Dynamic INtegrated Gap-filling and partitioning for OzFlux ( DINGO ) response to reviewer comments

The manuscript is well written and describes a useful system to automate gapfilling. I do not find anything radically new in the approach. Each step in the process is based on well known, and widely used approaches to gapfilling. The novelty, perhaps, is in the automation of the process that offers new ease of use and the potential for widespread application. The test for such a new automated process is the applicability and its wide usability. Many flux sites have such a process for their own site, and none (or few) considered publishing it. As long as the process is specific to only and nothing more than OzFlux, I do not think it makes an interesting scientific paper, and should be published as an OzFlux report or the DINGO software manual. This could be easily remedied, and I propose that the authors will make the effort to generalize DINGO so that it could be applied by any flux site that desires to.

As a whole the paper was well received but the reviewers and editor felt that the paper is rather technical with not strong scientific or methodological advancements (a lot of re-implementations of existing and published tools, not real validations).The editor also stated that at the same time the authors have an important paper in the context of the OzFlux Special Issue, because part of what this community did in the last years and for this reason the editor thinks that should be included.After discussion the editor felt that the paper should be re-organized as Technical note.We have still addressed the comments made by the reviewers that are relevant (see below).
In addition the editor also raised a specific question "(already raised by a reviewer but not yet clear): authors say that "'GPP is equated to zero at night but is not forced to zero during the day.Since GPP is the difference between measured net ecosystem exchange and estimate ecosystem respiration, it incorporates random error …….While these estimates are therefore unphysical, the effect of their removal is to filter the positive domain of the random error distribution, thereby converting random error to systematic error.'".I completely agree with the explanation but forcing GPP to zero nighttime is something that hidden this effect of the random errors in the EC data.For this reason it would be probably better to leave the positive and negative GPP in the night to better help a user to understand that there is an uncertainty in the data (and have a quantification).".Although the random error at nighttime can be positive or negative the absolute value of GPP could also be significantly non-zero.This could occur at nighttime due to biases in the measurement of NEP and calculation of Re and therefore GPP would have similar biases that compound.So the absolute value of GPP at night includes a bias and random error.Since we know physiologically that GPP must be zero we feel that the most robust GPP at nightime would be to force it to zero".We add this discussion to the text.

Anonymous Referee #1
Received and published: 16 June 2016 The manuscript is well written and describes a useful system to automate gapfilling.I do not find anything radically new in the approach.Each step in the process is based on well known, and widely used approaches to gapfilling.The novelty, perhaps, is in the automation of the process that offers new ease of use and the potential for widespread application.The test for such a new automated process is the applicability and its wide usability.Many flux sites have such a process for their own site, and none (or few) considered publishing it.As long as the process is specific to only and nothing more than OzFlux, I do not think it makes an interesting scientific paper, and should be published as an OzFlux report or the DINGO software manual.This could be easily remedied, and I propose that the authors will make the effort to generalize DINGO so that it could be applied by any flux site that desires to.

•
The reviewer is correct in pointing out there are a few novel scientific outcomes from the paper.However, we feel that the paper makes an important contribution to the OzFlux special issue of which this paper is a part.It forms part of an important contribution that documents the OzFlux network in terms of the overall vision (special issue overview paper) through to the different processing methodologies that can be applied (OzFlux QC and DINGO).Therefore in the context of the special issue we feel this paper is well positioned and in this case.We recognise that the paper is more of a technical nature and as such it could be submitted as a technical note in the journal if this is possible.
Remembering in the context of OzFlux and the special issue that dingo is defined as the dyamic INtegrated Gap filling and partitioning for Ozflux (DINGO).The OzFlux QC paper in the special issue is in this identical situation, however, in that paper it was suggested to remove any performance analysis results.
You write in the abstract that "DINGO was developed for Australian data but the frame work is applicable to any flux data or regional network".However, it seems to be hardcoded for the Australian data sources and therefore will technically no run for other sites.As there is no novelty to the formulation, I see one of the big advantages in the automation of the work process and particularly in the data retrieval for the gap-filling, which is usually a cumbersome and laborious process.Currently, when gap-filling the forcing data the software only reads Australian met data and Australian high resolution meteorological reanalysis.Is it possible to allow a generic, user provided, table with data from n met stations nearby?Each user can prepare the data from their local met stations but the process of data selection and correlation with the flux-site met obs can be done by DINGO.Can DINGO be made to access the ECMWF reanalysis?It is coarser than the Australian one, but global and would allow users worldwide to have an alternative to local met stations.
• We agree that it would be a worthy goal to enable DINGO to be utilised globally but this is a massive amount of effort to develop and test this and would therefore be part of a new suite of tools (not DINGO any more).We think that just by enabling DINGO to be applied globally does not add extra scientific value (as suggested was needed by the reviewer), however, it would certainly make it more usable to the community.We suggest that is beyond the scope of the current paper but would be willing to work towards this in a new platform that could be developed with others in the flux net community.

•
Thinking ahead to a possible global version of this tool in future work, ECMWF would be a great source of gridded climate data.This would need to be input in conjunction with local meteorology data from local weather stations.Most of the time in the Australian flux network the correlations between flux tower data are better with local weather stations and therefore this is the preferred source of data for gap filling.However, access to global weather station data may be rather difficult.
For precipitation, which is very poorly generalized from one met station to the other (and particularly the time of precipitation) I suggest that even for the OzFlux data processing there could be a great advantage in using satellite-derived precipitation from TRMM or the new GPM products.
• This is a really great suggestion from the reviewers and at the time of producing DINGO, TRMM was the main satellite precipitation product, however, it is not suitable for higher latitudes and does not cover all of Australia.We could consider including global precipitation data from GMP in a future global version of this platform.
Similarly, the soil moisture model can offer the option to be driven by MODIS LAI and ECMWF as an alternative to the strictly Australian data sources.That should be easy as MODIS is extracted anyhow for the albedo gapfilling.
• At present the soil moisture model is run off-line with AWAP gridded meteorological data.This is performed every 6-12 months.The soil moisture model is not currently part of the code.However, there is a great opportunity to embed the soil moisture model into the code and therefore it could be run using this type of global data.
There is nothing new in incorporating footprint and estimating the uncertainty of the gap-filling method.In fact, it is irresponsible not to do so.These are features that you say will be available in V13, but I see no purpose in publishing an application to a well known method that is not done yet.However, you say that version 13 is due July 2016.This is a few weeks from now, and before the revision will be due.Can you make this paper about the complete and fully functional version 13?

•
Version 13 of dingo is now complete so now update the manuscript accordingly and this strengthens the paper considerably as it includes the important uncertainty calculation contributions.We have added this and a figure showing its output.
Other comments: Why are you using different meteorology to drive the soil water model (meteorology from the Bureau of Meteorology's Australian Water Availability Project) than the station and other met data already processed and gapfilled for the station location, if station data is missing?
• Again, currently the soil moisture model is run off-line with AWAP gridded meteorological data.This is performed every 6-12 months.The soil moisture model is not currently part of the code.In a new are version of the code we could embed the model into DINGO and therefore we could use the gap filled meteorological data to drive the model.
The u* filtering section should come before the ANN gap-filling section as u* filtering creates most of the gaps that then need to be filled.

