the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 11 Oct 2026
TreeMap 2020, 2022, and 2023: Improved species range mapping in a tree-level forest dataset for the conterminous United States
Abstract. TreeMap is a tree-level model of the forests of the conterminous United States (CONUS) that provides a spatially continuous, 30 m resolution representation of forest structure and composition. It uses a random forests imputation approach to assign the most similar Forest Inventory and Analysis (FIA) plot to each forested pixel in gridded LANDFIRE data based on a suite of predictor variables that include vegetation, topographic, biophysical, and disturbance characteristics. By linking every mapped pixel to an FIA plot identifier for all forested areas in CONUS, TreeMap bridges localized inventory measurements with national-scale analyses and provides high-resolution estimates of forest structure and composition. In this update, TreeMap 2020, 2022, and 2023 incorporate climatic variables from Daymet in place of LANDFIRE biophysical variables. Further, we refined the independent validation by limiting the analysis to FIA plots where heterogeneity was administrative rather than ecological, excluding plots previously used to validate TreeMap versions that spanned heterogeneous forest conditions. TreeMap 2020, 2022, and 2023 also implement a new process constraining the pool of FIA plots available for imputation within each LANDFIRE zone to species documented within that zone and its adjacent zones, thereby reducing implausible species assignments across CONUS. TreeMap 2020, 2022, and 2023 are publicly available at https://doi.org/10.2737/RDS-2025-0031 (Zimmer et al., 2025) and https://doi.org/10.2737/RDS-2025-0032 (Houtman et al., 2025), and https://doi.org/10.2737/RDS-2026-0038 (Zimmer et al., 2026), respectively.
- Preprint
(1908 KB) - Metadata XML
-
Supplement
(522 KB) - BibTeX
- EndNote
Status: open (until 02 Dec 2026)
- RC1: 'Comment on essd-2026-633', Anonymous Referee #1, 12 Oct 2026 reply
Data sets
TreeMap 2020 CONUS: A tree-level model of the forests of the conterminous United States circa 2020 Scott N. Zimmer et al. https://doi.org/10.2737/RDS-2025-0031
TreeMap 2022 CONUS: A tree-level model of the forests of the conterminous United States circa 2022 Rachel M. Houtman et al. https://doi.org/10.2737/RDS-2025-0032
TreeMap 2023 CONUS: A tree-level model of the forests of the conterminous United States circa 2023 Scott N. Zimmer et al. https://doi.org/10.2737/RDS-2026-0038
Model code and software
gtac-treemap Lila S. T. Leatherman et al. https://github.com/firelab/gtac-treemap
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 345 | 95 | 28 | 468 | 47 | 34 | 26 |
- HTML: 345
- PDF: 95
- XML: 28
- Total: 468
- Supplement: 47
- BibTeX: 34
- EndNote: 26
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General
This manuscript documents three new vintages of TreeMap (2020, 2022, and 2023), the 30 m random forest imputation of FIA plot identifiers across the forested extent of the conterminous United States. TreeMap is one of the very few national products that carries a complete tree list to every pixel, and earlier versions are already embedded in fire risk, snag hazard, and carbon applications, so a citable description of the current vintages is useful to the ESSD community. The data and code are openly available, the workflow is clearly described, and the species ride-along fix addresses a real problem that users of earlier versions have encountered. The 20-seed stochasticity analysis is a welcome addition.
My principal concern is that the accuracy claims rest mostly on metrics that measure consistency with LANDFIRE rather than agreement with the ground. EVC, EVH, and EVT-GP enter the target data as predictors and are also the response variables (L199-201, Table 1), so out-of-bag accuracies of 99.5 to 100% and target layer accuracies of 96 to 99% (Table 2) largely confirm that the imputed plot carries the same LANDFIRE class as the pixel it was assigned to. However, the only ground-referenced evidence is Table 3, where exact EVC agreement is 44.5 to 49.9%, EVT-GP agreement is 47.0 to 61.0%, and EVC within 10% of the plot value is 48.6 to 52.7%, mostly below TreeMap 2016. The Conclusions (L584-586) nonetheless state that validation confirmed high accuracy across all four variables. I would ask the authors to present the OOB and target-layer results as internal consistency, lead with the independent results, and align the Conclusions with Table 3.
The independent validation also needs strengthening before it can support comparisons across versions. The 2016 values come from 2,937 plots, while the new vintages use 362 to 509 plots selected under a stricter rule, so the declines discussed at L330-354 cannot be attributed to LANDFIRE as the text suggests, since the validation set changed at the same time. At n = 362, the 95% interval on a proportion near 0.5 is roughly ±5 percentage points, and no intervals are reported. The at-least-one-matching-pixel metric also lacks a chance baseline. A 40.2 m radius encloses about 5,080 m², or five to six pixels, so the probability of at least one match under random assignment from the zone class frequencies is well above zero and should be reported beside the observed rate. The cleanest test of the ride-along fix is already within reach: the 2020 models, with and without the fix, evaluated against the same validation plots.
