Northern Chilean Patagonia is located in the south of Chile and represents the northernmost part of Chilean Patagonia (41–56° S). It spans from 41.4 to 46.7° S and from 72 to 74° W (Fig. 1). It is delimited to the east by the continent, to the north-west by the Island of Chiloé, and to the south-west by the Chonos Archipelago, composed of more than a thousand islands and playing a crucial role in reducing wind stress and mitigating waves intensity before they reach the mainland (Pinilla et al., 2019a). To the west of the sea, the continental shelf drops sharply, with depths increasing rapidly to over 2000 m due to the presence of the Peru-Chile Trench. The Northern Patagonian Fjords encompass diverse basins, channels and fjords. From north to south, the key areas include the large semi-enclosed gulf of Reloncaví Sound and the Reloncaví Fjord, which flows into the sound. To the south of Reloncaví Sound lies the Ancud Gulf and the Corcovado Gulf, separated by the Desertores Islands, forming a natural barrier separating the deeper waters of Reloncaví Sound and Ancud Gulf from the shallower Corcovado Gulf. Ancud and Corcovado Gulfs are regions more influenced by oceanic waters entering through the Guafo Mouth (Calvete and Sobarzo, 2011; Pérez-Santos et al., 2014; Linford et al., 2023), with some fjords extending eastward into the continent. South of the Corcovado Gulf lies the narrow canal of Moraleda, which transitions into the Elefantes Channel, forming an extensive and complex network of channels and fjords (Soto-Riquelme et al., 2023; Linford et al., 2024). This particular geography makes it considered as a semi-enclosed sea, with the main connections to the ocean passing between Chiloé Island and Chonos Archipelago, a path called the Guafo Mouth (Sievers and Silva, 2008). The bathymetry in the Northern Patagonian Fjords is very complex due to its numerous microbasins and islands. Although the bathymetry is mainly shallower than 400 m, certain microbasins can be exceptionally deep, reaching for example 600 m in the Jacaf channel (Schneider et al., 2014; Pinilla et al., 2020).

To construct the climatology, we gathered all available in situ temperature data to obtain a comprehensive dataset that best represents the region both spatially and temporally. The data come from various online repositories and direct field measurements, and were acquired using a range of sampling techniques, including CTD casts, drifting buoys and anchored moorings. Further details on data provenance are provided in Table 1 and in the supplementary material. Only observations located within the study area (75.1° W, 72.0° W and 41.4° S, 46.7° S) were retained, encompassing both surface and depth observations. All datasets were homogenised prior to merging, and variables were standardised to consistent units (°C for temperature and metres for depth). Data selection involved retaining only values flagged as valid when quality indicators were available, which means observations classified by the respective providers as “good”, “valid” or “probably good”, whereas observations flagged as “bad”, “probably bad”, “suspect” or otherwise invalid were excluded. Data considered “obviously wrong” (such as temperatures above 30 °C) were also excluded. In addition, for each major basin, the mean and standard deviation of temperature were calculated from the in situ observations, and any observation lying outside the range of mean ± 5 standard deviations were discarded. This relatively wide tolerance was chosen to preserve realistic extreme values while removing measurements that were physically implausible or likely to result from sensor errors (Garcia et al., 2024; Hoteit et al., 2009). Duplicate or near-duplicate observations were eliminated when the following criteria were simultaneously met: a horizontal distance of less than 0.01° (about 1 km), a vertical distance of less than 0.1 m, a sampling time difference of less than one day (given that many datasets have daily rather than hourly precision), and a temperature difference smaller than ± 0.1 °C.

In total, 3.3 million observations were used for the construction of the climatology, representing 5880 unique stations (Fig. 2A). The complete dataset spans the period from 1948 to September 2023, although spatial, vertical and temporal coverage are uneven (Fig. 2). Vertically, the measurements extend from surface to 400 m. Given that the climatology forms the core of our analysis, it was essential to use a reliable and representative dataset. The extensive spatial and temporal distribution of in situ measurements makes them well-suited for this purpose.

