Loading...
Loading...
Contributors:
Dataset URI:
10.5281/zenodo.22803968Analysis code, manuscript figures, 30 m map rasters, forest-frame rasters, findings memos, and derived aggregated estimates supporting a standalone manuscript comparing late-successional and old-growth (LSOG) forest classification across two inventory infrastructures sharing the Acadian forest type: Maine (USDA Forest Inventory and Analysis, true coordinates held server-side) and the Canadian province of New Brunswick (the public Continuous Land Inventory, a 2 km by 2 km systematic design-based grid).The deposit isolates how much cross-jurisdiction disagreement between LSOG maps is attributable to coordinate precision, to the structural instrument (scorecard) used, and to the forest frame, versus genuine differences in forest condition. It includes the New Brunswick design-based estimator comparison and small-cell suppression review, the instrument-sensitivity analysis, the small area estimation chain (CORE4 and the adopted CORE5 instrument, saeczi), the pooled cross-border shade-tolerance composition surface at 30 m in EPSG:3979 built on AlphaEarth embeddings with its bootstrap uncertainty raster, the two forest-frame rasters (M1 ESA WorldCover 2021 tree cover and M3 hybrid forest land use) with the frame-decision memo, all manuscript and appendix figures, the R and Python code, and the SLURM job files. The manuscript and Appendix S1 files are the September 16, 2026 working draft (edit pass 3).Version 1.5.0 follows version 1.3.0 directly and adds the 10 m CORE4 structural core build. A version 1.4.0 was staged on September 8, 2026 and never published; its corrections, described below, travel in this version. The ten 10 m rasters are deposited in the companion record (concept DOI 10.5281/zenodo.22410163, version 1.1.0). This record adds core4-10m-build_2026-09-16.zip (the border seam table, the build and calibration summaries, the pooled share intervals, both Figure 7 PDFs and their display summaries, SHA-256 sums for the rasters, the producing scripts, and a README) and the jurisdiction-calibrated Figure 7 PNG. The layers map a conservative CORE4 structural core: the probability that a 10 m forest cell meets the CORE4 rule at three of four structural axes. It is not a wall-to-wall old-growth map. The jurisdiction-calibrated layers are the primary product and the pooled layers are the uncalibrated sensitivity. The calibration is a logit intercept, a = -0.539 [-0.802, -0.340] for Maine with a New Brunswick offset b_NB = 0.646 [0.302, 1.024], giving calibrated forest shares of 4.63% [3.90, 5.65] for Maine and 5.53% [4.53, 7.50] for New Brunswick. It leaves a disclosed step at the border: within 500 m of the border the calibrated probability jumps by 0.0289 from Maine to New Brunswick, the jurisdiction step is 0.031 (0.008 to 0.061 across 20 bootstrap draws), and the part not explained by the intercept is -0.0019 against a tolerance of 0.019. Each pixel carries a probability (0 to 1, stored times 10000), a bootstrap SD and an extrapolation flag, in EPSG:3979 at 10 m. The layers must not be summed outside the forest frame or used to call single stands. The manuscript and Appendix S1 are replaced with the September 16, 2026 pass 3 draft, whose PDFs were rendered with LibreOffice in a cloud container. Two duplicate figure PNGs (the uncomposited Figure 3 panels, whose PDF remains, and the Figure S1 PNG, byte-identical to an image embedded in the main text docx) move into outputs_archive_2026-09-05.zip, holding the record at the 100-file per-deposition ceiling.Version 1.1.0 (September 6, 2026) adds the lsogscore R package. The source tarball lsogscore_0.1.0.tar.gz was built with R CMD build from commit d77cc52 of the source repository at https://github.com/holoros/lsog-ne, and its provenance and verification record lsogscore_PROVENANCE.md is deposited as a loose file so it can be read on the record page. The package is one tested implementation of every LSOG scoring instrument used in the analysis, seven in all, namely the Maine published six-axis card, the New Brunswick published five-axis card, the strict four-axis conjunctive rule, the CORE4 comparable instrument, both CORE5 forms, and the two small-area-estimation label cards. It replaces seven scattered R and Python implementations in two unit systems with a single threshold registry in which every threshold is stored in the units its source code used, carries the source file and line it came from, and is converted at exactly one boundary, so a jurisdiction difference stays visible rather than being silently harmonized. It is the first written definition of CORE4, whose axis columns had previously been carried precomputed in a table with no committed builder. The test suite holds 98 assertions, all passing under R 4.5, including oracle tests that reproduce the New Brunswick label card end to end from raw plot attributes on 10,681 of 10,681 plots on every axis and on the total, the Maine six-axis totals on 3,071 of 3,071 plots, and