A compiled dataset of Greenland weather station observations and reconstructed surface air temperature

Wang, Shujing; Wang, Yetang; Zhang, Yulun; Zhai, Zhaosheng; Hou, Shugui

doi:10.5194/essd-2026-449

Preprints

https://doi.org/10.5194/essd-2026-449

Preprints

02 Sep 2026

| 02 Sep 2026

Status: this preprint is currently under review for the journal ESSD.

A compiled dataset of Greenland weather station observations and reconstructed surface air temperature

Shujing Wang, Yetang Wang, Yulun Zhang, Zhaosheng Zhai, and Shugui Hou

Abstract. Climate variability is a main driver of mass imbalance of the Greenland Ice Sheet. However, quantitative assessments of climate change in this region are constrained by the sparseness, heterogeneity, and discontinuity of in situ measurements, imposing significant challenges to reliable modelling and future projections. To address this, we compile all available weather station observations over Greenland from 1941 to present into a single and standardized dataset at both daily and monthly temporal resolutions. This dataset includes seven climate variables: temperature, wind speed, wind direction, relative humidity, radiation, precipitation, and pressure, which is accessible at https://doi.org/10.11888/Atmos.tpdc.303490 (Wang and Wang, 2026). All data were harmonized and processed by a uniform quality control procedure, with the outliers flagged. For each station, detailed metadata covering source, location, and data availability are supplied. The dataset facilitates a thorough analysis of multi-variable climate characteristics in Greenland and can be used for the validation of global reanalysis products, regional climate models and remote sensing retrievals as well as data assimilation and climate reconstruction. Furthermore, we use the temperature observations from the compiled dataset together with a digital elevation model and ERA5 reanalysis to reconstruct a high-resolution (0.1°) monthly surface air temperature (SAT) dataset for Greenland (1950–2024), based on the optimized method identified after comparing five distinct machine learning models. The resulting SAT reconstruction exhibit a great improvement over ERA5, reducing the root mean square error (RMSE) by more than 60 % and increasing squared correlation coefficient (R²) by 5 %. The high accuracy of this reconstruction is further confirmed by spatial cross-validation across 11 distinct regional divisions. The reconstruction is also publicly available at the Third Pole Environment Data Center (https://doi.org/10.11888/Cryos.tpdc.303026, Wang and Wang, 2025).

Received: 09 Jun 2026 – Discussion started: 02 Sep 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Shujing Wang, Yetang Wang, Yulun Zhang, Zhaosheng Zhai, and Shugui Hou

Status: final response (author comments only)

RC1:
'Comment on essd-2026-449', Anonymous Referee #1, 05 Sep 2026
Summary

The paper presents a new compilation of Greenland near-surface observations from various stations and uses these to construct a monthly surface-air-temperature reconstruction spanning from 1950 to 2024 on a 0.1° grid. The authors harmonize and compile the rather fragmented observations into a single data set.

The authors train several machine learning models using the station data and ERA5-Land to reconstruct the Greenland-wide temperature. They report substantially lower errors than ERA5 and use their reconstructed product to examine spatial climatologies and long-term temperature trends.
The paper is discussing an important and timely topic. The paper is generally easy to follow and the language is mostly clear but needs to be more scientifically accurate in some parts.
However, I have several major issues with the paper that need to be addressed before being considered for publication.
Major issues
I have a major issue with the train/test/validation split. First, I don't understand why you choose a 86/14% split since these numbers seem relatilvely arbitrary. Why not the more common 80/20 split? But more problematic is the random split. Given the nature of the observations, it is inevitable that observations from the same station, adjacent months, and adjacent years will almost certainly occur in both training and testing. This makes the train-validation split not independent, leading to overconfidence in the metrics. It would be better if you split the training/validation data using distinct periods e.g. from 19XX-20XX. In this sense, the data that you call unseen is not really unseen but very likely highly correlated to the training data.

