The study area of this paper consists of two northcentral Yukon valleys, approximately 120 km to the north-east of Dawson City, around Distincta Peak (1790 m a.s.l.), the dominant point in the subrange of mountains. These valleys will be referred to as WS01 and WS02 (for weather station 1 and 2), and correspond to valleys labelled as SVWS and NVWS, respectively, in . Those sites are easily accessible by road, being located in the immediate vicinity of the Dempster Highway. They were shown to be prone to intense, frequent, and long-lasting inversion events .

The two valley floors are separated by approximately 8 km, with WS01 to the south of Distincta Peak, and WS02 to the north. The former is an approximately 5.5 km-long V-shaped valley, with elevations ranging from 950 m at the outlet and 1700 m on the closest highest ridge. The latter is an approximately 6.5 km-long U-shaped valley, with elevations comprised between 1040–1790 m. The valleys' outlet is oriented towards the southeast for WS01, and to the north for WS02.

The network of air temperature loggers consists of 24 locations in valley WS01, and 26 in WS02. The locations are chosen to sample the widest range of elevation, slope, aspect angle, and degree of valleyness .

The model calibration of this paper will be tested on the dataset presented in , where five weather stations are distributed within a 25 km radius from the Dawson airport, and with elevations ranging from 370 to 1200 m.

Finally, the temperature inversion model will also be applied to various sites across northern Canada, in the Yukon and Northwest Territories. We selected 13 ECCC/NAVCAN weather stations (Fig. , see Table  for coordinates) because they span a wide spatial area and a variety of topographic conditions, from river delta to mountainous terrain.

Then, we report the distribution of elevations in a radius of 50 km around each station in Fig. , together with all sites from the Dempster and Dawson areas.

The 50 locations scattered across the two valleys of the study area are instrumented with passively-ventilated air temperature loggers. Air temperature data was retrieved at the end of the 2024 summer during fieldwork.

Observational data for the Dawson area is taken from , and was originally provided by the Yukon Geological Survey. Finally, the air temperature data for the Dawson airport and for the 10 locations described in Table  is managed by Environment and Climate Change Canada (ECCC) and is downloaded with the R-package weathercan .

Reanalysis data products are accessible globally for several decades. They are consistent and gapless in both time and space, allowing researchers to study uninstrumented areas or to provide additional data where observations might be available. In this study, we use hourly ERA5 data from the European Centre for Medium-Range Weather Forecasts (ECMWF) since 1940 and 6-hourly JRA-3Q data from the Japan Meteorological Agency (JMA) (downscaled hourly) since 1947. The former is available on a regular 0.25°×0.25° grid in latitude and longitude, while the latter uses a TL479 Gaussian grid instead with a spatial resolution of 0.375°×0.375°. We download data divided into three categories: constant (geopotential, land-sea mask), single level (2 m temperature), and pressure levels (air temperature).

Note that at the latitude of the study area, the reanalysis grid cell approximately corresponds to a 46km×20km cell for ERA5, and 59km×30km for JRA-3Q. We downscale all reanalysis products to the stations' locations and elevations while accounting for local topography using GlobSim. Furthermore, we acquire pressure-level air temperature Tpl for a full column above each valley from the highest station upwards, at each available pressure level. For both products, we also interpolate single-level 2 m surface temperature Tsur. This is now automated with the addition of the new DReaMIT module into GlobSim.

The relationship between hypsometric position and elevation depends on the area from which the hypsometric position is calculated, and the chosen weighting scheme. We find that the best fit is obtained when no weighting is performed (see Appendix ), which simplifies the approach.

We calibrate the model from Eq. () with a subset of the full dataset (see Table ), where a least-squared fitting method is applied. The input data needed to calibrate the model are the hypsometric position on one side, which is calculated from a DEM, and reanalysis data on the other side (we use both ERA5 and JRA-3Q). The best fit parameters and their uncertainty are reported in Table . Note that we also fit the model in the absence of the reanalysis bias correction term β(t). This reduces the model performances but increases its versatility and applicability in certain cases presented in the Discussion (Sect. ).

The temperature inversion model REDCAPP makes explicit use of both hypsometric position h and MRVBF index v. We find that the Akaike and Bayesian information criteria (AIC and BIC, respectively) do not improve when explicitly accounting for v in the model, which motivates the functional form adopted in Eq. (), where the temperature inversions only depend on the hypsometric position h.