CORE4 on 13,708 of 13,708 rows. The package reads, writes, and requires no plot coordinate, and every file in it was swept for coordinate headers and restricted tokens before deposit. Adding the two files brings the record to exactly the 100-file per-deposition ceiling, so the package README travels inside the tarball rather than as a loose file. All 98 files of version 1.0.0 are carried forward unchanged apart from the README, CITATION.cff, manifest, and exclusion list, which were revised to describe this version.No FIA true plot coordinates, no per-plot records, and no per-parcel New Brunswick licensee (J.D. Irving, Acadian Timber) records are included; all tables are aggregated statistics that passed disclosure review, and the README documents the gates applied. The retired, single-jurisdiction Maine-only tol_score_ba surface (maine-tolscore-ba-surface_DATA_2026-09-03.tif) that is superseded by the pooled cross-border masked surface (Figure 5) was removed from this record on September 5, 2026 to fit the Zenodo per-record storage quota; the full-resolution raster remains available on OSC Cardinal and by request, and the exclusion is documented in the README and in reverse_gate_exclusions.txt.On September 5, 2026 this deposit was also split in two because the Zenodo storage constraint on this account was diagnosed as account-wide rather than per-record. The two large pooled cross-border rasters (surface and uncertainty, 668 MB and 720 MB, EPSG:3979, 30 m) are deposited separately as a companion Zenodo record, published September 5, 2026 as DOI 10.5281/zenodo.22410164, and remain available in full resolution on OSC Cardinal and by request in the interim. Removing the two giant rasters did not resolve the upload blocker for this record: a resume upload failed again at exactly 100 files attached (61.4 MB), and a controlled test on the new companion record (2 files, 1.39 GB, zero errors) showed the real constraint is a per-deposition 100-file ceiling, not an account-wide byte quota. This record's 208-file manifest could not fit under that ceiling as a flat list of individual files. On September 5, 2026 (third pass) the stale 100-file subset was deleted in full and the record was rebuilt: the 30 findings memos and 79 aggregated output files, which alone accounted for 109 of the 208 entries, were each consolidated into a single zip archive (memos_archive_2026-09-05.zip, outputs_archive_2026-09-05.zip), verified beforehand to reproduce every original file byte-for-byte and name-for-name. The PNG form of five manuscript figures (Figures 2 and 6, Appendix Figures S2, S3, and S5) was the one deliberate content exclusion made to clear the ceiling with margin; the PDF vector form of each of those figures is retained in this record and the PNGs remain available on OSC Cardinal and by request. The resulting 98-file manifest was uploaded and published as version 1.0.0 of this record. See the README's status note and memos/2026-09-05b_zenodo-split-publish_FINDINGS.md for the full file-count-ceiling diagnosis.Version 1.2.0 (September 7, 2026) corrects three items in the manuscript and Appendix S1 of version 1.1.0 (DOI 10.5281/zenodo.22559132) and adds one supporting-data archive. A lattice and plot-coverage check found that the per-state LCMS tiles used throughout the prior strict four-axis correction covered only 78 to 97 percent of each state's land area, because a gdalwarp destination was sized to the tile's corner rather than its full extent; rebuilding the tiles at full extent moves Maine's strict four-axis rate from 2.5% [1.9, 3.1] to 2.7% [2.1, 3.3] (469,043 of 17,418,749 ac, 100.0% tile coverage), with New Hampshire, Vermont, and New York moving inside their prior intervals. A balsam fir (Abies balsamea (L.) Mill., SPCD 12) composition-axis sensitivity was recomputed on the full-extent tiles: Maine moves from the adopted 2.7% [2.1, 3.3] to 3.0% [2.4, 3.6] with balsam fir added to the shade-tolerant species list (10 of 87 passing conditions flip); the production 10-code species list without balsam fir remains the adopted definition, and this sensitivity is flagged, not adopted. A CORE4 (four-axis, no age/diameter species variant) rate is entered for New Brunswick's public CLI release for the first time, cut>=3 4.61% [3.86, 5.37] (n=10,681, 38 holders, holder-cluster bootstrap), against a licensed (non-public) New Brunswick delivery rate of 9.73%, consistent with the public release's exclusion of the two licensed freeholds; the Maine side of the same CORE4 sampling table gives 4.44% [3.76, 5.23]. All three corrections were re-derived directly from their source CSVs and propagated into the manuscript, Appendix S1 (Table S4), and this metadata; the manuscript and Appendix S1 docx/pdf pairs are replaced with the corrected September 7, 2026 bytes (same filenames, same August 27, 2026 assembly-date convention; still labelled a working draft). The five source CSVs and their two producing R scripts (the balsam fir probe and the CORE4 public-CLI estimator; the