Volumetric Impact Characterization of the 2025 Palisades and Eaton Fires Using Aerial LiDAR

Highlights

What are the main findings?

• A multi-temporal aerial LiDAR differencing pipeline enables sub-meter volumetric damage characterization across approximately 152 km² of wildland–urban interface fire.
• Cell-based segmentation of pre- and post-fire point clouds yields bottom-up carbon emission estimates of $255\pm 61$ kt C (Palisades) and $162\pm 38$ kt C (Eaton), with the Eaton estimate corroborated by an independent network of particulate sensors.

What are the implications of the main findings?

• Cell-based LiDAR differencing at 0.5 m resolution recovers per-structure envelope volumes consistent with California Department of Forestry and Fire Protection (CAL FIRE) damage inventories for both fires, validating the threshold-only change-detection approach.
• Building combustion dominates the Eaton fire carbon budget, while vegetation combustion dominates the Palisades fire, a spatial decomposition unavailable from atmospheric or satellite methods alone.

Abstract

In January 2025, the Palisades and Eaton fires overtook large swaths of Los Angeles County, covering a combined area of approximately 152 km², composed of diverse coastal, urban, and forested environments. Aerial Light Detection and Ranging (LiDAR) surveys were commissioned directly following these fires, and compared against previously unreleased foundational LiDAR surveys captured in 2023 and 2024. The timeliness of these surveys presents a unique opportunity to approach large-scale damage characterization metrologically at sub-meter resolution. Cell-based height differencing across 367 million change-detected cells (on a 0.5 m grid) identifies 49.5 km² of vegetation loss and 1.66 km² of building footprint destruction in the Palisades fire, and 24.3 km² of vegetation loss and 1.48 km² of building footprint destruction in the Eaton fire. From the resulting volumetric loss inventory, we derive bottom-up carbon emission estimates of $255\pm 61$ kt C for the Palisades fire and $162\pm 38$ kt C for the Eaton fire. The Eaton estimate agrees to within 6% of an independent atmospheric inversion estimate derived from ground-based sensor networks, well within the propagated uncertainty of either method, providing an independent cross-validation, at the total-emission level, between LiDAR-based and atmospheric-inversion approaches to wildland–urban interface fire emissions. This paper details the segmentation and characterization methodology, and coincides with ALERTCalifornia’s public release of all described raw and derivative datasets.

1. Introduction

On 7 January 2025, two catastrophic wildfires ignited on opposite sides of the Los Angeles basin (Figure 1). The Palisades fire erupted in the Santa Monica Mountains above coastal Pacific Palisades at approximately 10:30 a.m., while the inland Eaton fire began that evening in Eaton Canyon near Altadena [1]. Driven by severe Santa Ana wind conditions (strong, dry offshore katabatic winds originating from high-pressure air masses over the Great Basin and Mojave Desert, with gusts reaching up to 160 km/h), both fires expanded with extraordinary speed [2,3,4]. The National Weather Service had issued a Particularly Dangerous Situation Red Flag Warning the day prior [5]. By full containment on 31 January, the fires had collectively burned almost 152 km², destroyed more than 16,000 structures, damaged approximately 2000 more, and claimed at least 30 lives [1]. The Palisades fire’s emissions were largely blown into the ocean, while the emissions impact of the Eaton fire was spread over the city to the south (Figure 2).

The early rates of spread were extreme. Fanned by the Santa Ana gusts over fuels desiccated by the driest nine-month period on record in Southern California, the Palisades fire grew from roughly 20 to 200 acres (8 to 81 ha) within twenty minutes of ignition, exceeded 1200  acres (∼5 km²) within hours, and reached approximately 15,000 acres within the first 24 h; the Eaton fire, igniting that evening, reached several hundred acres by midnight. At final containment, the Palisades fire had burned 23,448 acres (∼95 km²) and the Eaton fire 14,021 acres (∼57 km²). The combination of wind-driven spread, long-range ember cast, and dense wildland–urban interface fabric produced very high burn severity within the developed areas, where structure-to-structure ignition dominated and destroyed the majority of the 6837 (Palisades) and 9414 (Eaton) structures later confirmed by CAL FIRE. It is this multi-meter structural collapse that the present study quantifies.

Beyond the immediate destruction, wildland–urban interface (WUI) fires generate cascading environmental consequences. Burned structures release complex mixtures of heavy metals, polycyclic aromatic hydrocarbons, asbestos, and other toxicants that infiltrate soil, water systems, and the marine environment. In addition to the 30 direct fatalities, Paglino et al. [6] estimated approximately 440 excess deaths attributable to smoke exposure and healthcare disruption between 5 January and 1 February 2025. Casey et al. [7] found an approximately 35% increase in virtual care-seeking for cardiovascular, respiratory, and neuropsychiatric conditions among highly exposed Kaiser Permanente Southern California members. The removal of vegetation further compounds these impacts by dramatically increasing erosion potential and debris flow hazard, potentially leading to a series of collapses in the area. The San Gabriel Mountains foothills beneath the Eaton fire burn scar have been classified at “high to very high risk” for post-fire debris flow events [8].

Light Detection and Ranging (LiDAR) technology has emerged as an essential tool for disaster assessment, offering capabilities that far exceed traditional remote sensing approaches. LiDAR systems emit laser pulses and measure return times to generate highly accurate three-dimensional point clouds of terrain, vegetation, and built infrastructure. Unlike passive optical sensors, LiDAR can penetrate vegetation canopy to capture ground surface elevation, detect subtle structural changes, and characterize vertical vegetation structure with sub-meter precision, making it particularly valuable for post-disaster damage quantification [9].

ALERTCalifornia, a multidisciplinary public safety and research program at the University of California San Diego that manages a network of over 1263 public-facing wildfire camera feeds (as of June 2026) and has commissioned the largest LiDAR mapping campaigns to date covering California’s highest-risk fire zones [10], sponsored the collection of high-resolution airborne LiDAR data by NV5 on 21–22 January 2025, at 16 points per square meter [11]. Pre-fire baseline LiDAR data collected by NV5 in 2023–2024 through the United States Geological Survey (USGS) 3D Elevation Program (3DEP) in partnership with the Los Angeles Region Imagery Acquisition Consortium (LARIAC) provides the temporal comparison necessary for robust change detection [12]. This paper presents a methodology for the cell-based segmentation and characterization of these LiDAR point clouds to quantify fire-related changes across bare earth, vegetation, and built infrastructure, offering a metrological approach to damage assessment. The accompanying public release of all raw and derivative datasets enables broader application in ongoing post-fire investigations and future wildfire research.

2. Background

2.1. Prior 3D Survey Efforts

Brigham et al. [13] published topographic differencing products through the OpenTopography Community Dataspace comparing the older pre-fire LARIAC 2016 LiDAR against the 1 m resolution post-fire ALERTCalifornia 2025 data. Digital Surface Model (DSM) differencing revealed clear vegetation and structural loss, while Digital Terrain Model (DTM) differencing, focused specifically on the comparison of ground features, showed more modest sub-meter signals attributable to erosion and sediment reworking. A critical limitation is the eight-year baseline between acquisitions, which introduces landscape change unrelated to the fires.

The National Science Foundation (NSF)-sponsored Geotechnical Extreme Events Reconnaissance (GEER) Association conducted an extensive field campaign [14], deploying aerial and ground-based remote sensing platforms, soil and ash sampling, infiltrometer testing, and geophysical instrumentation across both burn areas between January and May 2025. Findings documented widespread erosion, debris flows, and infrastructure damage, though a quantitative analysis of these datasets has yet to be released.

2.2. CAL FIRE Damage Inspection Surveys

The CAL FIRE Damage Inspection (DINS) program provides the authoritative post-fire inventory of structural damage for California wildfires. DINS teams, composed of trained assessors deployed by CAL FIRE in coordination with local building officials, conduct systematic ground-based surveys of every structure within and adjacent to the fire perimeter, classifying each as destroyed, damaged (major, minor, or affected), or undamaged through a direct visual inspection of the structure and its immediate surroundings. Surveys typically begin within days of containment and continue over several weeks as access permits. The resulting parcel-level database serves as the basis for insurance claims, Federal Emergency Management Agency (FEMA) assistance eligibility, debris removal planning, and official casualty and loss reporting [1].

The final DINS totals for the Palisades and Eaton fires are summarized in

Table 1. CAL FIRE DINS damage and casualty totals.

Damage Type	Eaton	Palisades
Structures Damaged	1074	973
Structures Destroyed	9414	6837
Confirmed Civilian Fatalities	18	12

. These counts serve as an independent reference for the per-structure envelope volume validation described in Section 4.9.

