{"id":"fa4ca34a-4732-4a5f-b5ef-1ab46b31302d","arxiv_id":"2507.15045","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Continuous two-slope fits to ERA5 land temperature data find statistically significant trend changes in about half of global land area, with many changes clustered around 1980.","lead":"A team at the Max Planck Institute for the Physics of Complex Systems fit two connected straight lines to 72 years of ERA5 land temperature data at every global grid point to estimate when local warming trends changed. They find that about half of the land area shows a significant shift, with many regions accelerating their warming around the 1980s.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1980±10 change-point signal is never checked against ERA5's known 1979 observing-system discontinuity, so the headline timing may be a reanalysis artifact rather than a climate signal.","rationale":"The reader's model-misspecification worry is real and is explicitly acknowledged by the authors (Fig. 10), but I do not think it is the most load-bearing: a smooth or multi-regime series would still place a two-segment break near the period of fastest acceleration, so the qualitative 'changes around 1980' could survive. The reanalysis discontinuity is more directly fatal: the headline year is in the immediate vicinity of the 1979 change in ERA5's observing system, and the paper's stated reason for keeping 1950 data is to include the very change point it detects. This is a confound that can be resolved by an independent station-based check. I therefore keep the verdict CONDITIONAL, but the condition should be a robustness analysis against the 1979 assimilation break, in addition to the multiple-testing and uncertainty-map revisions the reader requested.","tokens_in":16189,"tokens_out":9072,"duration_ms":107477,"concrete_test":"Re-run the full dual-linear analysis on an independent 1° land grid from station-based products (e.g., Berkeley Earth or CRUTEM5) over 1950-2021, and compare the change-year histogram (Fig. 8), the 50% land-area fraction, and spatial maps to the ERA5 results. If the station-based histogram reproduces the 1980±10 peak within the paper's ±8 year detection error, the ERA5 observing system is not the driver; if the peak shifts, weakens, or disappears, the headline timing is contaminated. As a secondary check, fit the model to ERA5 with an explicit break indicator at 1979 and test whether the pre-1979 residuals show the same structure.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that many land grid points switch to a stronger warming trend around 1980±10. ERA5's assimilation undergoes a major discontinuity around 1979 when satellite radiances enter, which can imprint spurious breaks on local temperature series. The paper explicitly names this risk in Sec. VI ('safer to start this analysis with the year 1979') but nevertheless analyzes 1950-2021, justifying the choice by the fact that the global-mean ERA5 series itself shows a change point in the 1970s (Sec. IV). That justification is circular if the same discontinuity is what creates the change point. No independent station-based validation is provided, so the timing peak in Fig. 8 cannot be separated from a data-processing break. This is a correctness risk, not a disagreement with consensus: real 1970s climate shifts exist, but reanalysis fields in data-sparse regions are exactly where the paper detects many of its early/late change points. If this concern lands, the abstract's '1980±10' and the tipping-point question lose their evidential basis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies a continuous two-slope (piecewise-linear with a single change point) least-squares model to annual ERA5 2-meter temperature time series from 1950 to 2021 on roughly 20,000 land grid points. The two linear segments are constrained to merge at the change point, and the change year is selected by minimizing the RMS error over a grid of candidate years (1960-2010). The model is compared against a single linear trend using BIC, AIC, and a null-hypothesis test based on the absolute slope difference. The authors report that about half of the global land area shows a statistically significant trend change, produce global maps of the detected change years and segment slopes, and find that many change points cluster around 1980±10, which they connect to the question of whether climate tipping points have been passed.","tokens_in":16318,"tokens_out":3405,"duration_ms":40420,"significance":"If the central empirical claim holds, the paper provides a useful global inventory of local warming-trend accelerations and decelerations, with maps that could inform regional climate adaptation. The method has some genuine strengths: the constrained least-squares solution is derived in closed form in Appendix A, and the synthetic experiments in Section III give an honest assessment of the detection uncertainty (for realistic noise, 80% of change points are within ±8 years). The comparison of BIC and null-hypothesis testing on synthetic data is also a good practice. However, the headline result (the 1980±10 concentration of change points and the '50% of land area' claim) is currently not robustly established, because the analysis starts in 1950 despite the well-known 1979 satellite-assimilation