•
We suggest that it is not appropriate to produce a new platform that is globally applicable for this paper and is well outside the scope of the special issue.However, we think it is worthwhile discussing the potential challenges and opportunities for making the tool globally available and would add this to a future development section.

•
One of the very novel features in DINGO is the uncertainty calculations in version 13 and we will include them in the new version of the manuscript.Most flux tower sites will produce some estimates of random and model uncertainty and these are usually reported separately.In DINGO v13 the uncertainties are calculated for random, model and ustar uncertainty and these uncertainties are combined to provide a total uncertainty for the site, which has not been done to date.We will include this to enhance the scientific value of the paper.

•
We're not sure what the referee is referring to with respect to the u* filtering as section 4 is on friction velocity which precedes section 5 on gap filling.Also we clearly state "Once the data have been u* filtered, they are used to train an ANN (see Section 2.2) using nighttime data only with inputs of Sws, Ts, Ta and EVI as known drivers of ecosystem respiration (Migliavacca et al., 2010)".

GENERAL COMMENTS
This manuscript describes a dynamic integrated gap filling and partitioning tool (DINGO) developed for standardized processing of OzFlux station data.The tool is designed to gap fill both meteorological variables and fluxes, besides it partitions the net flux into its components (Gross Primary Productivity and Ecosystem Respiration).Although it raises the important issue of a standardized post processing for meteorological and eddy flux data at network level, unfortunately the manuscript doesn't introduce important novelty aspects, and it sounds more like a technical note than a scientific contribution.It provides a very detailed description of the tool, but lacks of analysis and interpretation.Besides, most of the figures included in the manuscript are standard diagnostic or result plots provided by the tool itself and not the result of an in depth analysis of the tool behavior, performance or applicability.Some important aspects that should be faced and deepened are not included, as for example: the evaluation of the tool performances and their dependence on gap percentage, length and distribution, a comparison with other existing gap-filling and partitioning tools, an analysis of the effects of the gap filling on annual sums... Results of the uncertainty estimation would also enrich the work, together with some performance analysis at site/PFT level.A technical observation: u* filtering should be applied before gap filling since it introduces gaps in the time series.For the above mentioned reasons, I do not think that the manuscript, in its present form, is suitable for publication in Biogeosciences.
• Also see comment responses to referee one.In addition, we also recognise that the paper is more of a technical note and we leave it to the editors discretion as to whether the paper should be submitted as a technical note to Biogeosciences.However, as mentioned above the paper makes an important contribution to the OzFlux special issue of which this paper is a part.It forms part of an important contribution that documents the OzFlux network in terms of the overall vision (special issue overview paper) through to the different processing methodologies that can be applied (OzFlux QC and DINGO).The OzFlux QC is very similar in scope and technical detail to this paper.We agree that the performance of the tool is important and we are planning a separate manuscript to address this across both OzFluxQC and DINGO processing tools as well as comparison with other available platforms such as EdiProc.
• We actually do perform u* filtering before gap filling, so we will clarify the wording in the text to make this apparent.We state that "Once the data have been u* filtered, they are used to train an ANN (see Section 2.2) using nighttime data only with inputs of Sws, Ts, Ta and EVI as known drivers of ecosystem respiration (Migliavacca et al., 2010)."

Anonymous Referee #3
Received and published: 17 June 2016 The manuscript presents an overview of the software framework and methods for automated gap-filling and partitioning of OzFlux tower network data.Overall, the manuscript does a nice job presenting this overview with the right amount of detail and rationale, and is well-written.The gap-filling and partitioning methods employed are generally state-of-the-art community standards.That said, there is not much new information here.Perhaps the greatest novelty lies in the gap-filling of ancillary drivers, where to my knowledge the only other recent work on this topic is that of Vuichard and Papale (2015).Although that work is not referenced, the present manuscript expands upon it to incorporate data from nearby weather stations and satellite remote sensing observations, with additional soil-and radiation-tailored gap-filling methods.
I have two major recommendations: 1) The manuscript could be made more impactful by including an analysis of performance of the gap-filling and partitioning results across all or part of the OzFlux network.This would demonstrate the utility/flexibility of the framework, its ability to address the grand science challenges identified in the introduction, and highlight areas for continued improvement.
2) The paper would benefit from a discussion of how the methods employed in this paper conform to or push the envelope of current community practice.This would clearly communicate the significance of the software suite and its novel contributions.

•
We thank the reviewers for their highly useful comments.We were not aware of the Vuichard and Papale (2015) reference so we thank you for that.We point to the comments in response to referee one and two.Given the comments made reviewer one and two we suggest that such a network wide analysis of the performance of gap filling tools is best suited to a separate paper that includes both of the major tools used in the OzFlux network (OzFluxQC and DINGO).That paper would also include some discussion of current community practice and how we can advance new or best practice.

Specific comments:
-The representation of meteorological quantities should make more effort to be consistent with community standards.I have found Reifsnyder et al. (1991) particularly useful for this purpose, presenting the symbols, units, and notation for use in the journal Agricultural and Forest Meteorology.Conforming to these where possible/practical will improve the readability and reach of the manuscript.Also, both Fre (pg.3, line 31) and ER (Fig. 1) are used in the manuscript to represent ecosystem respiration, and both Fc (pg.6, line 14) and NEE (Fig. 1) are used to represent net ecosystem exchange of CO2.
• Thanks again for the Reifsnyder et al. (1991) paper this will be a useful reference source in the future.
Unfortunately the terminology this paper derives from the OzFlux terminology used in the initial processing in OzFlux QC and this has transferred to this paper to be consistent.The terminology is defined there and has been published in the OzFlux QC special issue paper so we will maintain a consistent convention.We will however ensure that the use of terms are consistent throughout the manuscript.
-Pg. 5, line 17-20 & Pg. 6, line 9-11: Is there a threshold correlation below which the data are not gap-filled?A discussion of this topic is warranted.

•
There is currently no threshold, below which the data are not gap filled but this does raise a possible issue.We looked back through the correlation values for each of the meteorological variables across a range of sites and across them the 'best' correlation was never less than 0.5 (which is more than acceptable).The DINGO tool does take the three nearest weather stations as well as the gridded meteorology to do the correlations and so amongst all of those sources then ranks the 'best' correlations.As a result the 'best' correlation is always adequate.However, the 'worst' stations can have correlations of less than 0.1.We have added a description of this to the text.
-Pg. 5, line 17-20: Linear regression is a good start for using nearby station data for gap-filling.However, even from Fig. 2a (which looks to be a very clean example), the best correlation has a distinct non-linear component at the low end of values.Using this relationship to gap-fill the time series then extends the data into minimums not actually observed at the site, which has the potential to influence down-stream modeled physiological responses.Perhaps the manuscript could address this topic as a discussion point, and/or include a few different fitting functions in future iterations (maybe piecewise linear fits?).

