Multi-method late-successional and old-growth (LSOG) forest mapping uncertainty for Maine's unorganized townships

Version 1.10.0 (this version): redoes the drivers analysis with the completed CONUS harvest-probability model (HCS v5, 240 m), replacing the earlier slope accessibility proxy. The result reverses the proxy-based inference: mapped LSOG falls on higher, not lower, harvest-probability ground. The mapped-LSOG share rises from 13 percent in the lowest harvest-probability tercile to 34 percent in the highest, and mapped-LSOG pixels carry higher mean annual harvest probability (0.024 vs 0.021) and disturbance probability (0.128 vs 0.111) than non-LSOG. The likely mechanism is that the structure that reads as LSOG (large trees, high stocking) is also what makes a stand merchantable, so the mapped resource is not shielded by inaccessibility and if anything sits where modeled harvest pressure is greatest. Section 9 and Supporting Information Fig. S5 are updated accordingly; this supersedes the slope-proxy statements in earlier versions. Version 1.9.0: answers a reviewer question on the footprint of the trend analysis. The design-based older-forest trend is statewide; clipping the FIA estimate to the published map exact area-of-interest polygon (the unorganized townships, about 4.28 million ha; rFIA spatial estimation) shows older forest is increasing within that footprint too, not declining: stand age >= 120 yr rises from 3.1 to 4.0 percent and large-tree basal area from 9.1 to 9.7 percent over 2003 to 2024, all weighted-regression slopes positive with 95 percent intervals excluding zero (Supporting Information Table S10). The statewide increase is therefore not masking a within-UT decline. Also adds a no-regrets prioritization citation (Schloss et al. 2024) supporting the ensemble-agreement recommendation. Comment remains 16 pages. Version 1.8.0: adds an independent third-party cross-map check using the U.S. Forest Service national old-growth inventory (Pelz et al. 2023). On the FIA plots where both apply (National Forest System land across Maine, New Hampshire, New York, and Vermont, n about 925), the field-structure classification and the Pelz old-growth classification agree only weakly (Cohen kappa 0.25 for late-successional-plus-old-growth, 0.12 for any-LSOG; Supporting Information Table S9). Because Pelz et al. is authored by others, this addresses the objection that maps disagree only against their own variants and corroborates the cross-map disagreement. Also adds a carbon-stakes line: in Maine, mean live-tree carbon rises from about 39 Mg per hectare in non-LSOG stands to 72, 91, and 139 across the transitioning, late-successional, and old-growth classes, so the definitional choice moves attributed carbon several fold. Layout refit to 16 pages. Version 1.7.0: adds Table 8, a synthesis that places every LSOG estimate for the same Maine forest side by side (Hagan broad airborne-LiDAR canopy class about 20 percent, FIA large-tree basal area 12.5 percent, four-axis true LSOG 6.5 percent, strict age-based old growth 3.9 percent), making the roughly fivefold definitional spread explicit and reinforcing that the honest reporting unit is a range with definitions attached rather than a single headline number. Minor layout pass; the Comment remains 16 pages. This dataset quantifies the uncertainty in mapping late-successional and old-growth (LSOG) forest across the approximately 4.2 million hectares of Maine's unorganized townships, and tests whether LSOG is rapidly disappearing. Three to four independent, credible mapping methods are compared on a common 100 m grid: (M1) a reproduction of the Hagan et al. (2026) airborne-LiDAR canopy random forest, rebuilt from their public Zenodo deposit; (M2) a logistic model of the FIA field-structure LSOG class on Potapov (GEDI-calibrated) canopy height; (M3) a direct canopy-height threshold; and (M4) the FIA structural class imputed to every pixel via USFS TreeMap (2016, 2020, 2022). Materials behind the formal Ecosphere Comment on Hagan et al. (2026): a cross-validated accuracy assessment (AUC) of each mapping approach on the original authors' own training plots; an FIA design-based estimate of older forest with sampling-error confidence intervals (older forest about 3.9 percent [3.3 to 4.6] and rising, including on private commercial timberland); a threshold-sensitivity sweep and a 20-seed reproduction ensemble; an ownership-resolved breakdown; a hex-scale (8 km) summary of cross-method disagreement; and the Comment manuscript and Supporting Information. Version 1.6.0: adds a first Landsat continuity layer (Hansen Global Forest Change, 2001-2023): 24 percent of canopy-mapped LSOG shows stand-replacing disturbance since 2001, failing strict continuity even over this short window, demonstrating that the continuity axis is mappable. Tempers the FIA framing per state-biometrician feedback: FIA is unbiased for area but imprecise at its plot density, so it is the baseline for how much exists while a wall-to-wall map is complementary for locating patches the sparse plots may miss. Adds the ROC AUC to the threshold figure and a copy/layout pass. Earlier: revised after a three-perspective simulated peer review (remote sensing, forest biometrics, and an ecologist in the original author's voice; included). Refinements: the terrain/disturbance overlay now states it uses the airborne Hagan map (independent of the TreeMap-derived disturbance product) with slope as a first-order accessibility proxy; the four-axis classification thresholds are flagged as tunable with the dead-wood (snags-only, no coarse woody debris) and continuity (treatment-code, permissive) caveats, so 6.5 percent is an upper bound; the multidimensionality result is noted as conditional on a purposive training sample; and the broad-class breadth is reframed as an objective-dependent choice with continuity treated as a weighted axis rather than a disqualifier. Earlier: adds an overlay of the LSOG map with the satellite disturbance record and terrain over the Maine AOI (n = 3.8 million pixels). LSOG concentrates on steeper, less accessible ground (median slope 5.3 vs 3.0 degrees) and more disturbance-prone settings; the mapped-LSOG share rises from 12 percent on the gentlest terrain to 37 percent on the steepest and from 17 to 35 percent across the disturbance gradient. This reframes risk: much mapped LSOG persists where harvest is costly or impractical and where natural disturbance dominates, so