Generating High Resolution LAI Based on a Modified FSDAF Model
Round 1
Reviewer 1 Report
1) Experiments design mix two sources of error: LAI estimation using SVR regression, and improved FSDAF method proposed. Since the main idea is to demonstrate method improvement, it should be better to use the same type of both high resolution images (either LANDSAT or Sentinel) to estimate LAI at t1 and t2 times, to estimate the effect of the method improvement.
2) Sample size of 40 samples is too small to make supported conclusions about the method improvement. Since we have a lot of pixels to compare in HR images, authors should use bootstrap repeatedly randomly selecting 1,000 or more pixels to estimate both the mean improvement and it's variability (SD, 95% CI, etc.)
3) In the method description authors use some mixture of notations from two papers [14, 15], and rename some variable. It makes it hard to understand description, even having both source articles at hand. For example, in Formulae (4)-(5) the same value is denoted as deltaLAI(x_k,y_k) and deltaLAI(x_ij,y_ij),
(line 116): variable C seems to be incorrectly defined: in "C is the number of categories in the mixed pixel, and S is the total number of the endmember in the pixel" both C and S are the same value;
in Formulae (4)-(10) variables L(x_i,y_i) and E_he(x_ij,y_ij) are not defines (also it seems to be the same value).
4) Some typos:
(line 141): in formula (3) omega isn't support vector, but x_i are.
(line 201) in Fig.1 "Landsat LAI at t2" will be "Landsat LAI at t1"
Author Response
Response to Reviewer 1 Comments
Point 1: Experiments design mix two sources of error: LAI estimation using SVR regression, and improved FSDAF method proposed. Since the main idea is to demonstrate method improvement, it should be better to use the same type of both high resolution images (either LANDSAT or Sentinel) to estimate LAI at t1 and t2 times, to estimate the effect of the method improvement.
Response 1: We are very grateful to the reviewer for the valuable comments and suggestions on our manuscript. We added additional information in the Discussion section to address the questions, including:
We compared Landsat LAI data at the predicted time to the predicted LAI data estimated by the original and the improved FSDAF method, respectively, and the effect of the method improvement is estimated. The LAI estimation using SVR regressionat start time is the input data of the two fusion models, so we estimate it by true value (Sentinel as true value). A new Figure 7 that illustrates the correlation betweenthe predicted LAI using different fusion models and Landsat LAIis added in the revised manuscript.
Between revised lines 394-399:
Improved FSDAF method
(b) FSDAF method
Fig.7 Correlation between the predicted LAI using different fusion models and Landsat LAI
(red line is the 1:1 line)
Under the 4.2,we added new sentences (lines 369-372), which read as “Similarly, the correlation between the predicted LAI on August 5 and Landsat LAI retrieved by SVR on August 2 was also shown in Fig.7. The R of improved FSDAF method at Area A and Area B are 0.78644 and 0.80624, which are higher than those of FSDAF model (0.69236 and 0.68601)”.
Point 2: Sample size of 40 samples is too small to make supported conclusions about the method improvement. Since we have a lot of pixels to compare in HR images, authors should use bootstrap repeatedly randomly selecting 1,000 or more pixels to estimate both the mean improvement and it's variability (SD, 95% CI, etc.)
Response 2: We appreciate the valuable comments. We added additional information in the Discussion section to address the questions, including:
We selected 1000 pixels randomly as the samples toestimate the accuracy of Landsat LAI inversion using Sentinel-2A LAI.The new Figure 5 is shown in the revised manuscript.
Between revised lines 342-345:
(a) (b)
Fig. 5 The scatter plots of Landsat LAI inversion and Sentinel-2A LAI for different sites. (a) Area A, (b)Area B. (red line is the 1:1 line)
We use bootstrap repeatedly randomly selecting 1000 pixels to estimate the method improvement. The revised Figure 6 that illustrates the correlation between the predicted LAI using different fusion models and Sentinel LAI is added in the revised manuscript.
Between revised lines 388-393:
(a) Improved FSDAF method
(b) FSDAF method
Fig.6 Correlation between the predicted LAI using different fusion models and Sentinel-2 LAI
(red line is the 1:1 line)
3.We added a new Table 2 to illustrate the variability of each image.
Between revised lines 401-402:
Table 2. Comparison of the standard difference (SD) of LAI data from different images
Test Area |
Landsat by SVR(SD) |
Sentinel-2 (SD) |
Improved FSDAF(SD) |
FSDAF(SD) |
|
Area A |
0.65052 |
1.05779 |
0.98252 |
1.18621 |
|
Area B |
0.76216 |
0.91251 |
0.73485 |
0.99363 |
Point 3: In the method description authors use some mixture of notations from two papers [14, 15], and rename some variable. It makes it hard to understand description, even having both source articles at hand. For example, in Formulae (4)-(5) the same value is denoted as delta LAI(x_k,y_k) and delta LAI(x_ij,y_ij),
(line 116): variable C seems to be incorrectly defined: in "C is the number of categories in the mixed pixel, and S is the total number of the endmember in the pixel" both C and S are the same value;
in Formulae (4)-(10) variables L(x_i,y_i) and E_he(x_ij,y_ij) are not defines (also it seems to be the same value).