2.3. Independent Emissions Estimates

Two quantitative emissions estimates were published prior to this work. Vannucci et al. [15] conducted a multimodal atmospheric analysis of the Eaton fire, integrating satellite imaging, meteorological models, advection–diffusion models, and dense ground-level air quality observations, and estimated approximately 153 million kilograms of carbon emitted. The Palisades plume was carried offshore by Santa Ana winds and was not captured by their in-basin sensor network (Figure 2). Li et al. [16] provided a satellite-based emissions inventory for both fires using the Visible Infrared Imaging Radiometer Suite (VIIRS), the Geostationary Operational Environmental Satellite 18 Advanced Baseline Imager (GOES-18 ABI), MODIS, the TROPOspheric Monitoring Instrument (TROPOMI), and Tropospheric Emissions: Monitoring of Pollution (TEMPO), estimating total particulate matter (TPM), carbon monoxide (CO), and nitrogen oxides (NO_x) emissions per active-burn day. Both studies found that residential-area fire intensity significantly exceeded concurrent vegetation fire intensity, with emission coefficients distinct from those of typical chaparral wildfires.

Air quality studies documented severe impacts. Schollaert et al. [17] reported daily average fine particulate matter (PM_2.5) concentrations reaching 101.7 μg m⁻³ at the downtown Los Angeles monitor. The Centers for Disease Control and Prevention (CDC) ASCENT network recorded an approximately 110-fold increase in PM_2.5 lead levels [18], underscoring the distinct toxicological profile of WUI fire emissions.

2.4. Vegetation Communities and Biomass

Chaparral shrublands are the dominant wildland vegetation type in the fire-affected areas. Schrader-Patton and Underwood [19] developed biomass estimates for Southern California chaparral using random forest modeling, reporting mean above-ground biomass densities of 2.3–3.5 kg/m² for the Southern California national forests. The Palisades fire burned through a mosaic of coastal sage scrub on lower slopes (typically <1.5 m height) and dense chaparral on higher slopes, with oak woodlands in intervening canyons. The Eaton fire area includes lower montane chaparral transitioning at higher elevations into bigcone Douglas-fir and pine forest, with substantial overlap of prior Station fire (2009) and Bobcat fire (2020) burn scars.

2.5. LiDAR Methodologies for Fire Damage Assessment

García et al. [20] established foundational methods for quantifying biomass consumption from megafires using airborne LiDAR and Landsat imagery. Xu and Greenberg [21] produced LiDAR-derived above-ground biomass estimates with spatially explicit uncertainty for California forests. Antoine et al. [22] demonstrated high-resolution satellite stereo imagery for elevation change mapping of the Eaton fire, providing a complementary approach to airborne LiDAR. For debris flow hazards, Guilinger et al. [23] established methods for predicting post-fire sediment yields using LiDAR differencing, and Chen et al. [24] deployed unpiloted aerial vehicle (UAV) LiDAR to map topographic change after post-fire rainfall events in the Eaton fire area.

3. Datasets and Tools

3.1. Source LiDAR Datasets

This study leverages three aerial LiDAR acquisitions spanning the fire-affected areas of Los Angeles County (

Table 2. Source LiDAR datasets.

USGS Dataset Name	Acquisition Dates	Mean Density
CA_LosAngeles_1_B23	8 January 2023–7 January 2024	8 pts/m2
CA_LAPostWildfire_Eaton_C25	21 January 2025–22 January 2025	16 pts/m2
CA_LAPostWildfire_Palisades_C25	21 January 2025	16 pts/m2

). The pre-fire baseline is drawn from the CA_LosAngeles_ 1_B23 dataset, collected by NV5 Geospatial between January 2023 and January 2024 under the USGS 3D Elevation Program (3DEP) in partnership with the Los Angeles Region Imagery Acquisition Consortium (LARIAC). This dataset, acquired at Quality Level 1 with a mean density of 8 pts/m², had not been publicly released at the time of the fires; its availability for this study was made possible through a data-sharing agreement between ALERTCalifornia and the USGS.

The post-fire datasets (CA_LAPostWildfire_Eaton_C25 and CA_LAPostWildfire_ Palisades_C25) were collected by NV5 Geospatial on 21–22 January 2025, sponsored by ALERTCalifornia and UC San Diego, using a Riegl VQ-1560iiS dual-channel waveform LiDAR sensor flown at 3048 m above ground level (AGL) aboard a Cessna Caravan at Quality Level 1 with a mean density of 16 pts/m² [11]. The rapid turnaround (14 days from ignition to acquisition) was critical for capturing the post-fire landscape before the first major winter storm on 25 January 2025, which triggered substantial debris flows and sediment redistribution across both burn areas.

Acquisition timing and seasonality. The pre-fire baseline (LA County B23, USGS 3DEP/LARIAC) was acquired in 41 lifts between 8 January 2023 and 7 January 2024; the post-fire Palisades data were acquired on 21 January 2025 and the Eaton data on 21–22 January 2025. Rather than relying on this project-level window alone, the per-tile pre-fire acquisition dates were reconstructed directly from the LAS GPS-time field of the retained tiles (Appendix C, script 09_acquisition_dates.py). Within the Palisades footprint, acquisition is dominated by a single day, 3 December 2023 (64% of points, pooled across zones), with the remainder collected on 4 and 8 December 2023 and a smaller northeastern extent on 15 and 22 October 2023; overall, the Palisades pre-fire tiles were collected between 15 October and 8 December 2023, 13–15 months before the 7 January 2025 ignition. Within the Eaton footprint, acquisition is effectively a single flight, with more than 99.99% of points collected on 7 October 2023, approximately 15 months before ignition. Urban and wildland zones within each fire were acquired on the same dates, so no differential temporal offset exists between the two zones used in the segmentation. Both epochs fall within the winter acquisition window. Because both are winter (leaf-off for deciduous elements) acquisitions of a landscape dominated by evergreen chaparral and oak, canopy phenology and herbaceous ground cover are broadly matched between epochs, and phenological mismatch is small relative to the fire signal. Any residual seasonal difference in canopy state or ground moisture is absorbed into the vegetation-pathway uncertainty and further suppressed by the 20 cm detection threshold, which was set above typical seasonal ground-moisture and low-herbaceous variation.

Figure 3 shows the 2023 (Figure 3a) and post-fire 2025 (Figure 3b) classified LiDAR data for the Pacific Palisades Castellammare neighborhood. Vegetation is shown in green, buildings in orange, ground in brown, and unclassified points in grey. A clear difference is visible as buildings and vegetation disappear post-fire. A difference map (Figure 3c) shows the change by height above ground, the method providing the basis for our volumetric analysis. Because both the pre-fire and post-fire datasets were collected by the same contractor using sensors from the same instrument family, and were delivered in a common projected coordinate system (Universal Transverse Mercator (UTM) Zone 11N, NAD83(2011), NAVD88 GEOID18), they are inherently co-registered, eliminating the need for post hoc horizontal alignment that would otherwise introduce additional uncertainty into the change detection workflow. This represents a significant advantage over the Brigham et al. [13] analysis, which relied on the older 2016 LARIAC dataset (Section 3.3) with an eight-year temporal baseline and different acquisition specifications.

The post-fire data requires a non-vegetated vertical accuracy (NVA) of 19.6 cm at 95% confidence (root mean square error (RMSE_z) of 10 cm), tested against 22 independent checkpoints [25].

3.2. Software Tools

Point-cloud processing was performed primarily using the LAStools (version 2.0.4) suite [26], which provides optimized batch processing and parallelization for large aerial LiDAR datasets. Visualization, inspection, and validation were performed using the Potree web-based point-cloud viewer [27], which enables a granular inspection of very large datasets and side-by-side comparison of multiple epochs. CloudCompare was used for supplementary point-cloud manipulation.

3.3. The 2016 LARIAC Dataset

A third LiDAR epoch, the 2016 LARIAC dataset [28], was also acquired for this study and provides the broadest temporal baseline for the fire areas. However, this dataset exhibits substantially higher noise levels and lower point density than the 2023 acquisition, requiring additional preprocessing that is beyond the scope of the current analysis. The high noise is especially problematic as it intersects with the mountains at an elevation of about 397 m within the Palisades fire boundary, and leaves overlap-related gaps in the mountains of the Eaton fire boundary. Brigham et al. [13] used this dataset through OpenTopography for their topographic differencing products, though the methods used to accommodate its noise characteristics are not fully documented. The incorporation of the cleaned and processed 2016 dataset into the multi-temporal analysis will be the subject of future work, which will enable a characterization of landscape change (construction, vegetation growth, grading) during the 2016–2023 interval. Though this dataset was not used in analysis, the clipped region is included in the curated dataset in Appendix A.

4. Methodology

We present a methodology for post-fire volumetric change detection using multi-temporal aerial LiDAR (Figure 4). Rather than performing a full semantic segmentation of both scenes, a task complicated by the scale of the datasets and the heterogeneity of the urban–wildland interface, we adopt a differencing-first approach: the pre-fire and post-fire point clouds are thinned to a common grid, and only cells exhibiting negative height differences exceeding a minimum threshold of $-20$ cm are retained for classification and volumetric summation. This threshold, set at approximately $3\sigma$ above the combined vertical noise floor (Section 5.4), ensures that the inventory captures genuine material loss while excluding noise-level fluctuations. The resulting pipeline prioritizes volumetric accuracy over scene completeness. It quantifies what was destroyed rather than attempting to label everything that remains. The retained volumes are subsequently aggregated by classification and height class and converted to emitted carbon through the summarization and emissions pipeline (Figure 5).