We present a comparative analysis on the full northcentral Yukon valleys WS01 and WS02 dataset, which aims to put the performance of the new hypsometric model in context by comparing it to various other models in Fig. . The first three comparisons are essentially variants of the same framework, differing in calibration and the topographic variable used, whereas the hypsometric DReaMIT formulation represents the substantive model development in the last column. The first model (“Dawson”) is the inversion model proposed in , with their calibrated parameters. The second column, labelled “Absolute elevation”, uses the same model but with parameters specifically calibrated for the combined WS01 and WS02 valleys. The third column uses relative elevation with respect to valley bottom instead of absolute elevation, and the final column is the hypsometric model of this study. The “Dawson” model and parameters perform very poorly in this study area with a coefficient of correlation of 0.17 for ERA5 but even as low as -0.60 for JRA-3Q, confirming what was already known, i.e. it is a model that needs to be calibrated locally, which is done in the second column with much better agreement (R2=0.76 and 0.73 respectively). However, there is a clear distinction between the valleys, with values from WS01 above those of WS02. This is even more visible when looking at relative elevation, where the decoupling between the two valleys becomes so large that the model is clearly inappropriate. However, when turning to the more intricate variable transformation of elevation that hypsometry is (fourth column), we see that both valleys align, and a single model can now be applied confidently in different geographical locations. The correlation coefficient of the mean effective α over the full dataset with respect to the modelled α is equal to R2=0.93 in the case of ERA5, and 0.87 for JRA-3Q. This visually demonstrates why absolute and relative elevation do not fully spatialize the relationship, while hypsometry largely collapses the two valleys onto a common functional form.

Finally, we also perform single-valley calibration, and demonstrate the stability of the results when applied to the other valley in Appendix . This constitutes a first step towards demonstrating the spatialization power of the model.

The new surface-based inversion model proposed in this study spatializes the model of while making the calculation of the REDCAPP lapse rate dynamic. The factor that controls the strength of the temperature inversion events, α, is modified by considering hypsometric position instead of elevation. With this choice, we successfully unify the functional form of α as a function of a variable transformation of elevation for dissimilar valleys (Fig. ). A single calibration can now be performed simultaneously for the two valleys (WS01 and WS02) of the study area, and the resulting model is reliable in both valleys. To show that the model is indeed robust against spatialization, we perform cross-valley calibrations and find it to be reliable there too (Fig. ).

Another improvement compared to the “Dawson” model is the exponentiation of the elevation variable, instead of the linear form of α. The functional form of the proposed αexp in Eq. () reduces to the linear form of for locations with h≪1, while allowing for the representation of stronger inversions when h≲1, a frequently observed phenomenon.

SBIs in high-latitude valleys often exhibit non-linear lapse-rate profiles , where strong temperature stratification near the surface transitions to weakly stratified or isothermal layers above . This pattern was observed in the subarctic valleys used to calibrate this model . During an SBI, lapse rates typically decayed exponentially within the first roughly 150 m where they often exceeded 25 °Ckm-1 within the lowest 60 m and then approached isothermal above 150 m. In extreme cases, there was as much as 8 °C temperature increase over 19 m between the valley bottom and adjacent slope site .

The exponential function therefore better captures the observed behaviour of mean inversion strength (last column of Fig. ), while being safe from unwanted divergence given the boundedness of the hypsometric position (h∈[0,1]). Nonetheless, the correction is still slightly underestimating the inversion strength for the points with higher hypsometry (last column of Fig. ). For the readers who might be surprised not to find a better fit of the mean effective α, we would like to reiterate that the fit presented there is the full model calibrated with hourly data in both valleys, and not a direct fit of mean yearly effective α.

The model described by Eq. () was tested across a wide geographical range in Yukon and Northwest Territories, encompassing topographic diversity (Fig. ). The hypsometric model is more versatile than any of the reanalysis temperatures, even far from the calibration study area, which proves the success of this new formulation. Model performance in Whitehorse, Haines Junction, and MacMillan Pass, which are locations that experience weaker and more transient inversions than the subarctic valleys along the Dempster Highway , remained surprisingly strong. This suggests that while the model was calibrated for more continental Dempster valleys, it captures the underlying structure of SBIs well, and that its transferability extends to regions with comparable hypsometric ranges. In these settings, the model appears to represent topographic constraints on cold-air pooling processes that develop within landscapes characterized by large hypsometric range. Although the frequency and intensity of SBIs are less at Whitehorse, MacMillan Pass, and Haines Junction, their structure likely resembles those of the Dempster sites.