•
The reviewer raises a good point and one which we have previously considered.It would be easy to improve the statistical fit by increasing the complexity of the model from linear to piecewise linear to polynomial to machine learning.However, we feel that without any a priori reason for thinking that the relationship should be non-linear, then it is a slippery slope to go down to chase the best statistical fit particularly when this could result in non-physically realistic relationships between the two variables.We have added the following text "It would be easy to improve the statistical fit by increasing the complexity of the model from linear to piecewise linear to polynomial to machine learning.Vuichard and Papale (2015) have shown the limits of a linear regression correction method when applying a bias correction on meteorological fields for which the bias between flux tower and gridded meterology (ERA-I) data did not vary linearly.Nevertheless, we feel that without any a priori reason for thinking that the relationship should be nonlinear, then it may be unwise to chase the best statistical fit particularly when this could result in non-physically realistic relationships between the two variables." -Pg. 8, line 9-12: It would be informative to show some stats addressing how well this procedure replicates actual variation in solar radiation under cloudy conditions across the OzFlux network (opposed to using nearby BoM site data).What about using the diurnal average approach employed for estimating incoming longwave radiation under cloudy conditions (pg. 9, line 6-8)? • The Bureau of Meteorology insolation data we used are daily estimates as described in the manuscript -there are few ground-based BOM monitoring sites with half-hourly measurements of insolation (less than 10 across Australia), and none within a reasonable distance of the flux towers in the OzFlux network.The methodology described in the paper is simply a means of first correcting the daily estimates of insolation using the tower observations (where available), then downscaling these to halfhourly values.

•
The BOM satellite-and model-derived estimates compare very favourably with daily cumulative insolation measured at the sites (r 2 = 0.9+), and this could easily be included in the manuscript, but is generally discussed in the literature cited.Cloudiness is more problematic when downscaling from daily to (e.g.) half-hourly interval.The smooth insolation curve produced by the described algorithms is unlikely to reliably be realistic relative to ground-based measurements of insolation (whichdepending on cloud type -would be much more temporally variable) simply because we have no information about temporal variations on this scale.

•
We could instead use the approach taken with incoming longwave radiation, but if we averaged the diel cycle over multiple days with similar insolation, here we would approach a smooth curve anyway (and the remaining variation would be meaningless because the amount of cloud cover over short periods is inherently unpredictable, at least deterministically).Moreover, since a cloudy summer day may have daily insolation levels similar to a clear winter day, the ranking algorithm may group and average days from very different times of year.This would have the effect of smearing sunrise and sunset times undesirably, and generally reducing the accuracy of insolation estimates generally because of the differing path length (and corresponding extinction) of the solar beam through the atmosphere at different times of year.
-Pg. 10, line 4-5: How was the ANN hidden layer architecture of 24 and 16 nodes arrived at?Appropriate model complexity can significantly impact ANN model performance.Recent related works using ANNs for gap-filling have tested for ideal architectures (Papale and Valentini, 2003;Knox et al. 2015;Baldocchi et al., 2016).
• There is a compromise in the ANN model between statistical performance, training and over fitting.We simply used trial and error (through a systematic change in the range of parameters in the ANN model).Importantly we examined the performance of the neural network model across three different temporal ranges (diel, seasonal and annual).Many ANN combinations can achieve a good statistical performance across the mean range but be very poor at capturing the diurnal variation, for example.We have optimised the parameters to these three scales.Will have added this to the text.
-Pg. 10.ANN procedure: How are the data split into training and testing/validation sets?

•
The data are split into 80% training and 20% testing.This has been added to the text.
• What is meant is that simply, once the Fc timeseries has been filtered for low u* values it is gap filled using ANN Fre model and the resultant output is labelled as Fc_ustar.We have clarified this in the text.
-Pg. 12, line 16: What does "ve" mean?(as in +ve, -ve) • Positive and negative.We have expanded this in the text.
-Pg. 12, line 17-22: This explanation is confusing.Recommend rephrasing.Consider explaining that forcing GPP to zero at night removes positive and negative random error equally, but forcing any positive GPP values to zero during daytime would bias results because only negative random errors would remain.Also, I am not sure what is meant by "no random error in the Re ANN calculation", since random error is relevant in any model fit to real data.

•
Will have amended to read: 'GPP is equated to zero at night but is not forced to zero during the day.
Since GPP is the difference between measured net ecosystem exchange and estimate ecosystem respiration, it incorporates random error that is superimposed on the measurements (and potentially also systematic error in the model), and may be correspondingly higher or lower than 'true' GPP.As a result, some GPP estimates may switch to positive sign when the signal:noise ratio is low (e.g.early morning / later afternoon).While these estimates are therefore unphysical, the effect of their removal is to filter the positive domain of the random error distribution, thereby converting random error to systematic error.' -Pg.14, line 16-27: Since the ANN is employed for gap-filling, why not use the model residuals as an estimate of the random error?The model residuals of high performance gap-filling algorithms such as the ANN provide a good, if not conservative estimate of the random uncertainty (Moffat et al. 2007, Richardson et al. 2008).The daily differencing approach is muc more conservative as it includes natural environmental variability as a result of variation in the flux footprint.This would help alleviate some of the double-counting of uncertainty mentioned in a later paragraph (pg.15, line 7-12).
• Explicitly accounting for and interpreting the effects of radial variations in the flux source area requires relatively complex site-specific analysis that is presently considered to be beyond the scope of the development of this tool.Thus we would prefer to use a more conservative approach to uncertainty analysis that does include such effects, though we could be explicit in our reasoning in the manuscript.• Will be amended.
-Pg. 14, line 18-21: Why must random error be calculated over bins of u*? (rather than over bins of (or a regression with) flux magnitude) -

•
Will be amended to read: 'Since random error is heteroschedastic, its magnitude (expressed by the standard deviation or σ[δ], since the mean is expected to be zero) must be expressed as a function of the mean flux magnitude.'Section 2.6 Uncertainties: This section would benefit from a discussion on the uncertainty propagated along the entire from a discussion on the uncertainty propagated along the entire processing chain described in the manuscript (i.e. using gap-filled ancillary drivers value to gap-fill fluxes).
• Agreed.This is an area of expansion of the uncertainty estimation planned for the future.
-Pg. 17, line 16-18: What is the suggested turnaround time from data collection to quality output from the procedure outlined in the manuscript?

•
Dingo will process the level 3 data from the OzFlux QC through to fully gap filled and partitioned data in less than 30 minutes on a desktop PC.When the Barr et al.Methodology is used for computing u* this procedure adds another hour or so depending on the length of the dataset and the number of iterations in the bootstrapping.In addition, the uncertainty code adds another hour or so depending on the number of iterations chosen and the length of the dataset.This can be added to the manuscript.
Technical corrections -All bulleted lists: In some cases the lack of sentence case following the bullet point makes sense.However, in many cases the sentence following the bullet is a standalone sentence and should use sentence case.
• Have corrected this.
-Pg. 13, line 26: I am confused by the sentence fragment "Once complete daily time series for the parameter estimates,. ..".What about this: "Once daily estimates for the parameters are generated,. .."