4.1. Source Datasets and Positional Accuracy

The pre-fire dataset (CA_LosAngeles_1_B23) was acquired in 2023 at Quality Level 1 specifications (Figure 3a). The post-fire datasets (CA_LAPostWildfire_Eaton_C25 and CA_LAPostWildfire_Palisades_C25) were acquired on 21–22 January 2025 at Quality Level 1 by NV5 Geospatial using a Riegl VQ-1560iiS dual-channel waveform LiDAR sensor flown at 3048 m AGL aboard a Cessna Caravan. Both datasets were delivered in UTM Zone 11N, NAD83(2011), with NAVD88 (GEOID18) vertical datum, and were already co-registered to a common projected coordinate system, eliminating the need for post hoc horizontal alignment.

The post-fire dataset achieved a tested NVA of 7.27 cm at the 95% confidence level against 22 independent checkpoints (RMSE_z = 3.71 cm), well within the required 19.6 cm specification [25]. The vegetated vertical accuracy (VVA) was 29.84 cm at the 95th percentile. Relative vertical accuracy averaged 1.9 cm RMSE_z. The pre-fire dataset was independently tested at a non-vegetated vertical accuracy of RMSE_z = 5.46 cm (against ground-classified LAS) in the project accuracy report [29] (Figure 3b).

Because change detection computes differences between two independent surfaces, the combined vertical uncertainty ${\sigma }_{\mathrm{diff}}$ propagates as the root sum of squares of the pre- and post-fire non-vegetated vertical accuracies, ${\sigma }_{\mathrm{pre}}$ and ${\sigma }_{\mathrm{post}}$ (the RMSE_z values of 5.46 cm and 3.71 cm reported above):

$${\sigma }_{\mathrm{diff}}=\sqrt{{\sigma }_{\mathrm{pre}}^2+{\sigma }_{\mathrm{post}}^2}\approx \sqrt{5.{46}^2+3.{71}^2}\approx 6.60\mathrm{cm}$$

(1)

At 95% confidence ($1.96{\sigma }_{\mathrm{diff}}$), this yields approximately 12.9 cm on non-vegetated surfaces, informing the selection of the change detection threshold and grid resolution (Figure 3c).

4.2. Data Preprocessing

Ground classification. Ground points were re-classified using the LAStools lasground64 tool, which implements the Progressive TIN Densification (PTD) algorithm [30], with default parameters. This normalized the ground classification method between the two datasets, as the original classification parameters were unknown.

Height normalization. The lasheight tool computed the height of every non-ground point above the interpolated ground surface, stored in decimeters in the user_data field, serving as the basis for building story estimation and vegetation height categorization.

Object classification. Above-ground points were classified into buildings and high vegetation using lasclassify with default parameters. The tool groups above-ground returns into candidate clusters and evaluates the local surface roughness of each cluster within a 2 m neighborhood, then assigns a class according to two rules. First, a cluster must stand at least 2 m above the interpolated ground surface to be considered, which suppresses low anthropogenic clutter. Second, clusters whose returns fit a local plane to within $\sigma \le 0.1$ m are assigned class 6 (Building), according to the American Society for Photogrammetry and Remote Sensing (ASPRS) classification scheme, reflecting the planar geometry of rooftops. Clusters with rugged geometry ($\sigma >0.4$ m) are assigned ASPRS class 5 (High Vegetation), reflecting the irregular structure of canopy. Returns that satisfy neither rule are left unclassified. This planarity-based convention has known limitations: complex or steeply pitched roofs can be misclassified as vegetation, vegetation overhanging a building edge is captured within the building class, low objects below the 2 m height cutoff remain unclassified, and natural features such as steep slopes with flat crests or trees with locally planar canopies can be misclassified as buildings.

4.3. Urban–Wildland Segmentation

The classified point clouds were segmented into urban and wildland zones. This segmentation was necessary because unclassified above-ground points carry different semantic meaning in each context: in wildland areas, they predominantly represent shrub and herbaceous vegetation, whereas, in urban areas, they correspond to a heterogeneous mix of anthropogenic objects (fences, vehicles, retaining walls, utility poles). Existing WUI mapping products, such as those derived from LANDFIRE vegetation layers and building footprint databases [31], operate at 30 m spatial resolution, too coarse for the point-level segmentation required here, and producing blocky boundaries that mix built and unbuilt parcels at the fine scale.

Starting from the 30 m boundary WUI layer, boundaries were refined by a visual inspection of the classified point cloud within the Potree viewer [27], with the objective of isolating areas that are overwhelmingly composed of built infrastructure from those that are overwhelmingly vegetated wildland. The resulting boundary polygon was applied using lasboundary to split each tile into urban and wildland subsets. The delineation polygons are included in the public data release (Appendix A) to enable reproducibility and sensitivity analysis. We note that, because the boundary separates predominantly built from predominantly vegetated zones, rather than attempting to resolve individual parcels at the WUI fringe, the sensitivity of total volumetric estimates to boundary placement is limited. Small shifts in the boundary redistribute volume between the urban and wildland reporting bins without changing the aggregate volume; the emissions total nonetheless retains a limited residual dependence on boundary placement through two zone-dependent accounting rules (buildings counted in the urban zone, sub-5 m vegetation in the wildland zone), quantified by the sensitivity analysis in Section 5.4.3.

4.4. Point-Cloud Thinning and Grid Resolution

Both subsets were thinned to a regular 0.5 m × 0.5 m grid using lasthin, retaining the highest point per cell. All original point attributes were preserved. The 0.5 m resolution standardizes comparison across epochs: at 4 cells/m², both datasets populate most cells with at least one return, minimizing data gaps from the density disparity. This resolution also corresponds to the DEM ground sample distance of both project deliverables [25]. Retaining the highest point per cell biases the surface toward the canopy top or rooftop, a deliberate choice for volumetric estimation. The pre-fire dataset’s lower density introduces a conservative bias: the selected “highest point” may underestimate the true canopy maximum, thereby underestimating vegetation losses.

4.5. Cell-Based Height Differencing

Pre-fire and post-fire tiles were compared on a tile-by-tile basis. Both point clouds were binned onto a shared 0.5 m grid, and within each cell the 2025 height above ground was subtracted from the 2023 value to obtain the signed height difference. A minimum threshold of 20 cm was applied. At the 95% confidence level, the propagated vertical accuracy of the difference surface is approximately 12.9 cm on non-vegetated surfaces (Equation (1)). The 20 cm threshold sits at approximately $3\sigma$ (6.60 cm) above the per-cell noise floor, about 1.5× the 95% detection level, to ensure that retained cells reflect genuine physical change rather than measurement uncertainty, co-registration residuals, or seasonal ground-moisture variation. Cells occupied in only one epoch were excluded. Results were written to per-tile CSV files recording coordinates, elevations, normalized heights, classifications from both epochs, and the signed height difference.

Non-fire contributions to height change. Height decreases between epochs could in principle arise from processes other than combustion (structure demolition, grading, seasonal vegetation change, or erosion) rather than fire loss. Several factors bound these contributions. The analysis is spatially restricted to the fire perimeter; the 14-day ignition-to-acquisition window and the pre-first-storm timing of the post-fire flight (Section 3.1) precede both large-scale debris removal and storm-driven sediment redistribution; both epochs are winter acquisitions, limiting seasonal vegetation change (Section 3.1); and demolition within a two-week window is negligible at landscape scale, as organized debris removal had not yet begun. Ground-surface change is sub-threshold on non-vegetated surfaces: the independent DTM differencing of Brigham et al. [13] over the same fires recovered only modest, largely sub-20 cm ground signals attributable to erosion and sediment reworking. The residual non-fire contribution to the reported losses is therefore small relative to the multi-meter combustion signal that dominates the inventory. Two features of the method further limit non-fire contamination over the longer pre-fire baseline gap. First, the inventory counts only detected height loss (2023 higher than 2025) and does not credit height gain, so new construction or vegetation growth between the baseline survey and the fire cannot generate a false loss cell; where such gains occur, they are simply not counted. Second, the analysis operates on above-ground normalized heights rather than the ground surface itself, so grading and other bare-earth reworking, which alter ground elevation without removing above-ground structure or canopy, do not enter the loss tally. The residual pre-fire non-fire signal is bounded empirically by the inter-epoch co-registration check (Section 5.4.5): on stable control points spanning the same baseline gap, height change unrelated to the fires is small and centered near the noise floor.

4.6. Edge Artifacts at Building Boundaries

Multi-temporal differencing introduces systematic artifacts along building edges from the mixed-pixel phenomenon [32]: when a LiDAR pulse straddles a rooftop-ground boundary, the returned range averages between the two surfaces. Because the pre- and post-fire acquisitions differ in altitude and scan geometry, the spatial distribution of these mixed-pixel returns differs between epochs, generating a ring of false elevation differences along building perimeters.

