Spatial Estimation of Soil Organic Carbon Content Utilizing PlanetScope, Sentinel-2, and Sentinel-1 Data
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper titled "Spatial Estimation of Soil Organic Carbon Content Utilizing PlanetScope, Sentinel-2, and Sentinel-1 data" aims to estimate soil organic carbon (SOC) content in the Wuling Mountain region of Southwest China. By leveraging spectral, radar, and topographic variables from multi-temporal optical satellite images, synthetic aperture radar images, and digital elevation models, the study employs the eXtreme Gradient Boosting algorithm to model SOC content at various spatial resolutions. The research found that medium resolution (20m) models generally provided better results, with combined datasets yielding the most accurate predictions.
I recommend the authors compare their accuracy with existing methods of estimation of soil organic carbon content by drawing a table listing the indicators (e.g., Coefficient of Determination (R2), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Akaike Information Criterion (AIC)).
I recommend the authors using existing methods (other than XGBoost) to redo their work and validate if their conclusions are still supported.
Figure 1 Add basic cartographic elements: compass, latitude and longitude.
Discussion section:
Add detailed figures for states such as "higher accuracy" (Line 610), "accuracy has shown a decline" (Line 625), "improve the prediction accuracy" (Line 619).
3) should be a paragraph. (Line 615)
All the websites should include "access on..."
Author Response
The paper titled "Spatial Estimation of Soil Organic Carbon Content Utilizing PlanetScope, Sentinel-2, and Sentinel-1 data" aims to estimate soil organic carbon (SOC) content in the Wuling Mountain region of Southwest China. By leveraging spectral, radar, and topographic variables from multi-temporal optical satellite images, synthetic aperture radar images, and digital elevation models, the study employs the eXtreme Gradient Boosting algorithm to model SOC content at various spatial resolutions. The research found that medium resolution (20m) models generally provided better results, with combined datasets yielding the most accurate predictions.
Response: Thank you very much for your thorough review and constructive feedback on our manuscript. We have incorporated your suggestions by conducting new model validations and revising the content accordingly. These modifications have not only enhanced the rigor of our research but also improved the reliability of our results. We believe that the revised version now meets the high standards of Remote Sensing and we hope these improvements will be satisfactory to you. Once again, we greatly appreciate your valuable time and professional guidance.
The point-to-point responses to the comments were given below.
Comments 1: I recommend the authors compare their accuracy with existing methods of estimation of soil organic carbon content by drawing a table listing the indicators (e.g., Coefficient of Determination (R2), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Akaike Information Criterion (AIC)). I recommend the authors using existing methods (other than XGBoost) to redo their work and validate if their conclusions are still supported.
Response 1: Yes, we have compared the performance of XGBoost with Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Support Vector Regression (SVR). The results of RF, GBT and SVR are shown in Tables 1-3. According the R2, MAE, RMSE, and AIC, XGBoost outperformed other models. For paddy field, XGBoost’s Model F performs the best ((R2=0.699, MAE=0.114%, RMSE=0.148%, AIC=508.820) followed by RF (Model F: R2=0.533, MAE=0.146%, RMSE=0.185%, AIC=524.246), GBDT (Model F: R2=0.639, MAE=0.131%, RMSE=0.162%, AIC=515.166), and SVR (Model D: R² = 0.240, MAE = 0.163% , RMSE = 0.234% , AIC = 333.304). For dry land, XGBoost performs the best (Model E: R2=0.673, MAE=0.107%, RMSE=0.135% ,AIC=453.246) followed by RF (Model E: R2=0.358 , MAE=0.158%,RMSE=0.189%, AIC=475.479), GBDT (Model E: R2=0.493, MAE=0.136%,RMSE=0.168%, AIC=467.689), and SVR (Model C: R2=0.118, MAE=0.163%, RMSE=0.206%, AIC=272.967). For the entire area, XGBoost performs the best (Model E: R2=0.339, MAE=0.268%, RMSE=0.349%, AIC=610.782) followed by RF (Model E: R2=0.243, MAE=0.259%, RMSE=0.375%, AIC=624.469), GBDT (Model E: R2=0.213, MAE=0.280%, RMSE=0.381%, AIC=622.491), and SVR (Model D: R² = 0.069, MAE = 0.311% , RMSE = 0.415% , AIC = 473.784).
Table 1 Performance results of Random Forest in predicting SOC based on different combinations of environmental variables at different modeling resolutions. The most accurate results are shown in bold.