•
The referees suggestion is clearer and have amended this in the text.
-Labels in figures need cleaning up (e.g.Fig. 4 y-axis -what is CABLE?, Fig. 5 y-axis -make this human readable, Fig. 9b legend -what is the difference between Fc and Fc?, use superscript and subscripts where appropriate)

•
We have cleaned up the figures to have appropriate units.We have removed the titles from the plots in each figure.CABLE is the other name for the BIOS model used for soil gap filling but we have removed this from the figure.

•
In the legend of figure 9 the two colours are as stated in the caption "weekly plot with period of more than 30% missing data shown in light blue".We will tidy this up and make the legend clear.
-The manuscript needs thorough editing for typographical errors

Will do
called the Dynamic INtegrated Gap-filling and partitioning for OzFlux (DINGO) that enables gap-filling and partitioning of the primary fluxes into ecosystem respiration and gross primary productivity and subsequently provides diagnostics and results.
We outline the processing pathways and methodologies that are applied in DINGO (v12av13) to OzFlux data including 1) gapfilling of meteorological and other drivers; 2) gap-filling of fluxes using artificial neural networks; 3) the u* threshold determination; 4) partitioning into ecosystem respiration and gross primary productivity; 5) random, model and u* uncertainties and 56) diagnostic, footprint calculation, summary and results outputs.Opportunities remain for DINGO to incorporate robust measurements of uncertainty for application in land management and carbon accounting.In addition, footprint information is crucial in understanding and interpreting the scale and spatial influence of flux measurements.Both these features are scheduled for the next release (v13) but are detailed here.DINGO was developed for Australian data but the framework is applicable to any flux data or regional network.Quality data from robust systems like DINGO ensure the utility and uptake of the flux data and facilitates synergies between flux, remote sensing and modelling.

Introduction
OzFlux is the regional Australian and New Zealand flux tower network that aims to provide a continental-scale national research facility to monitor and assess Australia's terrestrial biosphere and climate for improved predictions (Beringer et al., 2016).High quality and reliable data are a crucial foundation in achieving the objectives of the OzFlux network (Beringer et al., 2016) and underpin the process understanding needed to: 1) support sound management of natural resources including water, carbon and nutrient resources for environmental and production benefits; 2) monitor, assess, predict and respond to climate change and variability; 3) improve weather and environmental information and prediction; 4) support disaster management and early warning systems needed to meet Australia's priorities in national security; and 5) ensure that Earth system models used to underpin Australia's policies and commitments to international treaties adequately represent Australian terrestrial ecosystem processes (Beringer et al., 2016).Beringer et al. (2016) provide an overview of the evolution, design and current status of OzFlux as well as a brief summary of the instrumentation and data collection that forms the backbone of the network.A detailed description of the quality control and post-processing of eddy covariance data using OzFluxQC and the data pathway to curation is given by Isaac et al. (2016).
In summary from Beringer et al. (2016), most sites have data loggers that provide the average (usually over 30 minutes) covariances that are processed through 6 levels using the OzFluxQC standard software processing scripts.Levels 1, 2 and 3 are as follows; L1 -the raw data as received from the flux tower, L2 -quality-controlled data, and L3 -post-processed, corrected but not gap-filled data.Quality control measures by OzFluxQC are applied at L2 and comprise checks for plausible value ranges, spike detection and removal, manual exclusion of date and time ranges and diagnostic checks for all quantities used in the flux correction calculations.The quality checks make use of the diagnostic information from the sonic anemometer and the infra-red gas analyser.For sites calculating fluxes from the averaged covariances, post-processing includes 2-dimensional coordinate rotation, low-and high-pass frequency correction, conversion of virtual heat flux to sensible heat flux and application of the WPL correction to the latent heat and CO2 fluxes (see Burba (2013) for general description of the data processing pathways).Steps performed at L3 include the correction of the ground heat flux for storage in the layer above the heat flux plates (Mayocchi and Bristow, 1995) and correction of the CO2 flux data for storage in the canopy (where available).
OzFlux data are available at http://data.ozflux.org.au/.OzFlux sites submit their data to FLUXNET at L3.
Given the international need by the community for standardised data processing to enable effective comparison across biomes and to understand inter-annual variability (Papale et al., 2006), we have developed a software tool to address this need.In this paper we describe the development and testing of the Dynamic INtegrated Gap-filling and partitioning for Ozflux (DINGO) system that utilises the L3 data from OzFluxQC to gap-fill and partition the fluxes in ecosystem respiration (Fre) and gross primary productivity (GPP) and subsequently provides diagnostics and results.This paper is not intended to be a thorough review of the data processing but the application of standard techniques in DINGO for the flux community.DINGO is a research version for OzFlux data whereas the OzFluxQC system (which has many similar features to DINGO) is considered an operational version.Quality data from robust systems like DINGO ensure the utility and uptake of the flux data and facilitates synergies between flux, remote sensing, modelling and canopy physiological studies.We conclude by looking ahead at the future direction of the DINGO system.

Approach
The overall approach used in DINGO is to take the L3 OzFluxQC data, which has gaps from data processing (data excluded due to values out of range, spike detection or manual exclusion of date and time ranges) or from site issues (instrument or power failure, herbivores, fire, eagles nests, cows, lightning, PI on sabbatical, etc.) and gap-fill and partition the data using a variety of data sources (Fig. S1).DINGO is programmed in python 2.7 and is currently at version 12a 13 and publically available on GitHub (https://github.com/jberinge/DINGO12https://github.com/jberinge/DINGO13).It should be noted that v13 is scheduled to be released in July 2016 and will include uncertainty and footprint analysis and these are documented in sections 2.6 and 2.7, respectively.It is designed to work with OzFlux data produced in NetCDF format by the OzFluxQC (Isaac et al., 2016) and draws on Australian AWS data but could be adapted for other data sources across other flux sites.The primary interface for the user is through a text based control file that has information on site characteristics (name, latitude and longitude, the frequency of the flux measurements (30 or 60 minutes) and elevation), file paths (to the OzFluxQC NetCDF files and other ancillary data inputs), data processing options and data plotting and output formats.In general, prior to the processing steps below, any gaps in fluxes or meteorological quantities of less than two hours are filled by DINGO using linear interpolation.The pathway for processing is shown in Fig. S1 and each aspect step is outlined in the supplemental material for technical note: Dynamic INtegrated Gap-filling and partitioning for OzFlux (DINGO)below through each section.