During development, we tested a morphological opening filter within a dilated building mask [33]. Diagnostic analysis showed that typical edge artifacts are 1–2 cells wide and sub-meter in magnitude. However, the morphological filter systematically removed approximately 97% of building-class loss cells, including genuinely destroyed structures, because the erosion step attenuated the sparse signal below the detection threshold. The per-structure envelope cross-check (Section 4.9) confirmed that the filter produced per-structure mean envelope volumes well below the expected residential range of 400–1500 m³, while the threshold-only pipeline produced means within that range (1013 m³ Palisades, 551 m³ Eaton).

We therefore adopted the threshold-only methodology. Mixed-pixel edge artifacts are retained as a residual systematic uncertainty; diagnostic analysis across all urban tiles indicates that the typical edge artifact is 1–2 cells wide with sub-meter magnitude, contributing an estimated +5% bias to building-class volume, well below the dominant ±19% uncertainty in the mass-density conversion factor (Section 5.4). This uncertainty was confirmed by a manual test sampling 75 destroyed buildings within the Palisades dataset, finding 13,817 erroneous points out of a total of 525,688 building classified points (+2.6%) in the associated pre-fire layer. The favorable per-structure envelope validation (Section 5.3) confirms that this residual contamination is small relative to the genuine multi-meter losses in destroyed building footprints.

4.7. Manual Classification Refinement

The change-detected results were subjected to manual inspection and classification refinement within the Potree viewer using a modified segmentation and export system with a 3D volumetric brush. Major corrections included the reclassification of rock outcrops and flat tree canopies erroneously identified as buildings, and the refinement of tree–building interfaces in urban areas. While this substantially improves reliability, residual misclassification is expected, as many points are sure to have escaped manual review in a dataset of this scale. The correction criteria were geometric and targeted the two dominant failure modes of the planarity-based classifier: building-class clusters whose planar surfaces corresponded to bare rock outcrops or locally flat tree canopies were reclassified, as were points along tree–building interfaces where overhanging canopy had been absorbed into the building envelope. Inspection was concentrated on the building-classified clusters and the urban/wildland fringe, where these failure modes are concentrated, rather than applied exhaustively to every point in the scene. The scale of the pre-refinement contamination is indicated by the edge-artifact test in Section 4.6 (2.6% of sampled building points), and the residual post-refinement error is quantified independently by the stratified accuracy assessment in Section 5.4.4, which found five misclassifications among 650 retained loss cells (overall accuracy 99.2%). Four of these are the spurious wildland building features that this refinement specifically targets. The available editing logs for the Palisades urban difference record 654,435 point reclassifications, of which 633,917 were mixed-pixel points along building edges (the cross-pixel edge fuzz discussed above) flagged as low-noise and withheld from the analysis, while 10,702 and 8985 were reclassified to vegetation and building, respectively, and 831 to unclassified. Comparable logs for the Eaton fire and the wildland zones were not systematically retained, so a complete before/after class-volume accounting across the full study area is not possible; the net effect of refinement on the reported quantities is instead bounded by the classification accuracy assessment (Section 5.4.4), which finds a post-refinement residual error below 1%. These counts were tabulated post hoc from the retained Palisades urban edit log, provided in the code repository (Appendix C); the reported classifications themselves were not altered by this tabulation.

4.8. Volume Loss Summarization

Volume loss was computed from cells with negative height differences exceeding the 20 cm threshold. The volume per cell was

$$V_{\mathrm{cell}}=\left|\mathrm{\Delta }h\right|\mathrm{m}\times 0.25{\mathrm{m}}^2$$

(2)

Classification was taken from the 2023 pre-fire epoch. Cells were assigned to the following categories:

Buildings (ASPRS class 6), stratified by story count, following Wu et al. [34]: single-story (<4 m), two-story (4–8 m), three-or-more stories (≥8 m).

Vegetation (ASPRS class 5), stratified by height: under 1 m, 1–5 m, and over 5 m. The 5 m threshold separates chaparral and sage scrub communities, which in the Santa Monica Mountains typically reach maximum heights of 3–4 m [35], from oak woodland canopy, which generally begins at 5–8 m. In the emissions calculation, the less than 1 m and 1–5 m vegetation classes include only wildland-zone volumes; urban-zone vegetation in these classes is excluded because the automated classifier performs poorly below 5 m in developed areas, where anthropogenic objects (fence lines, landscaping features, retaining walls, pitched-roof edges) are frequently misclassified as vegetation. Urban vegetation in the greater than 5 m class is included, as above-ground features exceeding 5 m in residential neighborhoods are overwhelmingly true canopy (Section 6.2).

Unclassified (all other classes), stratified by the same height breaks. All unclassified categories are excluded from the emissions calculation. In wildland areas, the less than 1 m unclassified category, which dominates by cell count, exhibits average per-cell height differences of approximately 12 cm, close to the 20 cm detection threshold. This suggests that a substantial fraction of these cells may reflect ground-classification inconsistencies between epochs. Some points are near-ground in both datasets but are classified differently by the PTD algorithm due to differences in point density and local surface roughness, rather than genuine vegetation loss. Because genuine low vegetation and ground-classification artifacts cannot be reliably separated without manual inspection, the conservative choice is to exclude the entire category. For the wildland 1–5 m unclassified category, which more plausibly represents real shrub vegetation missed by the classifier, the excluded volumes are well under 1% of the total vegetation emissions, representing 376,252 m³ (Palisades) and 252,814 m³ (Eaton). Even if these were entirely attributed to vegetation, the resulting additional carbon would be approximately 0.3 kt C and 0.2 kt C respectively. In urban areas, unclassified points represent a heterogeneous mix of anthropogenic objects (parked vehicles, fences, retaining walls, landscaping, utility infrastructure) whose combustion characteristics vary widely and cannot be assigned a single mass-density or carbon-fraction factor without individual object identification. Full attribution would require semantic segmentation (e.g., RandLA-Net, PointNet++), proposed as future work (Section 6.4). The exclusion of all unclassified categories produces a conservative bias in the reported emissions totals, bounded at approximately 2–3% for vegetation and unquantified but likely modest for urban anthropogenic objects, with vehicles representing the most significant omission.

4.9. Visualization and Validation Procedure

The volumetric outputs are derived from a chain of operations in which errors at any stage propagate into the final totals. We implemented a two-stage validation: visual inspection of intermediate products in Potree, and programmatic cross-checking against independent estimates.

Visual inspection was performed at three pipeline stages: post-classification (correcting misclassifications via the 3D volumetric brush), post-differencing, and pre-publication end-to-end inspection in the ALERTCalifornia Digital Atlas Archive, built upon the OpenHeritage3D.org framework [36], feeding into a globally unified digital twin system [37] (Figure 6).

Programmatic cross-validation included two checks:

Building envelope volume vs. CAL FIRE structure inventory. Total building-class volume divided by the DINS destroyed-structure count produces per-structure mean envelope volumes. For Los Angeles area residential construction (mean footprint ∼200 m², one to two stories), the expected range is approximately 400–1500 m³. Results outside this range triggered pipeline investigation.

Vegetation volume vs. published areal biomass. Wildland vegetation volume was compared against expected values derived from Schrader-Patton and Underwood [19] areal biomass densities divided by community bulk densities.

These checks proved decisive: the per-structure cross-check revealed that the morphological-filter pipeline (Section 4.6) produced volumes outside the expected range, leading to the adoption of the threshold-only methodology. The vegetation cross-check confirmed that bulk-density factors recovered biomass densities consistent with published values.

4.10. Translation of Volumetric Loss to Carbon Emissions

Because combustion behavior differs fundamentally between built structures and vegetation, the two are converted to emitted carbon through distinct pathways (Figure 5).

4.10.1. Building Emissions: Story-Stratified Floor Area Approach

The LiDAR-derived envelope volume includes pitched-roof void space and geometric variability that does not correspond to additional combustible material. Combustible mass scales most reliably with gross floor area [38,39].

Floor area inference. Destroyed floor area was inferred from per-class envelope volume using representative mean envelope heights:

Single-story (∼4 m): $A_{\mathrm{floor}}=V_{\sin \mathrm{gle}}/\left(4\mathrm{m}\right)$. Two-story (∼6 m, two floors): $A_{\mathrm{floor}}=V_{\mathrm{two}}/\left(3\mathrm{m}\right)$. Three-or-more-story (∼10 m, ≥3 floors): A_floor = 0.3 V_three+.

Combustible mass density. The central value of ${\sigma }_m=195$ kg/m² follows from the Hartwell et al. [39] Environmental Protection Agency (EPA) WUI inventory, which catalogued total combustible mass (structural materials, fixtures, furnishings, and contents) of an average single-family home at 36,600 kg for a 187.3 m² reference home. This is consistent with Xie et al. [40]’s contents-only fire load of ∼35 kg/m² when the structural materials’ contribution is accounted for. Central: 195 kg/m². Range: 140–290 kg/m².

