Late-successional and old-growth forest across the Northeast: definition, mapping, temporal (in)stability, and the consolidated regional product archive

Erratum, August 2026: the New Brunswick raster products in earlier versions of this record contain Maine, not New Brunswick. All New Brunswick raster layers published in earlier versions of this record are withdrawn and replaced. Those layers were masked with the US National Land Cover Database, a product of the conterminous United States that carries no data in Canada. Because the reprojection destination was initialised to zero rather than to nodata, every Canadian cell failed the forest test silently instead of raising an error, and the only cells that survived were the eastern Maine portion of the New Brunswick tile window. The affected files reported 4,948,535 ha of New Brunswick forest spanning longitude 70.07 W to 66.96 W and never reached eastern New Brunswick. The corrected layers take their forest mask from the 2020 Land Cover of Canada (Natural Resources Canada, NALCMS classes 1, 2, 5 and 6) within the Statistics Canada 2021 provincial boundary. They report 6,138,879 ha of New Brunswick forest, 86.2 percent of the provincial land area, spanning 69.05 W to 63.77 W, and each was verified to fall inside the province before release. Two figures change materially for anyone who used the superseded prioritization layer. The prioritized area falls from 21,028 ha to 2,124 ha. The high late-successional and old-growth share becomes an estimate rather than a top tercile, which was a definitional third carrying no information about prevalence. Report the share only with its denominator attached. The corrected forest denominator is 6,138,879 ha at 30 m against the superseded 4,948,535 ha, about a quarter larger, so the same threshold gives a different percentage on it. The design-based estimates table deposited in earlier versions is also withdrawn. It carried four-state values retired on 29 July 2026, and it is no longer shipped in this record; it remains downloadable at the version DOIs of 1.x and 2.0.0 as the audit trail. Fifteen files leave the record in this version, and eight of those are not part of the correction. Zenodo caps a record at 100 files. Version 4.3.0 held exactly 100 and 5.0.0 adds 15 new basenames, so 15 files had to leave for the corrected material to fit. Seven are the withdrawals the correction required: the four 10 m demonstration tiles, the Maine 10 m file under its Maine name, the 30 m New Brunswick change layer, and the retired design-based estimates table. The other eight were chosen because each is superseded rather than because it was convenient. LSOG_probability_10m_forestmasked.tif, v6_aoi_prob_mosaic.tif, v6_prob_mosaic.tif and v6_aoi_prob_100m.tif are v6-era development mosaics on a northern Maine window in EPSG:32619, a different grid and masking from every delivered product, and were never part of the published product set. The four CRSF technical report files are older builds of two documents whose current builds are deposited in the dedicated technical reports record, concept DOI 10.5281/zenodo.21684054, published 29 July 2026; their sizes agree within one percent and every checksum differs, which is what identifies them as earlier builds rather than copies. Every withdrawn file remains permanently downloadable at the version DOIs of the releases that carried it, so this changes what the current version ships and deletes nothing from the record's history. Two New Brunswick layers are withdrawn with no replacement in kind. LSOG_probability_10m_NB.tif was a 0.30 by 0.25 degree test tile over the upper Saint John valley, about 22.7 by 27.6 km, published under a provincial name; there is no provincial 10 m New Brunswick surface and the corrected 30 m product is the finest resolution that genuinely covers the province. LSOG_change_2018_2024_NB.tif at 30 m is replaced by a 150 m version, because the decline axis is a five-cell smooth on a 30 m grid, that is a 150 m window, so the 30 m encoding was always spurious precision. The United States layers in this record are unaffected. They were independently verified to fall within their intended states, because the land cover mask covers them correctly. Anyone holding a New Brunswick layer from an earlier version should discard it and any figure derived from it. About this record. This record is the