The inherently heterogeneous nature of in situ data results in variable sampling density through time and space. In particular, Comau Fjord was disproportionately sampled due to the long-term monitoring efforts of the Huinay Scientific Centre, which accounts for approximately 76 % of the total samples (Table 2). However, the other fjords and regions have also been relatively well sampled, though they represent a much smaller proportion of the total dataset (Table 2). Additionally, three anchored buoys were deployed in Reloncaví Fjord, Reloncaví Sound and Puyuhuapi Fjord from 2010 onwards (Table 1), contributing significantly to the increased number of observations during the last decade in these basins. As a result, the temporal distribution of observations is strongly skewed towards recent decades: 99 % of the measurements were collected after 1990 and 55 % after 2010 (if considering only non-mooring data). The resulting climatology should therefore be interpreted as a climatological reconstruction based predominantly on recent observations (last three decades), rather than as an equally weighted 75-year climatological mean. Nevertheless, pre-1990 observations, though fewer in number, remain valuable as they include more than 1000 unique measurements that extend the time span of the climatological record back to 1948. Importantly, despite this temporal heterogeneity largely driven by a few intensively monitored sites, the overall spatial coverage remains extensive across the region (Fig. 3). Vertically, sampling is also heterogeneous, with approximately 70 % of all observations located within the upper 25 m of the water column (Fig. 2D).

To create a continuous field of the climatology of sea temperature in Northern Patagonia from sparse in situ data, we employed DIVAnd (Data-Interpolating Variational Analysis in n dimensions). DIVAnd is a spatio-temporal interpolation tool for geophysical variables that utilises scattered, spatially and temporally irregularly distributed data to create a continuous field.

via variational interpolation, taking the bathymetry into account (Troupin et al., 2010; Barth et al., 2014). DIVAnd is an extension of DIVA (Data Interpolating Variational Analysis) for fields of more than 2 dimensions (Beckers et al., 2014; Brasseur and Haus, 1991; Troupin et al., 2012). This tool has been widely used for the construction of climatologies of many oceanic variables, from sparse data to continuous fields (Barth et al., 2021; Shahzadi et al., 2021a, b; Simoncelli et al., 2020). It is particularly well-suited for our study area since it can handle both spatial, temporal and vertical dimensions, allowing the interpolation of data in 2D (longitude × latitude), 3D (time or depth) or 4D (time and depth). The method is based on a variational inverse approach that minimises a cost function balancing two objectives: the analysed field should remain close to the available observations and it should exhibit a degree of spatial smoothness (Brasseur and Haus, 1991). Unlike a pure interpolation, DIVAnd does not force the resulting field to pass exactly through each observation, as measurements can be subject to errors and may not fully represent the environmental conditions (Janjić et al., 2018). The cost function is detailed below in Eq. (1): 1J(φ)=∑j=1Ndμjdj-φxj2+φ-φb2 where φ is a scalar ocean field defined in our case as the temperature, and dj are the Nd measurements of the field φ at the location xj for a given time, with their associated weight μj. φb is a background estimate of the field, representing a first guess of the field to be interpolated. The norm .2 also involves spatial/temporal derivatives of the field to penalise abrupt variations. The computation was performed using DIVAnd v2.7.12 available as a package in the Julia programming language and in conjunction with Jupyter Notebooks (https://github.com/gher-uliege/DIVAnd.jl, last access: 25 July 2022). The variable to be interpolate in this study was the temperature, with the objective of building a 4D climatology. The required input consists of a dataset of potentially irregularly distributed observations, with associated coordinates (longitude and latitude), depth and date.

To generate a fine-scale spatio-temporal climatology from in situ surface and subsurface data, we first computed a background field, which is a first guess representation of the temperature distribution that served as the basis for the subsequent monthly climatology (Barth et al., 2021). This background field was produced at a seasonal temporal resolution by grouping months into four standard seasons: January–February–March (austral summer), April–May–June (autumn), July–August–September (winter), October–November–December (spring). For each season, only the in situ observations corresponding to the relevant months were used. This approach mitigates the uneven temporal coverage and spatial gaps in the dataset, particularly in regions and periods with limited sampling, and provides a representative mean seasonal state of the temperature field. The result is an averaged representation of the study area over the time period 1948–2023 with seasonal temporal resolution.

The monthly climatology was then derived by refining the background fields. For each month, the corresponding seasonal grid was used as the background field, and all available in situ observations from that month were incorporated to enhance spatial and temporal detail. This two-step approach ensured that the monthly climatology remained dynamically and statistically consistent with broader seasonal patterns while resolving finer-scale monthly variability. The outcome is a set of twelve fields, corresponding to a monthly climatology (monthly temporal resolution), each representing the mean temperature state for every grid cell of the study area for the period 1948–2023. Both the seasonal background and the monthly climatology share identical vertical and horizontal resolutions and were processed following the same methodological framework.