Combustion completeness. Central: $c_b=0.80$. Range: 0.70–0.90 [38,41,42].

Carbon mass fraction. Central: $f_{C,b}=0.45$. Range: 0.40–0.50 [43,44].

Building carbon emissions:

$$C_{\mathrm{building}}=A_{\mathrm{floor},\mathrm{total}}\times {\sigma }_m\times c_b\times f_{C,b}$$

(3)

4.10.2. Vegetation Emissions: Volume-Based Biomass Approach

For vegetation, the LiDAR-derived column volume is a more direct proxy for combustible biomass than floor-area-equivalent approaches, because shrub and tree biomass occupies the full vertical extent of the canopy column. Bulk density values were assigned by height class (

Table 3. Vegetation bulk density assignments.

Height Class	Dominant Community	$\mathit{\rho }$ (kg/m3)
<1 m	Coastal sage scrub/herbaceous	0.9
1–5 m	Mixed/chamise chaparral	2.0
>5 m	Oak woodland/mixed canopy	3.0

).

These values derive from areal biomass estimates by Schrader-Patton and Underwood [19] for Southern California chaparral national forests (2.3–3.5 kg/m² above-ground), divided by typical canopy heights, cross-referenced with Countryman and Philpot [45] and Xu and Greenberg [21].

Combustion completeness: chaparral (<5 m), $c_v=0.90$; oak woodland (>5 m), $c_v=0.65$ [43,46]. Carbon fraction: $f_{C,v}=0.47$. Range: 0.44–0.50 [43,47].

$$C_{\mathrm{veg}}=\sum _i\left(V_i\times {\rho }_i\times c_{v,i}\times f_{C,v}\right)$$

(4)

4.10.3. Uncertainty Propagation

For a product $Y=X_1·X_2\cdots X_n$, relative uncertainty propagates as

$${\left({\displaystyle \frac{{\sigma }_Y}{Y}}\right)}^2={\displaystyle \sum _{i=1}^n}{\left({\displaystyle \frac{{\sigma }_{X_i}}{X_i}}\right)}^2$$

(5)

Building pathway: $\pm 29\%$ ($1\sigma$), from volume-to-area (12.5%), mass density (19.2%), combustion completeness (12.5%), and carbon fraction (11.1%). Vegetation pathway: $\pm 40\%$ ($1\sigma$), from volume (20%), bulk density (30%), combustion completeness (15%), and carbon fraction (6.4%). Total emitted carbon $C_{\mathrm{total}}=C_{\mathrm{building}}+C_{\mathrm{veg}}$, with uncertainty propagated as the root sum of squares of the independent pathway uncertainties.

Figure 7 decomposes this budget by parameter, varying each factor by $\pm 1\sigma$ about its central value. The two dominant terms are the vegetation bulk density and the building combustible mass density (±49 kt C each on the 418 kt C grand total), confirming that these are the parameters to which the total is most sensitive. Their applicability to the specific building stock and vegetation communities therefore warrants comment. The building mass density derives from the Hartwell et al. [39] single-family-home average and may under-represent the larger, frequently two- and three-story custom residences of Pacific Palisades relative to the predominantly single-story bungalow and ranch stock of Altadena; the wide 140–290 kg/m² range is intended to span this variation, and the ∼6% agreement of the bottom-up Eaton total with the independent atmospheric inversion (Section 6.1) bounds the aggregate emission magnitude for that fire, though only jointly with the other conversion parameters. The vegetation bulk densities were assigned by height class from Southern California chaparral and oak-woodland values and are most applicable to the dominant shrub and woodland communities; sparse coastal sage on lower Palisades slopes and mixed-conifer stands at higher Eaton elevations are less well-represented, motivating the wide ±30% range adopted for this factor.

5. Results

5.1. Volumetric Loss Totals

The pipeline’s per-class volumetric loss totals were aggregated across the urban and wildland subsets of each fire. The total detected volume loss was approximately 145.7 million m³ for the Palisades fire and 75.7 million m³ for the Eaton fire. Vegetation-class cells account for the large majority of this volume in both fires (roughly 92% for Palisades and 89% for Eaton), with building-class cells contributing a small volumetric fraction whose disproportionate significance for carbon emissions emerges in the following sections. The full per-category, per-zone breakdown, segmented by urban and wildland zone, is provided in Appendix B.

The total area of detected loss across all categories (computed as cell count × 0.25 m²) was approximately 60.2 km² for the Palisades fire and 31.5 km² for the Eaton fire. Of this, vegetation-class cells account for 49.5 km² and 24.3 km² of detected vegetation loss for the Palisades and Eaton fires respectively, while building-class cells account for 1.66 km² and 1.48 km² of detected building footprint loss. Dividing the building footprint areas by the DINS destroyed-structure counts yields per-structure mean footprints of 242 m² (Palisades) and 157 m² (Eaton), consistent with typical Los Angeles residential construction and with the per-structure envelope volume contrasts reported in Section 5.3.

The building volumetric losses reported in Section 5.1 were converted to gross destroyed floor area using the per-story envelope-height ratios described in Section 4.10.1, then to combusted carbon using the central conversion factors ${\sigma }_m=195$ kg/m², $c_b=0.80$, and $f_{C,b}=0.45$ (

Table 4. Destroyed gross floor area and building carbon emissions. Per-structure mean computed using DINS counts (6837 Palisades, 9414 Eaton). Uncertainty: $\pm 29\%$ ($1\sigma$).

Fire	1-Story (m2)	2-Story (m2)	3+ (m2)	Total (m2)	Per-Struct. (m2)	C (kt)
Palisades	484,935	1,327,481	301,666	2,114,082	309	$148\pm 43$
Eaton	644,982	751,200	105,336	1,501,518	160	$105\pm 30$

).

The per-structure mean floor areas of 309 m² (Palisades) and 160 m² (Eaton) bracket the broader range of Los Angeles area residential construction and are consistent with the contrasting housing stock of the two fire footprints. Pacific Palisades is characterized by a higher proportion of two-story and three-or-more-story homes with larger footprints, while Altadena is dominated by single-family ranch and bungalow construction with smaller per-structure floor areas. The factor-of-two difference in per-structure floor area between the two fires emerges directly from the LiDAR-derived envelope volumes and corresponds well to assessor records of mean parcel improvement square footage in the two communities.

5.2. Vegetation Emissions

The vegetation volumetric losses were converted to combusted carbon through the bulk-density and combustion-completeness factors described in Section 4.10.2 (

Table 5. Vegetation carbon emissions by fire footprint. Under-1 m and 1–5 m classes include wildland-zone volumes only; the over-5 m class includes both wildland and urban zones. Uncertainty: $\pm 40\%$ ($1\sigma$).

Fire	<1 m (m3)	1–5 m (m3)	>5 m (m3)	C (kt)	$\pm 1\mathit{\sigma }$
Palisades	463	91,452,885	32,315,777	107	43
Eaton	623	33,685,476	31,319,781	57	23

). Wildland vegetation accounts for the substantial majority of vegetation loss in both fires. For the less than 1 m and 1–5 m height classes, only wildland-zone vegetation is included in the emissions calculation, because the automated vegetation classifier performs poorly in urban settings at these heights: fence lines, landscaping, retaining walls, and pitched-roof edges are frequently misclassified as vegetation, and the combustion characteristics of this heterogeneous mix cannot be assigned a single bulk-density or combustion-completeness factor. Urban vegetation in the greater than 5 m class is included, as above-ground features exceeding 5 m in urban areas are overwhelmingly true canopy (mature oaks, sycamores, deodars, and ornamental trees) rather than anthropogenic objects, and can be assigned the same oak-woodland conversion factors used in the wildland zone.

A notable qualitative observation from the urban zones is the high survival rate of mature trees within neighborhoods where surrounding buildings were completely destroyed. A visual inspection of the post-fire point cloud reveals many parcels in which the building envelope has been entirely consumed while adjacent trees retained substantial canopy structure. This pattern, which is consistent across both fire footprints, suggests that mature urban trees, particularly the deep-rooted, thick-barked species common in these neighborhoods, are substantially more resistant to structure-fire radiation and convective heating than the structures themselves. The survival of urban canopy complicates the application of wildland combustion-completeness factors in urban zones below 5 m, where the mix of surviving trees, partially consumed ornamental plantings, and misclassified anthropogenic objects cannot be reliably decomposed without per-object semantic segmentation. The exclusion of the under-5 m urban vegetation introduces a conservative bias in the reported emissions totals; even if this volume were entirely attributed to consumed vegetation, the additional carbon would be approximately 9 kt C (Palisades) and 2 kt C (Eaton), modest relative to the total.