current version of the consolidated Northeast late-successional and old-growth archive, published under concept DOI 10.5281/zenodo.20614496. It holds the data, analysis code, figures and derived raster products for Maine, New Hampshire, Vermont, New York, Massachusetts and New Brunswick. Earlier separate records, including 10.5281/zenodo.20947390, 10.5281/zenodo.21070969 and 10.5281/zenodo.21070977, were consolidated into this lineage and are retained unchanged for provenance. Cite the concept DOI 10.5281/zenodo.20614496, which always resolves to the current version. The study defines LSOG through a multi-axis structural label built on U.S. Forest Inventory and Analysis (FIA) data, produces design-based four-state estimates (Maine, New Hampshire, Vermont, New York) with a New Brunswick provincial-inventory extension, maps the probability of LSOG from AlphaEarth satellite embeddings, and assesses the temporal stability of those embedding-based predictions. Headline findings. (1) Design-based FIA estimates place any-LSOG at 14.66% of forestland in Maine (95% CI 13.4 to 15.91), 27.80% in New Hampshire (25.01 to 30.60), 28.96% in Vermont (25.77 to 32.14), and 30.50% in New York (28.83 to 32.18) on FIA evaluation x2301, which spans 2019-2023 in Maine and 2017-2023 in the other three states; the LS+OG structural subset is 2.07% in Maine (1.57 to 2.58). Maine is clearly the lowest of the four: its point estimate is 13.1 percentage points below New Hampshire's and its interval clears theirs by 9.1 points. New Hampshire, Vermont and New York have overlapping intervals and are not separable from one another. These come from the adopted table headline_resolution_20260729_designbased.csv, whose producer jobA_design_based_rfia.R is deposited from version 5.0.0 onward. (2) A single-time LSOG classifier trained on the multi-axis structural label discriminates at AUC 0.83 to 0.84 under spatial cross-validation (temporal_stability_by_year.csv). (3) Canopy-height resolution and AlphaEarth-plus-canopy-height fusion tests show only marginal gains over AlphaEarth alone: on the Maine true-coordinate set the fusion gain is -0.0016 in AUC with a 95% H3-block bootstrap interval of -0.004 to 0.0004 and 5.9% of resamples positive (fusion_delta_ci.csv). (4) AlphaEarth embedding vectors are highly stable year to year (median cosine similarity 0.9789 to 0.9845 by region, 0.9817 pooled), and a classifier refit independently each year shows no temporal decay (pooled per-year AUC 0.6409 to 0.6513 on the region-scale table region_temporal_by_region.csv). Year-to-year embedding change does not usefully track ground-recorded disturbance at the plot scale: the change-detection AUC runs 0.3636 to 0.5119 across regions and is below 0.5 in every region of region_embedding_stability.csv, meaning disturbed plots showed slightly LESS embedding change than undisturbed ones rather than more. Benchmarked instead against LCMS-mapped heavy disturbance in Maine, the same embedding change discriminates it at AUC 0.85 on 8,502 plots, so the embedding does register stand-replacing disturbance; what it does not recover is the magnitude or sign of net LSOG class change. LSOG trends should therefore rest on design-based FIA rather than on differencing satellite embeddings between years. New Brunswick, corrected products. Four 30 m products (probability with its across-tree uncertainty band, grade index, calibrated grade, per-class probability) on a forest mask of 6,138,879 ha, mean P(any-LSOG) 0.1685, with 15.02% of forest at or above the delivered threshold P >= 0.33, which is 922,175 ha. A 150 m prioritization set (prioritization, 2018 to 2024 change, probability) on the SAME 30 m forest mask block reduced 5 by 5, with a 150 m cell counted as forest when at least half of its 25 underlying 30 m cells are, giving 2,808,348 cells and 6,318,783 ha, and 13.0% of forest at or above 0.33 which is 820,638 ha, 2,124 ha prioritized, and decline over 3.4% of forest. The two forest area figures reconcile: 6,138,879 ha is the true forest area, the sum of the per-cell forest fraction, and 6,318,783 ha counts every majority-forest 150 m cell as wholly forest, so the 179,904 ha gap is majority-rule quantization at mask edges rather than a second mask. Mean probability is 0.1685 at both resolutions, which is the forest-area-weighted mean; the 0.167 quoted for the 150 m layer in earlier versions of