the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
EUPollMap: The European atlas of contemporary pollen distribution maps derived from an integrated Kriging interpolation approach
Gregoire Mariethoz
Manuel Chevalier
Abstract. Modern and fossil pollen data are widely used in paleoenvironmental research to characterise past environmental changes in a given location. However, their discrete and discontinuous nature can limit the inferences that can be made from them. In contrasts, deriving continuous spatial maps of the pollen presence from pointbased datasets would enable more robust regional characterization of such past changes. To address this problem, we propose a comprehensive collection of European pollen presence maps including 194 pollen taxa derived from the interpolation of pollen data from the Eurasian Modern Pollen Database (EMPD v2) restricted to the EuroMediterranean Basin. To do so, we developed an automatic Krigingbased interpolation workflow to select an optimal geostatistical model describing the spatial variability for each taxon. The output of the interpolation model consists in a series of multivariate predictive maps of Europe at 25km scale, showing the occurrence probability of pollen taxa, the predicted presence based on diverse probability thresholds, and the interpolation uncertainty for each taxon. Visual inspections of the maps and systematic crossvalidation tests showed that the ensemble of predictions is reliable even in datascarce regions, with a relatively low uncertainty, and robust to complex and nonstationary pollen distributions. The maps, freely distributed as GeoTIFF files, are proposed as a readytouse tool for spatial paleoenvironmental characterization. Since the interpolation model only uses the coordinates of the observation to spatialise the data, similar maps could also be derived for fossil pollen records, thus enabling the spatial characterization of past changes, and possibly, their subsequent use for quantitative paleoclimate reconstructions.
 Preprint
(6078 KB)  Metadata XML

Supplement
(15655 KB)  BibTeX
 EndNote
Fabio Oriani et al.
Status: final response (author comments only)

RC1: 'Comment on essd2022364', Anonymous Referee #1, 13 Sep 2023
This paper introduces a Kriging interpolation method aimed at generating comprehensive pollen presence maps for the EuroMediterranean Basin. The subject matter of this paper holds considerable interest and is undoubtedly of value to the journal's readership. The paper is generally wellstructured, providing all necessary information effectively. I particularly commend the authors for their clear and accessible explanation of the Kriging technique, which aids readers unfamiliar with this field.
While I don't have specific comments on the paper, I do have a couple of suggestions for the authors to consider:

In lines 159162, the authors make an assumption that "any sample location where a taxon has not been observed corresponds to an absence datum for that taxon." It would be beneficial to emphasize the implications of this assumption and its support in the existing literature.

In lines 195196, the sentence reads, "The variogram range is approximately 1000 km, corresponding to the variogram range and representing the average correlation distance of the data.". This sentence could be rephrased for clarity and repetitions.
Citation: https://doi.org/10.5194/essd2022364RC1 
AC2: 'Reply on RC1', Fabio Oriani, 29 Sep 2023
We thank the reviewer for the useful suggestions. see our answers below
https://doi.org/10.5194/essd2022364AC1Citation: https://doi.org/10.5194/essd2022364AC2