Palisades vegetation emissions exceed Eaton’s despite a smaller burned area and structure count, reflecting the greater chaparral coverage and density characteristic of the Santa Monica Mountains relative to the urban-margin foothills of the San Gabriel Mountains, where the Eaton fire footprint included a higher proportion of developed parcels and shorter wildland fringe.

5.3. Validation Against External Estimates

The validation procedures described in Section 4.9 were applied to the production pipeline outputs.

Per-structure envelope volume. Dividing total building-class volume by the DINS-reported destroyed-structure count produces an inferred mean per-structure envelope volume (

Table 6. Per-structure envelope volume validation.

Fire	Building Vol (m3)	DINS Count	Per-Struct. (m3)	In Range?
Palisades	6,927,736	6837	1013	Yes
Eaton	5,184,649	9414	551	Yes

).

Both fires pass this validation. The factor-of-two contrast in per-structure means is itself a physically meaningful result: the Palisades mean of 1013 m³ lies in the upper portion of the expected range, consistent with the larger and frequently two- or three-story residential stock of Pacific Palisades, while the Eaton mean of 551 m³ lies in the lower-middle of the range, consistent with the single-story bungalow and ranch typology that dominates Altadena. The agreement between the LiDAR-inferred per-structure envelopes and the DINS-anchored expectation confirms that the threshold-only differencing pipeline recovers the correct order-of-magnitude signal and that residual mixed-pixel edge contamination is small.

Per-area wildland vegetation volume. Both fires fall within expected ranges of ∼1.2–2.0 m³/m² for chaparral and ∼0.5–1.2 m³/m² for mixed oak woodland, derived from Schrader-Patton and Underwood [19].

Cross-validation against Vannucci et al. atmospheric inversion. The bottom-up Eaton total of $162\pm 38$ kt C (building + vegetation) agrees to within 6% of the Vannucci et al. [15] estimate of 153 kt C, well within the ±20–30% propagated uncertainty of either method. These two methods share no upstream inputs, calibration constants, or modeling pipeline: one measures consumed volume at the source, while the other measures carbon-species enhancement in the downwind atmosphere. The convergence lends substantial confidence to the magnitudes reported by both; because the two totals depend on different parameter sets, the agreement constrains the aggregate emission magnitude rather than validating the Hartwell et al. [39] mass-density value of 195 kg/m² individually, and compensating errors among parameters cannot be excluded.

A Palisades cross-check is unavailable because the plume was carried offshore (Figure 2). However, the Li et al. [16] satellite-based analysis provides complementary constraints for both fires. Their derived emission coefficients for CO, TPM, and NO_x, distinct from typical chaparral values and indicative of an urban fuel signature, are qualitatively consistent with our finding that building combustion constitutes a dominant fraction of the Eaton carbon budget. A direct quantitative comparison of the Palisades LiDAR estimate with Li et al. total carbon values requires species-specific conversion factors and integration across their per-overpass-day temporal resolution, which we identify as an opportunity for future cross-method reconciliation. The Palisades bottom-up estimate of $255\pm 61$ kt C therefore stands as the most direct total-carbon emissions estimate currently available for that fire.

5.4. Uncertainty and Interpretation

5.4.1. Sources of Systematic Uncertainty

Per-cell random uncertainty (∼0.013 m³) is negligible when aggregated across millions of cells. The principal systematic sources are as follows:

(a)

Classification error. Misidentification between building, vegetation, and unclassified categories redistributes volume between reporting bins but does not affect the total across all categories. The impact is greatest for buildings, where pitched-roof mis-classification shifts volume to the vegetation total. Manual refinement reduced but did not eliminate this error. The residual classification error is quantified directly by the stratified accuracy assessment of Section 5.4.4.

(b)

Mixed-pixel edge effects. Residual edge artifacts surviving the 20 cm threshold contribute a small positive bias to urban volume estimates, bounded at approximately +5% of building volume, well below the dominant ±19% mass-density uncertainty.

(c)

Sampling bias from density disparity. The pre-fire dataset’s lower density produces a conservative underestimation of vegetation volumes, as the “highest point” may fall below the true canopy top.

(d)

Height threshold boundary effects. Points near category boundaries (4 m, 5 m, 8 m) may be misassigned to adjacent story or height classes, but the ambiguity zone (∼±10 cm) is narrow relative to bin widths.

(e)

Detection threshold exclusion. The 20 cm threshold excludes genuine small-magnitude losses, representing a systematic underestimate that is minor for buildings and tall vegetation but may be more significant for the under 1 m category. The sensitivity of the reported totals to this threshold is quantified in Section 5.4.2.

(f)

Exclusion of unclassified categories and urban under-5 m vegetation from emissions. All unclassified points are excluded from the carbon emissions calculation (Section 4.8). In wildland areas, the excluded volume is bounded: even if the entire wildland unclassified 1–5 m category were to be attributed to vegetation, the additional carbon would be less than 1% of the vegetation total for either fire. The under-1 m wildland unclassified category is larger by cell count but carries the highest contamination risk from ground-classification artifacts and contributes negligible carbon at the low bulk densities of herbaceous cover. In urban areas, the excluded unclassified volume (590,430 m³ Palisades; 161,234 m³ Eaton in the under-1 m bin) includes an unquantified contribution from combusted vehicles, landscaping, and accessory structures. Urban vegetation below 5 m is also excluded due to unreliable classification and the observed high survival rate of urban canopy (Section 6.2); if this volume were entirely attributed to consumed vegetation, the additional carbon would be approximately 9 kt C (Palisades) and 2 kt C (Eaton). The combined effect of all exclusions is a conservative bias bounded at approximately 4% for the vegetation pathway and unquantified but likely modest for the urban anthropogenic pathway, with vehicles representing the most significant omission. As an order-of-magnitude bound on that largest term, attributing one to two destroyed vehicles per destroyed structure at a representative 0.15–0.3 t of combustible carbon each (roughly 300 kg of combustible material at a carbon fraction of ∼0.6 and near-complete combustion) yields on the order of 1–3 kt C per fire, below 2% of either fire’s reported total and within the propagated uncertainty.

(g)

Grid thinning and resolution. Binning to one highest point per 0.5 m cell discards sub-cell structure, and the “highest point” rule biases retained heights upward; because this bias is common to the pre- and post-fire surfaces, it largely cancels in the difference. Re-binning at coarser or finer resolution rescales per-cell volumes but not the aggregate signal, which is dominated by multi-meter losses (Section 5.4.2).

(h)

Vegetation phenology and seasonality. Both epochs were acquired in the winter (Section 3.1), limiting phenological mismatch across the predominantly evergreen chaparral and oak canopy; residual seasonal differences in canopy state and ground moisture are absorbed into the vegetation-pathway uncertainty and are further suppressed by the 20 cm threshold, which exceeds typical seasonal ground-moisture variation.

(i)

Manual refinement uncertainty. The manual reclassification (Section 4.7) is targeted rather than exhaustive at full point resolution, so residual misclassification remains; its magnitude is bounded directly by the accuracy assessment (Section 5.4.4), which finds a post-refinement error rate below 1%.

5.4.2. Threshold Sensitivity

To assess robustness to the 20 cm detection threshold, the change-detection pipeline was re-run with the differencing floor lowered to 10 cm (the decimeter resolution of the stored height field, approximately $1.5\sigma$ of the propagated vertical noise), and the volumetric and carbon inventories were recomputed at 10, 20, 30, 40, and 50 cm (

Table 7. Threshold sensitivity of carbon and detected loss area. Carbon totals vary <1% across the 10–50 cm sweep, far within the ±23–24% ($1\sigma$) conversion-factor uncertainty; detected loss area is threshold-sensitive.

Fire	Thr. (cm)	Building (kt C)	Veg. (kt C)	Total (kt C)	Loss Area (km2)
Palisades	10	148.7	107.0	255.7	66.95
	20	148.4	107.0	255.4	60.22
	30	148.1	106.9	255.0	57.23
	40	148.0	106.7	254.7	55.33
	50	147.8	106.6	254.4	53.82
Eaton	10	105.5	57.3	162.8	35.32
	20	105.4	57.2	162.6	31.49
	30	105.3	57.1	162.4	29.40
	40	105.2	56.9	162.1	27.92
	50	105.2	56.7	161.9	26.69

, Figure 8). Total emitted carbon varied by less than 1% across this range for both fires, far within the ±23–24% ($1\sigma$) uncertainty of the conversion factors, because the emissions-bearing categories (destroyed building floor area and tall-canopy volume) comprise multi-meter losses that are insensitive to threshold placement. Detected loss area, by contrast, varied by 22–27% across the sweep, as it includes near-threshold cells dominated by ground-classification artifacts in the excluded unclassified categories. This behavior motivates the conservative ∼$3\sigma$ (20 cm) threshold for area reporting while confirming that the reported carbon totals are effectively threshold-independent. As a reproducibility check, the 20 cm inventory recovered from the lowered-floor superset reproduces the per-class totals of

Table A1. Palisades urban subset volume loss totals.

