the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 07 Aug 2026
A cold-season temperature reconstruction for central Japan since 259 CE from a two-millennia Kiso Chamaecyparis ring-width chronology
Abstract. We present an annually dated reconstruction of previous November–April (pNov–Apr) temperature in central Japan from an expanded, provenance-aware Kiso Chamaecyparis ring-width chronology. The numerical chronology spans more than two millennia, but formal climatic interpretation is conservatively limited to 259 CE onward because replication and common-signal strength are lower and less continuous earlier. The primary model relates current-year ring-width index to a regional Chubu rural/background temperature anomaly series incorporating eight stations over 1912–2023 (n = 112), yielding r = 0.416, adjusted R2 = 0.165, and RMSE = 0.789 °C. In seven annual split-period tests predefined in the frozen workflow, reduction of error (RE) and coefficient of efficiency (CEff) are positive in six. RE and CEff are also positive in all four formally evaluable G21 high-pass splits. No centered G21 or G51 low-pass split is formally evaluable under the leakage-safe boundary and effective-sample-size criteria; the smoothed curves are therefore used as persistence summaries rather than as independently verified low-frequency reconstructions. Synthetic tests recover an imposed 51-year signal with median r = 0.972, gain slope = 0.909, and zero median phase lag. Rule-based screening identifies two sustained G21-tail candidates: 1999–2015 CE is retained with conditional support of 1.000 and adequate interval-specific SSS, whereas 1458–1473 CE is not retained because its support is 0.744, below the 0.800 threshold. Spatial correlations and comparisons with documentary snowfall, sea-surface temperature, and circulation indices support a regional winter-climate interpretation but are not independent validation or attribution. The reconstruction provides directly tested annual and sub-21-year information together with a cautiously interpretable view of multidecadal persistence since 259 CE.
- Preprint
(4064 KB) - Metadata XML
-
Supplement
(1008 KB) - BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on egusphere-2026-4506', Anonymous Referee #1, 13 Sep 2026
-
AC1: 'Reply on RC1', Hitoshi Yonenobu, 28 Sep 2026
We thank Referee 1 for the constructive assessment and helpful suggestions. We have completed additional sensitivity and comparison analyses and prepared corresponding manuscript revisions. Below we respond in the order of the comments. Figures A1-A8 and Tables A1-A3 in the accompanying response supplement provide the additional evidence for both referee replies. This supplement contains response material, not the revised manuscript.
1. Calibration strength and reconstructed amplitude
We agree that the modest calibration strength requires clearer discussion of both unexplained variability and reconstructed amplitude. Temperature is the response variable in our regression, whereas the ring-width index (RWI) is the predictor. The ordinary R² is 0.1727 and adjusted R² is 0.1652. Thus, approximately 17.3% of annual instrumental target variance is explained in the calibration sample; the remaining fraction is not a causal partition of variance within the tree-ring chronology.
Possible contributors to imperfect correspondence include moisture availability, climate outside the selected season, biological carry-over, competition and disturbance, microsite variation, and differences between forest growth conditions and a regional temperature target. These are plausible influences, not effects partitioned quantitatively by this analysis. The revised Discussion will make this distinction explicit.
The amplitude consequence is substantial. For this single-predictor ordinary least-squares model with an intercept, the ratio of fitted to observed standard deviation in the calibration sample equals |r|. It is 0.4156 over 1912-2023: fitted annual temperatures retain about 41.6% of observed standard deviation and 17.3% of observed variance. This is a diagnostic of variance compression, not a correction factor for historical anomalies or smoothed variability. Individual reconstructed values are not guaranteed lower bounds on actual anomalies.
Positive RE and CEff in six of seven annual splits and all four evaluable G21 high-pass splits support limited predictive information within the tested instrumental period. They do not establish full amplitude recovery or response stationarity before that period. No low-pass split is formally evaluable. Predictive intervals include resampled residual variability, but cannot recover its actual historical sequence. The revised text will distinguish these limits from uncertainty in the fitted conditional mean.
2. Sawara inclusion and changing species contribution
The inclusion of Sawara follows regional dendrochronological practice, in which Sawara has been crossdated with Hinoki; the Sawara series in this study were dated against the Hinoki chronology. This supplies a practical chronological basis for combining the material, without establishing identical climate responses. We therefore removed all Sawara series and rebuilt the chronology under the same standardization, juvenile-ring exclusion, cohort assignment, cluster-balanced aggregation, seasonal target, and verification splits. All historical, archaeological, and buried-wood components were retained.
The raw living-tree inventory contains 316 Hinoki and 22 Sawara series (22/338, 6.5%). One Hinoki series had already been excluded in the original standardization because its fitted spline was non-positive. The analytical living-tree counts are therefore 337 combined and 315 after Sawara exclusion.
Over 1912-2023, the combined and Hinoki-only chronologies correlate at r = 0.9806. Exclusion changes the pNov-Apr target correlation from 0.4156 to 0.4057, adjusted R² from 0.1652 to 0.1570, and RMSE from 0.7894 to 0.7933 °C. Annual RE and CEff remain positive in the same six of seven splits, and both remain positive in all four evaluable G21 high-pass splits. Across eleven fixed seasonal windows, the response ranks correlate at Spearman rho = 0.973. MAM gives the highest correlation in both versions; pNov-Apr remains the original cold-season target and was not reselected to maximize correlation. Figure A1 summarizes this sensitivity.
The 6.5% inventory fraction is not a constant annual contribution. Sawara contributes during 1647-2024, and its series-count fraction reaches 15/31 (48.4%) in 2023. Combined versus Hinoki-only chronology correlation is 0.9980 over 1912-2001 but 0.8739 over 2002-2023. For verification over 2002-2023, CEff decreases from 0.1263 to 0.0548 after exclusion. Thus, the recent low-replication tail is more sensitive. Counts are not the effective weights of the cluster-balanced biweight chronology.
These results support retaining the combined chronology, with explicit endpoint qualifications. They do not establish identical species-specific responses or a stationary climatic response over two millennia. Species exclusion is a separate deterministic sensitivity analysis; species-composition effects are not included in the reported uncertainty ensemble, and historical interval screening was not rerun after exclusion.
3. Biological interpretation of the cold-season response
We agree that the association requires a fuller biological discussion. It does not imply continuous cambial activity throughout November-April. The revised Discussion will distinguish possible effects on subsequent cambial activity, spring carbon uptake, and overwintering water relations from processes actually measured in this study.
Localized stem cooling suppressed cambial activity in an experiment including Chamaecyparis obtusa (Begum et al., 2016), although the experiment concerned the active growing season. In a central-Japan forest containing Hinoki and Sugi, winter temperature and snowmelt timing were associated with spring photosynthetic capacity (Saitoh et al., 2010). These studies provide physiological context for temperature sensitivity and delayed effects, rather than direct evidence of the mechanism controlling Kiso ring width.
Frozen soil or stems may restrict water supply while foliage continues losing water, as described for alpine conifers by Mayr et al. (2006). Snowmelt infiltration may affect spring water availability. Snow also insulates roots: Schaberg et al. (2008) documented freezing injury in yellow-cedar seedlings under reduced soil protection. The latter studies are analogies from other conifers. Reduced snow cover could offset potential benefits of warmer conditions.
We have additionally examined monthly temperature and precipitation responses in addressing Referee 2 (Figs. A5 and A7). These analyses help delimit climatic associations but do not identify causal pathways. Concurrent observations of cambial phenology, soil moisture and temperature, root injury, and carbon storage are unavailable. The revised wording will therefore present these mechanisms as hypotheses and retain the empirical interpretation of the pNov-Apr association.
4. Role and endpoint of the 1999-2015 interval
We agree that this is not a new pre-instrumental paleoclimatic finding. Its role is to illustrate application of the same screening rules across the reconstruction, including the calibration era. Both identified candidates remain reported: the modern upper-tail candidate is retained, whereas the 1458-1473 lower-tail candidate fails the conditional-support threshold (0.744 versus 0.800).
The modern interval is not independent validation of warming-trend recovery, because it overlaps calibration and formal low-pass verification is not evaluable. Conditional support of 1.000 concerns that fixed interval under the stated ensemble; it is not a probability of historical uniqueness or warming attribution.
The dates are centered G21 years. Crucially, 2015 is the final complete center available from annual predictions through 2025, not evidence that the warm period ended in 2015. G21 center year t uses t-10 through t+10. The 2024-2025 predictions are excluded from calibration and verification but contribute to the terminal smoothed values. We will make this distinction explicit in the Abstract, Methods, Results, Discussion, and Conclusions. Ending the annual record in 2023 would instead leave 2013 as the last complete G21 center and would require separate interval screening; that sensitivity has not been claimed.
5. Detrended spatial correlations
We have calculated and checked the four-panel detrended map corresponding to Fig. 5 (Fig. A3). Within each pair of series, we select complete overlapping years, remove separate least-squares linear trends against calendar year, and correlate the residuals. The minimum overlap remains 50 years; the available paired periods are retained rather than replaced with one common period.