RC2: 'Comment on essd2022364', Anonymous Referee #2, 20 Sep 2023
The authors created EUPollMap  pollen occurrence probability maps of 194 pollen taxa for Europe at 25 km spatial resolution using Indicator (Ordinary) Kriging as an spatial interpolation method. The dataset is a valuable source and can serve for various analysis and as input for modeling of various environmental variables. The manuscript is wellorganized and well describes the dataset, methodology and validation. I acknowledge the authors for making this research reproducible by opening the code and the data in standard GeoTiff format.
I have one major recommendation on dataset validation and several minor comments/recommendations for the authors.
Major comment:
At first, I didn't quite understand how you validated the dataset from Section 2.3 Validation strategy. I recommend you to better explain how you validated the data. Later on, I understood that you removed 50% of the data so that the predictions and validation set are within the same range and validated the maps for 10 bins (00.09, 0.100.19, …). I’m generally wondering why you haven’t used e.g. stratified 10fold cross validation for each of the 194 variables? If you take stratified samples you will not have any problem with splitting the data. I think that won’t take you much time and it would be good to know the accuracy (R2, RMSE, …) for each of the variables for someone who wants to use the data. You can also analyze which variable has higher or lower accuracy and why.
Minor comments:
Line 66: I suggest replacing “variogram” to “semivariogram”. Most researchers use the word “variogram”, but in Kriging we actually use semivariogram (based on semivariance).
Line 68: change “at coordinates x” to “at spatial locations x”  if these were coordinates, you will have (x, y).
Line 70: change “any known values” to “any pair of observations”
Line 74: “using a leastsquare approach”  there are several more approaches to fit the semivariogram, but this one is most common. So change to “mostly using a leastsquare approach”.
Line 76: change “suitable for most situations” to “suitable for the spatially correlated data: …”
Line 80: “a smooth variation among adjacent data (zero lag)”  I wouldn't say that because the Kriging prediction is always “smooth”. It actually means that there is no measurement error and the prediction error at observation locations will be zero.
Line 80: “intersection the yaxis at positive values”  it's always positive (check the experimental semivariogram eq.).
Line 81: “indicates continuities in adjacent data, a phenomenon called nugget effect.”  this is actually not true. Nugget represents the variation between the observations at short distances (it can be seen as the measurement error or observational dataset variation). If you have some value for the nugget,it mustn’t always be the nugget effect. You can have a large nugget but there is still spatial correlation between the observations. The nugget effect is when you have nugget and there is no visible spatial correlation between the observations (you cannot fit any of the mathematical models you mentioned in Line 77). Please rephrase this.
Lines 88 and 90: Give some reference for OK and UK. When you say OK, you mean Indicator Kriging  mention this also
Line 91: “No sensible differences in the resulting interpolations were observed”  explain this. Is it because that there is no correlation with the elevation (you can give some accuracy metric, e.g. R2 from the linear regression) or it can be correlation, but there is not much spatial correlation left between the residuals.  I see now that you explained this in the Conclusions and Perspectives section, but still recommend that you explain it here too.
Line 91: “OK allowed including all data”  what do you mean by this? Is there no elevation for all observation locations?
Line 98: “and indicates the uncertainty in space of the estimated probability”  the Kriging variance is actually a metric that shows the estimation precision at a specific location and is dependent on the number of samples and spatial distribution of the samples, and is independent of the values of the samples. Please rephrase this.
Line 104: add “experimental semivariogram”
Line 106: change “e.g. with” to “such as”
Line 111: “impose flat or monotonicpositive model functions”  explain why  because we expect this if there is spatial correlation
Line 114: For step one, were there any cases when there was no spatial correlation? If so, how did you handle this?
Line 118: “Solve the Kriging system”  change to “Kriging prediction”
Line 124: “visually inspected”  how? I suppose you couldn’t do this for all 194 maps.
Line 127: “removing part”  say here that you split the data on train and test dataset.
Line 227: “as the mean of all taxa variance maps”  you take the average, bu you can also take min and max to check where and for which variable the models perform best and worst (see the major comment , the cross validation for each variable)
Line 237: You refer here to supplemental 1 with the accuracy for each of the variables  you can mention which one performs the best, which one the worst and why…
Citation: https://doi.org/10.5194/essd2022364RC2 
AC3: 'Reply on RC2', Fabio Oriani, 29 Sep 2023
We thank the reviewer for the useful suggestions. see our answers below
https://doi.org/10.5194/essd2022364AC1Citation: https://doi.org/10.5194/essd2022364AC3