a single flat loss rate overstates near-term harvest risk to the inaccessible majority. The CONUS harvest-probability model does not cover Maine's unorganized townships, so slope serves as the accessibility proxy. Earlier: adds the constructive endpoint, a repeatable multi-axis classification of true LSOG anchored in FIA. True LSOG must satisfy four axes (live structure, dead wood, compositional maturity, and temporal continuity), each a deterministic function of FIA records; the design-based area meeting all four is 6.5 percent [5.6, 7.4] of Maine forestland (between strict age-based old growth at 3.9 percent and broad canopy LSOG near 20 percent). The Comment shows which axis is limiting, an objective-to-threshold table, and the path from definition to map (LiDAR for live structure, the satellite disturbance record for continuity, plots or imputation for dead wood and composition), calibrated to FIA. Earlier: deepened the conceptual core around the question 'what is true LSOG?'. LSOG is framed as a multidimensional condition (live structure, dead wood, compositional maturity, temporal continuity); on the training plots a dominant stand-development gradient explains 59 percent of variance but dead wood forms a partially independent axis (r about 0.5), and canopy sensors are nearly blind to dead wood and to disturbance history, the last compromised in Maine by pervasive legacy skid trails even in canopy-classified LSOG. Adds a correlation/dimensionality figure, references (Franklin and Van Pelt 2004; Wirth et al. 2009), and moves the Maine-scale ensemble to the Supporting Information. Earlier: emphasized that a single-date classification is a snapshot that cannot credit future LSOG from ongoing conservation and natural ingrowth (the transitioning class, 15.8 percent of the area, becomes late-successional in 25 to 50 years per the original authors; reserved lands mature into LSOG), and added a brief synthesis of the complementary toolset (airborne LiDAR, spaceborne GEDI and canopy-height products, FIA design-based inventory, TreeMap imputation, and growth-and-yield models) that together can represent the forward trajectory a snapshot cannot. Earlier: added the threshold-independent ROC AUC (0.979) to the cutoff-sensitivity figure and variation metrics throughout; a figure of older forest over time under four definitions and thresholds; and a refined, robust New England multi-definition ensemble LSOG map summarized at the hexagon scale (built from the FIA condition imputed to each TreeMap pixel; structural class, stand age >= 120 yr, and large-tree basal area), with an uncertainty layer. The New England ensemble is presented hexagonally because TreeMap is an imputed product (finer resolution would overstate its precision) and is cross-consistent with the design-based state estimates. Across the region only 8.9 percent of forested area reaches high agreement and 24 percent is high-uncertainty. Earlier: incorporated a second simulated peer review (included). The prototype ensemble layer is now described as an unweighted agreement (membership) score rather than a calibrated probability; the threshold-sensitivity figure is clarified as computed on the labelled plots (training prevalence, not landscape area); FIA reserved status is framed as a conservative lower bound (formally reserved); and notes were added on the forest-type dependence of the structural threshold and on the Rhode Island age trend. Earlier: the Comment was expanded to 14 pages with new supporting evidence: an out-of-bag confusion matrix and a probability-threshold sensitivity figure; a multi-measure temporal panel (FIA stand age, FIA large-tree basal area, and the TreeMap-imputed LSOG classification over time); a New England regional-context section showing older forest stable to increasing in all six states by design-based FIA; a current-protection analysis (only ~12% of Maine older forest is reserved, almost all public; the resource is overwhelmingly private and unprotected); and a demonstrated prototype multi-method ensemble LSOG probability with a companion per-pixel uncertainty layer (archived as 100 m rasters), shown as hex and bivariate maps. Earlier: document layout finalized (captions placed above tables and below figures, tables kept from breaking across pages, white space reduced) and a new section of concrete recommendations for future research added. Earlier: the manuscript was stress-tested, with every quantity re-verified against the source data, and a single orphan reference (the rapid field protocol, Shamgochian et al. 2025) was cited in the text; the verification is documented in the included audit memo (Deep_review_Hagan_vs_Comment). Earlier in the 1.2.x line: authorship corrected to the sole author; the framing broadened (all maps and models are wrong, the question is whether they are useful) and the tone softened; the Comment now credits the original authors' Limitations section, future-work proposals, ingrowth caveat, and citation of the national FIA and GEDI LSOG literature, and adds an explicit treatment of remote-sensing noise and GEDI geolocation error (Shannon et al. 2024), with the design-based FIA estimate as the sole reference. The conclusion names three practices: benchmark against a design-based inventory (FIA), cross-compare against an independent product rather than rely on a single remote-sensing map, and report accuracy and uncertainty on every quantity. The proprietary Seven Islands comparison was removed. A document build error that had frozen the manuscript PDF at an earlier draft in versions 1.2.0 through 1.2.2 is fixed here, so this is the first version to contain the complete, current Comment and Supporting Information. Headline findings. Credible methods disagree by roughly 2.8 times on how much LSOG exists and on the location of most LSOG hectares, while agreeing closely on the rare, well-defined old-growth core. Top-priority protected sets selected by different maps overlap on only 16 to 30 percent of the ground. The design-based FIA estimate and TreeMap imputation both show older forest stable to increasing rather than rapidly declining; the apparent loss reported elsewhere is a gross harvest flux, not a net stock decline. Contents. Derived 100 m GeoTIFFs, summary tables, analysis R scripts, quick-look figures, the Ecosphere Comment manuscript and Supporting Information, the fairness-audit memo, and a full methods-and-findings report (PDF). No FIA plot coordinates are included.