For both products (background and climatology), the configuration of DIVAnd was designed to balance data fidelity and spatial smoothness, ensuring a robust and physically consistent representation of the temperature field. To reduce the influence of spatially clustered measurements, the DIVAnd weighting function was applied with a 0.1° (∼ 10 km) threshold for cluster identification. The interpolation smoothness was tuned using correlation length scales of 50 km horizontally and 50 m vertically, which define the spatial influence of the observations. The noise-to-signal ratio, representing the error variance relative to the background error variance, was set to 0.01 and scaled by the previously assigned weights. Parameter values were selected to optimise performance metrics obtained from independent validation tests. The final gridded fields were produced with a 0.008° (approximately 900 m) horizontal resolution, reaching our computational limits, representing a horizontal grid of size 357×751, covering the Northern Chilean Patagonian fjords between 75.1 to 72° W and 47 to 41° S. Bathymetric constraints were derived from GEBCO dataset which has a spatial resolution of 30 arcsec (https://www.gebco.net/data_and_products/gridded_bathymetry_data/, last access: 6 December 2021). As approximately 70 % of the observations were collected within the upper 25 m, climatologies were computed on a non-uniform vertical grid to reduce near-surface bias. Temperature was interpolated onto 32 unevenly spaced depth levels ranging from the surface to 400 m, with finer spacing near the surface (0, 2.5, 5, 7.5, 10, 12.5, 15, 17.5, 20, 25, 30, 35, 40, 45, 50, 60, 70, 80, 90, 100, 125, 150, 175, 200, 225, 250, 275, 300, 325, 350, 375, 400 m). The DIVAnd interpolation resulted in a monthly climatology at 32 different depths with a resolution of approximately 900 m, effectively resolving all fjords and channels from the northern Patagonia.

To assess the performance of the monthly climatology and validate it, 5 % of the in situ observations were put aside from the reconstruction process and used for independent validation (see Table 3 and Tables S1–S3 for details). This procedure was performed separately for four validation depths: the surface, 30, 100 and 200 m. The validation was carried out over depth intervals centred on each target depth rather than at the exact depth itself, in order to retain a sufficient number of independent observations: 0–1 m for the surface, 25–35 m for 30 m, 90–110 m for 100 m, and 180–220 m for 200 m. These intervals are narrower near the surface, where temperature varies more strongly with depth, and wider at depth, where the vertical temperature gradient is weaker and observations are sparser. At each depth interval, the withheld observations were used to evaluate the reconstructed temperatures independently of the data used for interpolation. The subsets were selected randomly across space and time to ensure representative coverage of the study area. However, for the anchored moorings, random temporal sequences of varying lengths were withheld instead of isolated points, in order to provide a more robust validation of the reconstructed temporal variability. Validation was performed separately for each monthly field by comparing the reconstructed and observed temperatures. Statistical performance was quantified using the correlation, the bias, the root mean square error (RMSE), the centred root mean square error (CRMSE), and the standard deviation. Together these metrics evaluate both the strength of the relationship and the magnitude of residual errors between independent observations and the reconstructed field (climatology). Validation was not extended below 200 m because the number of available in situ observations decreases substantially with depth. Withholding 5 % of the observations at these depths would have further reduced the already sparse data available for reconstruction, potentially compromising the interpolation. Moreover, the limited number of independent observations would have made the resulting validation statistics less representative and dependable. To further assess the spatial consistency of the reconstruction and account for the uneven distribution of observations across the study area, regional validation was performed for the surface layer. The same statistical metrics as described above were calculated separately for five main regions: Reloncaví Sound, Ancud-Corcovado Gulfs, Comau Fjord, Chonos Archipelago-Moraleda region, and the Puyuhuapi-Aysén Fjords.

DIVAnd also provides a relative analysis error field associated with the interpolated temperature, which quantifies the remaining uncertainty of the analysis relative to the background error variance allowing the user to assess the spatial and vertical confidence of the climatology (Beckers et al., 2014). This error field is adjusted according to the weighting function and the noise-to-signal ratio chosen prior to the interpolation in DIVAnd. The relative error ranges from 0 to 1, with lower values indicating that the analysis is more strongly constrained by the observations resulting in high confidence, while higher values indicate greater remaining uncertainty, meaning that the analysis relies more heavily on the background field (i.e. first guess) and is less constrained by the observations. The relative analysis error is influenced by the spatial and temporal distribution of the observations, their associated uncertainties, the specific correlation length scales, and the geometry of the interpolation domain, including physical boundaries. It therefore provides an additional measure of the spatial and vertical observational constraint of the reconstructed climatology.