tile-coverage funnel CSVs are shipped as compiled reference tables citing their Cardinal job provenance, since their producing script is resident on Cardinal and not yet committed to this deposit's script set) are added as v120-corrections-data_DATA_2026-09-07.zip. Adding this archive holds the record at exactly the 100-file per-deposition ceiling: the September 2, 2026 predecessor PNG of the instrument-sensitivity figure (fig_instrument_sensitivity_2026-09-02.png), an already-superseded raster duplicate of a PDF vector form retained in the record, was dropped to make room. No other file was removed and no science content beyond that one duplicate PNG was dropped. No FIA true plot coordinate, no per-plot record, and no per-parcel New Brunswick licensee record is present in the new archive or the corrected manuscript files; both were swept before deposit, see the README.Version 1.3.0 (September 8, 2026) adds the lead graphical abstract that closes Gate 1 (open since v1.0.0), and corrects Appendix Table S7 and its companion Section 4.5 paragraph of the version 1.2.0 manuscript (DOI 10.5281/zenodo.22646840) by adding 95 percent bootstrap intervals to four surface-validation quantities that had carried point estimates only. The four intervals, 2,000 plot-level resamples, seed 20260908, computed on the production (no-jurisdiction-covariate) out-of-fold predictions: within-jurisdiction spatially blocked cross-validation R-squared, Maine 0.2904 [0.2600, 0.3202] and New Brunswick 0.3369 [0.3221, 0.3505]; attenuation slope (observed regressed on predicted, all plots) 1.117 [1.088, 1.145]; attenuation slope, top decile of observed values (tol_score_ba >= 4.541, n = 1,347), 0.029 [0.007, 0.050]. The two R-squared values were independently reproduced from the raw out-of-fold prediction files before the interval was accepted. The two attenuation slopes were not, and version 1.4.0 supersedes them with 1.116 [1.089, 1.143] and 0.028 [0.005, 0.051], the values the archived pipeline actually returns. The manuscript and Appendix S1 docx/pdf pairs are replaced with the corrected September 8, 2026 bytes (same filenames, same August 27, 2026 assembly-date convention; still labelled a working draft); the main text page count held at 34. The new lead file, 00_menb-lsog_graphical-abstract_2026-09-08.png (four panels: jurisdiction and instrument overview, the Maine strict four-axis v1.2.0 correction, the CORE4 public-versus-licensed provenance finding, and the lsogscore test-suite summary; 2550 by 2160 px, 300 dpi), is named in default_preview and sorts first in the manifest by its 00_ prefix. Adding it holds the record at exactly the 100-file per-deposition ceiling: fig11_pooled_tolscore_surface_2026-09-04.png, the already-superseded unmasked predecessor of Figure 5, was dropped; its PDF vector form is retained unchanged, the same convention used for the September 2026 figure drops in prior versions. The graphical abstract's builder script is kept in the project working tree, not deposited, since depositing it too would require a second file drop to hold the ceiling; this is stated as a gap rather than silently omitted. No FIA true plot coordinate, no per-plot record, and no per-parcel New Brunswick licensee record is present in the corrected manuscript files or the graphical abstract; both were swept before deposit, see the README.Version 1.4.0 (staged September 8, 2026, never published; carried in version 1.5.0) corrects the two version 1.3.0 attenuation slopes above, which did not reproduce, adds four Appendix Table S7 rows and a Section 4.6 paragraph reporting the shade-tolerance surface's calibration in the region the CORE5 axis actually operates in, corrects a wrong statement of regression direction in the Methods of the version 1.3.0 manuscript (DOI 10.5281/zenodo.22650397), and adds the nine-row learner and predictor-set comparison behind the Section 4.5 slope range as a new Appendix S1 Section 7 and Table S9, with both attenuation directions and bootstrap intervals. Methods 3.10 of version 1.3.0 stated that the attenuation diagnostic was checked by regressing observed rate on predicted rate. That contradicts both the analysis code, which fits lm(p ~ o), and Section 4.5, which reports the range as predicted-on-observed. The clause is corrected and a sentence is added naming the direction the adoption rule uses, since the workspace contains scripts using each convention and the direction is therefore not inferable. Table S9 reports, for all nine combinations on the same 1,847 New Brunswick held-out plots, total and systematic error, both regression slopes with 95 percent nonparametric case-resampling bootstrap intervals (B = 40,000, seed 20260908), the squared Pearson correlation, and Nash-Sutcliffe efficiency. Reporting both directions matters because they order the learners differently: every predicted-on-observed interval excludes one with the highest upper bound at 0.533, while all three single-objective observed-on-predicted intervals straddle one, at 0.929 