Real-Time Forecast of SMAP L3 Soil Moisture Using Spatial–Temporal Deep Learning Model with Data Integration

Round 1

Reviewer 1 Report

The manuscript scope is very interesting. I encourage innovative methods on capturing the spatial-temporal features of timeseries data. However, I have several concerns, which I think must be addressed properly before this manuscript is suitable for publication.

The special comments can be found in the attached PDF file.

Comments for author File: Comments.pdf

Author Response

The manuscript scope is very interesting. I encourage innovative methods on capturing the spatial-temporal features of timeseries data. However, I have several concerns, which I think must be addressed properly before this manuscript is suitable for publication.

Responds：

Thanks for the acknowledgement of the novelty. We tried our best to improve the paper based on authors’ inputs and the comments.

 

Comment#1: Please give the line numbers to convenient for writing the review comments.

Responds：

Thank you for your helpful advice, we have given the line numbers in the manuscript.

 

Comment#2:These sentences are too long-winded and need deep summarization and simplification. What is your purpose here?

Responds：

We appreciate your feedback and recommendations. The logic of this paragraph has been adjusted in the amended version of the manuscript. We first gave some examples of the three categories of model, and then give comments on them. The updated phrase is as follows:

Many researchers [12-17] used deep learning (DL) models to forecast SM because of their ability to learn nonlinear mappings, automatically extract features and build dy-namic systems. DL models can be divided into three categories: spatial models [18], temporal models [19, 20], and spatial-temporal models [21]. Spatial models, such as the Convolutional Neural Network (CNN) were proposed by LeCun and Bengio (1995) [18]. CNNs typically comprise of a convolutional layer, a pooling layer, and a fully connected layer, with the convolutional and pooling layers alternated. Convolutional layers can extract spatial characteristics. As a result, the grid of SM and meteorological time series data can be taken as CNN-processed images. For example, Wang et al. [22] used CNN to predict SM content using near infrared spectroscopy. Hegazi et al. [23] used Sentinel-1 images to train a CNN-based algorithm to estimate SM content over agricultural areas. Temporal models include the Long Short-Term Memory (LSTM) and the Gated Recur-rent Unit (GRU). LSTM and GRU can successfully correlate contextual information when processing serial data and are frequently employed in Earth system research for time series data prediction. For example, Fang et al. [12, 13] employed LSTM to predict surface SM based on climate forcing and soil texture. In addition, Filipovi et al. [24] forecasted the second layer SM using LSTM. Spatial-temporal models, such as the Convolutional LSTM (ConvLSTM) [21], which substituted matrix multiplication with a convolution operation of each gate in an LSTM cell. The model can capture and use both spatial and temporal correlations, making it an effective tool for forecasting spa-tial-temporal variables. For example, Li et al. [15] shown that ConvLSTM outperformed independent CNN or LSTM in terms of SM forecast accuracy. A et al. [25] utilized ConvLSTM to assess the root zone SM. Li et al [16] suggested an attention mecha-nism-based ConvLSTM for SM prediction and demonstrated the relevance of temporal and spatial correlation on model performance. It is well known that the state of each pixel's SM at each time step depends not only on its own historical observation, but also on the state of its neighboring pixels at the current time. In addition, it may also be di-rectly influenced by the historical state of neighboring pixels, as well as by the time se-ries data of meteorological forcing variables and static geographical attributes. For spa-tial models, spatial features are extracted by convolutional layers, however, "flattening" loses spatial autocorrelation and its inter-grid order, and its forecasting in the time di-mension cannot introduce long-term memory structure. The time series model is used to build mapping relationships between time series of variables at a particular point, although there are significant issues such missing spatial information. As a result, the spatial-temporal models have the advantage of concurrently capturing spatial and temporal changes for SM forecast.

In addition, both LSTM and GRU retain important features through various gate functions, which ensures that the vital information for SM prediction will not be lost in long-term propagation. GRU has one less gate function than LSTM, so the number of parameters of GRU is less and it is faster than LSTM, but the accuracy is similar. ConvGRU based on the combination of GRU and CNN, by superimposing convolutional operations on different regions, is able to obtain temporal relationships and spatial features to better predict their future change patterns based on information mining.