Many oceanographic and ecological applications require daily climatologies to capture short-term physical or biological variability more accurately (e.g. Hemming et al., 2020; Hobday et al., 2016). We therefore derived a daily climatology from the monthly fields described in the previous section. Because the number of observations decreases with depth, the daily-derived climatology was limited to the upper 100 m on the same 14 vertical levels as the monthly climatology (i.e. 0, 5, 10, 15, 20, 25, 30, 40, 50, 60, 70, 80, 90 and 100 m). This choice avoids extending the temporal interpolation to deeper layers where observational support is more limited. To convert the monthly climatology into a daily time-scale product, each monthly field was expanded over the corresponding number of days in that month (for instance, the January field was repeated 31 times, February 29 times, March 31 times, etc.) This produced a continuous sequence of 366 daily fields with the same spatial characteristics as the monthly fields. To smooth the resulting daily time series and ensure a physically realistic and gradual transition between months, a 90 d centred moving average was then applied to each grid cell. This temporal filter mitigates abrupt discontinuities at month boundaries and yields a continuous, smoothed representation of the annual temperature cycle. The choice of a 90 d window balances temporal smoothness and the preservation of the seasonal signal, consistent with similar approaches in constructing high-frequency climatologies.

The resulting dataset (later referred to as the “daily-derived climatology” in the manuscript) provides one representative value per day of the year for each grid cell, ensuring internal consistency with the monthly product while allowing finer temporal analyses and comparison with daily observational or model data. This approach preserves the large-scale seasonal and monthly variability of the original climatology, without attempting to reproduce intra-monthly variability that is not resolved by the available in situ data in most of the basins.

Both climatology fields are distributed as NetCDF3-Classic files using geographical coordinates referenced to the WGS84 coordinate reference system. The files are available online through the repository indicated in the Data Availability section.

To validate the monthly climatology, 5 % of the in situ data was set aside and excluded from the reconstruction process at four targeted depths (see Methods). These independent subsets were then used to evaluate the accuracy of the monthly climatology through statistical comparisons. Overall, the results indicate that the monthly climatology reproduces the spatial and seasonal variability of surface temperature with reasonable accuracy (Tables 4 and S4–S6 in the Supplement).

In the surface layer, biases between the interpolated field and the observations are small throughout the year (ranging from -0.139 to +0.214 °C) and RMSE values rarely exceed 1 °C, suggesting a robust representation of the mean state. Correlation coefficients exceed 0.7 from June to September, when surface waters are the coldest and show less variation, while lower correlations and higher errors are found in summer months, reflecting stronger small-scale variability and stratification. January and April display the lowest correlations (0.159 and 0.171, respectively), consistent with the limited number of available in situ observations (99 and 161 respectively, compared to an average of 485 per month). The CRMSE is close to the RMSE for all months, indicating that most of the errors arise from spatial variability rather than from a systematic bias. The standard deviation of the climatology is consistently smaller than that of the in situ data, meaning that the reconstructed fields are smoother than the original observations, which is an expected feature for a climatological product.

When comparing the interpolation errors with the natural variability of the system, RMSE values are generally smaller than the standard deviation of the in situ data during the austral summer, indicating that the climatology successfully captures the large-scale temperature patterns. In contrast, during winter months, RMSE slightly exceeds the standard deviation of the observations, showing that even small absolute errors can represent a significant fraction of the reduced thermal variability typical of this period. Since RMSE quantifies the difference between the monthly climatology and instantaneous in situ measurements, it also encompasses sub-seasonal and interannual variability. Consequently, even a perfect climatology derived from an unlimited number of observations would not achieve a zero RMSE or a correlation of one.