to 1.032 baseline, 0.967 to 1.064 alongside and 0.914 to 1.009 replace. Appendix Section 7 answers why the adopted learner is nonetheless the multi-objective fit, and a paragraph added to Section 6 reconciles the opposite direction used for the Table S7 attenuation slopes so the two tables are not read as contradicting each other. This closes one of the nine point-only estimates inventoried as open in version 1.3.0. Two cautions are recorded in the appendix text. The R2 column of the deposited step 5c comparison table is Nash-Sutcliffe efficiency and not a squared Pearson correlation, so the identity converting one slope direction into the other cannot be applied through it, and doing so returns 0.611 for the baseline multi-objective fit against its measured 0.720. Interval endpoints are Monte Carlo estimates reproducing to about 0.001 across bootstrap seeds even at B = 40,000, so their third decimal carries resolution rather than exactness. Every value in Table S9 is rounded once from full precision after an independent re-derivation on OSC Cardinal reproduced all nine point estimates in both directions and closed the slope-product identity to 4.4e-16; that re-derivation also corrected three cells of an earlier build that had been rounded twice. Two Appendix Table S7 values are also corrected and the low end of the tolerance surface is validated for the first time. The production column of Table S7 printed an overall observed-on-predicted attenuation slope of 1.117 with intercept -0.426 and a top-decile slope of 0.029. Re-derivation from the same out-of-fold prediction file returns 1.116 with intercept -0.423 and 0.028, and the jurisdiction-aware column reproduces exactly, so the failure was specific to one column and main text Section 4.6, which quoted both figures, is corrected with it. Separately, the only diagnostic previously reported for that surface conditions on the top decile of observed tolerance while the CORE5 penalty fires at or below 2.75, the opposite end of the same scale. A mirror diagnostic on the 1,597 plots below that cutoff finds predicted tolerance running 1.1116 units high and recovers 13.53% of them at a specificity of 0.9912, a Cohen's kappa of 0.1934 [0.1697, 0.2164] and a predicted low-tolerance share of 2.3756% against an observed 11.8560%. The appendix claim that the surface supports classification at the 2.75 threshold is therefore withdrawn, the Table S7 low-tolerance shares are relabelled as lower bounds, and the Maine against New Brunswick tree-cover-frame comparison is retracted because detection differs between the jurisdictions. The CORE5 headline rates are unaffected, since that penalty is applied to each plot's measured tol_score_ba and not to the surface. Six further corrections of record travel with this version, namely a CORE4 sweep variant quoted from the wrong Table S6 row (33.92% for row X8_c2 where the CORE4 row X9_c2 reads 33.70%), a threshold-convergence tolerance that holds for two of its three comparisons and not the third (0.013 becomes 0.017 against the pooled cutoff), three broken internal cross-references in Appendix S1, and the removal of a Legaard, Simons-Legaard and Weiskittel (2020) citation from an observed-on-predicted regression, since that paper defines its slope as predicted on observed. The staged version replaced the manuscript and Appendix S1 docx/pdf pairs with the corrected September 8, 2026 bytes (same filenames, same August 27, 2026 assembly-date convention; still labelled a working draft), which version 1.5.0 in turn replaces with the pass 3 draft, and old Appendix Sections 7 and 8 are renumbered to 8 and 9 to make room, which is safe because the only cross-reference to an appendix section number in either document is to Section 3. In that staged render the main text ran to 35 pages, one more than version 1.3.0, the added page being the Section 4.6 calibration paragraph, measured on the Cardinal apptainer LibreOffice build. Appendix page count is renderer-dependent and is not quoted here, since three LibreOffice builds gave three different counts for byte-identical input. That staged version added and removed no file. The per-plot prediction files behind Table S9 remain on OSC Cardinal and are not deposited, in the same treatment given the version 1.3.0 Table S7 bootstrap inputs; the table itself is an aggregated nine-row summary. No FIA true plot coordinate, no per-plot record, and no per-parcel New Brunswick licensee record is present in the corrected manuscript files.
Keyword(s):
Acadian forest; AlphaEarth; coordinate precision; design-based estimation; forest frame.
MeSH Term(s):
Grants and Funding:
No Grants or Funding Data Available
Associated Publication(s):
No Associated Publication Data Available
Supplementary Data:
No Supplementary Data Available
Related Dataset(s):
The National Library of Medicine (NLM) provides information about biomedical datasets. Inclusion on an NLM Web page does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: NLM Web Policies (nih.gov)