For Kiso versus JMA stations, the mean correlation across all 151 eligible stations decreases from 0.496 to 0.305, with 149 positive detrended correlations. Across all 2,900 archived CRU TS cells, the mean decreases from 0.440 to 0.199. Within the displayed map extent, the corresponding CRU means are 0.480 and 0.240 for 1,640 cells. The map therefore shows predominantly positive regional covariation remaining after linear detrending. Table A3 also gives approximate significance diagnostics, added in response to Referee 2.
Linear detrending does not isolate exclusively interannual variability: nonlinear and slower variations may remain. We will therefore describe this result as covariation beyond fitted linear trends, not proof of an exclusively interannual signal or independent validation. The instrumental products represent related climate information, and neighboring stations and grid cells are not independent replicates.
6. Isotope and documentary comparisons
We have prepared the requested time-series comparison (Fig. A4). It aligns Kiso and the archived central-Japan oxygen-isotope anomaly series associated with Nakatsuka et al. (2020) over 259-2005 CE, and Kiso and Kawanishi DJF/FMA snowfall ratios over 1831-1980 CE. Each variable retains its own units and axis. Annual values accompany consistently calculated centered Gaussian G21 curves, with incomplete endpoint windows omitted. We have not shifted, inverted, or selected favorable subperiods to enhance agreement.
Kiso-isotope annual correlation is weak (r = 0.098; n = 1,747), and the smoothed histories do not show close one-to-one correspondence. The documentary annual correlations reproduce the inverse associations reported previously (DJF: r = -0.259; FMA: r = -0.210; n = 150). These comparisons make differences as well as correspondence visible.
The isotope record concerns mainly summer hydroclimate, Kiso targets cold-season temperature, and the documentary ratios describe local snowfall occurrence. Their complementarity therefore means evidence about different aspects of climate, rather than mutual validation of a single temperature history. Smoothed associations remain descriptive. The revised manuscript and Supplement will document the source series, units, overlapping periods, and filtering choices.
7. Minor comments
Line 164: The replacement wording is “Assigning SSS windows whose center years fall within a candidate interval”. Its role remains a non-decisional comparison; the primary replication rule is unchanged.
Line 350: We checked the revised source and rendered text; the wording is “central Japan”, without a spurious hyphen.
Lines 415-416: The internal “v5.3” identifier has been removed from the main-text and Supplement references to Table 3. Versioned provenance remains in the analytical archive.
Lines 473-475: The dynamically generated access-date suffix in the Bunn et al. reference has been removed. The publication year remains 2025 and the CRAN URL is retained.
Figure 3(c): We have redrawn the panel from the unchanged archived verification statistics, increased the upper plotting space, and separated annotations from the bars (Fig. A2). The previously clipped and overlapping labels are now visible.
We also identified and clarified mixed filtering conventions in the older external-context comparisons. Some archived G11/G21 labels denote arithmetic moving averages rather than Gaussian filters; the revised descriptions will distinguish these from formal Gaussian verification and the new consistently filtered comparison in Fig. A4. In addition, the Methods audit requested by Referee 2 revealed a mismatch in the earlier synthetic-test baseline. We have corrected and rerun that diagnostic; Table A2 and our RC2 reply document the correction. This does not change the chronology used for calibration or the Sawara-exclusion results above.
References added or used in this response
- Begum et al. (2016), Annals of Botany, 117, 465-477. https://doi.org/10.1093/aob/mcv181
- Saitoh et al. (2010), Journal of Plant Research, 123, 473-483. https://doi.org/10.1007/s10265-009-0308-7
- Mayr et al. (2006), Ecology, 87, 3175-3185. https://doi.org/10.1890/0012-9658(2006)87[3175:FDICAT]2.0.CO;2
- Schaberg et al. (2008), Global Change Biology, 14, 1282-1293. https://doi.org/10.1111/j.1365-2486.2008.01577.x
- Nakatsuka et al. (2020), Climate of the Past, 16, 2153-2172. https://doi.org/10.5194/cp-16-2153-2020
- Nakatsuka et al. (2020), archived oxygen-isotope data, NOAA/NCEI. https://doi.org/10.25921/crxm-hs43
-
AC1: 'Reply on RC1', Hitoshi Yonenobu, 28 Sep 2026
-
RC2: 'Comment on egusphere-2026-4506', Anonymous Referee #2, 15 Sep 2026
General comments
This study attempts to reconstruct previous November–April (pNov–Apr) temperature over nearly two millennia in central Japan using TRW proxy. The topic is interesting and valuable. I also appreciate the authors’ careful treatment of reconstruction skill, frequency dependence, signal retention, and uncertainty. The manuscript is transparent about what the statistical tests can and cannot support.
My main concern is whether the current evidence is sufficient to support that the full chronology since 259 CE can be used to do a reliable pNov–Apr temperature reconstruction. The calibration and verification results show that the chronology contains some cold-season temperature information during the instrumental period, but these tests cover only about the last century. It is not clear whether the TRW–temperature relationship remained stable over the preceding ~1700 years. This issue is particularly important because the long chronology combines material from different sites, source histories, and, to some extent, species, whereas the calibration is based on the modern part of the chronology. And also the adjusted R² is only 0.165, which is rather low for a TRW-based temperature reconstruction.
Therefore, I think the main issue is not the rigor of the tests themselves, but whether the available tests are sufficient to support applying the modern TRW–temperature relationship to the full pre-instrumental period. I encourage the authors to address this point more directly and, where possible, provide additional tests of the stability of the chronology and its climate response. If this cannot be tested with the available data, the interpretation of the pre-instrumental reconstruction may need to be further narrowed.
Specific comments
- The statistical relationship between TRW and pNov–Apr temperature is clear, but the biological basis of this relationship remains unclear. The statement that “the defensible inference is statistical” is reasonable, but a better discussion of the possible mechanism would strengthen the reconstruction. I suggest examining the monthly response of TRW to winter temperature, precipitation, drought as well as growing-season temperature, precipitation, soil moisture, and drought et al. This could help clarify whether the winter signal is direct or partly related to other climate variables or carry-over effects. Possible links through snowpack, soil temperature, frost, cambial reactivation, or stored resources could also be discussed.
- Stability of the TRW–temperature relationship is the main concern for the long reconstruction. The current calibration and verification tests show skill during the instrumental period, but they do not tell us whether the same relationship existed throughout the pre-instrumental period. Could the authors examine this using other stationarity tests? It would also be useful to compare the climate response of earlier and later parts of the chronology where replication is sufficient. If the pre-instrumental relationship cannot be tested directly, I suggest making a clearer distinction between the period with tested reconstruction skill and the much longer period for which the chronology is interpreted as a climate record.
- The chronology combines living trees, archaeological wood, historical timbers, and buried wood from different locations. The living chronology also contains a small Sawara component. This raises questions about whether the climate signal in the earlier chronology is comparable to that in the modern calibration material. I suggest examining correlations among chronologies from different sites and, where possible, between species during their overlapping periods. A sensitivity test excluding the Sawara samples would also be useful if the sample size allows. These analyses could help show whether the long chronology has a common regional signal similar to that used for the instrumental calibration.
- The comparisons with documentary snowfall, SST, and circulation indices could provide useful supporting evidence, especially for the pre-instrumental period. I suggest showing these comparisons more clearly, preferably with figures, and discussing what they add to the reconstruction. At the same time, it would be useful to keep a clear distinction between agreement with independent climate records and true independent validation of the reconstruction.
- Please define how the effective sample size is calculated and explain the choice of neff ≥ 8. Since this criterion determines whether the low-frequency verification can be evaluated, its basis should be stated clearly in the Methods.
- The predefined split periods are a strength of the study, but the choice of E25, L25, E45, L45, E56, and L56 could be explained more clearly. If the data allow, testing a few additional split configurations would help assess how sensitive the RE and CE results are to the choice of calibration and verification periods. For example, E25 and L25 for the period 1974-2023.
- The synthetic input–recovery tests are important for interpreting the reconstruction results. Their design and evaluation should be described in the Methods rather than only in the Results.
Technical corrections
- Please define G21 and G51 before their first use.
- The sentence “A long chronology is therefore not, by itself, a reliable climate reconstruction...” is not necessary to present, and could be removed.
- The paragraph beginning “Ring width and stable isotopes need not respond...” is also not necessary, and can be removed.
- Some text in the legend of Fig. 3c appears to be cut off; please check.
- I suggest to show the statistical significance levels in Fig. 5.
Overall, I appreciate the authors’ careful and transparent approach to reconstruction testing. My main concern is the gap between the period in which reconstruction skill is directly tested and the much longer period over which the reconstruction is interpreted. The current results provide evidence for a cold-season temperature signal during the instrumental period, but it remains unclear how well this relationship applies to the earlier chronology, given its changing sources, sites, and species. I therefore encourage the authors to either add analyses that address the stability of the proxy–climate relationship and chronology composition, or further qualify the interpretation of the pre-instrumental record. In particular, the manuscript would benefit from a clear distinction between what is directly supported by the instrumental-period tests and what is inferred from the extended chronology.
