Review of: A global long-term (1981–2019) daily land surface radiation budget

product from AVHRR satellite data using a residual convolutional neural network

by

Jianglei Xu, Shunlin Liang, and Bo Jiang

General Comments

This paper describes a machine-learning convolutional neural network (CNN) method to generate a high-resolution global surface net radiation (Rn) product from AVHRR data with the support of ground observations, reanalysis, and transformed MODIS data (GLASS). Special tools are used to better align point measurements of Rn with the larger-spatial-scale reanalysis and the GLASS Rn products for a spatially representative ground truth. Another component seeks to upscale data within the AVHRR footprint to recover finer-scale surface detail that is smeared out by the atmosphere. Getting finer resolution from coarse resolution data is not easily accepted without clear explanation. You explain that “spatial adjacent effect” is accomplished by applying reanalysis data and angular information of the satellite measurement and solar position. The spatial adjacent effect seems to be the novel advantage of the CNN method over other existing methods, but the explanation lacks detail and examples to help the reader better understand the upscaling process. Does it use reanalysis vertical profiles to correct the path of upwelling radiation to the satellite? Or is it some kind of statistical approach?

Overall, this new method produces a 5 km resolution daily product of Rn over the globe that seemingly outperforms comparable methods. Although it has problems over difficult regions such as the Tibetan Plateau, the Sahara, and snow-covered surfaces that need to be addressed, the overall result is impressive. The authors conduct several spatial and temporal comparisons to their accepted ground measurements and to other global Rn products to prove its worth. Except for a few places on the globe and the last few years of the 39-year dataset, they show good performance. Given that this data has high spatial resolution at a daily time scale, it should be useful tool for climate studies after its flaws are addressed.

Specific comments:

1. 14 Include the temporal resolution of the Rn estimates here.

1. 20-24 The statement beginning with “Inter-comparisons with three…” is not true. In section 4.3.3 you state: “The validation results in Fig. 8 and Table 7 for the ice/snow surface cover type further confirm that GLASS Rn product may offer a better performance in Greenland region.”

1. 40 “radiation” is not needed in front of “radiometers”

1. 132-161 This section describes the instruments used in the various networks which range from good thermopile pyranometers and pyrgeometers to not-so-good net radiometers. You only provide performance measures for the thermopile pyranometers, which are generally good. You don’t provide any performance information on the net radiometers, which are notoriously bad, especially the REBS model. According to Table 2, net radiometers dominate your observational dataset. You should provide performance measures of the net radiometers. Also, Table 2 is incomplete. For some you specify “Eppley PIR,” and others just “Eppley.”

1. 138 By “thermal effect,” are you referring to the thermal offset of single black detector pyranometers? If so, there are references for this measurement error.

1. 169. What does “along with inverse navigation to relate a specific Earth location to each sensor’s instantaneous field of view” mean?

1. 180-194 In this description of the GLASS product, Rn is estimated from downward shortwave radiation, and other variables using multiple MARS learners. Where do the input data come from?

1. 217 The last phrase of this sentence “the diurnal variation of daily surface Rn.” does not make sense.

1. 285 What does “when deeper networks converge” mean?

1. 323 This sentence does not make sense. Do you mean “Reliable and unreliable sites from each observation network, separated by a threshold ETC-derived correlation coefficient of 0.9, are listed in Table 5”?

1. 334 - 337 Please include references for the ARM, SURFRAD, BSRN, and FluxNet networks.

1. 374 I assume the color bar represents a normalized count scaled to the most frequent count. Regardless, explain the color bar in the caption.

1. 414 “under snow and ice surfaces” ? Perhaps use “for snow and ice surfaces” ?

1. 453 Change phrasing to “…especially clouds that have significant impacts on shortwave…”

1. 454 Where do cloud optical thickness (COT) and cloud water vapor (CWV) data come from?

1. 462 The sentence beginning with “Therefore, the performance…” does not make sense. Perhaps the end of that sentence should read: “…is comparable with regard to the accuracy of their Rn retrievals.”

1. 512 – 519 I don’t understand your Figure 11. The AVHRR and GLASS Rn’s as a function of COT are nearly on top of each other, yet the bias plotted on the same charts is significant. What am I missing here? Regardless, in the caption please define the bias and what the shading represents.