AC3: 'Reply on RC2', Fabio Oriani, 29 Sep 2023

AC1: 'Authors answer to the reviewers, essd2022364', Fabio Oriani, 29 Sep 2023
Dear Editorial Board,
We thank the editors and the reviewers for positively receiving our contribution. We have carefully examined the useful suggestions, which we address in the following.
In addition to the edits suggested by the reviewers, in the revised version of the dataset, we will exclude the use of Gaussian and power models from the pool of possible variogram models since they are not appropriate for use with Indicator Kriging (Chilès et Delfiner, 2012 p.102).
Reference:
Chiles, JeanPaul, and Pierre Delfiner. Geostatistics: modeling spatial uncertainty. Vol. 713. John Wiley & Sons, 2012.
RC1: https://doi.org/10.5194/essd2022364RC1
This paper introduces a Kriging interpolation method aimed at generating comprehensive pollen presence maps for the EuroMediterranean Basin. The subject matter of this paper holds considerable interest and is undoubtedly of value to the journal's readership. The paper is generally wellstructured, providing all necessary information effectively. I particularly commend the authors for their clear and accessible explanation of the Kriging technique, which aids readers unfamiliar with this field.
While I don't have specific comments on the paper, I do have a couple of suggestions for the authors to consider:
 In lines 159162, the authors make an assumption that "any sample location where a taxon has not been observed corresponds to an absence datum for that taxon." It would be beneficial to emphasize the implications of this assumption and its support in the existing literature.
We thank the reviewer for the valuable suggestion, the following paragraph will be added at lines 159162:
Determining the true/proper absence of a pollen taxon can only be done with extensive vegetation surveys, which is unpractical at the European scale. Moreover, such surveys cannot be done for fossil observations. Therefore, we chose to analyze the EMPD v2 dataset as we would analyze the fossil records. In addition, we believe our assumption to be coherent with the observation probabilities of the taxa. For taxa that produce large quantities of pollen grains (grasses, pines), low percentages already represent longdistance transport, which means that the taxon is not available in the surroundings of the collection sites (Lysitsyna et al. 2011). By extension, assuming that their nonobservation is proof of absence is therefore reasonable. On the other hand, rare taxa or lowpollinating taxa are more difficult to observe in both modern and fossil settings. It is common to observe them in one sample and not in the neighboring one. Using Kriging at the regional target scale for this study, this problem is mitigated since the presence is assessed as a continuous probability variable, computed as a weighted mean from multiple neighbor presence/absence data.”
Reference:
Olga V. Lisitsyna, Thomas Giesecke, Sheila Hicks, Exploring pollen percentage threshold values as an indication for the regional presence of major European trees, Review of Palaeobotany and Palynology, Volume 166, Issues 3–4, 2011, Pages 311324, https://doi.org/10.1016/j.revpalbo.2011.06.004.
 In lines 195196, the sentence reads, "The variogram range is approximately 1000 km, corresponding to the variogram range and representing the average correlation distance of the data.". This sentence could be rephrased for clarity and repetitions.
We thank the reviewer for spotting the error, the sentence will be rephrased as “The variogram range is approximately 1000 km, representing the average correlation distance of the data.”
RC2: https://doi.org/10.5194/essd2022364RC2
The authors created EUPollMap  pollen occurrence probability maps of 194 pollen taxa for Europe at 25 km spatial resolution using Indicator (Ordinary) Kriging as an spatial interpolation method. The dataset is a valuable source and can serve for various analysis and as input for modeling of various environmental variables. The manuscript is wellorganized and well describes the dataset, methodology and validation. I acknowledge the authors for making this research reproducible by opening the code and the data in standard GeoTiff format.
I have one major recommendation on dataset validation and several minor comments/recommendations for the authors.
Major comment:
At first, I didn't quite understand how you validated the dataset from Section 2.3 Validation strategy. I recommend you to better explain how you validated the data. Later on, I understood that you removed 50% of the data so that the predictions and validation set are within the same range and validated the maps for 10 bins (00.09, 0.100.19, …). I’m generally wondering why you haven’t used e.g. stratified 10fold cross validation for each of the 194 variables? If you take stratified samples you will not have any problem with splitting the data. I think that won’t take you much time and it would be good to know the accuracy (R2, RMSE, …) for each of the variables for someone who wants to use the data. You can also analyze which variable has higher or lower accuracy and why.