Model |
Land Use Type |
|||||||||||||||
|
Paddy field |
Dry land |
Total area |
|||||||||||||
|
R2 |
MAE (%) |
RMSE (%) |
AIC |
R2 |
MAE (%) |
RMSE (%) |
AIC |
R2 |
MAE (%) |
RMSE (%) |
AIC |
|
|||
Model A |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
0.401 |
0.155 |
0.209 |
228.950 |
0.069 |
0.178 |
0.217 |
228.515 |
0.084 |
0.295 |
0.411 |
376.667 |
|
||
|
20m |
0.419 |
0.160 |
0.206 |
227.892 |
0.327 |
0.155 |
0.194 |
211.006 |
0.128 |
0.292 |
0.401 |
373.374 |
|
||
|
30m |
0.398 |
0.162 |
0.210 |
229.142 |
0.222 |
0.145 |
0.192 |
222.914 |
0.0942 |
0.297 |
0.409 |
375.912 |
|
||
|
80m |
0.328 |
0.188 |
0.222 |
232.966 |
0.087 |
0.161 |
0.215 |
227.877 |
-0.038 |
0.299 |
0.366 |
373.813 |
|
||
Model B |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
0.380 |
0.177 |
0.213 |
278.163 |
0.182 |
0.159 |
0.203 |
272.250 |
0.081 |
0.289 |
0.412 |
424.921 |
|
||
|
20m |
0.437 |
0.161 |
0.203 |
274.811 |
0.304 |
0.151 |
0.197 |
270.132 |
0.065 |
0.297 |
0.416 |
426.011 |
|
||
|
30m |
0.418 |
0.162 |
0.206 |
275.952 |
0.109 |
0.162 |
0.212 |
275.097 |
0.065 |
0.311 |
0.416 |
426.027 |
|
||
|
80m |
0.340 |
0.184 |
0.220 |
280.358 |
0.063 |
0.158 |
0.212 |
274.953 |
-0.025 |
0.299 |
0.363 |
420.887 |
|
||
Model C |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
-0.144 |
0.210 |
0.289 |
299.608 |
0.128 |
0.168 |
0.210 |
274.364 |
0.132 |
0.299 |
0.400 |
421.076 |
|
||
|
20m |
0.414 |
0.150 |
0.207 |
276.199 |
0.199 |
0.171 |
0.211 |
274.756 |
0.202 |
0.277 |
0.38 |
415.397 |
|
||
|
30m |
0.191 |
0.181 |
0.243 |
287.472 |
0.173 |
0.158 |
0.204 |
272.611 |
0.138 |
0.296 |
0.399 |
420.615 |
|
||
|
80m |
0.079 |
0.191 |
0.260 |
292.022 |
0.211 |
0.149 |
0.194 |
269.279 |
-0.075 |
0.302 |
0.372 |
424.295 |
|
||
Model D |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
-0.135 |
0.221 |
0.288 |
347.337 |
0.255 |
0.146 |
0.194 |
317.169 |
0.098 |
0.300 |
0.408 |
471.644 |
|
||
|
20m |
0.310 |
0.161 |
0.225 |
329.921 |
0.231 |
0.165 |
0.207 |
321.428 |
0.115 |
0.292 |
0.404 |
470.380 |
|
||
|
30m |
-0.041 |
0.206 |
0.276 |
344.294 |
0.186 |
0.155 |
0.203 |
321.102 |
0.104 |
0.303 |
0.407 |
471.154 |
|
||
|
80m |
0.095 |
0.187 |
0.257 |
339.418 |
0.054 |
0.156 |
0.213 |
323.274 |
-0.081 |
0.309 |
0.373 |
472.681 |
|
||
Model E |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
0.336 |
0.159 |
0.220 |
488.577 |
0.139 |
0.176 |
0.209 |
481.958 |
0.094 |
0.292 |
0.409 |
631.936 |
|
||
|
20m |
0.470 |
0.145 |
0.197 |
480.662 |
0.358 |
0.158 |
0.189 |
475.479 |
0.243 |
0.259 |
0.375 |
624.469 |
|
||
|
30m |
0.366 |
0.167 |
0.215 |
486.939 |
0.207 |
0.163 |
0.200 |
479.231 |
0.132 |
0.284 |
0.401 |
629.108 |
|
||
|
80m |
0.217 |
0.196 |
0.239 |
494.337 |
0.199 |
0.145 |
0.196 |
477.792 |
0.005 |
0.303 |
0.426 |
634.808 |
|
||
Model F |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
0.309 |
0.175 |
0.225 |
537.983 |
0.212 |
0.163 |
0.199 |
527.009 |
0106 |
0.295 |
0.407 |
679.089 |
|
||
|
20m |
0.533 |
0.146 |
0.185 |
524.246 |
0.316 |
0.157 |
0.195 |
525.568 |
0.147 |
0.280 |
0.371 |
672.369 |
|
||
|
30m |
0.244 |
0.179 |
0.235 |
541.129 |
0.174 |
0.160 |
0.204 |
528.582 |
0.112 |
0.284 |
0.399 |
676.240 |
|
||
|
80m |
0.227 |
0.186 |
0.238 |
541.908 |
0.159 |
0.152 |
0.201 |
527.370 |
-0.026 |
0.295 |
0.363 |
676.6635 |
|
||
Notes: Model A, PlanetScope+DEM; Model B, PlanetScope+Sentinel-1+DEM; Model C, Sentinel-2+DEM; Model D, Sentinel-2+Sentinel-1++DEM; Model E, PlanetScope+Sentinel-2+DEM; Model F, PlanetScope+Sentinel-2+Sentinel-1+DEM.