However, the model's transferability has limits. First of all, the inclusion of a reanalysis bias correction term β creates issues. Indeed, reanalysis bias is not uniform in time (see Fig. 7 of ), but more importantly, it varies in space while the model explicitly assumes a spatially-independent β(t). Such a term increases model performances in a region of applicability, but is not appropriate for applications elsewhere. Instead, for such locations, we recommend either recalibrating β locally if nearby weather stations exist, or using the model calibrated without the β term (right column of Table ), which increases the model's versatility. We showed that model performances are only minimally affected in the absence of reanalysis bias correction, demonstrating that the spatial scale at which the model can be applied is not limited by the scale over which the reanalysis bias varies. Second, the model performed poorly for certain ECCC stations, including Tulita, Old Crow, and Norman Wells. As shown in Fig. , these locations correspond to those with higher hypsometry, while not in a mountain area. The topography for these areas is mainly flat (the elevation distribution peaks around the elevation of the lowland), and hence, any slight elevation variation has a disproportionate effect on the hypsometry. This is a limiting case of the model. Note that for Old Crow and Tulita, the calculated value of h is close to 1, but the stations are nowhere close to being at the bottom of a deep valley. This constitutes an extreme topographic situation where the model might not be suitable. These exceptions highlight the limits of the domain: in extremely flat, river-dominated terrain, hypsometry becomes unstable and river-generated warm anomalies overwhelm the topographic signal the model relies on. In this context, the inclusion of the MRVBF in a more intricate model might mitigate this undesirable effect by better distinguishing flat from incised terrain.

A related limitation arises in generally flat areas with a low range in hypsometry. Attempts to incorporate a modified variable, a bounded version of the hypsometric range h⋅dz, which accounts for the topographic potential to sustain inversions did not substantially improve performance. Similar challenges have been reported in other lapse rate-based downscaling studies, where the absence of pronounced topographic relief often reduces the prediction of elevation-derived parameters . In such environments, microtopographic indices such as MRVBF used in REDCAPP may provide complementary information on the structure of SBIs and resulting elevational patterns of air temperature. This metric, or something similar, could be used as a mechanism, to limit hypsometry closer to 0 in these flat areas.

Our results indicate a clear domain limit: stations situated on low-relief floodplains on the leeward side of mountain ranges (e.g. Norman Wells, Tulita, Old Crow) exhibit degraded performance relative to incised valleys. There, the dominant controls on near-surface temperature are not primarily terrain-controlled cold-air pooling. In these locations, river influence, floodplain energy balance, weak relief, and possible reanalysis/grid elevation artefacts may dominate over the hypsometric signal. In such settings, the gravitational pooling that sustains surface-based inversions (SBIs) is weak because slopes and basin geometry are minimal; small uncertainties in hypsometry are therefore amplified, especially when the seasonal temperature contrast DT is large. Leeward positioning further reduces inversion persistence via mechanically-driven mixing and episodic flushing , while riverine processes (late freeze-up, open water, advective heat) can imprint non-orographic thermal anomalies during transition seasons . Together, these factors yield a near-surface temperature structure that is not primarily topography-controlled, and our hypsometry-based correction can over-adjust. We therefore recommend caution for leeward floodplain sites, particularly those proximal to major rivers. In this configuration, the better-performing air temperature is Tsur. This framing clarifies that the model performs as intended in incised, terrain-controlled valleys, while transferability is limited where floodplain and river energy budgets dominate boundary-layer thermodynamics.

We focused some effort into determining the best radius around each point location to compute the hypsometric position, and concluded that even though there might exist intermediate radii for specific configurations, the most robust option was to consider a radius of 50 km, with no weighting. Recent studies have shown that the interplay between large-scale synoptic conditions and local radiative and cold-air drainage processes is scale-dependent, with larger valley systems having a more complex interaction between the two mechanisms . This suggests that the radius that performs best empirically (50 km) is of the order of the physical scale at which local drainage and synoptic influences control SBI strength, providing a physical rationale for the effective use scale for hypsometric position used in our model.

We showed the existence of a hierarchy of scales, with a local minimum in model performance reached for local valley scales (roughly 5 km radius for the study area), but a global and potentially location-independent best fit at larger scales (Appendix ). However, the physical scales controlling cold-air pooling and inversion persistence in the western Canadian Arctic may differ from those in other regions, and thus the relationships should be interpreted cautiously. Moreover, considering larger areas in the hypsometry calculation is yet to be tested for various topographic conditions across the western Canadian Arctic.

In subarctic environments, SBIs are sustained over periods of days to weeks rather than hours. During these events, cold-air pooling reflects not only local slope drainage (acting over scales between 1–10 km) but also the cumulative redistribution of cold air across interconnected valley systems. Consequently, the relevant topographic control extends beyond the immediate valley and may be better represented by landscape-scale hypsometric position than by local relief alone. We have used a radius arising from the data available and satisfying these approximate characteristic scales, while potentially interacting with the effective resolution of the reanalysis grid cells.