this record is the unweighted mean over majority-forest cells and is a different estimator, not a disagreement between the products. Provincial inventory context from the New Brunswick Continuous Land Inventory, 9,042 plots across 38 holders, post-stratified by Commission for Environmental Cooperation North America Level III ecoregion with a holder-cluster bootstrap: any-LSOG 31.47% of forest area (95% CI 29.17 to 33.31), robust to the weighting (five configurations spanning the published, corrected gap-filled and corrected strict weights, and both treatments of the ecoregion coverage gap, span only 31.47 to 31.65). That point estimate moved twice on 6 August 2026: a published 31.49% became 31.65% when the provincial ecoregion area weights were recomputed on the corrected forest mask, and then 31.47% when the plot set was made consistent with those weights. The corrected weights cover 100% of the 6,138,879 ha provincial forest mask, so the plot set must cover the same 100%; the 303 plots that fall inside the province and the forest mask but on no Commission for Environmental Cooperation Level III polygon are assigned to their nearest ecoregion by exact Euclidean distance (184 to Acadian Plains and Hills, 119 to Maritime Lowlands, none to the Highlands), the same rule used to gap-fill the 274,092 orphan hectares on the area side. That fill is a nearest-neighbour imputation and is stated as an assumption rather than as an observation, because the imputed ground differs measurably from its host: the 119 plots folded into Maritime Lowlands run 18.49% any-LSOG against a host stratum of 30.40%, while the 184 folded into Acadian Plains and Hills run 34.24% against a host of 32.21% and are indistinguishable from it. Dropping those plots and using the strict weights instead gives 31.61% and old growth 0.97% on 8,739 plots over a 5,864,787 ha basis. The corrected weight table nb_ecoregion_weights_corrected_20260806.csv and the weights recomputation record ecoregion_weights_report_20260806.md are deposited from version 5.0.0 onward. Area-weighted old growth is 0.95% and NO interval is quoted for it anywhere in this record, because the sensitivity of that figure to the choice of post-stratification weights exceeds any interval that could be put on it: the envelope on the corrected weights runs 0.95 to 1.41%, a spread of 0.47 percentage points against a bootstrap CI half-width of 0.39. An interval attached to an old-growth percentage in an older table or figure should be removed rather than revised. The CLI is not a probability sample and carries no inclusion probabilities. Consolidated regional products. This lineage carries the full multi-state raster product set for the Acadian and northern New England region, so that all openly available LSOG data, code and products live under one concept DOI. Products span Maine, New Hampshire, Vermont, New York, Massachusetts and New Brunswick: per-state LSOG probability at 10 m and 30 m, calibrated LSOG grade and grade index at 30 m, per-class probability at 30 m, 2018 to 2024 change layers, and 150 m conservation-prioritization surfaces, together with the regional probability mosaic. Associated outputs. These results and products underpin two CRSF technical reports, a four-state regional report and a Maine-focused report (Weiskittel, Woodall and D'Amato 2026, archived at 10.5281/zenodo.21684054), and two peer-reviewed manuscripts in review: a methods paper (Ecological Applications) and a Comment on Hagan et al. 2026 (Ecosphere). Attribution. Contains information licensed under the Open Government Licence, Canada. 2020 Land Cover of Canada, Natural Resources Canada, Canada Centre for Remote Sensing. Provincial boundary from Statistics Canada, 2021 Census cartographic boundary files. Canada's National Forest Inventory is acknowledged as a source, per the NFI Data Use Agreement. Restricted data. FIA true plot coordinates, New Brunswick MagPlot per-plot records, and J.D. Irving ownership layers and per-parcel outputs are NOT included and remain restricted; only aggregated summaries (H3 hexagon and state-level tables), analysis code, figures, and the derived raster products are provided. The lead file 00_LSOG_NE_graphical_abstract.png is a graphical abstract summarising the study and the correction. Known limitations, including the deposited tables that have no deposited producer, are listed in README.md.