Category	Cells	Volume Lost (m3)
Building Single Story	2,830,510	1,939,741
Building Two Story	3,160,903	3,982,442
Building Three Story Plus	595,310	1,005,554
Vegetation Under 1 m	1435	199
Vegetation 1–5 m	19,402,628	10,365,456
Vegetation Over 5 m	8,458,535	8,731,963
Unclassified Under 1 m	5,182,815	590,430
Unclassified 1–5 m	295,254	73,420
Unclassified Over 5 m	712	464
TOTAL	39,928,102	26,689,669

,

Table A2. Palisades wildland subset volume loss totals.

Category	Cells	Volume Lost (m3)
Building Single Story	23,849	13,282
Building Two Story	16,234	19,802
Building Three Story Plus	4235	8139
Vegetation Under 1 m	2537	463
Vegetation 1–5 m	148,408,982	91,452,885
Vegetation Over 5 m	21,909,331	23,583,814
Unclassified Under 1 m	29,048,349	3,538,497
Unclassified 1–5 m	1,526,628	376,252
Unclassified Over 5 m	1	2
TOTAL	200,940,146	118,993,136

,

Table A3. Eaton urban subset volume loss totals.

Category	Cells	Volume Lost (m3)
Building Single Story	3,759,672	2,579,927
Building Two Story	1,963,242	2,253,601
Building Three Story Plus	187,978	351,121
Vegetation Under 1 m	169	28
Vegetation 1–5 m	5,144,629	2,629,305
Vegetation Over 5 m	7,817,847	11,423,900
Unclassified Under 1 m	1,375,869	161,234
Unclassified 1–5 m	88,460	19,849
Unclassified Over 5 m	1	3
TOTAL	20,337,867	19,418,968

and

Table A4. Eaton wildland subset volume loss totals.

Category	Cells	Volume Lost (m3)
Building Single Story	7458	4054
Building Two Story	6156	5662
Building Three Story Plus	4804	6193
Vegetation Under 1 m	3642	623
Vegetation 1–5 m	64,206,002	33,685,476
Vegetation Over 5 m	19,948,319	19,895,881
Unclassified Under 1 m	20,343,855	2,429,133
Unclassified 1–5 m	1,104,715	252,814
Unclassified Over 5 m	113	209
TOTAL	105,625,064	56,280,045

exactly.

5.4.3. Boundary Placement Sensitivity

The urban/wildland boundary was hand-refined in Potree (Section 4.3), and its placement affects the carbon total through two zone-dependent accounting rules: buildings are counted only in the urban zone, and vegetation below 5 m only in the wildland zone. Vegetation above 5 m and unclassified points are treated identically in both zones and are therefore unaffected by the boundary. To bound this sensitivity, the genuine urban/wildland interface (the portion of the boundary interior to the fire, holding the coastline and fire-perimeter edges fixed) was offset inward and outward by up to ±50 m, and the loss cells were re-tagged and the carbon recomputed (

Table 8. Urban/wildland boundary-placement sensitivity. The interface is offset in/out by up to ±50 m (positive = urban grows into wildland; coastline and fire-perimeter edges held fixed). Total carbon varies by less than 1.4% for either fire, within the ±23–24% ($1\sigma$) conversion-factor uncertainty.

Fire	Offset (m)	Building (kt C)	Veg. (kt C)	Total (kt C)
Palisades	$-50$	146.4	109.0	255.5
	$-25$	147.9	108.2	256.1
	$-10$	148.2	107.5	255.8
	0	148.4	107.0	255.4
	$+10$	148.5	106.4	254.8
	$+25$	148.5	105.3	253.8
	$+50$	148.5	103.6	252.1
Eaton	$-50$	104.2	57.5	161.7
	$-25$	105.2	57.4	162.5
	$-10$	105.4	57.3	162.6
	0	105.4	57.2	162.6
	$+10$	105.4	57.1	162.5
	$+25$	105.4	57.0	162.4
	$+50$	105.4	56.8	162.2

, Figure 9). Total emitted carbon varied by less than 1.4% for either fire across the full ±50 m range, within the ±23–24% ($1\sigma$) conversion-factor uncertainty. The building and vegetation responses partially offset: expanding the urban zone adds fringe buildings to the building pathway while removing sub-5 m vegetation from the wildland pathway, so the net change in total carbon is smaller than either component change, and for the Eaton fire the two nearly cancel. The perturbed boundary-scenario polygons are included in the public data release (Appendix A).

5.4.4. Classification Accuracy

Because the 2023 classification determines the emission bin of every retained loss cell, its accuracy was assessed directly following the stratified good-practice protocol of Olofsson et al. [48]. A stratified random sample of 650 retained loss cells was drawn, stratified by mapped class (building, vegetation, unclassified) and zone (urban, wildland) following the sampling design summarized in

Table 9. Accuracy-assessment sampling design: per-stratum population, area weight, and sample size (pooled across both fires).

Class	Zone	Population (Cells)	Weight	Sample n
Building	Urban	12,497,621	0.034	100
Building	Wildland	62,736	0.0002	50
Vegetation	Urban	40,825,297	0.111	100
Vegetation	Wildland	254,478,813	0.694	175
Unclassified	Urban	6,943,051	0.019	100
Unclassified	Wildland	52,023,661	0.142	125
Total		366,831,179	1.000	650

; each cell was labeled to its true class by visual inspection against both LiDAR epochs and post-fire orthoimagery, blind to the mapped class, with all 650 samples inspected. The confusion matrix (

Table 10. Classification confusion matrix for the retained loss cells (2023 class assignment; $n=650$). Rows are the mapped class; columns are the reference (true) class. Raw overall accuracy 99.2%; area-weighted overall accuracy 99.9% (±0.2); $\kappa =0.997$. UA: user’s accuracy; PA: producer’s accuracy (raw sample counts).

Mapped	Reference			Total	UA
	Building	Vegetation	Unclassified
Building	146	2	2	150	97.3%
Vegetation	0	274	1	275	99.6%
Unclassified	0	0	225	225	100.0%
Total	146	276	228	650
PA	100.0%	99.3%	98.7%

) contains five misclassifications: four in the wildland building stratum (two rock outcrops and two trees adjacent to a water tower, mapped as building) and one in the urban vegetation stratum (post-fire structural debris mapped as vegetation).

The per-stratum sample sizes, populations, and weights are given in Table 9. The sample was drawn independently at random within each stratum by an exact two-pass index method over the full retained-cell population; the minimum separation between any two sampled cells was 1.44 m (Palisades) and 5.85 m (Eaton), so no two neighboring 0.5 m cells entered the sample and spatial autocorrelation between samples is not a concern.

The raw sample overall accuracy is $645/650=99.2\%$; the area-weighted estimator of Olofsson et al. [48] yields an overall accuracy of 99.9% (±0.2, 95% CI) and $\kappa =0.997$, the two differing because the errors concentrate in the negligible-weight wildland building stratum. Urban building classification (the pathway that governs the building carbon estimate) was 100% accurate (100/100); every building error occurred in the wildland zone, where such cells are spurious rock and canopy features already excluded from the emissions calculation (Section 4.3). Applying the error-adjusted area estimator to the same matrix (

Table 11. Error-adjusted loss areas for the retained inventory, following Olofsson et al. [48]. Adjusted areas differ from the mapped areas by less than 0.15 km².

Class	Mapped (km2)	Adjusted (km2)	95% CI (km2)
Building	3.14	3.14	<0.01
Vegetation	73.83	73.72	±0.20
Unclassified	14.74	14.84	±0.20

) changes the class loss areas by less than 0.15 km², with 95% confidence intervals of ±0.2 km², confirming that classification error contributes negligibly to the reported loss-area totals.

This agreement is higher than typical whole-scene airborne-LiDAR classification benchmarks (roughly 80–85% overall accuracy for multi-class ALS labeling tasks [49] and 90–95% for building/vegetation/ground schemes [50]), but is not directly comparable to them. The assessment is scoped to the change-detected cells that enter the inventory, a high-signal subset dominated by large, unambiguous geometric changes (collapsed roofs, lost canopy), and the 20 cm detection threshold removes the marginal, low-relief cells where classifiers most often fail; reported benchmark accuracy in ambiguous boundary regions falls to 70–73% for buildings and low vegetation [50], precisely the population excluded here. The assessment therefore quantifies the reliability of the classifications as used in the volumetric inventory rather than the accuracy of the underlying classifier over the full scene.

Two limitations follow from this scoping. First, the assessment is structurally blind to classification errors that are stable across both epochs, since a cell misclassified identically in 2023 and 2025 exhibits no height change and never enters the differenced set; the magnitude of any such population is bounded independently by the DINS building-count and biomass cross-validations (Section 4.9), which constrain the aggregate class balance and would not agree if stable misclassification was substantial. Second, with only five observed errors, the off-diagonal cells and producer’s-accuracy intervals rest on small counts and are best read as bounds rather than precise rates.