All the aforementioned deep learning-based SM prediction methods have some drawbacks, including the inability to handle large amounts of missing remote sensing data. The key to forecast SM is SM memory characteristics. Thus, using lagged SM in DL models is a common practice, but this is challenging to perform with remotely sensed data because of missing data. Previously, Fang et al. [14] proposed the LSTM with an adaptive data integration kernel (called DI_LSTM) model for training SMAP L3 SM with missing spatial coverage at a time step. That is, the lagged SM is added as an input variable to the input data. When lagged SM data is missing, the predicted value from the previous moment is added as an input variable to the input data. But, LSTM-based models are trained in a point-to-point manner, which allows them to completely understand SM temporal correlation but ignores the impact of SM’s spatial distribution on SM forecast.

 

Comment#3:This paragraph has only little effect and can be deleted.

Responds：Thanks for your suggestion, we have removed this paragraph from the manuscript.

 

Comment#4: Figure 4, 7, 8, and 10: The South China Sea were not drawn in these maps.

Responds：

The islands in the South China Sea are labeled in Figures 4, 7, 8, 10, Figures S1, S2, S3, and S5 in accordance with this comment.

 

Comment#5: 3.1.1 give the reference where your function is cited

Responds：

Thank you for your kind comments and helpful suggestions, which we have newly expressed in the manuscript as follows:

For each time step  , the update gate and reset gate of ConvGRU are formulated as follows (for further detail refer to [21]):

 

Comment#6: The citation of figures and tables should be closely arranged to the related texts rather than put them together at the first sentence.

Responds：Modified as suggested.

 

Comment#7: All captions of figures are too brief to show the enough information. What is the unit of SM in your figures?

Responds：We modified the captions as follows:

Figure 1. Inner structure of ConvGRU. X is the input and H is the hidden layer.

Figure 2. The network structure of DI_ConvGRU. X is the input, H is the hidden layer, and Y is the output.

Figure 3. The density scatter plot of the predicted SM and SMAP L3 SMobserved SM by DI_ConvGRU (a-c), interp_ConvGRU (d-f) and DI_LSTM (g-i) for 1-day, 2-day, and 3-day fore-casts. The blue line is the 1:1 line and the red line is the regression line.

Figure 4. during the test period from 1 April 2017 to 31 March 2018 in China. Shown are the Bias (a, d, g) for 1-, 2-, and 3-day forecasts, respectively, RMSE (b, e, h) and R (c, f, i) R.

Figure 5. Box plots of the Bias, RMSE, R and KGE of comparison between SMAP L3 SM and predicted values of DI_ConvGRU, interp_ConvGRU and DI_LSTM.

Figure 6. Improvements from interp_ConvGRU (or DI_LSTM) to DI_ConvGRU. RMSE improvements (a,c) were calculated as (RMSE(interp_ConvGRU (or DI_LSTM))- RMSE(DI_ConvGRU))/(RMSE(interp_ConvGRU (or DI_LSTM)))×100%, and  R improvements (b,d) were calculated as (R(DI_ConvGRU)- R(interp_ConvGRU (or DI_LSTM)))/(R(interp_ConvGRU (or DI_LSTM)))×100%.

Figure 7. (a) Randomly selected pixel locations in the Köppen-Geiger climate region map of China and (b) Forecast of SM time series in different climatic regions by different models (from April 1th 2017 to March 31th 2018).

Figure 8. Box plots of the Bias, RMSE, R and KGE of models trained with different datasets during the tested period. Those four models were trained using 1) ConvGRU with Climatic Forcing data (blue; including the lagged precipitation, temperature, radiation, humidity, and wind speed, where the lagged time is 1 day). 2) interp_ConvGRU with SM (lagged SMAP L3 SM gap filled by linearly interpolation) and Climatic Forcing data (purple); 3) DI_ConvGRU with SM (lagged SMAP L3 SM) and Climatic Forcing data (pink); and 4) DI_ConvGRU with SM (lagged SMAP L3 SM), Climatic Forcing and Static data (blue; including including sand, silt, clay content, bulk density, land cover type and DEM).

Table 1. Mean value of different metric of different climate regions, including soil moisture (SM, m3/m3), Coefficient of Variation (CV) and the lagged correlation of SM (lagged R), the Bias (m3/m3), root-mean-square error (RMSE, m3/m3) and Pearson’s Correlation Coefficient (R) of the DI_ConvGRU, and number of pixels.

 

Comment#8: Figure5: Unnecessary figure.