We thank the reviewer for bringing this point up, and we agree further clarification is needed. Stratified sampling and canonical error scores are not applicable here since the target continuous variable (the probability of pollen occurrence, divided into 10 bins in the reliability plot) is derived from a primary variable, which is the binary (0/1) pollen presence. So, in this case, we cannot stratify in 10 bins the data sampling because data only have 2 possible values (0/1). Error scores such as R2 or RMSE suffer from the same limitation. In this case, it would only be possible to apply them to the threshold maps derived by the probabilityofoccurrence field, then compared with 0/1 data. The error would not be a robust metric, since it would be the distance from arbitrary thresholds (see original manuscript lines 196199). Considering multiple thresholds, this metric would become similar to the reliability plot, with the difference that the latter is a real probabilistic metric. That is why, after considering these possible solutions and previous study cases (lines 125126 original manuscript), we think that the reliability plot is more appropriate for validation in this case.
We will expand the validation section 2.3 to better explain the test applied. The paragraph at lines 133135 will be rephrased with the following: “The way binary presence/absence data aggregate in space determines the estimated probability of occurrence. Some values are rarely found in the output probability map, therefore a high amount of validation data is needed for the reliability plot to be representative of all probability bins (Jolliffe and Stephenson, 2012). Moreover, the sampling for these data cannot be stratified according to the posterior probability values, which are not available a priori. To cope with this limitation, we randomly removed 50% of the data to approach stable statistical values for all bins.”
Minor comments:
We thank the reviewer for the suggestions below, all of them are considered positively and will be applied, see more details below.
Line 66: I suggest replacing “variogram” to “semivariogram”. Most researchers use the word “variogram”, but in Kriging we actually use semivariogram (based on semivariance).
We agree, the word is replaced everywhere.
Line 68: change “at coordinates x” to “at spatial locations x”  if these were coordinates, you will have (x, y).
Agreed.
Line 70: change “any known values” to “any pair of observations”
Agreed.
Line 74: “using a leastsquare approach”  there are several more approaches to fit the semivariogram, but this one is most common. So change to “mostly using a leastsquare approach”.
OK, the sentence is edited accordingly.
Line 76: change “suitable for most situations” to “suitable for the spatially correlated data: …”
Agreed.
Line 80: “a smooth variation among adjacent data (zero lag)”  I wouldn't say that because the Kriging prediction is always “smooth”. It actually means that there is no measurement error and the prediction error at observation locations will be zero.
Agreed, the sentence will be rephrased as “indicates that the variation among adjacent observations (zero lag) is close to zero”.
Line 80: “intersection the yaxis at positive values”  it's always positive (check the experimental semivariogram eq.).
Agreed, the sentence is corrected to “intersection at values larger than zero.”
Line 81: “indicates continuities in adjacent data, a phenomenon called nugget effect.”  this is actually not true. Nugget represents the variation between the observations at short distances (it can be seen as the measurement error or observational dataset variation). If you have some value for the nugget,it mustn’t always be the nugget effect. You can have a large nugget but there is still spatial correlation between the observations. The nugget effect is when you have nugget and there is no visible spatial correlation between the observations (you cannot fit any of the mathematical models you mentioned in Line 77). Please rephrase this.
We agree, the sentence is rephrased as “indicates the variation in adjacent data”.
Lines 88 and 90: Give some reference for OK and UK. When you say OK, you mean Indicator Kriging  mention this also
The OK and UK acronyms are defined at line 88. OK can be applied to Indicator Kriging, as explained at lines 9394.
Line 91: “No sensible differences in the resulting interpolations were observed”  explain this. Is it because that there is no correlation with the elevation (you can give some accuracy metric, e.g. R2 from the linear regression) or it can be correlation, but there is not much spatial correlation left between the residuals.  I see now that you explained this in the Conclusions and Perspectives section, but still recommend that you explain it here too.