5.4.5. Inter-Epoch Co-Registration

The agreement of coordinate system and vertical datum (Section 3.1) does not by itself demonstrate sub-decimeter relative vertical registration between the two epochs, and Section 4.2 identifies non-fire height change over the pre-fire baseline gap as a distinct concern. Both are addressed with the same measurement. Forty stable hard-surface control points (paved roads) were selected per fire, distributed across each study area and restricted to flat pavement away from curbs, driveways, and vegetation. For each control point, the nearest point in the 2023 and 2025 thinned point clouds was located independently (the same physical point need not exist in both epochs) and the two elevations differenced.

Table 12. Inter-epoch vertical co-registration on stable hard-surface control points. Bias is the mean (2025 minus 2023) elevation difference; RMSE is the root-mean-square difference; n is the number of the 40 sampled control points per fire that matched a cloud point within 1 m in both epochs.

Fire	n	Bias (cm)	95% CI (cm)	RMSE (cm)
Palisades	37	$-5.65$	$-6.77$ to $-4.53$	6.55
Eaton	27	$-6.11$	$-7.40$ to $-4.83$	6.89
Pooled	64	−5.84	−6.67 to −5.02	6.70

reports the resulting vertical offset (bias), root-mean-square error (RMSE), and 95% confidence interval, both per fire and pooled.

A small offset was detected: the pooled bias is statistically significant ($t=-14.2$, $p<0.001$), with the 2025 surface sitting on average 5.8 cm lower than the 2023 surface on stable, non-fire-affected ground. Because the reported height difference is defined as the 2023 value minus the 2025 value (Section 4.2), this offset biases stable cells toward a small apparent loss rather than away from one; however, at 5.8 cm, it remains below the 20 cm detection threshold on any single cell and cannot by itself generate a retained loss cell. Its magnitude is comparable to the propagated 1$\sigma$ vertical noise floor (6.60 cm, Equation (1)), and its effect on the population of near-threshold cells is already bounded by the threshold-sensitivity analysis (Section 5.4.2), which shows the reported carbon totals change by less than 1% across a 10–50 cm sweep. No additional vertical correction was applied to the production dataset: a fire-independent offset of this size is consistent with ordinary inter-survey calibration differences between two independently flown LiDAR acquisitions and does not indicate a registration failure. This measurement also directly bounds the non-fire background invoked in Section 4.2: on the same stable control points, ground-surface change unrelated to the fires is small and centered near the noise floor, consistent with the independent DTM comparison by Brigham et al. [13] cited there.

5.4.6. Recommended Uncertainty Bounds

Buildings (±10–15%): relatively reliable classification, manual refinement, and bounded edge effects. Vegetation (±15–25%): wider uncertainty from density asymmetry, canopy height estimation, and the 5 m chaparral/oak threshold. Unclassified (±25–35%): the most heterogeneous mix of surface types; should be treated as order-of-magnitude indicators.

All reported values should be understood as lower bounds on the true volume loss. The degree of underestimation varies by category, with vegetation losses most likely underrepresented and building losses most accurately captured.

6. Discussion

6.1. Comparison with Independent Emissions Estimates

The bottom-up Eaton estimate of $162\pm 38$ kt C and the Vannucci et al. [15] atmospheric inversion of 153 kt C converge to within 6% of each other, well within the propagated uncertainty of either method, providing an independent cross-validation at the total-emission level between these two methodological approaches for a major WUI fire. The methods share no upstream inputs: one measures consumed volume at the source, the other measures the enhancement of carbon-containing species in the downwind atmosphere. The agreement holds at the level of total emissions and does not isolate individual parameters: because the bottom-up total depends on the joint effect of the mass density, combustion completeness, and carbon fraction, compensating errors among these factors cannot be excluded, and the comparison should not be read as validating the 195 kg/m² mass density in isolation.

The three available approaches (atmospheric inversion, satellite emissions retrieval, and LiDAR bottom-up) measure overlapping but non-identical quantities. The atmospheric approach measures gaseous emissions within the sensor footprint; the satellite approach measures species-specific emission rates at the resolution of satellite overpasses; the LiDAR approach measures total consumed material volume, convertible to total combusted carbon through mass-density and combustion-completeness factors. Numerical agreement is therefore not expected in any simple sense, but convergence across independent methods strengthens confidence in each.

Qualitatively, all three approaches converge on the finding that structural combustion dominates total carbon emissions from the Eaton fire relative to vegetation combustion, a distinguishing characteristic of WUI fires. The LiDAR analysis independently identifies this asymmetry and additionally reveals the contrasting spatial structure: the Palisades fire is vegetation-dominated, while the Eaton fire is building-dominated. This per-class spatial decomposition is structurally unavailable from atmospheric or satellite methods.

6.2. Urban Vegetation Treatment

The treatment of vegetation within urban zones warrants separate discussion because of the contrasting behavior of trees and structures observed in the post-fire point cloud. Visual inspection across both fire footprints reveals a striking pattern: in many urban parcels where buildings were completely destroyed, adjacent mature trees retained substantial canopy structure. This high survival rate of urban canopy, particularly deep-rooted, thick-barked species such as coast live oak (Quercus agrifolia), California sycamore (Platanus racemosa), and deodar cedar (Cedrus deodara), suggests that mature urban trees are substantially more resistant to structure-fire radiation and convective heating than the buildings themselves.

This observation has direct methodological implications. In the wildland zone, the automated vegetation classifier performs reliably because the landscape is dominated by continuous canopy with minimal anthropogenic interference. In urban zones, however, the less than 5 m vegetation class is contaminated by a heterogeneous mix of fence lines, landscaping, retaining walls, partially consumed ornamental plantings, and pitched-roof edges misclassified as vegetation. The combustion characteristics of this mix cannot be assigned a single bulk-density or combustion-completeness factor without per-object semantic segmentation. Urban vegetation in the greater than 5 m class presents fewer classification ambiguities, as above-ground features exceeding 5 m in developed residential neighborhoods are overwhelmingly true canopy rather than anthropogenic objects, and can be reliably assigned the same oak-woodland conversion factors used in the wildland zone.

We therefore include urban vegetation above 5 m in the emissions calculation and exclude urban vegetation below 5 m. The excluded less than 5 m urban vegetation volume, 10.4 million m³ for the Palisades fire and 2.6 million m³ for the Eaton fire, would contribute approximately 9 kt C and 2 kt C respectively if entirely attributed to consumed vegetation, representing a conservative bias bounded at less than 4% of the total for either fire.

6.3. Implications for Urban WUI Fire Accounting

Wildfire emissions inventories used in regional and national models, including the Global Fire Emissions Database version 5 (GFED5) [51], CAL FIRE’s First Order Fire Effects Model, and the California Air Resources Board (CARB) Natural and Working Lands inventory, were developed for vegetation-dominated fires and rely on vegetation-specific emission factors [43,47]. Our findings, alongside Vannucci et al. [15], Li et al. [16], and Baliaka et al. [18], suggest that this underrepresentation is a leading-order effect for WUI events, not a minor correction.

Three observations bear directly on this question. First, the direct measurement of consumed building floor area at landscape scales is now feasible through multi-temporal LiDAR differencing, with uncertainty bounds (±29% on building carbon) compatible with regional emissions reporting, an advance over the structure-count-times-default-mass approach of legacy inventories. Second, the Hartwell et al. [39] mass density of 195 kg/m² is approximately five times the older CARB Emission Inventory Improvement Program (EIIP) default of ∼39 kg/m², which captured only movable contents. The adoption of the updated value would substantially increase estimated emissions from WUI events. Third, the Palisades–Eaton asymmetry (one fire vegetation-dominated, the other building-dominated) emerges from the spatially explicit LiDAR decomposition and supports the case for incorporating bottom-up LiDAR analysis alongside atmospheric and satellite approaches.

The same LiDAR-derived dataset supports downstream environmental flux estimates beyond carbon emissions, including ash mass, debris flow potential, and sediment loading to adjacent watersheds and the marine environment.

6.4. Future Work

Several extensions follow naturally from this work. Direct semantic segmentation using deep learning (e.g., RandLA-Net, PointNet++) would replace rule-based classification with a learned model trained on the manually refined point cloud, reducing misclassification uncertainty and enabling the attribution of currently unclassified urban objects. Multi-temporal extension to incorporate post-storm LiDAR acquisitions would isolate the storm-driven sediment budget from the fire signal and enable validation of debris flow predictions [23,24]. The integration and cleanup of the 2016 LARIAC baseline dataset would provide a third temporal epoch, enabling the characterization of pre-fire landscape change. The translation of the consumed-material inventory into downstream environmental fluxes (ash production, heavy metal mobilization, particulate carbon deposition) remains an open research direction, particularly given the Krichels et al. [52] finding that post-fire soil carbon losses persist for over four years.

The public release of all raw and derivative datasets (see the Data Availability Statement) enables independent analysis. As building-mass and vegetation-bulk-density conversion factors improve through forthcoming WUI debris characterization work, the underlying volumetric record will support reanalysis without re-acquisition.