the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Submitted as: development and technical paper 17 Feb 2021
Submitted as: development and technical paper | 17 Feb 2021
Climate model-informed deep learning of global soil moisture distribution
- 1Max Planck Institute for Chemistry, Hahn-Meitner-Weg 1, 55128 Mainz, Germany
- 2The Cyprus Institute, P.O. Box 27456, 1645 Nicosia, Cyprus
- 1Max Planck Institute for Chemistry, Hahn-Meitner-Weg 1, 55128 Mainz, Germany
- 2The Cyprus Institute, P.O. Box 27456, 1645 Nicosia, Cyprus
Abstract. We present a deep neural network (DNN) that produces accurate predictions of observed surface soil moisture, based on meteorological data from a climate model. The network was trained on daily satellite retrievals of soil moisture from the European Space Agency (ESA) Climate Change Initiative (CCI). The predictors precipitation, temperature and humidity were simulated with the ECHAM/MESSy atmospheric chemistry-climate model (EMAC). Our evaluation shows that predictions of the trained DNN are highly correlated with the observations, both, spatially and temporally, and free of bias. This offers an alternative for parametrisation schemes in climate models, especially in simulations that use, but may not focus on soil moisture, which we illustrate with the threshold wind speed for mineral dust emissions. Moreover, the DNN can provide proxies for missing values in satellite observations to produce realistic, comprehensive, high resolution global datasets. As the approach presented here could be similarly used for other variables and observations, the study is a proof of concept for basic but expedient machine learning techniques in climate modelling, which may motivate additional applications.
Klaus Klingmüller and Jos Lelieveld
Status: final response (author comments only)
-
RC1: 'Comment on gmd-2020-434', Anonymous Referee #1, 15 Mar 2021
The paper applies a deep neural network to build a relationship between 18 predictors (simulations of rain, surface temperature and humidity, location, seasonality, root depth etc.) and a predictand (soil moisture down to about 5 cm). The simulation data was produced by a global atmospheric chemistry-climate model EMAC nudged to reanalysis data. The predictand’s reference data was the ESA CCI Soil Moisture product.
The motivation for the application of a neural network was to replace EMAC’s soil moisture parameterization with a better one in a mineral dust emission parameterization. The study shall be seen as a proof of concept (line 195). Yes, it is, but a few issues should be clarified.
The application has very dry areas in its focus. I have in mind that the soil moisture satellite product is especially uncertain in these areas. This should be discussed a bit. The trained prediction is most uncertain in the most interesting regions (Fig. 4: Sahara, Gobi Desert etc.). Why? Quality of the satellite reference or a training period of only 8 years?
The DNN is built with 512 units and four hidden layers. This parameter selection should be motivated a bit. Of more concern is the DNN performance. With location and seasonality as predictors, I expect a high correlation between prediction and reference soil moisture. What is the benefit of using meteorology/climate simulation in the prediction?
Finally, it would be helpful to have short discussions on the applicability of the chosen approach in a changing climate and an alternative DNN training of EMAC parameters (avoiding two parametrizations predicting soil moisture).
-
AC1: 'Reply on RC1', Klaus Klingmueller, 17 May 2021
The comment was uploaded in the form of a supplement: https://gmd.copernicus.org/preprints/gmd-2020-434/gmd-2020-434-AC1-supplement.pdf
-
AC1: 'Reply on RC1', Klaus Klingmueller, 17 May 2021
-
RC2: 'Comment on gmd-2020-434', Anonymous Referee #2, 19 Mar 2021
This study explores DNN-based soil moisture predictions that are trained on remote sensing data and forced by a climate model, which are used to replace low-quality soil moisture predictions in that model for an improved prediction of mineral dust emissions.
Overall, the topic is relevant, the methodology novel and sound, and the manuscript concise, well structured, and was a pleasure to read. I just have a few concerns and suggestions which I hope the authors will consider before publication.
-) To me, hardware developments do not seem to be the most pressing issue and motivation for the presented study. I would recommend to start the introduction outright with the more relevant issue, i.e., the low-quality internal soil moisture predictions of the climate model which could potentially be replaced by a DNN-based module.
-) Maybe a better way to achieve independence between training and test data than doing a standard 80/20 splitting could be to take the characteristics of the ESA CCI SM product into account, that is, splitting the data at a time where there is a change in the underlying satellite instruments that are being merged. Or perhaps better yet: training the model on the "active-only" ESA CCI SM data set, and evaluating it on the "passive-only" data set (but still in a different period of course). That's probably too much to ask for this paper, but perhaps something to keep in mind for future studies.
-) What motivates the selection of the predictor variables? Are they taken from another study, or did you try out different combinations and evaluated variable importance? A little elaboration on that would be helpful.
-) On a related note: I am a bit concerned that the DNN predicts mostly a spatial and temporal climatology. The correlations of >0.9 are arguably unusually high for typical soil moisture data sets. This argument could be settled, for example, by showing anomaly correlations as well. As for the temporal correlations: The values of ~0.5 indeed appear to be more realistic, but I don't understand the reason for not showing the acutal values in Figure 4. A discussion on how these temporal correlations compare to values typically found for other data sets could also be beneficial. For example, Figure 7 in Dorigo et al. (2017) shows correlation coefficients of modeled versus ESA CCI SM soil moisture, which may serve as a good benchmark for putting the attained values into perspective. Another way to appease the reader may come from an elaboration on the selection of predicture variables, as noted above... For example, how would the predictions look if merely coordinates and the time of the year would be provided? Or if they would be omitted? Also, the time series shown in Fig. 5 do not really convince me. For example, it is pointed to "irregular features" in, for example, October 2014. But to me it seems that there are distinct dry-downs in the prediction in both October 2014 and October 2015, while this dry-down is only visible in the observation in 2014.