Citation: https://doi.org/10.5194/egusphere-2026-4506-RC2 -
AC2: 'Reply on RC2', Hitoshi Yonenobu, 28 Sep 2026
We thank Referee 2 for identifying the distinction between instrumental-period skill and pre-instrumental applicability as the central issue. We agree that neither successful crossdating, strong chronology replication, nor positive instrumental verification can establish that the temperature response remained unchanged throughout the preceding centuries. We have therefore carried out additional instrumental-period, chronology-composition, and verification-sensitivity analyses, and will narrow the interpretation of the pre-instrumental record accordingly. The accompanying response supplement presents the quantitative results; it is not the revised manuscript.
General comment: scope of the temperature reconstruction
The revised manuscript will explicitly distinguish (i) predictive information directly tested within 1912-2023 from (ii) a temperature-scaled interpretation of the earlier chronology that depends on transfer of the modern relationship. The 259 CE boundary reflects replication and common-signal criteria; it is not the beginning of a period with independently demonstrated temperature skill. Our uncertainty intervals remain conditional on the specified calibration and chronology ensemble and do not encompass all possible changes in climate response, provenance, or ecology.
We will replace the concluding Abstract statement with: “Instrumental-period tests support limited annual and G21 high-pass temperature information. Before the instrumental period, the temperature-scaled chronology provides a conditional interpretation of past variability, assuming transferability of the modern relationship; neither that assumption nor low-frequency temperature amplitudes is independently validated.” Corresponding qualifications will be included in the Methods, limitations, and Conclusions. Low calibration strength and conditional-mean variance compression will also be addressed as detailed in our RC1 response.
1. Monthly responses, other climate variables, and possible mechanisms
We have compared monthly regional temperature responses for 1912-2023 and its non-overlapping 56-year halves (Fig. A5), including both raw and linearly detrended correlations. Months span the preceding January to the current December, covering winter, the growing season, and potential carry-over associations. We also downloaded JMA Takayama monthly mean temperature and precipitation totals and applied the same comparisons (Fig. A7). Only unflagged numeric observations were accepted; flagged or missing months were not interpolated. These station data are additional explanatory diagnostics and do not replace the regional calibration target.
As a further descriptive check, we used fixed November-April and May-September windows and calculated the partial correlation between RWI and the regional pNov-Apr target after adjusting for Takayama winter precipitation, May-September precipitation, and May-September temperature. Across the 107 years with all required observations (1912-2022, with gaps), the unadjusted correlation is 0.440 and the partial correlation is 0.350. After separate linear detrending of all variables, these values are 0.408 and 0.366. Thus, the positive cold-season association is not wholly explained by these selected station covariates. This is not a causal test and does not rule out other moisture effects or collinearity; the 107-year result should not be compared directly with the 112-year calibration as if the samples were identical.
We do not have corresponding site-level soil moisture, soil temperature, snowpack, drought-index, cambial, or stored-carbon observations for a comparable analysis. We have not treated station precipitation as a substitute for these variables. This remaining limitation will be stated explicitly.
The revised biological discussion will consider temperature sensitivity of cambial activity, snowmelt and spring carbon uptake, frost drought, root insulation and freezing risk, and possible resource carry-over. Evidence involving Hinoki (Begum et al., 2016; Saitoh et al., 2010) will be distinguished from analogies in other conifers (Mayr et al., 2006; Schaberg et al., 2008). These pathways remain hypotheses; we do not infer active cambial growth throughout winter or identify one dominant mechanism.
2. Temporal stability of the temperature relationship
For 1912-1967 and 1968-2023, raw pNov-Apr correlations are 0.384 and 0.431, respectively, and fitted slopes are 2.204 and 1.691 °C per RWI unit (Table A1). A paired moving-block bootstrap, resampling independently within each half, gives a late-minus-early slope difference of -0.513 with a 95% percentile interval of -2.242 to 0.694 for five-year blocks (3,000 draws). The corresponding correlation-difference interval is -0.387 to 0.291. Three- and ten-year block sensitivities also include zero. These broad intervals do not demonstrate equivalence or stationary responses.
The detrended correlations are 0.414 in the first half and 0.186 in the second half. This weaker late-period association must be acknowledged even though the raw correlations are similar. Descriptive 31-year moving correlations also vary substantially (Fig. A6a). The revised discussion will therefore avoid claiming uniform instrumental-period sensitivity, much less stability over two millennia.
There is no directly observed pNov-Apr target throughout the pre-instrumental chronology. Comparing its early and late segments against instrumental temperature is consequently impossible. Overlap tests among material groups address continuity of shared growth variation (next response), but they cannot fill this climate-validation gap. We will state that the earlier temperature-scaled record is conditional on an unverified transferability assumption.
3. Different sites, material sources, and species
We used the same primary standardized individual-series indices to calculate disjoint source/location and material-group chronologies with Tukey biweight means. Comparison required at least five contributing series in each group in each paired year and at least 30 paired years. The 20IK bridge component assigned to D was omitted from these pairwise diagnostics so that its shared provenance was not treated as independent evidence. This omission concerns the diagnostic comparisons, not the production chronology.
The living cohorts A-B, A-C, and B-C correlate at 0.914, 0.900, and 0.861, respectively. Earlier links are also positive: living cohort C versus archaeological group D gives r = 0.506 (89 paired years, 1555-1643); D versus historical building group E gives r = 0.451 (116 years, 1057-1172); and E versus buried-wood group F gives r = 0.553 (58 available paired years within 610-670). Detrended correlations are 0.560, 0.451, and 0.538 for these three earlier links (Fig. A6b). Non-evaluable pairs are shown as NE in Fig. A6b and retained in the analytical CSV table rather than interpreted as zero correlations.
Among 171 possible source/location pairs, 85 meet the same replication and overlap criteria. Their raw correlations range from 0.216 to 0.874, with median 0.627. The identifiers include forest sites and timber-source or archaeological groups; they are not all known growth locations. These results concern an already crossdated collection, so they are evidence of shared growth variation within that collection, not independent tests of dating or temperature specificity.
Living Hinoki and Sawara group chronologies correlate at r = 0.643 over 280 qualifying years within 1744-2023 (detrended r = 0.625). Their separately aggregated instrumental temperature correlations differ (Hinoki: 0.433; Sawara: 0.274; 1912-2023). Differences in species, site composition, sample size, and tree history are confounded, so these are not estimates of an isolated species effect.
The complete Sawara-exclusion experiment is described in our RC1 reply and Fig. A1. Removal of all 22 Sawara series changes calibration r from 0.4156 to 0.4057 and preserves the positive results in the same six annual and four evaluable high-pass splits. Greater sensitivity in the recent low-replication tail is reported explicitly. Neither these tests nor the overlap correlations establish stable temperature sensitivity of all earlier material.
The input audit also identified three unobserved widths in two buried-wood series (one in 437 CE and two in 512-513 CE). These are omitted before the existing spline fit, which treats retained observations as equally spaced; the calendar labels of the resulting indices are preserved. As a separate sensitivity check, we excluded both entire series under the same standardization. The resulting chronology correlates at r = 0.9987 with the primary chronology over 259-2025; changes are confined to 340-689, with a maximum absolute difference of 0.0978 RWI. The instrumental predictor and calibration are exactly unchanged. This removal test does not quantify or bound the effect of a calendar-aware spline correction. We will explicitly document this limited gap-handling issue; no widths were imputed and the primary chronology and uncertainty ensemble remain unchanged.
4. External comparisons and independent validation
We have prepared the time-series comparison in Fig. A4: the archived central-Japan oxygen-isotope series over 259-2005 and Kawanishi snowfall ratios over 1831-1980 are displayed alongside Kiso with native units, aligned time axes, and consistent complete-window Gaussian smoothing. We will retain and clarify the existing documentary, SST, and circulation diagnostics in the manuscript Supplement and direct readers to them explicitly.
These comparisons offer limited contextual support, not independent quantitative validation of pre-instrumental pNov-Apr temperatures. The annual isotope correlation is weak (0.098), and documentary DJF/FMA correlations are modest and negative (-0.259/-0.210). The documentary overlap partly predates instrumental calibration, but concerns snowfall occurrence at another location. SST and circulation associations address related climate variability and likewise do not validate the reconstruction back to 259 CE. No subperiod or lag was selected to optimize the new comparison. The revised discussion will make these distinctions explicit.
5. Effective sample size and the threshold of eight
The implemented approximation is n_eff = n(1 - r1_obs × r1_pred)/(1 + r1_obs × r1_pred), where n is the number of retained verification pairs and r1_obs and r1_pred are lag-one correlations of the observed and predicted series at the evaluated scale. Lag-one correlations use consecutive values after incomplete transformed edges have been removed. The result is bounded to [2, n]. This is an AR(1) approximation associated with Bretherton et al. (1999), not an exact count of independent observations. The implementation returns n if an AR1 coefficient is undefined and rejects a non-positive denominator; such fallback cases do not determine the reported low-pass conclusion.