Response 3: In Formulae (4)-(5), we changed to because they are the same value. We redefined C and S. In Formulae (4)-(10), we supplemented the definations of and . The defination of is on line 201.
Between revised lines 174-186:
(4)
(5)
in which i is index of a low-resolution pixel, j is index of a high-resolution pixel in the corresponding low-resolution pixel, and is the coordinate index of the jth high resolution pixel within the ith low resolution pixel. k is the index of the similar pixel, which is the neighboring pixel with the same land cover type as the pixel at location . is the predicted high resolution LAI data at time , is the high resolution LAI data at time , and the weight of the similar pixel is [14]. represents the LAI changes between time and time of similar pixel. The change information of all similar pixels is weighted to obtain the total change value of the target pixel . The final estimate of all change is added to the initial LAI observation at to obtain the final prediction of the target pixel value at . In equation (5), is the changed LAI of class in high resolution data from time to time , and is the residual assigned to the high resolution pixel in the low resolution pixel.
Between revised lines 118-119: in which is the number of land cover categories in a mixed pixel, and is the total number of land cover categories in the whole MODIS image.
Between revised lines 201-202: is the residual between the observed high resolution LAI and the predicted LAI.
Between revised lines 206-208: is the temporal prediction error, is the equal error within a low resolution pixel when the landscape is heterogeneous.
Point 4: Some typos:
(line 141): in formula (3) omega isn't support vector, but x_i are.
(line 201) in Fig.1 "Landsat LAI at t2" will be "Landsat LAI at t1"
Response 4: The thoughtful suggestions are accepted. We changed the "Landsat LAI at " to "Landsat LAI at " in Fig.1. We modified the interpretations of omega and .
Between revised lines 144-145: where is the coefficient vector, represents the support vector, is the mapping function of and is the deviation of the function, respectively.
Author Response File: Author Response.pdf
Reviewer 2 Report
The methods for improved FSDAF to take advantage of high frequency of lower resolution MODIS and lower frequency higher resolution Landsat are presented. These are complicated, but the text is well written and the flowchart in Fig. 1 is very helpful. He authors present evidence that the improved classification gives higher resolution and somewhat better estimates of LAI.
Here are a few suggestions for revisions:
Figs 5 and 6 captions should mention that red line is the 1:1 line.
In Figure 4, LAI of some pixels went up in the improved FSDAF but down in the FSDAF (for example in upper left and lower middle of the images). Is there any indication as to why they went opposite direction in some areas (even though direction of change was the same for most parts of the image) and is it related to why the improved version performs better?
On line 348, I think the authors are actually referring to Fig 6, not Fig. 5.
Author Response
Response to Reviewer 2 Comments
Point 1: Figs 5 and 6 captions should mention that red line is the 1:1 line.
Response 1: At first, I am very grateful to the reviewer for valuable comments and suggestions on the manuscript 668491. The following is a revised explanation, please review it. We added the description of the red line in the captions of Fig.5 and Fig.6.
Between revised lines 344-345: Fig. 5 The scatter plots of Landsat LAI inversion and Sentinal-2A LAI for different sites. (a) Area A, (b)Area B. (red line is the 1:1 line)
Between revised lines 395-395: Fig.6 Correlation between the predicted LAI using different fusion models and Sentinel-2 LAI (red line is the 1:1 line)
Point 2: In Figure 4, LAI of some pixels went up in the improved FSDAF but down in the FSDAF (for example in upper left and lower middle of the images).Is there any indication as to why they went opposite direction in some areas (even though direction of change was the same for most parts of the image) and is it related to why the improved version performs better?
Response 2: We appreciate the comments. We added some explanations for this phenomenon.
Between revised lines 321-325: The predicted data on July 28 showed different patterns in the upper left corner and lower middle areas of the study area. LAI of some pixels went up in the improved FSDAF image, but decreased in the original FSDAF image. The predicted high-resolution LAI using the improved FSDAF model showed the obvious temporal differences by decomposing mixed pixels.
Point 3: On line 348, I think the authors are actually referring to Fig 6, not Fig. 5.
Response 3: The thoughtful suggestions are accepted. We changed the Fig. 5 changed to Fig 6.
Between revised lines 367-368: Fig.6 illustrates the correlation between the predicted LAI on August 5 and Sentinel-2A LAI product.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
I have no additional comments