Conclusions
The OzFlux network has been highly successful in generating standardised measurements and protocols that provide robust primary data.Only via transparent, advanced and consistent QA/QC will we ensure compatibility within the OzFlux network (Beringer et al. 2016) and with international databases (FLUXNET) (Papale et al., 2006), ensuring uptake by the broader scientific community.Through robust software systems such as OzFluxQC (Isaac et al., 2016) and DINGO we are able to ensure timely and quality gap-filling and partitioning of fluxes that in turn enables a significant uptake of the eddy covariance data for application to a range of research questions as exemplified in Beringer et al. (2016).This includes integration of eddy covariance and remote sensing datasets for the validation of satellite products (e.g.GPP and ET (Kanniah et al., 2009;Restrepo-Coupe et al., 2015)) and to aid the parameterisation of models that rely on remotely sensed data (e.g.GPP, ET, canopy conductance, and light use efficiency (LUE) (Barraza et al., 2014(Barraza et al., , 2015;;Glenn et al., 2011;Goerner et al., 2011)).In addition, OzFlux data have been instrumental in constraining a continent-wide assessment of terrestrial carbon and water cycles (Haverd et al., 2013) and featured in the development of new models (Haverd et al., 2007(Haverd et al., , 2009)).There is utility of the data to support carbon accounting type activities (Hutley et al. 2005), as demonstrated in research focussed on the conversion of savanna to pasture (Bristow et al. 2016).However, as for many applications like this, a measure of uncertainty of the fluxes are required (Schmidt et al., 2012) and we see this as a major opportunity for further development of DINGO.In addition, the integration of flux data must have due consideration for scale and the associate footprint of the flux measurements for correct interpretation.We also see this as a feature that would be useful for users and therefore we also plan to incorporate into V13.
Ultimately flux data are required to address the key ecosystem science questions of OzFlux (Beringer et al., 2016)  Pathway for processing is shown in Fig. S1 and each aspect is outlined below through each section below.

2S1.1.1 Gap-filling of meteorological drivers
In order to produce a continuous time series of fluxes and meteorological drivers for carbon accounting and other uses (Hutley et al., 2005) the first step is to generate a continuous time series of meteorological drivers that is in turn used for gap-filling of fluxes.In general, DINGO gap-fills missing meteorological data based on multiple data sources that include Bureau of Meteorology (BoM) operated automatic weather stations (AWS) and spatially gridded meteorological data at 0.1 degree resolution (Jones et al., 2009;Raupach et al., 2009).It then and chooses the 'best' source based on the best linear regression correlation with available site data.It uses BoM AWS data from nearby stations to create a constructed time series and then compares this with the same regression from gridded meteorological data.Given the relatively low density of AWS's across Australia, a station may not be representative of the flux tower location or may be too far away and therefore may not be well correlated with local meteorology.In this case the correlation will be low and if the gridded data has a better correlation it will be used in preference.This procedure is repeated independently for each meteorological variable (Fig. Fig. S1).Correlation values for each of the meteorological variables across a range of sites showed the 'best' correlation was never less than 0.5 (which is more than acceptable) and therefore there is no threshold, below which data are not gap filled.The general processing of AWS data is as follows: 1. search for the nearest BoM AWS stations and check to see if the station has data available for the same time period needed for the flux gap-filling and ensure that the AWS has the required meteorological variables,  4. the frequency at which to perform the correlation analysis between flux tower and AWS can be set to 'all' the available data or done by 'year' or 'month', and 5. then outputs and saves the 'best' AWS time series and creates plots.
The same process is repeated for the BoM spatially gridded data to obtain a correlation and linear regression with local meteorology.It would be easy to improve the statistical fit by increasing the complexity of the model from linear to piecewise linear to polynomial to machine learning.Vuichard and Papale (2015) have shown the limits of a linear regression correction method when applying a bias correction on meteorological fields for which the bias between flux tower and gridded meteorology (ERA-I) data did not vary linearly.Nevertheless, we feel that without any a priori reason for thinking that the relationship should be non-linear, then it may be unwise to chase the best statistical fit particularly when this could result in non-physically realistic relationships between the two variables.Gap-filling of the flux tower meteorological data then proceeds as follows: 1. the linear regression equation from the best available time series (based on the best correlation (AWS or gridded)) is used to adjust the time series to best match the flux site meteorology.The new 'correlated' variable is named {variable}_Corr (e.g.Ta_Corr).This variable is the best alternate meteorological time series adjusted by the linear regression with site meteorological data, 2. offset the best available time series by 30 minutes if it is from the BoM which has timestamps that are for the proceeding period, whereas flux tower dataloggers have timestamp at the end of the period, 3. missing meteorological data are then gap-filled using the {variable}_Corr time series and the new 'constructed' time series is saved as {variable}_Con (i.e.Ta_Con).An associated QC flag is generated {variable}_Con_QCFlag.If valid observations are present then they are used and the flag set to 1.If the data are gap-filled then flag set to 100, and 4. for each variable if at the end of all this there are still missing data then they are filled using a climatology approach using monthly averages of the diurnal time series for the variable and given a QC flag of 97.
As an example, for air temperature (Ta), we see the correlations of site Ta with the best nearby AWS (Fig. (8) Where h is the elevation above sea level (m).The extinction coefficient (k) is the only remaining unknown quantity in Eq. 1, and is optimised using the site observations.We first generated an upper envelope for daily insolation -taken to represent clear sky conditions -defined by selecting maxima for successive (non-overlapping) 14-day windows from the gap-filled daily time series (in the tropics it is necessary to exclude the wet season from this procedure, since substantial cloud mass routinely builds in the afternoons).The value of k is then optimised by minimising the sum of squares error between the envelope estimate and daily insolation totals calculated using eq. 1.The SciPy package of the Python programming language is used for the optimisation procedure.
Thus a clear sky estimate for each half-hourly period is obtained from the above procedure.Since no information about subdaily variations in cloudiness is available, here it is assumed that cloudiness is constant throughout the day.As such, the clear sky half-hourly estimate for a given day is simply reduced by the ratio of the daily total BOM-estimated insolation to the daily total calculated clear sky insolation.
While k is likely to vary seasonally, here a single value is used for the sake of simplicity.By using site data to constrain estimates of k, a more accurate diurnal course of Fsd is obtained.This is crucial, because as the energy source for photosynthesis, insolation is routinely used as a driver in models that estimate daytime carbon fluxes.The calculation of theoretical site clear sky estimates of solar radiation also allows an estimate of cloudiness to be obtained.In turn, this is an important term in the determination of incoming long wave radiation, as described below.