We agree, the explanation is reported earlier as suggested.
Line 91: “OK allowed including all data”  what do you mean by this? Is there no elevation for all observation locations?
We agree, the sentence is rephrased “OK allows including all data points with limited computation, while UK would require excessive computational burden. Also, elevation does not appear to be clearly correlated with the pollen presence at the scale of our dataset, thus not sensibly affecting the prediction when used as external drift in the Kriging model”.
Line 98: “and indicates the uncertainty in space of the estimated probability”  the Kriging variance is actually a metric that shows the estimation precision at a specific location and is dependent on the number of samples and spatial distribution of the samples, and is independent of the values of the samples. Please rephrase this.
We agree, the explanation of the variance field should be clarified. Variance independence from the data values is true, for a given covariance model (Goovaerts 1997, p. 179), while in this and most practical cases, the variogram model is fitted on the data. As a consequence, the OK variance field will also depend indirectly on the data values. We obtain different variance fields for different pollen species, presenting the same data amount and locations, only changing their 0/1 values. We rephrase the sentence as “and indicates the uncertainty of the prediction depending on the data amount, their spatial distribution, and the semivariogram model.”
Reference:
Goovaerts, P., 1997, Geostatistics for Natural Resources Evaluation, Oxford University Press.
Line 104: add “experimental semivariogram”
Agreed.
Line 106: change “e.g. with” to “such as”
Agreed.
Line 111: “impose flat or monotonicpositive model functions”  explain why  because we expect this if there is spatial correlation
We agree, this sentence is modified as “Monotonicity and a positive slope are expected in a model semivariogram, but, in case of a noisy experimental semivariogram (which is the case for some of the observed pollen taxa), unconstrained fitting can lead to a negative slope in the model. For this reason, we impose flat or monotonicpositive model functions ”.
Line 114: For step one, were there any cases when there was no spatial correlation? If so, how did you handle this?
We agree, we add the sentence at point two: “Lack of correlation in the data leads to a fitted semivariogram model tending to a constant function and a subsequent constant estimated mean and variance fields. ”
Line 118: “Solve the Kriging system”  change to “Kriging prediction”
Agreed.
Line 124: “visually inspected”  how? I suppose you couldn’t do this for all 194 maps.
We agree, we correct to “For a selection of common species for Europe, constituting examples of different spatial distributions (e.g., broad extent, rare species, discontinuous distributions), the interpolated probability surfaces and their variances are visually inspected and shown with the original observations. Moreover, for all species, we assessed the reliability of the interpolations with reliability plots (Murphy and Winkler, 1977), which are common quantitative and graphical representations in geo and atmospheric sciences (Bröcker and Smith, 2007; Allard et al., 2012)”.
Line 127: “removing part”  say here that you split the data on train and test dataset.
We agree.
Line 227: “as the mean of all taxa variance maps”  you take the average, bu you can also take min and max to check where and for which variable the models perform best and worst (see the major comment , the cross validation for each variable)
We agree, see next answer.
Line 237: You refer here to supplemental 1 with the accuracy for each of the variables  you can mention which one performs the best, which one the worst and why…
We haven’t detected any particular taxonomic group showing a worst/best performance, while we have observed that the worst cases are the ones showing a complex spatial heterogeneity (e.g. with isolated presence/absence points, specified at line 245) departing from local stationarity. This additional consideration will be added at line 248.
Citation: https://doi.org/10.5194/essd2022364AC1
Fabio Oriani et al.
Data sets
EUPollMap: The European atlas of contemporary pollen distribution maps Fabio Oriani, Gregoire Mariethoz, Manuel Chevalier https://doi.org/10.5281/zenodo.7259017
Model code and software
EUPollMap_scripts Fabio Oriani, Gregoire Mariethoz, Manuel Chevalier https://github.com/orianif/EUPollMap_scripts
Fabio Oriani et al.
Viewed
HTML  XML  Total  Supplement  BibTeX  EndNote  

349  111  29  489  34  12  10 
 HTML: 349
 PDF: 111
 XML: 29
 Total: 489
 Supplement: 34
 BibTeX: 12
 EndNote: 10
Viewed (geographical distribution)
Country  #  Views  % 

Total:  0 
HTML:  0 
PDF:  0 
XML:  0 
 1