This was an interesting bit of research for me to review. I understood the nature of the modelling in terms of having large national model and then adjusting it to make it more useful at local level (here across Tasmania) through using transfer learning. For research’s sake the authors also tested extension of Australian model to Tasmania, which as a null hypothesis would be clearly rejected. Local models better, but transfer learning, better still. There is merit in this work but my misunderstandings of it come from the data inputs, mainly the observational data. I think these data were not well described and there is little to go on about how they were sourced. Some of the inputs into the model could be better too and which are also publicly available. I made many comments and feel there is more work to do in places. Am not a fan of the introduction for example which just seems to be a disparate collection of things in much need of pulling together into a single narrative.  

Abstract. Last sentence needs re-wording

Line 40-60. Might worth recognizing that in Australia there are a number of spatio-temporal models of soil moisture that generate maps continentally. Take for example, The Bureau of Meteorology’s Australian Water Resources Assessment Landscape model (AWRA-L) version 6 (Frost et al. 2016). Another one is The National Soil Moisture Information Processing System (SMIPS; Stenson et al. 2021). Similarly model presented in Wimalathunge & Bishop (2019) is set up to run daily and uses inputs derived from Soil and Landscape Grid of Australia. It might also be helpful to mention a few of the popular water balance models that are out there too or incorporated into broad land surface models. 

Line 70. Not sure i follow about  "..to a model with similar tasks"

Line 78-81. Would be good to reference some reports here. Probably guessing these would be from Australian federal or Tasmanian government reporting on land assessments of threats and opportunities. 

Line 83. There has been little information share din intro about the nature of the Australia-wide data, or the nature of the modelling processes that would use these data. Deep learning is the modelling approach, there seems to be some associated of DL with SMAP, but there seems a gap in the narrative of how one goes from these features to using Australia-wide data to get daily soil moisture estimates. The model transfer bit seems well described, however. Not looking for detail here as this will come in the methods description, but the introduction seems to be a collection of loose ends and no clarity on what the intention of the work is.

Line 84: "we contribute to" ?

Table 1. Given the public availability of daily gridded temperature and rainfall data from silo (https://www.longpaddock.qld.gov.au/silo/about/climate-variables/) for all of Australia (5km res) what would be the reason to go with ERA5 data?

Table 1. The Searle et al 2022 reference pertains to just the AWC product. 

Table 1. Australia also has a publicly available post processed version of the SRTM. Processed in terms of vegetation removal and hydro logically corrected (https://ecat.ga.gov.au/geonetwork/srv/eng/catalog.search#/metadata/69888). Would have though this were a better dataset to use. 

Table 1. For the land use lander mapping, am not sure this is the correct reference citation for this data source.

Table 2. Not clear about number of data points. What information is at these points?

Lines 120-149. These sections would be helpful if authors guided readers to fundamental research on these concepts. Maybe there could be some figures used in these sections to illustrate too. General feeling here is that this is all quite jargony.

Line 150. Maybe a map showing distribution of data and soil moisture sensors used both in across Australia and then Tasmania for focus. 

Line 153-54. Bit unclear about these data: "Reference soil moisture data were measured using frequency-domain reflectometry sensors available at different soil depths between stations". When was this collected, by whom and at what depths and frequency.

Line 155. More detail needed about soil probe calibration. Saying based on bulk density information does not give enough info. Would be good to provide information about the type of sensors, general details.

Line 163. What is meant by daily soil moisture was averaged? What part of the data is being averaged?

Line 174. Explain the 'multiband image was calculated each day'

Line 175-78. Need some clarity about these reference data. This was mentioned above too in my comments.

Line 200-05. What soil moisture data is used in AU model?

Line 207. Are each of these individual models combined into one?

Figure 2. So the out "Tasmania Soil Moisture Maps" is a combination of AU, Tas and transfer models, or is it the transfer model output as it was determined to be the best model?

Line 239. Are these the soil moisture sensor data and the 'other' data? This is not predicted or are they observed? 

Figure 3. It is not clear what this data means or what can be interpreted from it. Just showing observed data all compiled together over the period of specified time does not provide anything too much informative.

Line 260-67. So what does one make of these data. Pointing out some distinctness between data is fine, but what else are the authors trying to say here?

Figure 5. It would be useful to plot time series of soil moisture sensor data with SMAP to see consistencies of data through time.

Figure 6 and 7. What is shown as correlation is actually Lin's Concordance correction coefficient? Or is it actually just correlation?