Calibration and verification are filtered independently. In addition to n_eff >= 8 in verification, each transformed segment must contain at least eight paired values. The threshold eight is a predefined pragmatic minimum, not a universal significance cutoff and not a threshold prescribed by Bretherton et al. We will state its status explicitly in Methods.
We tested effective-sample-size thresholds of 4, 6, 8, 10, and 12 while retaining the eight-observation minimum in each segment. For the seven original splits, all choices yield the same evaluability counts: seven annual, four G21 high-pass, zero G21 low-pass, and zero G51 low-pass. For the count-eligible G21 low-pass cases, n_eff is only 2.00-2.05. Thus, the low-pass conclusion does not depend narrowly on choosing eight. The revised Methods will give the equation, implementation, separate count criterion, and sensitivity result.
6. Split definitions and additional configurations
The original 25-year pair compares 1912-1936 with 1937-1961 in both directions; the 45-year pair compares 1912-1956 with 1957-2001; and the 56-year pair compares 1912-1967 with 1968-2023. These represent a legacy-equivalent interval, extension through the pre-modern-tail endpoint, and the full instrumental record, respectively. A seventh test calibrates on 1912-2001 and verifies on the low-replication 2002-2023 tail. We will spell out the years and rationale rather than rely on the E/L abbreviations.
As requested, we added the 1974-1998 versus 1999-2023 comparison in both directions. Forward RE and CEff are 0.400 and 0.038; reverse values are 0.333 and 0.036. We also shifted the 45-year pair to 1934-1978 versus 1979-2023; forward RE/CEff are 0.195/0.075 and reverse values 0.409/0.301 (Table A1). All four additional annual tests are positive, although the recent 25-year CEff values are small. We will identify these explicitly as review-stage sensitivity tests and retain all seven original tests, including the unfavorable one.
The 25-year segments leave only five complete centers after a 21-year centered filter, so their frequency-dependent tests remain non-evaluable under the eight-observation criterion. The additional 45-year high-pass tests are evaluable and positive; low-pass tests remain non-evaluable. Additional annual success therefore does not establish low-frequency predictive skill.
7. Synthetic input-recovery design and an audit correction
We agree that the design belongs in Methods. Auditing the earlier test revealed that its baseline covered 482-2025 and did not reproduce the final ABCDEF chronology: the legacy input reader interpreted the F cluster code as a logical value and excluded that buried-wood component. We have therefore rerun the test using the final input loader and verified that the unperturbed baseline reproduces each of the four frozen production chronology variants to numerical precision (Table A2).
For each variant, raw ring widths were multiplied by exp[0.05 × s(t)], where s(t) is a unit-amplitude sinusoid on a continuous elapsed-year axis. The transition from 1 BCE to 1 CE advances by one year, consistent with the absence of year zero in the historical labels. Periods are 11, 21, 31, 51, 75, 101, 150, and 200 years, each with phase fractions 0, 0.25, 0.5, and 0.75 of a cycle. The original standardization and aggregation were then repeated. The recovered signal is the injected-minus-unperturbed chronology divided by 0.05. Recovery correlation and the regression slope of recovered on imposed signal quantify shape and gain; the archived lag diagnostic searches -10 to +10 years for the largest absolute correlation. Results are summarized across the four phases separately for each variant.
The corrected primary 51-year test gives median recovery r = 0.974, median gain = 0.919, and median lag = 0 years. These replace the earlier values r = 0.972 and gain = 0.909. The correction concerns the diagnostic test, not the production chronology, temperature calibration, or annual verification.
The revised Methods and numerical summaries will use this corrected, reproducible experiment. It tests processing of a common imposed growth signal, not whether real temperature sensitivity was stable across source groups or centuries. It neither repairs low calibration strength nor independently validates historical temperature amplitudes.
Technical corrections
G21 and G51 will be defined at first use as centered Gaussian smooths with 21- and 51-year support. High-pass means the annual series minus its corresponding smooth. The unnecessary introductory sentence beginning “A long chronology is therefore not...” and the paragraph beginning “Ring width and stable isotopes need not respond...” will be removed; the necessary limitations will be stated specifically where their evidence is discussed.
Figure 3(c) has been redrawn to correct clipping and overlap (Fig. A2). For Fig. 5, we have prepared raw and detrended maps with significance markers (Figs. A8 and A3). Approximate two-sided correlation tests use the lag-one effective sample size above and a minimum n_eff of eight. Spatial lag-one correlations use only pairs separated by exactly one calendar year, without joining across missing years. Degrees of freedom are n_eff - 2 for raw correlations and n_eff - 3 after removing linear time as a shared covariate. We apply the Benjamini-Yekutieli adjustment separately within each displayed panel and exclude Takayama's self-correlation. Crosses mark adjusted p < 0.05. Table A3 reports both pointwise and adjusted counts. Because these are approximate AR1 tests, especially limited for trending raw series, we do not interpret the markers as proof of field-level significance, exclusively interannual response, or independent climate validation.
We appreciate the referee's emphasis on the calibration-to-history gap. The additional analyses delimit the evidence more clearly; they do not eliminate that gap. The revised interpretation will explicitly retain it as a central limitation.
References and additional data
- Bretherton et al. (1999), Journal of Climate, 12, 1990-2009. https://doi.org/10.1175/1520-0442(1999)012<1990:TENOSD>2.0.CO;2
- Benjamini and Yekutieli (2001), The Annals of Statistics, 29, 1165-1188. https://doi.org/10.1214/aos/1013699998
- JMA, Takayama monthly temperature and precipitation tables, accessed 28 September 2026: https://www.data.jma.go.jp/stats/etrn/view/monthly_s3.php?prec_no=52&block_no=47617&view=a1 and https://www.data.jma.go.jp/stats/etrn/view/monthly_s3.php?prec_no=52&block_no=47617&view=p5
- Physiological and isotope references are listed in our RC1 reply and the accompanying response supplement.
Status: closed
-
RC1: 'Comment on egusphere-2026-4506', Anonymous Referee #1, 13 Sep 2026
In this manuscript, the authors present a two-millennia cold-season temperature reconstruction for central Japan based on an extensive Kiso Chamaecyparis tree-ring width chronology. Combining wood from diverse sources and addressing the urban heat island (UHI) effect through careful target selection are strong methodological points. While the manuscript is well presented and results can provide valuable insights into the cold-season temperatures in central Japan, I has a few concerns regarding the methods, results, and interpretation.
Specific concerns are detailed below:
The primary calibration model yields an adjusted R2 of 0.165 (r = 0.416) for the period 1912–2023. While the cross-validation statistics (RE and CEff) are positive and validate the model’s predictive skill, an explained variance of ~16.5% is relatively low for a temperature reconstruction, implying that over 80% of the variance in the tree-ring chronology is driven by unexplained factors. The authors should expand the discussion on what might be driving this high proportion of unexplained variance and clearly state the implications on interpreting the amplitude of historical temperature anomalies.
In Section 2.1 (Lines 89-91), the authors state that 22 out of 338 living-tree series are Chamaecyparis pisifera (Sawara), which were combined with Chamaecyparis obtusa (Hinoki). The authors acknowledge that “species-specific climate responses were not estimated and this remains a limitation”. If the ratio of these two species changes significantly over the two-millennia chronology, it could introduce non-stationarity into the climate response. The authors should provide a stronger justification for why mixing them does not impact the long-term climatic signal.
The selection of the previous November–April target is well justified statistically. However, the biological discussion in Section 4.1 (Lines 274-279) is somewhat brief. Given that cambial activity in this region typically occurs in the warmer months, the authors should explicitly discuss the ecophysiological mechanisms through which cold-season temperature strongly impact the subsequent ring width in Kiso Chamaecyparis (e.g., specific references to winter desiccation, snowmelt infiltration, or root freezing).
In Section 3.6 and Table 3, the rule-based screening identifies the 1999–2015 CE interval as a sustained upper-tail feature. Because this period falls entirely within the modern instrumental era, its value as a paleoclimatic finding is limited. The authors should further clarify the utility of highlighting this interval, is it primarily to demonstrate that the proxy successfully captures the modern warming trend, or does it serve another purpose?
In Section 4.4 (Lines 343-344), the authors state that detrending reduces the mean station correlation from 0.496 to 0.305, indicating that shared slower variations (likely the modern warming trend) contribute significantly to the raw spatial footprint shown in Figure 5. To prove that the proxy truly captures a coherent regional climate signal at the interannual scale, rather than a global warming trend, the authors should provide a supplementary figure showing the detrended spatial correlation map.
In Sections 4.5 and 4.6, the manuscript discusses the reconstruction’s complementarity with documentary snowfall records and the central-Japan isotope hydroclimate record (Nakatsuka et al., 2020). However, this comparison is entirely textual. Given the paleoclimate context of central Japan, adding a time-series figure that visually compares this new cold-season temperature reconstruction with the documentary and isotope records (for the overlapping periods) would substantially elevate the paper’s impact and make the discussion much more convincing.