discontinuity in ERA5, because no multiple-testing correction is applied to the ~20,000 per-grid tests, because the change-year and slope maps carry no uncertainty information, and because the two-segment model is acknowledged to be misspecified in at least some regions.","major_comments":[{"comment":"The decision to analyze the full 1950-2021 period is justified in Section VI by the statement that 'the global mean surface temperature shows a change point in the 7th decade,' but this justification is circular if that global-mean change point is itself produced by the well-known 1979 satellite-data discontinuity in ERA5. The abstract's headline '1980 ±10' and the histogram in Fig. 8 peak in exactly the same period. Because no independent station-based validation is provided, the central empirical claim cannot be separated from a reanalysis artifact. Please repeat the analysis for the 1979-2021 subset, and validate the detected change years and slope changes against station-based gridded products such as Berkeley Earth or CRUTEM, or otherwise demonstrate that the spatial pattern of detected change points is not aligned with the spatial distribution of observation density changes.","section":"Section VI and Section IV"},{"comment":"The claim that 'the temperature in approximately 50% of the global land area has experienced a statistically significant change in trend over these years (in terms of 95% confidence)' is based on per-grid-point tests applied to roughly 20,000 series. Under the null hypothesis of no change, about 5% of tests would be expected to reject by chance even if no grid point had a real change; spatial correlation makes the effective number of independent tests smaller but still large. No multiple-testing correction (e.g., false discovery rate) or field-significance test is reported. The 50% figure and the 'many grid points' language in the abstract therefore overstate the evidence. Please provide a multiple-testing-adjusted assessment of the fraction of land area with significant trend changes.","section":"Section V"},{"comment":"The paper itself acknowledges model misspecification: Section VI and Fig. 10 show a Siberia example in which two comparable RMS minima exist and a three-segment fit is 'more appropriate,' and Section III shows that, under a no-change-point null, the procedure produces spurious change points preferentially near the series ends. This means that the detected 'change year' and the 'acceleration around 1980' summary can be artifacts of fitting a single breakpoint to a smoothly accelerating or nonlinear temperature evolution, rather than evidence of a discrete regime shift. The paper does not compare the two-segment model against a smooth alternative (e.g., a quadratic trend or LOESS) or against multi-change-point models for the full grid. Please add such a diagnostic, at least for representative regions, to show that the detected change points are not simply the piecewise-linear approximation to a continuous acceleration.","section":"Section VI and Section III"},{"comment":"The maps in Figs. 7 and 9 and the histogram in Fig. 8 present point estimates without any uncertainty quantification, even though Section III shows that the change year is estimated with an 80% interval of ±8 years for realistic noise levels and Section IV explicitly provides formulas (Eqs. 3-4) for trend uncertainties in the presence of short- and long-range correlations. The year-to-year fluctuations in Fig. 8 are described as 'much larger than statistical estimation errors,' but no confidence bands are shown. Without uncertainty measures, the reader cannot distinguish real spatial-temporal structure from estimation noise. Please provide per-grid confidence intervals or masks showing where the change year is well constrained, and add error bands to the cumulative distribution in Fig. 8.","section":"Section V and Section III"}],"minor_comments":[{"comment":"There are several typos: 'indentify' in the abstract, 'annomalies' and 'temperture' in Section I, and 'veryfied' and 'fector' in Section II. Please correct these.","section":"Abstract and Section I"},{"comment":"The notation in Eq. (3) is not fully defined in the text: the symbol d appears in Eq. (3) and is related to H later, but the text says 'd = H - 1/2'; Eq. (4) uses d in the hypergeometric function without explaining the parameter range or the role of ϕ. A brief definition of all symbols directly after Eq. (3) would improve readability.","section":"Section IV, Eqs. (3)-(4)"},{"comment":"In the paragraph after Fig. 6, the sentence 'more than 44% of them are located in Antarctica' is ambiguous because 'them' could refer to grid points where the single-linear model is preferred or to the total land area; please rephrase.","section":"Section V"},{"comment":"The manuscript does not state whether the code and the processed change-point maps will be made available, nor does it specify the exact ERA5 data version and download date beyond the CDS reference. Given that the paper is a data-driven statistical analysis, a data and code availability statement would strengthen reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is an honest, clearly written application of a standard method to a public dataset, and the resulting maps are a useful addition to the regional climate literature. The 1980±10 acceleration is probably a real feature of the data, but the paper overstates it—no multiple-testing correction across ~20,000 grid points, no uncertainty on the maps, and the tipping-point framing outruns what a two-segment fit can show.