1. 524 Please state how the difference is defined. Jan.- July or July – Jan.

1. 525 Do you mean “northern Australia” and “South America” ?

1. 545 “produced by NOAA” ? Should this read “replaced by NOAA”?

1. 555-560 The four timeseries in Fig. 12 after 2017 for all data sets may be well correlated but are obviously wrong and could not be used for climate studies.

            What does the shading represent in Fig. 12?

1. 557 Has the AVHRR calibration across all satellites been applied to the AVHRR data shown in Fig. 12?

1. 572 Do you mean 0.708?

1. 591 What is a wide overpass time?

1. 600 What Study?

Technical corrections:

1. 164 You don’t need “, respectively” in this sentence.

1. 425 “BSRN_DRA” site.

1. 483 “very low” not “vary low”

1. 543 1999-2000

I also attached my review in a file called "Earth Data Science Data review.pdf"

This paper produces a long-term dataset of net surface radiation from AVHRR TOA measurements and MERRA2 reanalysis spatial context using a neural network algorithm trained to surface observations from multiple global networks. One of the unique aspects of this dataset is that an extended triplet collocation method is used to determine when surface measurements should be representative of a 0.05 degree grid box and only reliable sites are used to train the model. The dataset performs very well, even better than CERES-SYN and GLASS satellite datasets, against surface observations except in cases with ice and snow. The method using spatial context is also quite effective at removing algorithm biases that impact other datasets due to incorrect assumptions about atmospheric conditions (e.g. clouds and aerosol).

The authors also identify the need for future work to improve the algorithm over ice/snow, fill in gaps when satellites were down, and suggest that temporal variability information could improve the algorithm further.

Major comments:

1. One of the real advantages I see with this dataset is the long record—since the dataset starts in 1981 and has an accuracy equal to or exceeding other satellite-based estimates, this extends observation-based estimates of the surface radiation budget significantly. That could be of significant value for long-term climate studies. The authors could highlight this advantage more strongly in the abstract and conclusions.
2. Evaluation and training are done against multiple networks, but some of these networks are interconnected, for example, some ARM and all SURFRAD sites are included in the BSRN. As I look at the list of sites in Table S1, it appears that some of these stations are included multiple times. For example, BSRN_DRA is the same station as SF_DRA because the SURFRAD Desert Rock stations is submitted to the BSRN global network. This is particularly a problem if any of the independent validation stations are also included in the training dataset. Please look into this duplication.
3. I am curious whether the results shown in Figure 7 reflect the fact that some of these networks are included in training the AVHRR dataset. It isn’t clear to me from the description whether training stations were also used in this analysis, or whether this only includes independent testing stations and stations that didn’t meet the reliability requirements. But even if these validation stations are independent from training data, the network of measurements around the ARM Southern Great Plains sites, for example, may be more similar to each other than a site that is located in a much different climate regime (e.g. independent sites ARM_E06 and ARM_E41 sites). That could lead to overfitting. It would be helpful to understand how independent this validation dataset is.
4. Figure 9 is really interesting! I found it quite compelling that the AVHRR dataset shows smaller biases and relationship to atmospheric factors than the GLASS dataset. That strongly suggests that the neural network approach is able to accurately capture variability related to the surface measurements in a more reliable way.
5. Does Figure 14 show local time? Please label for clarity.

Minor comments:

Line 50: “RT-based physical methods show a great generalization” I am not sure what this phrase means, please revise for clarity

Line 310: should it be: “data was then removed”?

Line 346: “for in surface radiation estimations.” Wording doesn’t seem quite right here.

Lines 360-361: This sentence is awkwardly written and should be revised. Changing consistently to consistent, and site to sites would improve readability.

Line 483: should be very instead of “vary”

Line 545: should “produced” be replaced?

Line 563: GLASS is misspelled GALSS

Line 572: I think that 7.08 must be 0.78. Please check.

Line 690: I think “satellite replacement works” should be satellite replacement work if you are referring to times when there is no satellite data because it the satellites are being worked on.

Line 697: should be “covered surfaces”.