S12.2.2 Incoming long wave radiation
Similar to insolation, incoming long wave radiation (Fld) is calculated in a two-step process.The first step is to estimate daily radiation totals.This s done by using the Stefan-Boltzmann relation to calculate daily average Fld from screen-level air temperature, as follows (Oke, 1987): Where ε is emissivity, σ is the Stefan-Boltzmann constant (5.672 x 10 -8 W m -2 K -1 ) and T is temperature (K).While Fld is received from all parts of the sky, measurements of water vapour (which influences ε) and temperature are generally only available at a discreet point near the surface.The clear sky formulation of Brutsaert ( 1975) is used to take this into account by making approximating assumptions about average lapse rates of temperature and water vapour profiles based on surface values (Crawford and Duchon, 1999), and has been found to perform well under clear conditions (for example see Duarte et al. (2006)).Its general form is as follows (Burman and Pochop, 1994): found in (Moffat et al. (2010) and Bryant and Shreeve (2002).There is a compromise in the ANN model between statistical performance, training and over fitting.We simply used trial and error (through a systematic change in the range of parameters in the ANN model).Importantly we examined the performance of the neural network model across three different temporal ranges (diel, seasonal and annual).Many ANN combinations can achieve a good statistical performance across the mean range but be very poor at capturing the diurnal variation, for example.We have optimised the parameters to these three scales.We implemented the ANN as follows: 1. gap-filling performed using the ffnet package (http://ffnet.sourceforge.net/)that is a fast and easy-to-use feed-forward neural network.The basic model we implemented is a multiple layer network with two layers of 24 and 16 nodes respectively, which creates a multilayer network that is fully connected.Training is done using a truncated Newton algorithm (TNC) to minimise a function with variables subject to bounds using gradient information.The type of ANN model and training algorithms can be altered in the code.We use 80% of the data for training and 20% for testing, 2. the ANN is passed a list of inputs and outputs (which can be defined) as well as a number of iterations that are configurable in the control file (about 500-1000 is satisfactory).Additional iterations takes longer and has a tendency to over fit the model to the data.For the energy balance, the targets are processed together (Fg, Fe and Fh) and the ANN inputs are incoming solar radiation (Fsd), vapour pressure deficit (VPD), soil moisture content (Sws), Ts, Ta, and MODIS EVI (250m resolution 16 day).For Fc the inputs are Fsd, VPD, Sws, Ts, WS and EVI.DINGO will add a storage term to Fc if a profile system is present and this is a user option.Finally, the inputs for Fre are Sws, Ts, Ta and EVI.The Fre ANN is trained only using data above a u* threshold (see Section 2.3).The performance of the ANN also increased when using EVI as this provides surrogate information of vegetation activity (i.e.LAI and growth) and we hypothesise that that this equivalent to a measure of autotrophic respiration, 3. the ANN returns a model predicted time series of data that is saved as {variable}_NN.The tower data is gap-filled using these time series and a new constructed time series is generated {variable}_Con.A QC flag is assigned as {variable}_Con_QCFlag = 1 if there is valid data from the tower, else = 99, 4. the gap-filling frequency is configurable and can be changed to either use all data at once (ALL) or divide it into groups to process each group independently.The grouping can be all, annual, monthly OR a variable that is categorical.For example a site may be under different management at different times and the ANN can be performed separately on those periods.At some sites in the arid zone (Cleverly et al., 2013) a single ANN for the entire period does not work well so there needs to be an option to break broken down the data into monthly chunks, and 5. the code module also outputs many diagnostic plots including general model performance (Fig.

S12.4 Friction velocity (u*) filtering
The eddy covariance technique is well known to underestimate turbulence fluxes of carbon dioxide under atmospherically stable conditions, particularly at night time where the surface can be decoupled from the measurements at a height above the canopy (Goulden et al., 1996).An excellent overview of this subject is given by Aubinet et al. (2012).This problem has been shown to impact fluxes across some Australian sites such as the old growth Mountain ash site (Kilinc et al., 2010) and a cool temperate eucalypt forest (van Gorsel et al., 2009).The primary technique to deal with this is to exclude data taken where the eddy covariance measurements is not representative of the true flux.Typically this occurs when u* values are below a critical threshold (Goulden et al., 1996).There are several ways to calculate the friction velocity (u * ) threshold as shown in Aubinet et al. (2012) and in DINGO we calculate the threshold based on both the procedures of (1) Reichstein et al. (Reichstein et al., 2005a) and ( 2) Barr et al. (2013).Alternatively the user may select their own constant value.The threshold used for subsequent filtering is user selectable but the threshold determined using Barr et al. (2013) is used by default.Whatever choice is made the resulting u* that is used is saved to the main file named 'ustar_used'.The two methods are implemented as follows: 1.For for the Reichstein et al. (Reichstein et al., 2005a) approach, the non-gap-filled data set is split into 6 temperature classes of the same sample size (according to quantiles) and for each temperature class the set is split into 20 u*classes according to Papale et al., (2006).The threshold is defined as the u*-class where the night-time flux reaches more than 95% of the average flux at the higher u*-classes.The threshold is only accepted if the temperature class and u* are not or only weakly correlated (|r| < 0.3).The final threshold is defined as the median of the thresholds of the (up-to) six temperature classes.This procedure is applied to the entire dataset, giving a maximum, but conservative u* threshold (Fig. Fig. S7).The maximum value is saved as 'ustar_Reich_max'.In addition, the u* threshold is calculated continuously using a 1 month moving window to account for seasonal variation of vegetation structure (Fig. Fig. S7) and saved as 'ustar_Reich_var', and.
2. The the Barr et al. (2013) approach uses a change point detection technique to objectively identify the best estimate and uncertainty range for the u* threshold.In brief, the method involves fitting a two-phase linear regression model to all possible change points within a nocturnal data sample (i.e. 2 ≤ c ≤ n-1), finding the change point that minimised the sum of squared error and establishing whether its performance was statistically significantly improved relative to a reduced form (no change point) null model.For each year, the data were divided into sequential samples of n = 10 3 , with 50% overlap between samples.Each sample was in turn divided into four temperature classes, ordered by u* then bin-averaged (n = 5 for each bin) to reduce the effects of random error.To increase the sample size, the data were bootstrapped (n = 10 3 ) by simply randomly sampling (with replacement) records from the original dataset and rerunning the analysis.This yielded a (Gaussian) distribution of u* thresholds, the mean and 95% confidence interval of which were taken as the best estimate and uncertainty of the u* threshold.As per Barr et al. (2013) we identified the dominant mode of the NEE dependency on u* (i.e.positive or negative slope below the change point), and rejected all thresholds from the non-dominant mode (in practice the negative dependency slope was very rare), and rejected any annual analysis where the number of valid change points (across all temperature strata and bootstraps) was less than 4000 or less than 20% of the total.Annual u* statistics are saved and the annual Barr u* threshold is written to the main file as 'ustar_Barr',.

