Reproducible LSOG mapping for Maine: design-based estimation, an FIA-anchored 10 m product, and an apples-to-apples comparison with the Hagan et al. (2026) LiDAR map

This version extends the prior deposit with a reproducible, FIA-anchored late-successional and old-growth (LSOG) mapping workflow for Maine and a fully like-for-like comparison with the published LiDAR classification of Hagan et al. (2026, Ecosphere). Contents. (1) Design-based FIA estimates of older-forest area with confidence intervals and the multi-axis structural classification. (2) An open, FIA-anchored 10 m LSOG probability product over the study area built from learned satellite embeddings (TESSERA), with a full-AOI mosaic, a Big Reed/Baxter showcase, and a resolution sweep (10-200 m) documenting the dependence of the estimate on map grain. (3) A faithful reproduction of the published LiDAR map from its own public archived inputs (random forest on the 463 training hectares applied to the public hectare-level LiDAR canopy statistics), reproducing the reported any-LSOG extent within ~2 percentage points (22.0% vs 19.7%). (4) An apples-to-apples comparison: at matched landscape prevalence the reproduced LiDAR map and the FIA-anchored embedding map agree on location only weakly (Cohen's kappa ~0.19), with a Seven Islands/Pingree comparison against the authors' own map at kappa 0.24. (5) A remote-sensing feature-fusion benchmark (TESSERA, AlphaEarth, Landsat/Potapov canopy height, Meta canopy height) and the documented-old-growth validation layer compiled from the 1983 and 1986 Maine Critical Areas Program inventories. Note. FIA plot coordinates are restricted and are not included; only derived rasters, aggregate tables, and code are released. v3.0.0 adds the completed full-study-area 10 m LSOG probability map (LSOG_probability_10m_fullstudyarea.tif; 29,083 x 29,162 px, ~848 million 10 m cells, EPSG:32619), the wall-to-wall fused TESSERA product over the entire Hagan study footprint, plus the refreshed technical report and briefing deck. v3.0.1 applies minor corrections from a final stress test of the technical report (a stale AUC value reconciled to the fusion benchmark, an any-LSOG class label, a citation-year consistency fix, and an AUC-range wording fix). Data products are unchanged. v3.0.1 applies minor corrections from a final stress test of the technical report (a stale AUC value reconciled to the fusion benchmark, an any-LSOG class label, a citation-year consistency fix, and an AUC-range wording fix). Data products are unchanged. v3.0.1 applies minor corrections from a final stress test of the technical report (a stale AUC value reconciled to the fusion benchmark, an any-LSOG class label, a citation-year consistency fix, and an AUC-range wording fix). Data products are unchanged. v3.0.1 applies minor corrections from a final stress test of the technical report (a stale AUC value reconciled to the fusion benchmark, an any-LSOG class label, a citation-year consistency fix, and an AUC-range wording fix). Data products are unchanged. v3.0.5 applies two source-fidelity refinements after a line-by-line re-review against Hagan et al. (2026). The reproduced classifier's any-LSOG figure (21.9%) is labelled as a fraction of forested hectares on the common comparison grid, alongside the published map's own any-LSOG of 19.7% of the full study area; and the manuscript describes mapped LSOG as 'lost' rather than 'harvested' to match the Global Forest Watch loss terminology. Data products are unchanged. v3.0.6 finalizes the technical report: confidence reported as High/Medium/Low with a defining legend, softened absolute language, an independent general-assessment framing (not prepared for any single organization) with a date, professional layout (centered table cells, justified body text, tables and captions kept together), and revised figures (Fig 5 with block-bootstrap 95% CIs; Fig 6 a two-panel comparison of our estimate with the reproduced Hagan model; Fig 7 expanded with a remote-sensing attribute-contribution panel; Fig 8 a forest-masked two-panel showing the reproduced Hagan map beside our 10 m map with county, Baxter, and Big Reed reference). Data products are unchanged. v3.0.7 adds final report refinements: green/amber/red shading of the High/Medium/Low confidence cells; a definition of the v5.1 structural classifier; a new Conclusions section with specific research, policy, and conservation suggestions; Figure 1 clipped to the Maine land area so the GEDI panel no longer extends beyond the state; and Figure 8 forest-masked with the FIA nearest-neighbor forest/non-forest mask (Wilson et al. 2012). Data products are unchanged. v3.0.8 adds the forest-masked wall-to-wall 10 m LSOG probability raster (LSOG_probability_10m_forestmasked.tif): the full 10 m surface with non-forest and open water set to no-data using the FIA nearest-neighbor forest/non-forest mask (Wilson et al. 2012, 2009 vintage; 473.4 of 848.1 million cells retained as forest). The full unmasked surface (LSOG_probability_10m_fullstudyarea.tif) is retained. Report Figure 8 caption updated accordingly.