effects rather than fire alone <ref:2610. p. 805614#pg16>. Future work could incorporate additional disturbance and management information such as USFS Aerial Detection Survey products to distinguish rapid fire effects from other processes <ref:2610.05614#pg16>. The Bayesian framework provides full uncertainty quantification, which is complementary to conventional burn-severity products <ref:2610. p. 005614#pg18>.

Table 1

Table 1: Posterior summaries for model parameters described in Section 2.2 indicates expected positive relationships between percent TCC and both live forest presence and AGB, as reflected by βp and βy, respectively <ref:2610. p. 805614#pg10>. Posterior estimates for the local- and landscape-scale spatial decay parameters imply effective spatial ranges of approximately 0. p. 84 km and 78.64 km, respectively <ref:2610. p. 005614#pg10>. Posterior estimates for λy,1 and λy,2 imply effective temporal ranges of approximately 266. p. 85 and 252.5 years, respectively <ref:2610. p. 805614#pg10>. The posterior median of αy is 0. p. 8230 which corresponds to approximately 77. p. 80% assigned to the local-scale component <ref:2610. p. 805614#pg10>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can estimate within-fire changes in live aboveground biomass (AGB) and forest mortality area across National Forest System lands by combining US Forest Inventory and Analysis plot measurements with annual 30 m tree canopy cover data in a Bayesian spatio-temporal modeling framework.

  2. The improved system can provide posterior predictive estimates of pre- and post-fire conditions at the pixel scale, allowing for within-fire spatial detail while retaining uncertainty quantification.

  3. The system can quantify three primary quantities of interest: total biomass loss (Mg), forest mortality area (ha), and wildfire severity (Mg/ha) using a structurally based definition that aligns with the modeling specification.

  4. The improved AI system can provide fire-level summaries by aggregating pixel-level estimates to individual fires and states, highlighting within-fire changes in forest structure rather than simple summaries of total fire-perimeter area.

  5. The system can identify significant spatial heterogeneity within fire perimeters, including large differences between total burned area, forest mortality area, and associated biomass loss.

  6. The improved AI system can assess the relative impact of different years and regions by comparing posterior summaries across time and state levels (e.g., noting that 2020 stands out as the most consequential year for both total AGB loss and forest mortality area).

Related papers