To assess whether the domain-wide validation results were representative of the different hydrographic environments and not disproportionately influenced by the uneven spatial distribution of observations, the same metrics were additionally calculated separately for the main regional basins. The regional validation further indicates that the reconstruction performance is broadly consistent across basins. RMSE values are higher in the Ancud-Corcovado and Puyuhuapi-Aysén regions in January (1.81 and 2 °C respectively) and February, but remain below 1 °C during the other months, following a pattern similar to the domain-wide statistics. The larger errors in the Puyuhuapi-Aysén region are consistent with the higher temperature variability of that region (standard deviation exceeding 2 °C in summer). The Reloncaví region shows a comparable seasonal pattern, with RMSE values below 1 °C from April to November and peak errors in January. In the Chonos-Moraleda region, RMSE values remain below 1 °C throughout the year, although these relatively low errors should be considered in relation to the comparatively low standard deviation of the observations in this region. Finally, Comau Fjord, accounting for the largest proportion of observations, shows relatively low RMSE values throughout the year (< 1 °C). Overall, the reconstruction errors remain consistent with the magnitude of natural temperature variability across the different regional environments, confirming that the good performance observed for the entire domain is also broadly maintained at the regional scale.

The subsurface validation also indicates good reconstruction performance (Table S4). At 30 m depth, correlation coefficients are generally higher than at the surface, ranging from 0.628 in February to 0.955 in September. This February minimum coincides with the largest positive bias (+0.254 °C) and the largest RMSE (0.441 °C). Biases remain small throughout the year (from -0.1 °C in June to +0.254 in February), and RMSE and CRMSE values remain below 0.5 °C. Contrary to surface, RMSE and CRMSE show on certain months a larger difference between RMSE and CRMSE. The standard deviation of the reconstructed climatology can be higher than the standard deviation of the in situ data used for the reconstruction, which contrasts with the surface behaviour, where the climatology was systematically smoother than the observations. It suggests that the reconstruction can show higher variability than the in situ observations. However, the RMSE values remain smaller than the standard deviations, indicating that the climatology captures a substantial fraction of the subsurface signal.

At 100 m depth, the validation statistics indicate that the reconstruction remains accurate, although with a different signature than at the surface (Table S5). Correlation is consistently high throughout the year, ranging from 0.855 in October to 0.996 in January, with a single notable exception in April, when the correlation drops to 0.537. It is important to note that both January and April coincide with few available in situ observations at this depth (respectively 155 and 197 observations, see Table S2). Biases are small and predominantly negative (from -0.055 °C in January to +0.016 °C in April), indicating a slight cold bias of the climatology relative to the observations.

As for the upper layers, CRMSE and RMSE show near-identical values, both remaining below 0.3 °C in all months and reaching a minimum RMSE in September. The standard deviation of the reconstructed climatology is always lower than that of the in situ observations, except in January, indicating that the subsurface fields are also smoother than the original observations, consistent with the expected behaviour of a climatology. The RMSE is also always lower than the standard deviation of the in situ data, showing that the climatology captures a large fraction of the subsurface temperature signal.

At 200 m depth, the validation results indicate generally good agreement between the reconstructed climatology and the independent observations, although data availability limits the calculation of validation statistics for some months. No reliable statistics could be obtained for January and April because, despite the availability of in situ observations at these depths, few of them coincided with locations where the climatology provided interpolated values. For the remaining months, correlation coefficients exceed 0.86, reaching 1 in May (however this result should be interpreted with caution because only five observations were available for validation). Biases remain small, ranging from -0.093 to 0.123 °C, and RMSE values remain below 0.35 °C. The standard deviation of the reconstructed fields is generally lower than, or comparable to, that of the in situ observations, suggesting that the reconstructed fields are smoother than the observed temperature variability. RMSE values are consistently lower than the standard deviation of the in situ observations, indicating that the reconstruction errors are smaller than the observed temperature variability.

Overall, the validation at the four targeted depths confirms that the generated climatology is consistent with the available observations. The errors remain of the same order as, or smaller than, the natural variability of the system, and the RMSE-CRMSE comparison indicates that spatial variability, rather than systematic bias, dominates the error budget at the surface, 100 and 200 m, whereas a small systematic component is present at 30 m in some months. The largest errors found at 30 m depth might be explained by the vertical influence of the highly variable upper layers. Lower performances are also observed during months of limited independent observations.

Altogether, the climatology provides a reliable baseline for representing surface temperature variability in Northern Chilean Patagonia from the surface to 200 m depth (greater depths were not analysed due to lack of independent observations). The complete climatology (both monthly and daily-derived) is freely available at 10.5281/zenodo.21959707 (Pujol et al., 2026) with two companion codes (MATLAB and Jupyter Notebook Python 3) for easy data handling and visualisation.