S12.5.1 ANN modelling of Fre
Once the data have been u* filtered, they are used to train an ANN (see Section 2.2) using nighttime data only with inputs of Sws, Ts, Ta and EVI as known drivers of ecosystem respiration (Migliavacca et al., 2010).Importantly, in DINGO we also only use flux data from the first 3 hours after sunset where the canopy is still coupled with the atmosphere, as shown in Van Gorsel et al. (2007).This makes the selection of data for the ANN model more conservative than using the entire nighttime period.This option is also user selectable.Nighttime, daytime and evening periods (first 3 hours after sunset) are determined from sunset and sunrise times calculated precisely from solar geometry accounting for elevation, latitude, longitude and day of year and saved as an integer 'day_night' (day=1, evening=2, night=3).
Once the Fc timeseries has been filtered for low u* values it is gap filled using ANN Fre model and the resultant output is labelled as Initially, the Fc timeseries has nighttime gaps from the u* threshold filtering and these are filled using the ANN Fre model with the u* corrected output labelled as 'Fc_ustar'.Next, the ANN model output trained using nightime data is then extrapolated to the daytime data to create a continuous timeseries of modelled Fre (Fre_NN) that is used to fill missing gaps in the nighttime data where data where either filtered out using the u* threshold or where missing.This creates a continuous time series of Fre (Fre_Con) that is a combination of valid observations and ANN model output.GPP is then calculated as the difference between Fc and Fre (GPP_Con) where our sign convention is -venegative is a net flux into ecosystem and +vepositive away from it.GPP is equated to zero at nighttime but it is not forced to zero at any other time.
This is because for any 30 minute value, the random error in Fc results in values of Fc being higher or lower than the true value.Due to this error, it is possible then to get values of GPP that have either a higher or lower random error component.GPP is equated to zero at night but is not forced to zero during the day.Since GPP is the difference between measured net ecosystem exchange and estimate ecosystem respiration, it incorporates random error that is superimposed on the measurements (and potentially also systematic error in the model), and may be correspondingly higher or lower than 'true' GPP.As a result, some GPP estimates may switch to positive sign when the signal:noise ratio is low (e.g.early morning / later afternoon).While these estimates are therefore unphysical, the effect of their removal is to filter the positive domain of the random error distribution, thereby converting random error to systematic error.'There is also no random error in the Re ANN calculation and if we force these GPP values to zero then there is a bias in the GPP integrals (e.g.daily totals) because we only remove the values that have a negative error.Forcing GPP to zero during the daytime leads to ever so slightly higher GPP Formatted: Font: 10 pt because lower values are cut off particularly in the evening and morning when GPP approaches zero.We feel that it is not appropriate to leave the positive and negative GPP in the night because although the random error at nighttime can be positive or negative the absolute value of GPP could also be significantly non-zero.This could occur at nighttime due to biases in the measurement of Fc and calculation of Fre and therefore GPP would have similar biases that compound.So the absolute value of GPP at night includes a bias and random error and since we know physiologically that GPP must be zero we feel that the most robust GPP at nightime is to force it to zero.

S12.5.2 Separation of Fc using a light response curve approach
An additional approach is also applied where a simple parametric approach to the imputation of Fc is undertaken in DINGO, and largely follows the approach of Lasslop et al. (2010).In this approach, Fc is modelled as the sum of gross primary production (GPP) and ecosystem respiration (Fre): A Michaelis Menten-type simple rectangular hyperbolic light response model (Ruimy et al., 1995) of modified form (as per Falge et al. (2001)) is used in conjunction with the Arrhenius-style temperature response function as proposed by Lloyd and Taylor (1994), such that GPP and Fre are replaced by: Where α is the initial slope of the photosynthetic light response, Q is photosynthetic photon flux density, and β is photosynthetic capacity at 2000 μmol photons m -2 s -1 , rb is the reference respiration at a reference temperature (Tref -here set to 10 o C), Eo is an activation energy parameter that determines the function's temperature sensitivity, and T0 is the temperature at which metabolic activity approaches zero.Lasslop et al. (2010) proposed an additional criterion -adopted here -to take account of the effect of vapour pressure deficit (VPD) on stomatal conductance -and thereby photosynthetic capacity -in which there is a non-linear decline in photosynthetic capacity once VPD exceeds a given threshold, as follows: As has been widely reported, even in isolation the unconstrained temperature response function is over-parameterised (Lloyd and Taylor, 1994;Reichstein et al., 2005b;Richardson and Hollinger, 2005).Thus T0 and Tref are fixed at -46.02 o C and 10 o C, respectively, as per Lloyd and Taylor (1994).Eo is fitted nocturnally using the data for each year, as per Reichstein et al. (Reichstein et al., 2005a).The default fitting window for the remaining parameters is 15 days and the default time step is 5 days (missing dates are then interpolated to generate a complete time series of daily parameter estimates), but both of these Formatted: Font: (Default) +Body (Times New Roman) Formatted: Font: (Default) +Body (Times New Roman), 10 pt Formatted: Font: (Default) +Body (Times New Roman), 10 pt can be configured (minimum of 1 day, maximum of 30 days in each case).The quality control scheme for parameter acceptance used by Lasslop et al. (2010) was adopted for DINGO.
The user can make two choices: i) optimise all remaining parameters (α, β, k, rb) using the daytime data alone, or; ii) optimise rb using nocturnal data and optimise light response parameters (α, β, k) using daytime data.The first choice may be selected where difficult nocturnal conditions limit the amount of data available for robust respiratory parameter estimation.However, simultaneous optimisation of all parameters may result in unrealistic estimates when signal magnitude is weak.Random error in eddy covariance data is heteroschedastic (i.e.random error magnitude increases with flux magnitude), and is non-zero at zero flux (Richardson and Hollinger, 2005).Thus signal: noise is likely to be lower for low productivity ecosystems, which are common in Australia due to soil moisture and nutrient limitation.As such, the second choice may be more appropriate for many sites.
Once daily estimates for the parameters are generatedOnce complete daily time series for the parameter estimates, eq.14 is used in conjunction with gap-filled driver data to calculate half-hourly values for Fc, GPP and Fre that are named as Fc_Lasslop, GPP_Lasslop and Fre_Lasslop, respectively.The drivers used are also user configurable -for example, either incident or absorbed photosynthetically active radiation can be used for GPP parameter estimation and calculation, and soil or air temperature (or a weighted combination thereof) can be used for Fre.