-) Having said that, for the application shown, it seems that this may not be relevant at all. The improvements in dust emission predictions shown in Figure 6 are quite remarkable, and it seems that the DNN-predicted values, even it they would be only a climatology with possibly a little short-term variability signal added on top, are doing already much better than the model-internal soil moisture representation. The taken approach of validating soil moisture predictions using a downstream application is a great way to show their actual utility, which is much more useful than the more common approach of simply looking at rather meaninglesss correlation values alone.
-) I'd change the title of Sec. 4 to singular since only a single application is shown.
-) I recommend to add a note of caution to the argument that the DNN could be used to fill gaps in the remotely sensed data. For example, tropical rain forest are masked out entirely in the ESA CCI SM. I am not sure if the DNN is able to properly learn the relation between the predictor variables and soil moisture in such a distinct regime if it is not represented in the training data set at all. A similar argument can probably made for the winter. If soil moisture cannot be retrieved because the soil is frozen or covered with snow, then I would not expect a DNN to properly turn precipitation, which is most likely in the form of snowfall, into accurate predictions of soil moisture.
-
AC2: 'Reply on RC2', Klaus Klingmueller, 17 May 2021
The comment was uploaded in the form of a supplement: https://gmd.copernicus.org/preprints/gmd-2020-434/gmd-2020-434-AC2-supplement.pdf
-
AC2: 'Reply on RC2', Klaus Klingmueller, 17 May 2021
-
RC3: 'Comment on gmd-2020-434', Anonymous Referee #3, 24 Mar 2021
General Comments:
This paper presents a deep neural network model that predicts global surface soil moisture from precipitation, temperature, and humidity outputs from a climate model. The model was trained on daily satellite retrievals of soil moisture. The authors suggest two uses for the model: 1) to provide modeled soil moisture inputs for related applications, and 2) to fill missing values in satellite retrievals. The authors demonstrate an application by simulating threshold surface friction velocity for mineral dust emission in the Arabian Peninsula and Mesopotamia.
I found this paper to be rigorous, complete, and convincing. While I have some questions, I think the overall quality is very good. The conclusions are well-founded, and the future research questions are well discussed.
This paper could use some help from an English language editor. Some sections of the paper are very well written, and some have grammatical and language flaws. Even so, the paper is easy to read and understand.
Specific Comments:
The authors present a simulation of threshold surface friction velocity for mineral dust emissions in the Arabian Peninsula and Mesopotamia as an application of the model. They simulate the threshold friction velocity using both observed and DNN-modeled soil moisture with good agreement. This leads me to ask: why use the DNN here at all? Why not just use the observations directly?
The authors say that Figure 6 shows “The results based on the observed and predicted soil moisture show good agreement and a strong seasonal cycle”. More detail should be given here. The model appears to overpredict the threshold surface friction velocity a bit in the summer. Then, “whereas the result based on the EMAC soil water has little variability” – they could also compare the DNN to EMAC soil water directly.
The authors should provide more information about the model selection criteria they used for the DNN input variables. They used a set of 18 input variables, all of which are intuitive. However, it would be interesting to see which of these input variables drive the predictive power of the model. This might be a particularly interesting question for the mineral dust application: what are the most important drivers of soil moisture in the Arabian Peninsula and Mesopotamia and what does this mean for vulnerability to dust storms? In regions where the temporal correlation is weaker, are the cos(2*pi*t/a) and sin(2*pi *t/a) terms dominating to impose the observed seasonal cycle?
Figure 5 shows a time series comparison of predicted and observed soil moisture at a single pixel during the test period. This pixel is located in Germany, where the model is reported to have strong temporal correlation (Fig. 4). What does the time series look like in a pixel with a poorer temporal correlation? What does is look like in a pixel in the poorly correlated region of the Arabian Peninsula?
Figure 7 shows the global distribution of observed and predicted volumetric soil moisture on two days in the training period. It would be very interesting to see similar plots for the test period.
Technical Corrections:
There is quite a bit of model evaluation work in the “Applications” section. In particular, I think that the presentation and some of the discussion of Figures 5 and 7 could be moved up.
-
AC3: 'Reply on RC3', Klaus Klingmueller, 17 May 2021
The comment was uploaded in the form of a supplement: https://gmd.copernicus.org/preprints/gmd-2020-434/gmd-2020-434-AC3-supplement.pdf
-
AC3: 'Reply on RC3', Klaus Klingmueller, 17 May 2021
Klaus Klingmüller and Jos Lelieveld
Klaus Klingmüller and Jos Lelieveld
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 381 | 154 | 14 | 549 | 4 | 2 |
- HTML: 381
- PDF: 154
- XML: 14
- Total: 549
- BibTeX: 4
- EndNote: 2
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1