\n\nWhat's new: a systematic global gridpoint change-point analysis of ERA5 land temperature with a continuous two-slope model, giving maps of change years and before/after slopes. That hasn't been done at this scale for ERA5. The synthetic calibration is informative—the ±8 year detection error is a useful number—and the authors are unusually candid about model misspecification, showing the Siberia case with two comparable minima and openly discussing the reanalysis pre-satellite issue. The least-squares derivation in the appendix is straightforward and reproducible with the public ERA5 data.\n\nThe soft spots are real but not fatal. The significance claim of ~50% of land area is per-grid with no adjustment for the ~20,000 simultaneous tests; their own synthetic results imply a 12% false-positive rate at H=0.65, which alone would light up thousands of grid points. The change-year and slope maps carry no uncertainty, even though the detection error is known to be ±8 years or worse. And the stress-test concern about the 1979 observing-system discontinuity deserves a direct answer: the paper acknowledges the risk but doesn't test it, and the justification via the global mean is partly circular. The NOAA global series showing a similar change point offers some cover, but that doesn't extend to local grid points in data-sparse regions. A sensitivity run starting in 1979, or a comparison of detected change years against a few hundred station series, would settle it.\n\nWho should read this: anyone working on regional warming trends or using reanalysis for change-point analysis. It deserves a serious referee—the method is clean and the results are potentially useful—but it needs revision on the statistics and a more careful discussion of what the data can and cannot say about tipping points.","headline":"Honest and useful map of regional trend changes, but the headline claims need multiple-testing control and a direct check of the reanalysis discontinuity.","tokens_in":16922,"tokens_out":3196,"would_cite":true,"duration_ms":33703,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P12"],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-slope model with one change point describes local land temperature evolution better than a single trend on about half of Earth's land, with the change clustering near 1980.","keywords":["change point detection","ERA5","warming trend","two-slope fit","model selection","BIC","local climate change","tipping points"],"falsifier":"Re-run the same grid-cell fit on ERA5 data restricted to 1979 onward; if the cluster of change points around 1980 disappears or shifts, the reported timing is an artifact of combining pre-satellite and satellite-era data.","tokens_in":15901,"feed_emoji":"🌡️","tokens_out":8897,"duration_ms":81876,"temperature":0.7,"pith_summary":"The paper asks whether local land temperature records from 1950 to 2021 are better described by one straight trend or by two straight trends that meet at a change year. Using ERA5 2-meter annual temperatures on a 1-degree grid, it fits a continuous two-slope model at every land point and compares it with a single linear fit using Akaike's and Bayesian information criteria plus a significance test. It finds that for roughly half of global land area, the two-slope model wins, and detected change points cluster around 1980 plus or minus 10 years, with many locations showing stronger warming after the change and some showing slowdown. If correct, this means single-trend summaries of local warming hide a widespread acceleration, and it raises the question of whether some local climates have already passed a tipping point.","feed_headline":"Half of Earth's land shifted warming pace around 1980","feed_subtitle":"Fitting two merging trend lines to 72 years of local temperatures shows where warming sped up and where it slowed.","key_machinery":"The central object is the continuous two-slope model: two linear segments constrained to meet at a single change point, so the model captures a change in warming trend rather than a jump in temperature. At each candidate change year the slopes and intercepts are obtained analytically by a constrained least-squares fit, and the selected change year is the one minimizing the root-mean-square error. Model selection between the dual-linear and single-linear descriptions is carried out with the Bayesian information criterion, with Akaike's criterion checked as an alternative, and a null-hypothesis test based on the absolute difference of the two slopes provides a 95% confidence decision.","core_discovery":"The paper argues that a continuous dual-linear fit, two straight lines that merge at a single change year, is the parsimonious description of local 2-meter temperature evolution on ERA5 land grid points from 1950 to 2021. Comparing this model with a single-trend fit by BIC and a 95% confidence hypothesis test, it finds that for roughly half of the global land area, area-weighted, the two-slope model is preferred. For the majority of those grid