Responds：Deleted

 

Comment#9: You can discuss on the applicable scope and conditions of DI_ConvGRU.

Responds：

We discussed this as follows:

It is expected that the DI_ConvGRU proposed in this work can be applied on re-mote sensing data of soil moisture other than SAMP. Furthermore, it also has potential to be applied on predictions of remote sensing variables with gaps other than soil moisture, which needs further study to verify its suitability. As this work shows that the performance of the DI_ConvGRU depends heavily on soil moisture memory effects represented by the lagged SM, there are questions to be answered whether this method can be useful for gap-filling of other variables such as leaf area index and how well its performance will be.

 

Comment#10: Figure 10(a) is not cited in the main text. Figure 10 This figure would be better put into the supplementary materials.

Responds：

According to this comment, we have removed Figure 10 (a) and relocated the remaining three figures in Figure 10 to the supplementary material as Figure S3.

 

Comment#11: Please further summarize and highlight your results to improve the depth of manuscript.

Responds：

We have revised the conclusion as follows:

In this work, we proposed a convolutional gated recursive unit with data integration (DI_ConvGRU) model for accurate and real-tim e SM prediction using SMAP L3 SM data. The model can capture the spatial and temporal correlations of time series SM and adapt to the irregular observations of SMAP SM for prediction. Comparisons were made with interp_ConvGRU (for verifying the role of DI terms) and DI_LSTM (for verifying whether the spatial-temporal model improves the prediction accuracy of SM) and the results showed that has improved the model performance in 74.88% and 68.99% of the regions according to RMSE compared with interp_ConvGRU and DI_LSTM, DI_ConvGRU, respectively. We analyzed and discussed the performance of the model in three aspects: the overall performance of the model, the model performance in different climatic regions, and the influence of different factors. The conclusions are as follows:

• DI_ConvGRU can not only better capture the spatio-temporal characteristics of SM, but also effectively handle the missing data in SMAP SM. In terms of prediction accuracy and convergence speed, the DI_ConvGRU model outperformed the other DL models (ConvGRU, LSTM, DI_LSTM and interp_ConvGRU), and it achieved good performance with a bias of 0.0132 m3/m3, an ubRMSE of 0.022 m3/m3 and an R of 0.977. And ConvGRU instead of LSTM has a greater impact on the model performance than linear interpolation with DI terms.
• Among the eight climate zones, the polar regions had the best prediction performance, and the tropical regions had the worst performance. We find that the prediction performance of the model is strongly related to the lagged R of the SM and the coefficient of variation of the SM. The spatial-temporal model's image-to-image training strategy collected not only information on the time series but also the spatial information of surrounding pixels, whereas the DI term-based model better captured the peaks.
• The lagged SM has the most significant impact on the model performance, followed by the DI term and static data. Error buildup may result from the linear interpolation-based DL model, while the DI-based model can successfully avoid it. Additionally, while the missing data rate for various places has little impact on the model's performance, the missing data rate in different seasons has some effects.

 

Comment#12: You have observed SM in Figure 3, but where have you discussed this dataset in detail? I am assuming that by 'observed' you do not mean SMAP L3 SM.

Responds：

We adjusted all of the "observed SM" in the manuscript to SMAP L3 SM in response to your comments and helpful ideas.

Reviewer 2 Report

Dear Authors,

Congratulations for all your work and results! As I could notice, your research is a long time one and, even for this, it's accurate and correct.

I appreciated the way you explained each term of the equations, in order to be easily understood by a non-profesional reader. Also, the graphs are complexe and they are presenting the amoun of original data.

Please pay attention just to the format - Table 1 is written with a higher fond, one part of Conclusion needs alignement.

Except this, all the best for you all! 

Author Response

Congratulations for all your work and results! As I could notice, your research is a long time one and, even for this, it's accurate and correct.

I appreciated the way you explained each term of the equations, in order to be easily understood by a non-profesional reader. Also, the graphs are complex, and they are presenting the amount of original data.

Responds：

Thanks for the acknowledgement of our work.

Comment#1: Please pay attention just to the format - Table 1 is written with a higher fond, one part of Conclusion needs alignment. Except this, all the best for you all! 

Responds：

We have modified the font of Table 1 and correct the alignment of the conclusion.

Reviewer 3 Report

The authors have selected a very timely and important study, and I find it to be very interesting. 

There are some things that need to be clarified, given that they are not precisely stated.