Figure 9. Variation in RMSE seems higher compared with concordance. Even at top of Tasmanian, there is high concordance but some of those sites also have high RMSE. This is a bit of an odd outcome and should warrant an interpretation. 

Table 4. It is interesting that model outcomes are good for irrigated land use given model only considered rainfall and not any other supplementary inputs from irrigation. So model is adjusting for this given the data from the stations in this land use? This aspect seems to be overlook in discussion. 

Figure 11. In western Tasmania, this are to my understanding has significant areas of shallow peats, that is peat thinly blanketing rock. Having estimates of SM for 30-60cm would therefore be unrealistic. Maybe this is affecting to model reliability? In any case, estimates of plant available water or simply hydrologically available water could be predicted quite a long way off from reality.  

Line 413. "both were.." meaning the for the transfer modelling?

Line 414. Not sure i understand this thing about 'memory'. Maybe explain more. Other than the fact that model included variables to capture the latency between soil moisture and rainfall, what other memory is captured here?

Line 465-70. Soil thickness to consider here too

References

Frost, A.J., Ramchurn, A., Smith, A., 2016. The Bureau’s Operational AWRA Landscape (AWRA-L) Model. Australian Bureau of Meteorology Technical Report.

Stenson, M., Searle, R., Malone, B. P., Sommer, A., Renzullo, L. J. and Di, H., 2021. Australia wide daily volumetric soil moisture estimates. Version 1.0 [Dataset]. Terrestrial Ecosystem Research Network, Canberra. https://doi.org/10.25901/b020-nm39.

Wimalathunge, N.S., Bishop, T.F.A., 2019. A space-time observation system for soil moisture in agricultural landscapes. Geoderma 344, 1-13.

The paper is well-written, although the data sets and the modelling approaches are not always clearly described. My main issue is the novelty and the real aim of the paper i.e. beyond a regional study.

The real aim of the paper should be reinforced in the introduction. As it stands the focus is on producing a soil moisture product for Tasmania, while the transfer learning is the scientific novelty. You mention in lines 70-76 that this technique has already been applied in other studies, but I can not find any details on the extrapolation methods. Therefore, it is difficult for the reader to understand what the main challenges are that remain to be resolved are. What is the novelty of the use of transfer learning in Tasmania? The requirements for the soil moisture product are not clearly outlined and therefore there are no clear specifications of the soil moisture product. Unfortunately, these issues are not treated in the discussion section either.

Line 78 Please provide some more details on the current land use change in Tasmania. It would also be good to explain the practical needs for soil moisture products that emanate from this land use change.

Table 1 The abbreviations in the ‘variable/bands’ column are not clear. What does ‘band’ refer to?

Line 155 Why do you need bulk density data? The probes will already provide volumetric soil moisture content. Or is the recorded soil moisture data expressed as gravimetric moisture content. Please explain in some more detail.

Line 171 How was the AWC calculated? Pedotransfer functions for AWC generally take SOC and clay content into account. Is not there a risk for redundancy?

Line 174 What does ‘mulitband image’ refer to? The collation of all co-variates or an image from a multispectral satellite?

Lines 203-204 The sentence is not entirely clear. Which variables were excluded? Why do you mention ‘including’ in the same sentence?

Line 213 Does the workflow of Fig. 2only describe the transfer learning in bullet point ‘c’ or all modelling scenarios? If the latter, please refer to Fig. 2 in line 200

Line 224 For model evaluation the R² is generally used. Why do you prefer the ‘r’?

Figure 4 Please explain in the caption what the numbers refer to e.g. AWC1.

Line 270-275 I do not understand the difference between the two correlation approaches. Is the first approach for all stations and the second per probe location? Can you indicate the link to panels a and b in Fig 5?

Figs 11 and 12 When you discuss the model performance (e.g. Table 4), the standard deviation is based on the prediction in a number of stations (e.g. 3 for the forest). I am not clear what the standard deviation means in the maps (Figs 11 and 12. I might have missed it.

Line 413 What does ‘both’ refer to?

Line 454 Do I understand it correctly that the R² of predicted vs observed for soil moisture in some areas of China was 0.2? If so, could you please comment on the practical use of a prediction with such low R² values?

Line 459 It is difficult to compare model performances based on the standard deviation. Uncertainty maps normally show the percentile interval range.

Line 465 What does ‘this area’ refer to?