The distinction between the validation and test sets is unclear and should be clarified. For the LSTM-CNN and Transformer-CNN, the data are split into training (86%), validation (4%), and test (10%) sets, whereas the other three models use only a training (86%) and test (14%) split. However, the manuscript subsequently states that all five models underwent iterative training and tuning until predefined performance criteria were reached. It is therefore unclear how the three models without a validation set were tuned.

If the reported test set was inspected or used to guide hyperparameter tuning, model development, or model selection, it no longer constitutes an independent test set and should instead be considered a validation set. A genuinely unseen test set should only be evaluated once model development and model selection are complete.

Relatedly, the manuscript refers to the combined validation and test data for the LSTM-CNN and Transformer-CNN as "unseen data." This terminology is misleading because the validation set is explicitly used for early stopping and therefore contributes to model development. Validation and test performance should be reported separately.

I therefore recommend that the authors clearly describe the role of each data subset and ensure that model tuning and model comparison are performed using training/validation data (or cross-validation), followed by evaluation on a separate, untouched test set.

If I understand correctly, you use the random split for finding the best performing ML model, here the ANN, and then only afterwards do a spatial cross-validation for only the ANN. However, given the flaw in the random split validation, it could be that another model/method shows better performance in the spatial testing but you never test this.

In Figure 11, you compare the ANN results against the ERA5 data. The label is "train data". I am not sure if this is an error because you should show the performance of the ANN on the test data not train data. Otherwise, the RMSE of the ANN results has no real value but seems to be one of the main results of the paper.

In section 5.2.4 I am not sure I can follow the procedure. You have 11 ANN models that were trained on different regions (which are in my opinion not mutually independent by the way since only one region was always excluded, so the other regions are shared across ANNs, right? 9 out of 11 regions are always shared with each ANN, making them highly correlated with each other). But then you input the whole data into each of the 11 ANNs and take the mean of all models as a result? In each case 10 out of 11 of the ANNs have been trained on the held-out region. The correct procedure would be to use only the unseen prediction of each respective ANN for constructing the data set.

The standard deviation is not really the model uncertainty.

ERA5 T2M is already highly correlated to the SAT (r=0.98), so I wonder how much do you even gain from using an ANN? Is a simple linear regression not enough? I strongly suggest the authors include a linear regression as a baseline and see how much improvement their ML models even give.

Why did you chose the cosine monthly encoding? This function maps different months to the same value, e.g., f(m=4)=0 and f(m=10)=0 as well.

I suggest the authors include error statistics by decade. Sicne the data is very sparse in the beginning, I wonder if the model performs well in recent time but not very well in earlier times.

Why is the difference between the reconstructed anomaly and ERA5 anomaly a bias? It's just a difference (Fig.12).

Specific comments:

L22 Shouldn't it be "a method" not "the method" since you didn't talk about it before

L23 I suggest you avoid terms like "great improvement"

L.25 missing "the"

L.35 There are more recent references than the Lenton papers, e.g. Boers et al. 2025, Bochow et al. 2023, Global Tipping Point Report 2025

L.52 I guess the "at" should be an "and"

L.53 Specifically instead of "specially"
I am missing some discussion on more recent temperature reconstructions from incomplete spatial data. There is quite some advancement in the last years using CNNs and generative methods. Some recent references are Kadow et al. 2020, Bochow et al. 2023, Qian et al. 2026
L.120 how did you merge the information?

L. 133f Do you actually aggregate precipitation as average or sum?
Section 5.1

It is not completely clear to me if you use ERA5 or ERA5-Land.
L. 332 I find it difficult to merge the term validation and testing data here. The validation data is indirectly involved in the training since you tune the model parameters using the validation set. I would suggest the authors keep the distinction between test and validation data clear throughout the manuscript.

L.340 Refers to a previous major comment. I wonder how and if the other variables besides T2M even improve the SAT reconstruction or if T2M is just enough, did the authors test this?