Minor comments:
Line 164: The phrasing “Center-year-within-interval assignment” is hard to read. Consider rephrasing (e.g., “Assigning the center year within an interval...”).
Line 350: Typo. “downstream central- Japan” should be “downstream central Japan”.
Line 415-416: The text mentions “documents the complete Table 3 v5.3 candidate”. The “v5.3” appears to be an internal versioning artifact. Please remove or correct it.
Lines 473-475: In the reference list for Bunn et al. (2025) regarding the dplR package, there is a typo at the end of the citation: last accessed: July 29, 2026, 2025. The presence of two years (“2026, 2025”) appears to be an artifact from citation management software. Please correct this.
Figure 3: In Figure 3(c), the text above the bars (specifically “primary 0.127” and “4/4 positive” for the G21 HP category) are partially clipped by the top border of the plot frame. Additionally, the label “carry-over 0.049” overlaps with the bars, making it difficult to read. Please adjust the formatting.
Citation: https://doi.org/10.5194/egusphere-2026-4506-RC1 -
AC1: 'Reply on RC1', Hitoshi Yonenobu, 28 Sep 2026
We thank Referee 1 for the constructive assessment and helpful suggestions. We have completed additional sensitivity and comparison analyses and prepared corresponding manuscript revisions. Below we respond in the order of the comments. Figures A1-A8 and Tables A1-A3 in the accompanying response supplement provide the additional evidence for both referee replies. This supplement contains response material, not the revised manuscript.
1. Calibration strength and reconstructed amplitude
We agree that the modest calibration strength requires clearer discussion of both unexplained variability and reconstructed amplitude. Temperature is the response variable in our regression, whereas the ring-width index (RWI) is the predictor. The ordinary R² is 0.1727 and adjusted R² is 0.1652. Thus, approximately 17.3% of annual instrumental target variance is explained in the calibration sample; the remaining fraction is not a causal partition of variance within the tree-ring chronology.
Possible contributors to imperfect correspondence include moisture availability, climate outside the selected season, biological carry-over, competition and disturbance, microsite variation, and differences between forest growth conditions and a regional temperature target. These are plausible influences, not effects partitioned quantitatively by this analysis. The revised Discussion will make this distinction explicit.
The amplitude consequence is substantial. For this single-predictor ordinary least-squares model with an intercept, the ratio of fitted to observed standard deviation in the calibration sample equals |r|. It is 0.4156 over 1912-2023: fitted annual temperatures retain about 41.6% of observed standard deviation and 17.3% of observed variance. This is a diagnostic of variance compression, not a correction factor for historical anomalies or smoothed variability. Individual reconstructed values are not guaranteed lower bounds on actual anomalies.
Positive RE and CEff in six of seven annual splits and all four evaluable G21 high-pass splits support limited predictive information within the tested instrumental period. They do not establish full amplitude recovery or response stationarity before that period. No low-pass split is formally evaluable. Predictive intervals include resampled residual variability, but cannot recover its actual historical sequence. The revised text will distinguish these limits from uncertainty in the fitted conditional mean.
2. Sawara inclusion and changing species contribution
The inclusion of Sawara follows regional dendrochronological practice, in which Sawara has been crossdated with Hinoki; the Sawara series in this study were dated against the Hinoki chronology. This supplies a practical chronological basis for combining the material, without establishing identical climate responses. We therefore removed all Sawara series and rebuilt the chronology under the same standardization, juvenile-ring exclusion, cohort assignment, cluster-balanced aggregation, seasonal target, and verification splits. All historical, archaeological, and buried-wood components were retained.
The raw living-tree inventory contains 316 Hinoki and 22 Sawara series (22/338, 6.5%). One Hinoki series had already been excluded in the original standardization because its fitted spline was non-positive. The analytical living-tree counts are therefore 337 combined and 315 after Sawara exclusion.
Over 1912-2023, the combined and Hinoki-only chronologies correlate at r = 0.9806. Exclusion changes the pNov-Apr target correlation from 0.4156 to 0.4057, adjusted R² from 0.1652 to 0.1570, and RMSE from 0.7894 to 0.7933 °C. Annual RE and CEff remain positive in the same six of seven splits, and both remain positive in all four evaluable G21 high-pass splits. Across eleven fixed seasonal windows, the response ranks correlate at Spearman rho = 0.973. MAM gives the highest correlation in both versions; pNov-Apr remains the original cold-season target and was not reselected to maximize correlation. Figure A1 summarizes this sensitivity.
The 6.5% inventory fraction is not a constant annual contribution. Sawara contributes during 1647-2024, and its series-count fraction reaches 15/31 (48.4%) in 2023. Combined versus Hinoki-only chronology correlation is 0.9980 over 1912-2001 but 0.8739 over 2002-2023. For verification over 2002-2023, CEff decreases from 0.1263 to 0.0548 after exclusion. Thus, the recent low-replication tail is more sensitive. Counts are not the effective weights of the cluster-balanced biweight chronology.
These results support retaining the combined chronology, with explicit endpoint qualifications. They do not establish identical species-specific responses or a stationary climatic response over two millennia. Species exclusion is a separate deterministic sensitivity analysis; species-composition effects are not included in the reported uncertainty ensemble, and historical interval screening was not rerun after exclusion.
3. Biological interpretation of the cold-season response
We agree that the association requires a fuller biological discussion. It does not imply continuous cambial activity throughout November-April. The revised Discussion will distinguish possible effects on subsequent cambial activity, spring carbon uptake, and overwintering water relations from processes actually measured in this study.
Localized stem cooling suppressed cambial activity in an experiment including Chamaecyparis obtusa (Begum et al., 2016), although the experiment concerned the active growing season. In a central-Japan forest containing Hinoki and Sugi, winter temperature and snowmelt timing were associated with spring photosynthetic capacity (Saitoh et al., 2010). These studies provide physiological context for temperature sensitivity and delayed effects, rather than direct evidence of the mechanism controlling Kiso ring width.
Frozen soil or stems may restrict water supply while foliage continues losing water, as described for alpine conifers by Mayr et al. (2006). Snowmelt infiltration may affect spring water availability. Snow also insulates roots: Schaberg et al. (2008) documented freezing injury in yellow-cedar seedlings under reduced soil protection. The latter studies are analogies from other conifers. Reduced snow cover could offset potential benefits of warmer conditions.
We have additionally examined monthly temperature and precipitation responses in addressing Referee 2 (Figs. A5 and A7). These analyses help delimit climatic associations but do not identify causal pathways. Concurrent observations of cambial phenology, soil moisture and temperature, root injury, and carbon storage are unavailable. The revised wording will therefore present these mechanisms as hypotheses and retain the empirical interpretation of the pNov-Apr association.
4. Role and endpoint of the 1999-2015 interval
We agree that this is not a new pre-instrumental paleoclimatic finding. Its role is to illustrate application of the same screening rules across the reconstruction, including the calibration era. Both identified candidates remain reported: the modern upper-tail candidate is retained, whereas the 1458-1473 lower-tail candidate fails the conditional-support threshold (0.744 versus 0.800).
The modern interval is not independent validation of warming-trend recovery, because it overlaps calibration and formal low-pass verification is not evaluable. Conditional support of 1.000 concerns that fixed interval under the stated ensemble; it is not a probability of historical uniqueness or warming attribution.
The dates are centered G21 years. Crucially, 2015 is the final complete center available from annual predictions through 2025, not evidence that the warm period ended in 2015. G21 center year t uses t-10 through t+10. The 2024-2025 predictions are excluded from calibration and verification but contribute to the terminal smoothed values. We will make this distinction explicit in the Abstract, Methods, Results, Discussion, and Conclusions. Ending the annual record in 2023 would instead leave 2013 as the last complete G21 center and would require separate interval screening; that sensitivity has not been claimed.
5. Detrended spatial correlations
We have calculated and checked the four-panel detrended map corresponding to Fig. 5 (Fig. A3). Within each pair of series, we select complete overlapping years, remove separate least-squares linear trends against calendar year, and correlate the residuals. The minimum overlap remains 50 years; the available paired periods are retained rather than replaced with one common period.
For Kiso versus JMA stations, the mean correlation across all 151 eligible stations decreases from 0.496 to 0.305, with 149 positive detrended correlations. Across all 2,900 archived CRU TS cells, the mean decreases from 0.440 to 0.199. Within the displayed map extent, the corresponding CRU means are 0.480 and 0.240 for 1,640 cells. The map therefore shows predominantly positive regional covariation remaining after linear detrending. Table A3 also gives approximate significance diagnostics, added in response to Referee 2.
Linear detrending does not isolate exclusively interannual variability: nonlinear and slower variations may remain. We will therefore describe this result as covariation beyond fitted linear trends, not proof of an exclusively interannual signal or independent validation. The instrumental products represent related climate information, and neighboring stations and grid cells are not independent replicates.