S12.6 Uncertanities
In version 12a and 13, the estimates of uncertainty include firstly a simple sensitivity to changing the chosen u* threshold above (130%) and below (80%, 60%, 40%, 20% and no threshold) the determined threshold.The carbon fluxes (GPP, Fc and Fre) are recalculated as above and annual sums given for each of the threshold manipulations and saved as a separate output.
In version 13, DINGO also calculates a subset of annual Fc uncertainties associated with: 1) u* threshold estimation error; 2) random error, and; 3) imputation error and these are described below: 1. the most straightforward approach to propagating u* threshold uncertainty to annual Fc is simply to run the gap-filling procedure for Fc after filtering the nocturnal data using the upper and lower bounds of the 95% confidence interval for the threshold derived from change point detection (see section 2.4).However, because a much larger proportion of nocturnal data is removed when filtering for the upper bound than for the lower bound, the proportions of observational and gap-filled data necessarily change, and the random and model error uncertainty contributions thus vary depending on the u* threshold.For this reason, a full Monte Carlo-style simulation (described below) is required to simultaneously account for all of the above error sources,.
Formatted: Tab stops: 3.4 cm, Left 2. we adopted the daily differencing approach of Hollinger and Richardson (2005) where δ is the random error derived from the daily differencing procedure.This method assumes that when differences in critical drivers are sufficiently small (<35 W m -2 for insolation, <3 o C for air temperature, and <1 m s -1 for wind speed), differences in estimates of Fc separated by 24 hours represent random error.Since random error is heteroschedastic, its magnitude (expressed by the standard deviation or σ[δ], since the mean is expected to be zero) must be expressed as a function of the mean flux magnitude.Since random error is heteroschedastic, its magnitude (expressed by the standard deviation or σ[δ], since the mean is expected to be zero) must be calculated for ranges of u*, and expressed as a function of the mean flux magnitude over that range.Thus a linear relationship between σ(δ) and flux magnitude is generated.Since the random error distribution for eddy covariance data is Laplacian rather than Gaussian, a random error estimate for a given datum can be generated from a Laplacian random error distribution with location parameter of zero and scaling parameter of σ(δ) / √2.To compound the uncertainty on each point to annual Fc, DINGO runs a Monte Carlo-style simulation: for each of 10 4 trials (for each year in the dataset), a realisation of random error is generated for the entire observational Fc time series, and the annual sum of the observational data is calculated.The 95% confidence interval for the resulting distribution of Fc sums is returned as the uncertainty due to random error, and.
3. DINGO adopts the approach of Keith et al. (2009), separately for day and night conditions, a sub-sample of 10 3 observations is randomly selected from the annual dataset.Gaps are then introduced into this dataset such that the proportion of missing data is equal to the observed proportion of missing data annually.The missing data is filled with model estimates, and the percentage difference between the complete observational and gap-filled time series' is calculated.As with random error, this procedure is run 10 4 times.The 95% confidence interval for the distribution of the percentage difference estimates is calculated, and these percentages are applied to the annual sum (for example, if the 95% CI is -5% to +5%, then a carbon sink of 500 gC m -2 a -1 has a model-induced uncertainty of ± 25 gC m -2 a - 1 ).
To combine uncertainties for random and model error alone, we assume their independence and sum in quadrature: Where εtot, εr and εm are combined total, random and model uncertainty, respectively.It must be emphasised that this is not strictly valid, since the model error uncertainty quantification above necessarily includes some effects of random error (since the observational input data includes random error).While it is possible to separate model and random error components (see Dragoni et al. (2007), for example), this requires an accurate estimate of random error.It has been noted elsewhere (Billesbach, 2011;Dragoni et al., 2007;Hollinger and Richardson, 2005) that the above method for deriving random error tends to overestimate, since some signal is likely to be included with noise during the differencing procedure, and wind direction is neglected (whereas the source-sink may be azimuthally variable).We argue that since this is likely to result in overestimation of error, the approach is conservative.

2 .Formatted
obtain the 10 closest sites and extract the data accounting for different data formats.DINGO uses the Pandas library (http://pandas.pydata.org/)predominately for data manipulation.Here we do time stamp management and all processing is done based on local time.DINGO then looks for duplicates and or missing data and deals with 'NaNs' and QC flags from AWS data, correlate each AWS site with the available flux tower meteorology and use data from the best three out of ten sites tested, Fig.S2a) and the final gap-filled Ta series.This is repeated for absolute humidity (Ah), wind speed (WS), atmospheric pressure (P) and precipitation.The plots provide excellent diagnostics as the flux tower data is compared against the best available meteorology and this can detect anomalies such as instrumentation errors or processing errors.As an example in fig.fig.S3we see that a range threshold had been set incorrectly in OzFluxQC by the user and that accidently excluded data with an absolute humidity below 5 g m -3 . =  −0.0001184 ℎ cos  + 0.50572(96.07995− 180  ⁄ ) −1.6364 Fig. S6a), monthly time series (Fig. Fig. S6b), a 30 minute timeseries data check (Fig. Fig. S6d) and a check to see if the ANN performs diurnally (Fig. Fig. S6c).

Figure S1 :
Figure S1: Overview of processing pathways in Dynamic INtegrated Gap-filling and partitioning for Ozflux (DINGO) from level 3 OzFluxQC data to gap-filled and partitioned outputs.Where BoM is the Australian Bureau of Meteorology, AWS is Automatic5

Figure S2 :
Figure S2: An example of gap-filling meteorological variable (air temperature (Ta -o C) in this case) using nearby Bureau of Meteorology automatic weather stations (AWS).Example is from the Calperum flux tower (see Beringer et al. (2016) for details) for 2010 to 2013. Figure illustrates a) the correlations of site Ta with the best nearby AWS including station ID and r 2 and b) the final gap-filled Ta series of 30 minute data and the original flux data with gaps.

3 ) 22 Figure S4 :
Figure S3: Diagnostic plot of absolute humidity (Ah -g m -3 ) showing the time series used for gap-filling (green) and noticeably the flux tower data from OzFluxQC level 3 that has data cut off due to incorrect setting of threshold for quality control.Following this 5

Figure S5 :
Figure S5: An example of ingestion and processing of MODIS satellite information.Here MOD13Q1 Enhanced Vegetation Index (EVI) for the 3x3 km cut-out around the tower is illustrated.The Modis eight day values are shown as circles and data is then interpolated (green line) and smoothed (red line) as detailed in Section 2.1.4.Example is from the Calperum flux tower (see Beringer et al. (2016) for details) for 2010 to 2013.5

Figure S7 :
Figure S7: The u* threshold is determined as given in Section 2.3.Shown here is u* determined using the procedure of Reichstein et al. (Reichstein et al., 2005a).The u* is shown for the 3 month moving windows (green line with standard error bars indicating annual variability).Also illustrated are the maximum u* thresholds over the entire period (conservative, blue line) and mean u* 5

Figure S8 : 5 28Figure 8S9 :
Figure S8: A frequency plot of trials is showing the annual uncertainty around the mean uptake for the component uncertainties (model, random and ustar) as well as the total uncertainty.The example given here is from the OzFlux Whroo site (AU-Whr) for 2013.
Figure 9S10: DINGO produces many diagnostic plots for all fluxes and meteorological variables (in this case net ecosystem exchange (Fc -µmol m -2 s -1 )) including a) amount of missing data and data not gap-filled (percent) and b) weekly plot (dark-blue) with periods 5

Figure 10S11 :
Figure 10S11: Example of summary fingerprint plot of a) net ecosystem exchange (Fc), b) ecosystem respiration (Fre) and c) gross primary production (GPP) all in units of µmol m -2 s -1 .Example is from the Calperum flux tower (see Beringer et al. (2016) for details) for 2010 to 2013.

Figure 11S12 :Figure 12S13 : 5 Timestep
Figure 11S12: Summary timeseries plot of daily means and a 30 day running mean (dark line) of net ecosystem exchange (Fc, yellow), ecosystem respiration (Fre, red) and c) gross primary production (GPP, green) all in units of µmol m -2 s -1 .Example is from the Calperum flux tower (see Beringer et al. (2016) for details) for 2010 to 2014.
that are focused on improved understanding of the responses of carbon and water cycles of Australian ecosystems to current climate and disturbance regimes as well as impacts of projected future changes to precipitation, temperature and atmospheric CO2 concentration.Key questions include 1) what are the key drivers of ecosystem productivity (carbon sinks) and greenhouse gas emissions; 2) how resilient is Australian ecosystem productivity to an increasingly variable and changing climate, and 3) what is the current water budget of the dominant Australian ecosystems and how will it change in the future?