L. 363 Here you make the distinction between validation and test data again but one sentence later, you refer to unseen data again.

L.366 How did you get your definition of "good"?

L.558 The numbers should be in the exponent
Table 2

Why do you exclude radiation values =0 Wm-2?

Figures
I suggest the authors extend the figure captions.

Fig. 12 It is not clear to me from the text how the anomalies are calculated and what the temporal reference is.
Citation: https://doi.org/10.5194/essd-2026-449-RC1
RC2:
'Comment on essd-2026-449', Anonymous Referee #2, 29 Sep 2026
Summary
This manuscript compiles and harmonizes Greenland near-surface air temperature observations from multiple station networks and uses them to develop a monthly 0.1° temperature reconstruction for 1950-2024. The study also analyzes the spatial climatology and long-term temperature variability of Greenland using the reconstructed product. Overall, the dataset has clear potential value, and the manuscript addresses a relevant topic. However, I have several questions about the reconstruction methodology and evaluation that I think should be clarified to strengthen the paper.
General comments:
I am a bit confused by the data splitting strategy. LSTM-CNN and Transformer-CNN use training, validation, and testing sets, while RandomForest, XGBoost, and ANN only use training and testing sets. Why are the models treated differently here? ANN, for example, can also use a validation set and early stopping. More importantly, later in the manuscript the authors say that the models were repeatedly trained and tuned, and that the best-performing model on unseen data was selected. What exactly is meant by unseen data here? If the test set was checked during tuning or used to select the final model, then it is no longer an independent testing set. Please explain clearly what data were used for training, tuning, early stopping, model selection, and final evaluation.

The manuscript says that 22 predictors are used, but I could not understand how these variables are arranged for the LSTM-CNN and Transformer-CNN. My understanding is that the dataset may simply be an N 22 table, where each row is one station-month sample and the 22 columns are predictors. If that is correct, then what is the time dimension for the LSTM and Transformer? And what exactly is the 1 D CNN convolving over? Please give the actual input shape for each model and explain how the sequences are constructed. For example, how many time steps are used, whether the 22 variables are present at each time step. If there is no explicit temporal sequence, then I do not really see the advantage of using LSTM or Transformer here. Likewise, if the CNN is operating across the 22 predictor variables, the physical meaning of local convolution across neighboring features is not obvious.

The five models seem to use quite different training strategies. LSTM-CNN and Transformer-CNN use different early stopping rules, ANN is trained with a fixed architecture and number of epochs, and the tree-based models appear to use fixed parameter settings. In that case, how can we know whether the differences in performance come from the architecture itself or simply from the way each model was tuned? I strongly suggest adding one table with the full training setup for every model: number of layers, hidden units, number of filters, kernel size, dropout, batch size, learning rate, optimizer, loss function, number of epochs, early stopping, and the main tree parameters. I am also curious why 250 neurons were chosen for the ANN, why the patience is 10 epochs for LSTM-CNN but 20 for Transformer-CNN, and whether any systematic parameter search was performed for RandomForest and XGBoost. These models can be quite sensitive to hyperparameter choices. Also, why were LSTM-CNN and Transformer-CNN chosen rather than comparing simpler LSTM-only or CNN-only models first? Without these simpler baselines, it is difficult to know which part of the hybrid architecture is actually helping.

For Table 3, what exactly do the authors mean by correlation? Is this Pearson correlation? If so, I do not think a low linear correlation is enough reason to remove a predictor such as longitude. These are nonlinear models, so a variable can still be useful through nonlinear effects or interactions with other predictors even if its individual linear correlation is small. I am also not sure how these correlations were calculated. Elevation, latitude, and longitude are static, while temperature, pressure, and wind vary in time. Another point is that ERA5-Land 2 m temperature is already extremely strongly correlated with observed SAT. I would therefore really like to see a simple ablation test. For example: T2m only, T2m+geographic information, T2m+the other meteorological variables, and the full 22-variable setup. Otherwise it is difficult to know how much information the additional predictors actually add. Also, why is only the cosine of month used? A cosine term alone does not uniquely represent all months. Usually both sine and cosine are used for cyclic month encoding.