6. Isotope and documentary comparisons
We have prepared the requested time-series comparison (Fig. A4). It aligns Kiso and the archived central-Japan oxygen-isotope anomaly series associated with Nakatsuka et al. (2020) over 259-2005 CE, and Kiso and Kawanishi DJF/FMA snowfall ratios over 1831-1980 CE. Each variable retains its own units and axis. Annual values accompany consistently calculated centered Gaussian G21 curves, with incomplete endpoint windows omitted. We have not shifted, inverted, or selected favorable subperiods to enhance agreement.
Kiso-isotope annual correlation is weak (r = 0.098; n = 1,747), and the smoothed histories do not show close one-to-one correspondence. The documentary annual correlations reproduce the inverse associations reported previously (DJF: r = -0.259; FMA: r = -0.210; n = 150). These comparisons make differences as well as correspondence visible.
The isotope record concerns mainly summer hydroclimate, Kiso targets cold-season temperature, and the documentary ratios describe local snowfall occurrence. Their complementarity therefore means evidence about different aspects of climate, rather than mutual validation of a single temperature history. Smoothed associations remain descriptive. The revised manuscript and Supplement will document the source series, units, overlapping periods, and filtering choices.
7. Minor comments
Line 164: The replacement wording is “Assigning SSS windows whose center years fall within a candidate interval”. Its role remains a non-decisional comparison; the primary replication rule is unchanged.
Line 350: We checked the revised source and rendered text; the wording is “central Japan”, without a spurious hyphen.
Lines 415-416: The internal “v5.3” identifier has been removed from the main-text and Supplement references to Table 3. Versioned provenance remains in the analytical archive.
Lines 473-475: The dynamically generated access-date suffix in the Bunn et al. reference has been removed. The publication year remains 2025 and the CRAN URL is retained.
Figure 3(c): We have redrawn the panel from the unchanged archived verification statistics, increased the upper plotting space, and separated annotations from the bars (Fig. A2). The previously clipped and overlapping labels are now visible.
We also identified and clarified mixed filtering conventions in the older external-context comparisons. Some archived G11/G21 labels denote arithmetic moving averages rather than Gaussian filters; the revised descriptions will distinguish these from formal Gaussian verification and the new consistently filtered comparison in Fig. A4. In addition, the Methods audit requested by Referee 2 revealed a mismatch in the earlier synthetic-test baseline. We have corrected and rerun that diagnostic; Table A2 and our RC2 reply document the correction. This does not change the chronology used for calibration or the Sawara-exclusion results above.
References added or used in this response
- Begum et al. (2016), Annals of Botany, 117, 465-477. https://doi.org/10.1093/aob/mcv181
- Saitoh et al. (2010), Journal of Plant Research, 123, 473-483. https://doi.org/10.1007/s10265-009-0308-7
- Mayr et al. (2006), Ecology, 87, 3175-3185. https://doi.org/10.1890/0012-9658(2006)87[3175:FDICAT]2.0.CO;2
- Schaberg et al. (2008), Global Change Biology, 14, 1282-1293. https://doi.org/10.1111/j.1365-2486.2008.01577.x
- Nakatsuka et al. (2020), Climate of the Past, 16, 2153-2172. https://doi.org/10.5194/cp-16-2153-2020
- Nakatsuka et al. (2020), archived oxygen-isotope data, NOAA/NCEI. https://doi.org/10.25921/crxm-hs43
-
AC1: 'Reply on RC1', Hitoshi Yonenobu, 28 Sep 2026
-
RC2: 'Comment on egusphere-2026-4506', Anonymous Referee #2, 15 Sep 2026
General comments
This study attempts to reconstruct previous November–April (pNov–Apr) temperature over nearly two millennia in central Japan using TRW proxy. The topic is interesting and valuable. I also appreciate the authors’ careful treatment of reconstruction skill, frequency dependence, signal retention, and uncertainty. The manuscript is transparent about what the statistical tests can and cannot support.
My main concern is whether the current evidence is sufficient to support that the full chronology since 259 CE can be used to do a reliable pNov–Apr temperature reconstruction. The calibration and verification results show that the chronology contains some cold-season temperature information during the instrumental period, but these tests cover only about the last century. It is not clear whether the TRW–temperature relationship remained stable over the preceding ~1700 years. This issue is particularly important because the long chronology combines material from different sites, source histories, and, to some extent, species, whereas the calibration is based on the modern part of the chronology. And also the adjusted R² is only 0.165, which is rather low for a TRW-based temperature reconstruction.
Therefore, I think the main issue is not the rigor of the tests themselves, but whether the available tests are sufficient to support applying the modern TRW–temperature relationship to the full pre-instrumental period. I encourage the authors to address this point more directly and, where possible, provide additional tests of the stability of the chronology and its climate response. If this cannot be tested with the available data, the interpretation of the pre-instrumental reconstruction may need to be further narrowed.
Specific comments
- The statistical relationship between TRW and pNov–Apr temperature is clear, but the biological basis of this relationship remains unclear. The statement that “the defensible inference is statistical” is reasonable, but a better discussion of the possible mechanism would strengthen the reconstruction. I suggest examining the monthly response of TRW to winter temperature, precipitation, drought as well as growing-season temperature, precipitation, soil moisture, and drought et al. This could help clarify whether the winter signal is direct or partly related to other climate variables or carry-over effects. Possible links through snowpack, soil temperature, frost, cambial reactivation, or stored resources could also be discussed.
- Stability of the TRW–temperature relationship is the main concern for the long reconstruction. The current calibration and verification tests show skill during the instrumental period, but they do not tell us whether the same relationship existed throughout the pre-instrumental period. Could the authors examine this using other stationarity tests? It would also be useful to compare the climate response of earlier and later parts of the chronology where replication is sufficient. If the pre-instrumental relationship cannot be tested directly, I suggest making a clearer distinction between the period with tested reconstruction skill and the much longer period for which the chronology is interpreted as a climate record.
- The chronology combines living trees, archaeological wood, historical timbers, and buried wood from different locations. The living chronology also contains a small Sawara component. This raises questions about whether the climate signal in the earlier chronology is comparable to that in the modern calibration material. I suggest examining correlations among chronologies from different sites and, where possible, between species during their overlapping periods. A sensitivity test excluding the Sawara samples would also be useful if the sample size allows. These analyses could help show whether the long chronology has a common regional signal similar to that used for the instrumental calibration.
- The comparisons with documentary snowfall, SST, and circulation indices could provide useful supporting evidence, especially for the pre-instrumental period. I suggest showing these comparisons more clearly, preferably with figures, and discussing what they add to the reconstruction. At the same time, it would be useful to keep a clear distinction between agreement with independent climate records and true independent validation of the reconstruction.
- Please define how the effective sample size is calculated and explain the choice of neff ≥ 8. Since this criterion determines whether the low-frequency verification can be evaluated, its basis should be stated clearly in the Methods.
- The predefined split periods are a strength of the study, but the choice of E25, L25, E45, L45, E56, and L56 could be explained more clearly. If the data allow, testing a few additional split configurations would help assess how sensitive the RE and CE results are to the choice of calibration and verification periods. For example, E25 and L25 for the period 1974-2023.
- The synthetic input–recovery tests are important for interpreting the reconstruction results. Their design and evaluation should be described in the Methods rather than only in the Results.
Technical corrections
- Please define G21 and G51 before their first use.
- The sentence “A long chronology is therefore not, by itself, a reliable climate reconstruction...” is not necessary to present, and could be removed.
- The paragraph beginning “Ring width and stable isotopes need not respond...” is also not necessary, and can be removed.
- Some text in the legend of Fig. 3c appears to be cut off; please check.
- I suggest to show the statistical significance levels in Fig. 5.
Overall, I appreciate the authors’ careful and transparent approach to reconstruction testing. My main concern is the gap between the period in which reconstruction skill is directly tested and the much longer period over which the reconstruction is interpreted. The current results provide evidence for a cold-season temperature signal during the instrumental period, but it remains unclear how well this relationship applies to the earlier chronology, given its changing sources, sites, and species. I therefore encourage the authors to either add analyses that address the stability of the proxy–climate relationship and chronology composition, or further qualify the interpretation of the pre-instrumental record. In particular, the manuscript would benefit from a clear distinction between what is directly supported by the instrumental-period tests and what is inferred from the extended chronology.
Citation: https://doi.org/10.5194/egusphere-2026-4506-RC2 -
AC2: 'Reply on RC2', Hitoshi Yonenobu, 28 Sep 2026
We thank Referee 2 for identifying the distinction between instrumental-period skill and pre-instrumental applicability as the central issue. We agree that neither successful crossdating, strong chronology replication, nor positive instrumental verification can establish that the temperature response remained unchanged throughout the preceding centuries. We have therefore carried out additional instrumental-period, chronology-composition, and verification-sensitivity analyses, and will narrow the interpretation of the pre-instrumental record accordingly. The accompanying response supplement presents the quantitative results; it is not the revised manuscript.
General comment: scope of the temperature reconstruction
The revised manuscript will explicitly distinguish (i) predictive information directly tested within 1912-2023 from (ii) a temperature-scaled interpretation of the earlier chronology that depends on transfer of the modern relationship. The 259 CE boundary reflects replication and common-signal criteria; it is not the beginning of a period with independently demonstrated temperature skill. Our uncertainty intervals remain conditional on the specified calibration and chronology ensemble and do not encompass all possible changes in climate response, provenance, or ecology.
