A deep learning hybrid predictive modeling (HPM) approach for estimating evapotranspiration and ecosystem respiration

Chen, Jiancong; Dafflon, Baptiste; Tran, Anh Phuong; Falco, Nicola; Hubbard, Susan S.

doi:https://doi.org/10.5194/hess-25-6041-2021

Articles | Volume 25, issue 11

https://doi.org/10.5194/hess-25-6041-2021

© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/hess-25-6041-2021

© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 25, issue 11

Research article

|

25 Nov 2021

Research article |

| 25 Nov 2021

A deep learning hybrid predictive modeling (HPM) approach for estimating evapotranspiration and ecosystem respiration

Jiancong Chen, Baptiste Dafflon, Anh Phuong Tran, Nicola Falco, and Susan S. Hubbard

Download

Final revised paper (published on 25 Nov 2021)
Supplement to the final revised paper
Preprint (discussion started on 21 Jul 2020)
Supplement to the preprint

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review of hess-2020-322', Anonymous Referee #1, 14 Aug 2020
- AC1: 'Response to comments by Anonymous Referee #1', Jiancong Chen, 29 Nov 2020
RC2: 'Review of “A Deep- Learning Hybrid-Predictive-Modeling Approach for Estimating Evapotranspiration and Ecosystem Respiration” by Jiancong Chen, Baptiste Dafflon, Anh Phuong Tran, Nicola Falco, and Susan S. Hubbard', Anonymous Referee #2, 23 Oct 2020
- AC2: 'Response to comments by Anonymous Referee #2', Jiancong Chen, 29 Nov 2020

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

ED: Reconsider after major revisions (further review by editor and referees) (05 Dec 2020) by Carlo De Michele

AR by Jiancong Chen on behalf of the Authors (11 Jan 2021) Author's response Manuscript

ED: Referee Nomination & Report Request started (19 Feb 2021) by Carlo De Michele

RR by Anonymous Referee #1 (19 Mar 2021)

Suggestions for revision or reasons for rejection

I’m not a modeler, but it seems like the model works really well/represents a broadly applicable advance. My remaining comments are generally focused on the setup, presentation, and interpretation of results.

General comments:
1) The first three paragraphs of my previous review (general comments) were not specifically addressed. In my experience, this is not typical for the reviewer response process, and is likely to be unacceptable to some reviewers and/or editors going forward.
In particular, I would still appreciate the authors responding to the third paragraph of my previous review, repeated here:

Relatedly, Figures 4-9 all show similar long-term timeseries data with scatterplots that lend themselves to similar interpretations in terms of R2 of MAE. These are useful, but perhaps they could be condensed and/or supplemented with other figure types that were more conducive to process-based interpretation. For example, I found Figures 11e and 11f fascinating insofar as they highlighted seasonal differences between vegetation types, but little explanation was provided to “unpack” these results (grasslands and shrublands not even mentioned). Likewise, Figures 12a and 12b present a rich opportunity to speak to differences between the biophysical controls on ET at the SNOTEL and East River sites. Some of the specific factors I’m left wondering about are differences in snow accumulation and melt between sites, evaporation versus transpiration, and heterotrophic versus autotrophic respiration. I understand that you don’t have all these measurements, but you’ve generated a lot of suggestive data that could be leveraged to push this field of research.

2) Some of the results and discussion require more nuanced and/or focused interpretation (see detailed comments below). At the same time, the manuscript is long and could be shortened/tightened in many places to more accurately present/highlight key results (details below).

Specific comments:
L10: The decision to focus on Reco could be set up better. In other words, why Reco instead of NEE or GPP and/or all three? I don’t necessarily have a problem with your decision to focus on Reco, but it must be clearly justified.

L17 “sites within sites”?

L21: Suggest adding “USA” here for the global audience.

L27ff: Please specify “air, soil, snow”, etc. whenever “temperature” is invoked. Lots of room for confusion here because most would expect ET to vary more with air temperature versus Reco that is more sensitive to soil temperature.