At one point the manuscript says that each model was trained independently ten times and that the average result was used for comparison. Later it says that all five models went through ten rounds of iterative training and tuning until they reached the required performance threshold. These sound like two different things. Were these ten independent runs with exactly the same hyperparameters, or did the authors change the model settings between runs? If the best run was chosen based on test performance, then that would again mean that the testing set was used for model selection.

The final reconstruction is a Greenland-wide 0.1° temperature field, but the training data are station observations. I think the manuscript needs to explain this transition much more clearly. How were ERA5-Land and DEM values extracted at each station during training? Was nearest-neighbor sampling used, bilinear interpolation, or something else? And after training, is the model simply applied independently to every 0.1° grid cell using the 22 local predictor values, after which all predicted grid cells are assembled into a map? If this is the case, please state it explicitly and give the actual input and output dimensions during reconstruction. This point matters because, if prediction is performed independently at each grid cell, then the ANN itself does not learn spatial relationships between neighboring grid cells.

The DEM is available at 500 m but is resampled to 0.1° to match ERA5-Land. I understand why the final temperature product is produced at 0.1°, but I am not sure why all of the finer topographic information has to be discarded. In Greenland, especially near the margins, elevation varies strongly within a 0.1° grid cell. Could the authors have used the high-resolution DEM more effectively, for example by including sub-grid elevation range, slope, or other terrain statistics rather than reducing everything to one coarse elevation value?

Finally, are the reported metrics calculated by pooling all stations and all months together? If so, the strong seasonal cycle may contribute a lot to the very high R². I would like to see performance by station or region, and ideally also for temperature anomalies after removing the seasonal cycle.
Citation: https://doi.org/10.5194/essd-2026-449-RC2

Shujing Wang, Yetang Wang, Yulun Zhang, Zhaosheng Zhai, and Shugui Hou

Data sets

A Greenland weather station observation dataset Shujing Wang and Yetang Wang https://doi.org/10.11888/Atmos.tpdc.303490

Monthly SAT at the resolution of 0.1°×0.1° across Greenland during 1950–2024 Shujing Wang and Yetang Wang https://doi.org/10.11888/Cryos.tpdc.303026

Model code and software

SAT_reconstruction_code-v1.0 Shujing Wang and Yetang Wang https://doi.org/10.5281/zenodo.21714043

Code-Related-to-Weather-Station-Observation-Datasets-v1.0 Shujing Wang and Yetang Wang https://doi.org/10.5281/zenodo.21713966

Shujing Wang, Yetang Wang, Yulun Zhang, Zhaosheng Zhai, and Shugui Hou

Viewed

Total article views: 557 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
307	187	63	557	79	64

HTML: 307
PDF: 187
XML: 63
Total: 557
BibTeX: 79
EndNote: 64

Views and downloads (calculated since 02 Sep 2026)

Month	HTML	PDF	XML	Total
Sep 2026	256	138	41	435
Oct 2026	51	49	22	122

Cumulative views and downloads (calculated since 02 Sep 2026)

Month	HTML	PDF	XML	Total
Sep 2026	256	138	41	435
Oct 2026	51	49	22	122

Viewed (geographical distribution)

Total article views: 557 (including HTML, PDF, and XML) Thereof 557 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 12 Oct 2026

Short summary

Greenland climate observations are sparse, heterogeneous, and fragmented. We compiled and standardized all available weather station records across Greenland from 1941 to the present, providing seven key climate variables and detailed station metadata. We also reconstructed a high-resolution monthly surface air temperature dataset for 1950–2024. These datasets provide valuable resources for climate research, model evaluation, and climate reconstruction.


Total:	0
HTML:	0
PDF:	0
XML:	0