We will replace the concluding Abstract statement with: “Instrumental-period tests support limited annual and G21 high-pass temperature information. Before the instrumental period, the temperature-scaled chronology provides a conditional interpretation of past variability, assuming transferability of the modern relationship; neither that assumption nor low-frequency temperature amplitudes is independently validated.” Corresponding qualifications will be included in the Methods, limitations, and Conclusions. Low calibration strength and conditional-mean variance compression will also be addressed as detailed in our RC1 response.
1. Monthly responses, other climate variables, and possible mechanisms
We have compared monthly regional temperature responses for 1912-2023 and its non-overlapping 56-year halves (Fig. A5), including both raw and linearly detrended correlations. Months span the preceding January to the current December, covering winter, the growing season, and potential carry-over associations. We also downloaded JMA Takayama monthly mean temperature and precipitation totals and applied the same comparisons (Fig. A7). Only unflagged numeric observations were accepted; flagged or missing months were not interpolated. These station data are additional explanatory diagnostics and do not replace the regional calibration target.
As a further descriptive check, we used fixed November-April and May-September windows and calculated the partial correlation between RWI and the regional pNov-Apr target after adjusting for Takayama winter precipitation, May-September precipitation, and May-September temperature. Across the 107 years with all required observations (1912-2022, with gaps), the unadjusted correlation is 0.440 and the partial correlation is 0.350. After separate linear detrending of all variables, these values are 0.408 and 0.366. Thus, the positive cold-season association is not wholly explained by these selected station covariates. This is not a causal test and does not rule out other moisture effects or collinearity; the 107-year result should not be compared directly with the 112-year calibration as if the samples were identical.
We do not have corresponding site-level soil moisture, soil temperature, snowpack, drought-index, cambial, or stored-carbon observations for a comparable analysis. We have not treated station precipitation as a substitute for these variables. This remaining limitation will be stated explicitly.
The revised biological discussion will consider temperature sensitivity of cambial activity, snowmelt and spring carbon uptake, frost drought, root insulation and freezing risk, and possible resource carry-over. Evidence involving Hinoki (Begum et al., 2016; Saitoh et al., 2010) will be distinguished from analogies in other conifers (Mayr et al., 2006; Schaberg et al., 2008). These pathways remain hypotheses; we do not infer active cambial growth throughout winter or identify one dominant mechanism.
2. Temporal stability of the temperature relationship
For 1912-1967 and 1968-2023, raw pNov-Apr correlations are 0.384 and 0.431, respectively, and fitted slopes are 2.204 and 1.691 °C per RWI unit (Table A1). A paired moving-block bootstrap, resampling independently within each half, gives a late-minus-early slope difference of -0.513 with a 95% percentile interval of -2.242 to 0.694 for five-year blocks (3,000 draws). The corresponding correlation-difference interval is -0.387 to 0.291. Three- and ten-year block sensitivities also include zero. These broad intervals do not demonstrate equivalence or stationary responses.
The detrended correlations are 0.414 in the first half and 0.186 in the second half. This weaker late-period association must be acknowledged even though the raw correlations are similar. Descriptive 31-year moving correlations also vary substantially (Fig. A6a). The revised discussion will therefore avoid claiming uniform instrumental-period sensitivity, much less stability over two millennia.
There is no directly observed pNov-Apr target throughout the pre-instrumental chronology. Comparing its early and late segments against instrumental temperature is consequently impossible. Overlap tests among material groups address continuity of shared growth variation (next response), but they cannot fill this climate-validation gap. We will state that the earlier temperature-scaled record is conditional on an unverified transferability assumption.
3. Different sites, material sources, and species
We used the same primary standardized individual-series indices to calculate disjoint source/location and material-group chronologies with Tukey biweight means. Comparison required at least five contributing series in each group in each paired year and at least 30 paired years. The 20IK bridge component assigned to D was omitted from these pairwise diagnostics so that its shared provenance was not treated as independent evidence. This omission concerns the diagnostic comparisons, not the production chronology.
The living cohorts A-B, A-C, and B-C correlate at 0.914, 0.900, and 0.861, respectively. Earlier links are also positive: living cohort C versus archaeological group D gives r = 0.506 (89 paired years, 1555-1643); D versus historical building group E gives r = 0.451 (116 years, 1057-1172); and E versus buried-wood group F gives r = 0.553 (58 available paired years within 610-670). Detrended correlations are 0.560, 0.451, and 0.538 for these three earlier links (Fig. A6b). Non-evaluable pairs are shown as NE in Fig. A6b and retained in the analytical CSV table rather than interpreted as zero correlations.
Among 171 possible source/location pairs, 85 meet the same replication and overlap criteria. Their raw correlations range from 0.216 to 0.874, with median 0.627. The identifiers include forest sites and timber-source or archaeological groups; they are not all known growth locations. These results concern an already crossdated collection, so they are evidence of shared growth variation within that collection, not independent tests of dating or temperature specificity.
Living Hinoki and Sawara group chronologies correlate at r = 0.643 over 280 qualifying years within 1744-2023 (detrended r = 0.625). Their separately aggregated instrumental temperature correlations differ (Hinoki: 0.433; Sawara: 0.274; 1912-2023). Differences in species, site composition, sample size, and tree history are confounded, so these are not estimates of an isolated species effect.
The complete Sawara-exclusion experiment is described in our RC1 reply and Fig. A1. Removal of all 22 Sawara series changes calibration r from 0.4156 to 0.4057 and preserves the positive results in the same six annual and four evaluable high-pass splits. Greater sensitivity in the recent low-replication tail is reported explicitly. Neither these tests nor the overlap correlations establish stable temperature sensitivity of all earlier material.
The input audit also identified three unobserved widths in two buried-wood series (one in 437 CE and two in 512-513 CE). These are omitted before the existing spline fit, which treats retained observations as equally spaced; the calendar labels of the resulting indices are preserved. As a separate sensitivity check, we excluded both entire series under the same standardization. The resulting chronology correlates at r = 0.9987 with the primary chronology over 259-2025; changes are confined to 340-689, with a maximum absolute difference of 0.0978 RWI. The instrumental predictor and calibration are exactly unchanged. This removal test does not quantify or bound the effect of a calendar-aware spline correction. We will explicitly document this limited gap-handling issue; no widths were imputed and the primary chronology and uncertainty ensemble remain unchanged.
4. External comparisons and independent validation
We have prepared the time-series comparison in Fig. A4: the archived central-Japan oxygen-isotope series over 259-2005 and Kawanishi snowfall ratios over 1831-1980 are displayed alongside Kiso with native units, aligned time axes, and consistent complete-window Gaussian smoothing. We will retain and clarify the existing documentary, SST, and circulation diagnostics in the manuscript Supplement and direct readers to them explicitly.
These comparisons offer limited contextual support, not independent quantitative validation of pre-instrumental pNov-Apr temperatures. The annual isotope correlation is weak (0.098), and documentary DJF/FMA correlations are modest and negative (-0.259/-0.210). The documentary overlap partly predates instrumental calibration, but concerns snowfall occurrence at another location. SST and circulation associations address related climate variability and likewise do not validate the reconstruction back to 259 CE. No subperiod or lag was selected to optimize the new comparison. The revised discussion will make these distinctions explicit.
5. Effective sample size and the threshold of eight
The implemented approximation is n_eff = n(1 - r1_obs × r1_pred)/(1 + r1_obs × r1_pred), where n is the number of retained verification pairs and r1_obs and r1_pred are lag-one correlations of the observed and predicted series at the evaluated scale. Lag-one correlations use consecutive values after incomplete transformed edges have been removed. The result is bounded to [2, n]. This is an AR(1) approximation associated with Bretherton et al. (1999), not an exact count of independent observations. The implementation returns n if an AR1 coefficient is undefined and rejects a non-positive denominator; such fallback cases do not determine the reported low-pass conclusion.
Calibration and verification are filtered independently. In addition to n_eff >= 8 in verification, each transformed segment must contain at least eight paired values. The threshold eight is a predefined pragmatic minimum, not a universal significance cutoff and not a threshold prescribed by Bretherton et al. We will state its status explicitly in Methods.
We tested effective-sample-size thresholds of 4, 6, 8, 10, and 12 while retaining the eight-observation minimum in each segment. For the seven original splits, all choices yield the same evaluability counts: seven annual, four G21 high-pass, zero G21 low-pass, and zero G51 low-pass. For the count-eligible G21 low-pass cases, n_eff is only 2.00-2.05. Thus, the low-pass conclusion does not depend narrowly on choosing eight. The revised Methods will give the equation, implementation, separate count criterion, and sensitivity result.