You have observed SM in Figure 3, but where have you discussed this dataset in detail? I am assuming that by 'observed' you do not mean SMAP L3 SM.

What is the unit of SM in your figures? It is best to show the unit of measure on the axis, where relevant. 

In your Results section, you present findings for two approaches - using China as a single study area and dividing China into climatic regions. Where have you discussed this in your Methodology? And why? What do you expect by the 2 different approaches? 

I feel that some parts of section 4.5 are more suited to go to the Methodology section. More precisely, the design aspect of the four experiments.

In Figure 4, column 1 (the bias for the 1-day, 2-day, and 3-day prediction), the bias on the top right corner (40 d - 50 d N, 120 d - 130 d E), the bias has shifted from positive to negative when you consider 1-day to 3-day prediction. Is there a scientific/numerical reason for this behaviour? This is not so prominent in other parts of the study area. 

I feel that 'Data Integration' as a keyword is more suited to your work than 'remote sensing'.

Author Response

The authors have selected a very timely and important study, and I find it to be very interesting. There are some things that need to be clarified, given that they are not precisely stated.

Responds:

Thanks for the acknowledgement of the novelty. We tried our best to improve the paper based on authors’ inputs and the comments.

 

Comment#1: What is the unit of SM in your figures? It is best to show the unit of measure on the axis, where relevant. 

Responds：

It is m3/m3. We indicated the units on the axes of all the figures.

 

Comment#2: You have observed SM in Figure 3, but where have you discussed this dataset in detail? I am assuming that by 'observed' you do not mean SMAP L3 SM.

Responds：

We adjusted all of the "observed SM" in the manuscript to SMAP L3 SM in response to your comments and helpful ideas.

 

Comment#3: In your Results section, you present findings for two approaches - using China as a single study area and dividing China into climatic regions. Where have you discussed this in your Methodology? And why? What do you expect by the 2 different approaches?

Responds：

We added the following content to answer these questions:

The performance of the DL models was evaluated for both the whole China and eight climate regions. The former was used to evaluate the overall performance of models, while the latter was used to reveal the differences of model performance in different regions and the possible reasons as soil moisture has different behavior and influencing factors in different climate [1].

 

Comment#4: I feel that some parts of section 4.5 are more suited to go to the Methodology section. More precisely, the design aspect of the four experiments.

Responds：

We move the design aspect of the four experiments to the Methodology section as follows:

To explore the sensitivity of predictions to different inputs, we designed four experiments to test the prediction performance of the DL models based on different input data for 1-, 2-, and 3-day forecasts. Experiment I was served as the baseline, in which the model was built using ConvGRU using only climate forcing as inputs. In Experiment II, Interp_ConvGRU was built using climate forcing and lagged SM as input, which was used to show the effect of adding lagged SM gap filled by linearly interpolation compared with Experiment I. It is not suitable to build a ConvGRU with lagged SM but without gap filling because of the frequent gaps in the original SMAP SM. As a result, we did not build ConvGRU with climate forcing and lagged SM. In Experiment III, DI_ConvGRU was built using climate forcing and lagged SM, which was used to show the effect of DI instead of the linear interpolation in Experiment II. In Experiment IV, DI_ConvGRU was built using climate forcing, lagged SM and static physiographic attributes, which was used to show the effect of static data compared with Experiment III.

 

Comment#5 In Figure 4, column 1 (the bias for the 1-day, 2-day, and 3-day prediction), the bias on the top right corner (40 d - 50 d N, 120 d - 130 d E), the bias has shifted from positive to negative when you consider 1-day to 3-day prediction. Is there a scientific/numerical reason for this behavior? This is not so prominent in other parts of the study area. 

Responds：

The reviewer has observed an interesting phenomena. If we take a look at the Figure S1 for interp_ConvGRU and Figure S2 for DI_LSTM, there is not any consistent phenomena for this region (40 d - 50 d N, 120 d - 130 d E). So, the bias of this region is model dependent and we do not think this behavior is related to the soil moisture itself but related to the model itself. As machine learning is Blackbox, we do not think we can find or guess the scientific/numerical reason behind it.

Round 2

Reviewer 1 Report

My previous concerns have been proper addressed.

I suggest the author do not submit the manuscript in modified mode, which is very mess and is uncomfortable to read. Please submit the manuscript with the marked revision by using the red color.