L34-35: Same comment as L10.

L129-131: Recent work by Chu et al. 2021 on the representativeness of statistical tower measurement footprints to surrounding areas may be relevant here.

L191-192: “Same” is not the equivalent of “too close to distinguish on the map”.

L483: Is it that earlier snowmelt triggers the onset of vegetation activity or that higher air temperatures trigger both snowmelt and the onset of vegetation activity?

L485-486: Can you speak to the synoptic meteorological conditions in 2012 versus 2015? Why choose these two years for comparison? Similarly, the comparison of March, April, and May between years is interesting, but what about the rest of the year? I’d be very interested in a similar post-monsoon analysis, potentially between years with strong and weak monsoons.

L492-497: Please edit this section to remove/acknowledge differences in NDVI that would be expected due to deciduous versus evergreen physiology. Some of this basic information currently comes across as results. I appreciate the attempt to relate these results back to processes, but this section needs refinement.

L517ff: What does the “1” syntax correspond to?

L525-526: Please be specific about the meaning of “drought” in this context. Is it simply meant to connote some limitation to ET and/or Reco? If so, can you justify the underlying expectation that these variables would be affected at the same moisture threshold? I’d also argue that “usually” is the wrong word here. Earlier snowmelt certainly “can” trigger summer drought, but this scenario is subject to modification by monsoon precipitation and other factors as the authors acknowledge in this sentence. See recent work by Knowles et al. 2020, Xu et al. 2020, and many references therein.

L583-596: I support this opportunity to discuss physiological differences between evergreen and deciduous vegetation, but simply citing Baldocchi et al. 2010 is insufficient. More thorough and nuanced discussion that incorporates foundational research on this topic is required.

L600-601: See comment on L525-526.

L604-605: This implies that growing season length determines snow water storage when in fact, it’s closer to the opposite i.e., air temperature and/or snow accumulation determine the onset of the growing season. See Lian et al. 2020 and Zhang et al. 2020 for examples of more recent work on this topic. Combining the Sloat et al. 2015, Wainwright et al., 2020, and Hu et al., 2010 references here also raises an important distinction. Whereas the Sloat and Wainwright references invoke fore-summer i.e., pre-monsoon drought, the Hu reference pertains to late summer drought i.e., after snowmelt water inputs have subsided. This distinction reflects the typical relative importance of snowmelt vs. monsoon precipitation at a given site e.g., snow-dominated sites may be susceptible to moisture limitation after the snowmelt pulse (late summer; Hu et al. 2010), whereas monsoon-dominated sites may be susceptible to moisture limitation before the onset of monsoon rains (early/fore-summer; Sloat et al 2015; Wainwright et al. 2020). Please establish the typical relative importance of snow versus monsoon precipitation at the East River site and how your results may be expected to change at sites where moisture availability is typically more or less affected by snowmelt versus monsoon precipitation.

L612: Hard to follow, I think “whereas” may be the wrong word here.

L629: “Microclimate” is misspelled.
Chu, H., Luo, X., Ouyang, Z., Chan, W. S., Dengel, S., Biraud, S. C., Torn, M. S., Metzger, S., Kumar, J., Arain, M. A., Arkebauer, T. J., Baldocchi, D., …, and Zona, D.: Representativeness of Eddy-Covariance flux footprints for areas surrounding AmeriFlux sites, Agricultural and Forest Meteorology, 301–302, 108350, https://doi.org/10.1016/j.agrformet.2021.108350, 2021.
Knowles, J. F., Scott, R. L., Biederman, J. A., Blanken, P. D., Burns, S. P., Dore, S., Kolb, T. E., Litvak, M. E., and Barron‐Gafford, G. A.: Montane forest productivity across a semiarid climatic gradient, Global Change Biology, 26, 6945–6958, https://doi.org/10.1111/gcb.15335, 2020.
Lian, X., Piao, S., Li, L. Z. X., Li, Y., Huntingford, C., Ciais, P., Cescatti, A., Janssens, I. A., Peñuelas, J., Buermann, W., Chen, A., Li, X., Myneni, R. B., Wang, X., Wang, Y., Yang, Y., Zeng, Z., Zhang, Y., and McVicar, T. R.: Summer soil drying exacerbated by earlier spring greening of northern vegetation, Science Advances, 6, eaax0255, https://doi.org/10.1126/sciadv.aax0255, 2020.
Xu, B., Arain, M. A., Black, T. A., Law, B. E., Pastorello, G. Z., and Chu, H.: Seasonal variability of forest sensitivity to heat and drought stresses: A synthesis based on carbon fluxes from North American forest ecosystems, Global Change Biology, 26, https://doi.org/10.1111/gcb.14843, 2020.
Zhang, Y., Parazoo, N. C., Williams, A. P., Zhou, S., and Gentine, P.: Large and projected strengthening moisture limitation on end-of-season photosynthesis, Proc Natl Acad Sci USA, 201914436, https://doi.org/10.1073/pnas.1914436117, 2020.