6. Split definitions and additional configurations
The original 25-year pair compares 1912-1936 with 1937-1961 in both directions; the 45-year pair compares 1912-1956 with 1957-2001; and the 56-year pair compares 1912-1967 with 1968-2023. These represent a legacy-equivalent interval, extension through the pre-modern-tail endpoint, and the full instrumental record, respectively. A seventh test calibrates on 1912-2001 and verifies on the low-replication 2002-2023 tail. We will spell out the years and rationale rather than rely on the E/L abbreviations.
As requested, we added the 1974-1998 versus 1999-2023 comparison in both directions. Forward RE and CEff are 0.400 and 0.038; reverse values are 0.333 and 0.036. We also shifted the 45-year pair to 1934-1978 versus 1979-2023; forward RE/CEff are 0.195/0.075 and reverse values 0.409/0.301 (Table A1). All four additional annual tests are positive, although the recent 25-year CEff values are small. We will identify these explicitly as review-stage sensitivity tests and retain all seven original tests, including the unfavorable one.
The 25-year segments leave only five complete centers after a 21-year centered filter, so their frequency-dependent tests remain non-evaluable under the eight-observation criterion. The additional 45-year high-pass tests are evaluable and positive; low-pass tests remain non-evaluable. Additional annual success therefore does not establish low-frequency predictive skill.
7. Synthetic input-recovery design and an audit correction
We agree that the design belongs in Methods. Auditing the earlier test revealed that its baseline covered 482-2025 and did not reproduce the final ABCDEF chronology: the legacy input reader interpreted the F cluster code as a logical value and excluded that buried-wood component. We have therefore rerun the test using the final input loader and verified that the unperturbed baseline reproduces each of the four frozen production chronology variants to numerical precision (Table A2).
For each variant, raw ring widths were multiplied by exp[0.05 × s(t)], where s(t) is a unit-amplitude sinusoid on a continuous elapsed-year axis. The transition from 1 BCE to 1 CE advances by one year, consistent with the absence of year zero in the historical labels. Periods are 11, 21, 31, 51, 75, 101, 150, and 200 years, each with phase fractions 0, 0.25, 0.5, and 0.75 of a cycle. The original standardization and aggregation were then repeated. The recovered signal is the injected-minus-unperturbed chronology divided by 0.05. Recovery correlation and the regression slope of recovered on imposed signal quantify shape and gain; the archived lag diagnostic searches -10 to +10 years for the largest absolute correlation. Results are summarized across the four phases separately for each variant.
The corrected primary 51-year test gives median recovery r = 0.974, median gain = 0.919, and median lag = 0 years. These replace the earlier values r = 0.972 and gain = 0.909. The correction concerns the diagnostic test, not the production chronology, temperature calibration, or annual verification.
The revised Methods and numerical summaries will use this corrected, reproducible experiment. It tests processing of a common imposed growth signal, not whether real temperature sensitivity was stable across source groups or centuries. It neither repairs low calibration strength nor independently validates historical temperature amplitudes.
Technical corrections
G21 and G51 will be defined at first use as centered Gaussian smooths with 21- and 51-year support. High-pass means the annual series minus its corresponding smooth. The unnecessary introductory sentence beginning “A long chronology is therefore not...” and the paragraph beginning “Ring width and stable isotopes need not respond...” will be removed; the necessary limitations will be stated specifically where their evidence is discussed.
Figure 3(c) has been redrawn to correct clipping and overlap (Fig. A2). For Fig. 5, we have prepared raw and detrended maps with significance markers (Figs. A8 and A3). Approximate two-sided correlation tests use the lag-one effective sample size above and a minimum n_eff of eight. Spatial lag-one correlations use only pairs separated by exactly one calendar year, without joining across missing years. Degrees of freedom are n_eff - 2 for raw correlations and n_eff - 3 after removing linear time as a shared covariate. We apply the Benjamini-Yekutieli adjustment separately within each displayed panel and exclude Takayama's self-correlation. Crosses mark adjusted p < 0.05. Table A3 reports both pointwise and adjusted counts. Because these are approximate AR1 tests, especially limited for trending raw series, we do not interpret the markers as proof of field-level significance, exclusively interannual response, or independent climate validation.
We appreciate the referee's emphasis on the calibration-to-history gap. The additional analyses delimit the evidence more clearly; they do not eliminate that gap. The revised interpretation will explicitly retain it as a central limitation.
References and additional data
- Bretherton et al. (1999), Journal of Climate, 12, 1990-2009. https://doi.org/10.1175/1520-0442(1999)012<1990:TENOSD>2.0.CO;2
- Benjamini and Yekutieli (2001), The Annals of Statistics, 29, 1165-1188. https://doi.org/10.1214/aos/1013699998
- JMA, Takayama monthly temperature and precipitation tables, accessed 28 September 2026: https://www.data.jma.go.jp/stats/etrn/view/monthly_s3.php?prec_no=52&block_no=47617&view=a1 and https://www.data.jma.go.jp/stats/etrn/view/monthly_s3.php?prec_no=52&block_no=47617&view=p5
- Physiological and isotope references are listed in our RC1 reply and the accompanying response supplement.
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 243 | 64 | 35 | 342 | 44 | 26 | 21 |
- HTML: 243
- PDF: 64
- XML: 35
- Total: 342
- Supplement: 44
- BibTeX: 26
- EndNote: 21
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
In this manuscript, the authors present a two-millennia cold-season temperature reconstruction for central Japan based on an extensive Kiso Chamaecyparis tree-ring width chronology. Combining wood from diverse sources and addressing the urban heat island (UHI) effect through careful target selection are strong methodological points. While the manuscript is well presented and results can provide valuable insights into the cold-season temperatures in central Japan, I has a few concerns regarding the methods, results, and interpretation.
Specific concerns are detailed below:
The primary calibration model yields an adjusted R2 of 0.165 (r = 0.416) for the period 1912–2023. While the cross-validation statistics (RE and CEff) are positive and validate the model’s predictive skill, an explained variance of ~16.5% is relatively low for a temperature reconstruction, implying that over 80% of the variance in the tree-ring chronology is driven by unexplained factors. The authors should expand the discussion on what might be driving this high proportion of unexplained variance and clearly state the implications on interpreting the amplitude of historical temperature anomalies.
In Section 2.1 (Lines 89-91), the authors state that 22 out of 338 living-tree series are Chamaecyparis pisifera (Sawara), which were combined with Chamaecyparis obtusa (Hinoki). The authors acknowledge that “species-specific climate responses were not estimated and this remains a limitation”. If the ratio of these two species changes significantly over the two-millennia chronology, it could introduce non-stationarity into the climate response. The authors should provide a stronger justification for why mixing them does not impact the long-term climatic signal.
The selection of the previous November–April target is well justified statistically. However, the biological discussion in Section 4.1 (Lines 274-279) is somewhat brief. Given that cambial activity in this region typically occurs in the warmer months, the authors should explicitly discuss the ecophysiological mechanisms through which cold-season temperature strongly impact the subsequent ring width in Kiso Chamaecyparis (e.g., specific references to winter desiccation, snowmelt infiltration, or root freezing).
In Section 3.6 and Table 3, the rule-based screening identifies the 1999–2015 CE interval as a sustained upper-tail feature. Because this period falls entirely within the modern instrumental era, its value as a paleoclimatic finding is limited. The authors should further clarify the utility of highlighting this interval, is it primarily to demonstrate that the proxy successfully captures the modern warming trend, or does it serve another purpose?
In Section 4.4 (Lines 343-344), the authors state that detrending reduces the mean station correlation from 0.496 to 0.305, indicating that shared slower variations (likely the modern warming trend) contribute significantly to the raw spatial footprint shown in Figure 5. To prove that the proxy truly captures a coherent regional climate signal at the interannual scale, rather than a global warming trend, the authors should provide a supplementary figure showing the detrended spatial correlation map.
In Sections 4.5 and 4.6, the manuscript discusses the reconstruction’s complementarity with documentary snowfall records and the central-Japan isotope hydroclimate record (Nakatsuka et al., 2020). However, this comparison is entirely textual. Given the paleoclimate context of central Japan, adding a time-series figure that visually compares this new cold-season temperature reconstruction with the documentary and isotope records (for the overlapping periods) would substantially elevate the paper’s impact and make the discussion much more convincing.
Minor comments:
Line 164: The phrasing “Center-year-within-interval assignment” is hard to read. Consider rephrasing (e.g., “Assigning the center year within an interval...”).
Line 350: Typo. “downstream central- Japan” should be “downstream central Japan”.
Line 415-416: The text mentions “documents the complete Table 3 v5.3 candidate”. The “v5.3” appears to be an internal versioning artifact. Please remove or correct it.
Lines 473-475: In the reference list for Bunn et al. (2025) regarding the dplR package, there is a typo at the end of the citation: last accessed: July 29, 2026, 2025. The presence of two years (“2026, 2025”) appears to be an artifact from citation management software. Please correct this.
Figure 3: In Figure 3(c), the text above the bars (specifically “primary 0.127” and “4/4 positive” for the G21 HP category) are partially clipped by the top border of the plot frame. Additionally, the label “carry-over 0.049” overlaps with the bars, making it difficult to read. Please adjust the formatting.