points the change year falls between about 1970 and 1990, clustering near 1980, and the fitted second slopes are mostly positive and often larger than the first, indicating accelerated warming in many regions while some regions slow down or cool. The same two-slope fit applied to global mean land temperature finds a change around 1976 to 1980, from slight cooling to strong warming.","pith_inferences":["The paper's own Siberia example implies that a three-segment model may be the better description there; extending the BIC comparison to three or more continuous segments would show how many apparent 1980 change points are actually two separate changes, such as one in the 1970s and one in the 1990s.","If the acceleration around 1980 is real, projections that anchor a single historical trend will tend to underestimate near-future local warming in the affected regions; that extrapolation is not made in the paper.","A checkable prediction following from the paper is that grid points with a larger second slope should mostly continue on that steeper slope in years after 2021 unless another change point appears."],"forward_implications":["About half of the global land area is better described by a two-slope fit than by a single linear trend, so one-trend descriptions of local warming are incomplete for those regions.","Detected change years lie mostly between 1970 and 1990, with a concentration near 1980, making those years candidate past tipping events in the paper's interpretation.","In many regions, especially northern land masses, the second slope is larger than the first, meaning local warming accelerated; in some regions the second slope is smaller or negative, meaning warming slowed or reversed.","The global mean land temperature series itself shows a change around 1976 to 1980, from a slightly negative to a strongly positive trend, consistent with the local picture."],"supporting_citations":[{"why":"Supplies the ERA5 2-meter temperature time series on land grid points from 1950 to 2021, the dataset the analysis is run on.","marker":"[14]"},{"why":"Describes how ERA5 is constructed by data assimilation, which the paper relies on for treating the reanalysis as a consistent long-term record.","marker":"[15]"},{"why":"Defines the Bayesian information criterion used to decide between the single-line and dual-line models.","marker":"[33]"},{"why":"Defines Akaike's information criterion, the alternative model-selection criterion the paper compares against BIC.","marker":"[1]"},{"why":"Provides the variance formula for trend estimates under short- and long-range correlations, used for error bars and the significance threshold.","marker":"[30]"},{"why":"Gives the lag-1 autocorrelation and long-range dependence estimates used to calibrate realistic noise in the significance tests.","marker":"[17]"}],"fun_headline_variants":["Half of Earth's land shifts warming pace around 1980","Two-slope fit finds global warming shift around 1980","Warming trends change for half of land areas by 1980","ERA5 reveals widespread 1980s shift in land warming"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each 72-year local temperature series is adequately described by exactly two straight-line segments that meet at one change year.","fun_headline_variants_meta":{"raw":{"variants":["Half of Earth's land shifts warming pace around 1980","Two-slope fit finds global warming shift around 1980","Warming trends change for half of land areas by 1980","ERA5 reveals widespread 1980s shift in land warming"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00048,"raw_usage":{"total_tokens":2362,"prompt_tokens":922,"completion_tokens":1440,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":1368}},"tokens_in":538,"tokens_out":1440,"duration_ms":14273,"temperature":1.0,"reasoning_tokens":1368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:42:34.701022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same grid-cell fit on ERA5 data restricted to 1979 onward; if the cluster of change points around 1980 disappears or shifts, the reported timing is an artifact of combining pre-satellite and satellite-era data.","supporting_citations":[{"cited_title":"The family of era5 datasets","cited_arxiv_id":null,"evidence_quote":"Supplies the ERA5 2-meter temperature time series on land grid points from 1950 to 2021, the dataset the analysis is run on."},{"cited_title":"Hersbach, B","cited_arxiv_id":null,"evidence_quote":"Describes how ERA5 is constructed by data assimilation, which the paper relies on for treating the reanalysis as a consistent long-term record."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Bayesian information criterion used to decide between the single-line and dual-line models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines Akaike's information criterion, the alternative model-selection criterion the paper compares against BIC."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the variance formula for trend estimates under short- and long-range correlations, used for error bars and the significance threshold."},{"cited_title":"Kassel and H","cited_arxiv_id":null,"evidence_quote":"Gives the lag-1 autocorrelation and long-range dependence estimates used to calibrate realistic noise in the significance tests."}],"review_version":1}