Table 2 Performance results of Gradient Boosting Decision Tree predicting SOC based on different combinations of environmental variables at different modeling resolutions. The most accurate results are shown in bold.
Model |
Land Use Type |
|||||||||||||||
|
Paddy field |
Dry land |
Total area |
|||||||||||||
|
R2 |
MAE (%) |
RMSE (%) |
AIC |
R2 |
MAE (%) |
RMSE (%) |
AIC |
R2 |
MAE (%) |
RMSE (%) |
AIC |
|
|||
Model A |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
0.554 |
0.137 |
0.181 |
218.616 |
0.149 |
0.160 |
0.203 |
223.969 |
0.038 |
0.309 |
0.422 |
379.939 |
|
||
|
20m |
0.510 |
0.145 |
0.189 |
221.939 |
0.332 |
0.163 |
0.193 |
220.789 |
0.152 |
0.285 |
0.396 |
371.479 |
|
||
|
30m |
0.449 |
0.161 |
0.201 |
226.007 |
0.203 |
0.148 |
0.201 |
223.400 |
0.059 |
0.299 |
0.417 |
378.432 |
|
||
|
80m |
0.237 |
0.204 |
0.237 |
237.680 |
0.077 |
0.156 |
0.210 |
226.448 |
-0.079 |
0.300 |
0.373 |
376.598 |
|
||
Model B |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
0.500 |
0.153 |
0.191 |
270.631 |
0.110 |
0.162 |
0.212 |
275.058 |
0.083 |
0.295 |
0.412 |
424.768 |
|
||
|
20m |
0.445 |
0.158 |
0.201 |
274.270 |
0.374 |
0.141 |
0.186 |
266.639 |
0.073 |
0.296 |
0.414 |
425.463 |
|
||
|
30m |
0.311 |
0.169 |
0.225 |
281.864 |
0.092 |
0.149 |
0.214 |
275.718 |
0.046 |
0.305 |
0.420 |
427.405 |
|
||
|
80m |
-0.007 |
0.198 |
0.271 |
295.134 |
-0.008 |
0.170 |
0.220 |
277.364 |
-0.085 |
0.301 |
0.374 |
424.925 |
|
||
Model C |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
0.049 |
0.191 |
0.264 |
293.148 |
0.137 |
0.158 |
0.209 |
274.015 |
0.153 |
0.300 |
0.396 |
419.404 |
|
||
|
20m |
0.247 |
0.173 |
0.235 |
284.961 |
0.187 |
0.177 |
0.213 |
275.246 |
0.178 |
0.287 |
0.389 |
417.433 |
|
||
|
30m |
0.164 |
0.177 |
0.247 |
288.627 |
0.213 |
0.150 |
0.199 |
270.992 |
0.120 |
0.302 |
0.403 |
422.013 |
|
||
|
80m |
0.089 |
0.182 |
0.258 |
291.609 |
0.137 |
0.150 |
0.203 |
272.228 |
-0.019 |
0.293 |
0.362 |
420.503 |
|
||
Model D |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
0.029 |
0.187 |
0.266 |
341.862 |
0.162 |
0.156 |
0.206 |
321.060 |
0.182 |
0.298 |
0.389 |
465.077 |
|
||
|
20m |
0.420 |
0.160 |
0.206 |
323.809 |
0.254 |
0.170 |
0.204 |
320.404 |
0.136 |
0.292 |
0.399 |
468.759 |
|
||
|
30m |
-0.135 |
0.204 |
0.288 |
347.332 |
0.197 |
0.158 |
0.202 |
319.666 |
0.105 |
0.310 |
0.407 |
471.093 |
|
||
|
80m |
0.089 |
0.182 |
0.258 |
339.609 |
0.116 |
0.154 |
0.206 |
321.018 |
-0.026 |
0.296 |
0.364 |