Hide

RR by Anonymous Referee #3 (21 Mar 2021)

ED: Publish subject to revisions (further review by editor and referees) (12 Apr 2021) by Carlo De Michele

AR by Jiancong Chen on behalf of the Authors (21 May 2021) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (26 May 2021) by Carlo De Michele

RR by Anonymous Referee #1 (28 May 2021)

RR by Anonymous Referee #4 (28 Sep 2021)

Suggestions for revision or reasons for rejection

This study uses a novel approach that combines process-based modeling and eddy covariance measurements with freely-available spatial datasets and machine learning to predict evapotranspiration and respiration fluxes at the ecosystem scale. The input training data are meteorological data and remote sensing NDVI, which are linked to measured (via eddy covariance, aka ‘data HPM’) or modeled (via Community Land Model, aka ‘mechanistic HPM’) daily ET and ecosystem respiration. The method used is a long short-term memory (LSTM) model - which is a type of recurrent neural network that has been used successfully for rainfall runoff and hydrological monitoring modeling. The authors demonstrated the use of both a data HPM and mechanistic HPM approach, and then used the latter method to model ET and Reco at several locations within the East River Watershed in southwestern Colorado. The authors then relate the resulting modeled fluxes to changes in biome and year-to-year weather patterns. They conclude that HPM is a feasible method for estimating spatial and temporal patterns of ET and Reco, and I agree with them. This paper is written logically, with very few typos and needs just a bit of clarification (see my comments below). This paper is an interesting study on using data and tools available to us to address the need for spatially- and temporally-dense water and carbon fluxes, especially in mountainous landscapes. I am not an expert in neural network modeling, but I found the methods were not too difficult to follow.

Line-by-line comments:

111: ‘multi-model and multi-data approaches…’
223: Shouldn’t ‘CLM predictions’ be added to this list of input features?
263: It is not clear to me why reconstruction of daily NDVI is necessary. Why can’t the NDVI measurements be used with the gaps in time coverage? As long as there are accompanying meteorological measurements for that day, it should be fine. Please briefly add a statement justifying why it was necessary to gap-fill NDVI data and how many data points and the percentage of total NDVI data points were modeled instead of measured.
501:’These models were trained using and validated against eddy covariance…’

Hide

ED: Publish subject to minor revisions (review by editor) (04 Oct 2021) by Carlo De Michele

AR by Jiancong Chen on behalf of the Authors (14 Oct 2021) Author's response Manuscript

ED: Publish as is (29 Oct 2021) by Carlo De Michele

Short summary

The novel hybrid predictive modeling (HPM) approach uses a long short-term memory recurrent neural network to estimate evapotranspiration (ET) and ecosystem respiration (R_eco) with only meteorological and remote-sensing inputs. We developed four use cases to demonstrate the applicability of HPM. The results indicate HPM is capable of providing ET and R_eco estimations in challenging mountainous systems and enhances our understanding of watershed dynamics at sparsely monitored watersheds.