Most users do not work with EVC or EVT-GP directly. They use attributes derived from the tree lists, such as basal area, trees per hectare, aboveground biomass, and forest type, and the Discussion recommends use at project, landscape, or regional scales (L438-440). Neither is evaluated. I would suggest two additions. First, add plot-level bias and RMSE, or an observed-vs.-predicted equivalence test, for a few tree-list attributes at the validation plots. The second, and more important for the recommended use, is a comparison of TreeMap totals against FIA design-based estimates and their sampling errors by state, LANDFIRE zone, and forest-type group. Agreement at that scale would justify that guidance, and divergence would show users where caution is needed.
I also disagree that uncertainty cannot be quantified (L473-474, L556-558). The 20-seed runs already contain what is needed for a pixel-level reliability layer, such as the share of seeds assigning the same plot or the variance of derived basal area across seeds. Only 65.33% of pixels kept the same plot with and without the fix (L326-327), suggesting pixel-level assignments are far less stable than the CV of target-layer accuracy implies. Releasing one such layer per vintage would help users more than enumerating error sources.
A revision would therefore need, in order, ground-referenced accuracy for tree-list attributes and an aggregated comparison against FIA estimates; validation metrics with intervals and a chance baseline, on a validation set held fixed when versions are compared; a CONUS-wide measure of out-of-range species assignments before and after the fix, since the title rests on one national forest chosen for familiarity (L277-279); a per-pixel reliability layer; and clearer documentation of what actually changes among the three vintages.
Specific
L66: Four 1/24 acre subplots total 1/6 acre (about 675 m²), consistent with L155-156 and L503. The 2/3 acre (2,698 m²) figure appears to be multiplied by four twice.
L158-162: Saplings (2.5 to 12.6 cm, or 1.0 to 4.9 in) are tallied on microplots and live trees ≥ 12.7 cm on subplots. Stems below 2.5 cm are seedlings rather than saplings.
Table 1 and L164-175: Please give the projection and units for latitude and longitude, and correct the northness and eastness reference source, which cites LANDFIRE Slope where Aspect is presumably meant. Please also state which inventory years make up the 69,800 reference plots, the FIADB version or download date (CN lookups are only reproducible against a fixed snapshot), and the distribution of plot measurement years relative to each landscape year, since the tree lists are not projected forward.
L193-195: Replacing the LANDFIRE biophysical variables with Daymet is presented as an improvement without evidence. Please report variable importance and, ideally, the accuracy of a 2020 model built with the previous set.
L233-240: Please explain the 40.2 m radius, since the outer subplot edges sit about 43.9 m from the plot center. Please also confirm that no validation plot occupies a reference-plot location from an earlier cycle.
L268-279: A CONUS-wide check is possible with independent species range maps, reporting the forested area assigned to plots that contain out-of-range species, by zone, with and without the fix. Please also report how many plots each zone loses and whether the fix disproportionately removes plots with rare species. The blue oak example (L455-462) shows that zone adjacency is a coarse constraint.
L297-298 and L337-338: Disturbance presence is assigned to Tables S9-S11, which overlap the EVT-GP tables, and no 2023 tables are cited. L337-338 says EVT-GP agreement for 2020 is lower than 2016, but Table 3 gives 53.4% against 51.8%. Differences at L332 and L348-349 are percentage points rather than percent.
Table 3: The caption says agreement used the mean TreeMap value, whereas the header and L238-240 describe at least one matching pixel.
L383-390 and Figures 4-5: The distance analysis is informative. Because it is computed from true coordinates, please state whether a per-pixel distance layer is released and confirm FIA has reviewed it for disclosure risk.
L519-521: Please report the share of forested pixels whose plot assignment changed between 2020 and 2022 and between 2022 and 2023, and state in the Abstract and data documentation that the vintages share model files and are not independent estimates for change detection.
Editorial
Please use SI units throughout with imperial values in parentheses where FIA convention requires (acres, inches, feet, and tons at L533, where 2,000 short tons per acre is about 4,480 Mg ha⁻¹, and please state whether this is dry biomass). Legends in Figures 2 and 5c are unreadable at print size. L445 reads 2002 for 2022, L202 reads “was training,” and the agreement is a Memorandum of Agreement at L171 and of Cooperation at L623. Breiman (2001), Cutler et al. (2007), Lobell et al. (2014), Thornton et al. (2021), Wilson et al. (2012, 2018), and Zimmer et al. (2026) are cited but not listed. Riley et al. 2022a and 2022b appear swapped between the text (L62, L93, L382) and the reference list, and Hudak et al. (2026) carries the same DOI as Peeler et al. (2023). Please verify each. The GitHub repository should be archived as a tagged release with a persistent identifier.