469.000 |
|
||
Model E |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
0.316 |
0.171 |
0.224 |
489.618 |
0.190 |
0.152 |
0.202 |
479.943 |
0.109 |
0.308 |
0.406 |
630.785 |
|
||
|
20m |
0.598 |
0.135 |
0.171 |
470.985 |
0.493 |
0.136 |
0.168 |
467.689 |
0.213 |
0.280 |
0.381 |
622.491 |
|
||
|
30m |
0.465 |
0.158 |
0.198 |
480.957 |
0.126 |
0.165 |
0.210 |
482.452 |
0.104 |
0.301 |
0.407 |
631.190 |
|
||
|
80m |
0.248 |
0.186 |
0.235 |
492.945 |
0.124 |
0.158 |
0.205 |
480.738 |
-0.053 |
0.308 |
0.368 |
630.784 |
|
||
Model F |
|
|
|
|
|
|
|
|
|
|
|
|
|
|||
|
3m |
0.286 |
0.175 |
0.229 |
539.126 |
0.247 |
0.143 |
0.195 |
525.544 |
0.139 |
0.289 |
0.398 |
676.442 |
|
||
|
20m |
0.639 |
0.131 |
0.162 |
515.166 |
0.447 |
0.145 |
0.175 |
518.538 |
0.135 |
0.287 |
0.399 |
676.788 |
|
||
|
30m |
0.429 |
0.162 |
0.204 |
531.313 |
0.125 |
0.164 |
0.207 |
529.549 |
0.091 |
0.295 |
0.410 |
680.158 |
|
||
|
80m |
0.165 |
0.192 |
0.247 |
544.579 |
0.090 |
0.154 |
0.209 |
529.973 |
-0.046 |
0.298 |
0.367 |
678.334 |
|
||
Notes: Model A, PlanetScope+DEM; Model B, PlanetScope+Sentinel-1+DEM; Model C, Sentinel-2+DEM; Model D, Sentinel-2+Sentinel-1++DEM; Model E, PlanetScope+Sentinel-2+DEM; Model F, PlanetScope+Sentinel-2+Sentinel-1+DEM.
Table 3 Performance results of Support Vector Regression predicting SOC based on different combinations of environmental variables at different modeling resolutions. The most accurate results are shown in bold.
Model |
Land Use Type |
||||||||||||||||
|
Paddy field |
Dry land |
Total area |
||||||||||||||
|
R2 |
MAE (%) |
RMSE (%) |
AIC |
R2 |
MAE (%) |
RMSE (%) |
AIC |
R2 |
MAE (%) |
RMSE (%) |
AIC |
|
||||
Model A |
|
|
|
|
|
|
|
|
|
|
|
|
|
||||
|
3m |
0.036 |
0.185 |
0.266 |
245.617 |
-0.010 |
0.169 |
0.226 |
231.213 |
-0.017 |
0.309 |
0.430 |
382.658 |
|
|||
|
20m |
0.033 |
0.185 |
0.266 |
245.741 |
0.012 |
0.178 |
0.235 |
233.710 |
0.001 |
0.310 |
0.429 |
382.453 |
|
|||
|
30m |
0.029 |
0.190 |
0.266 |
245.854 |
0.049 |
0.160 |
0.219 |
229.224 |
0.046 |
0.309 |
0.420 |
379.381 |
|
|||
|
80m |
0.029 |
0.195 |
0.267 |
245.867 |
0.020 |
0.160 |
0.217 |
228.443 |
0.010 |
0.285 |
0.357 |
370.405 |
|
|||
Model B |
|
|
|
|
|
|
|
|
|
|
|
|
|
||||
|
3m |
0.022 |
0.189 |
0.268 |
294.130 |
-0.006 |
0.168 |
0.226 |
279.087 |
-0.002 |
0.312 |
0.430 |
430.707 |
|
|||
|
20m |
0.030 |
0.187 |
0.266 |
293.821 |
-0.002 |
0.179 |
0.236 |
282.167 |
-0.009 |
0.312 |
0.431 |
431.124 |
|
|||
|
30m |
0.017 |
0.188 |
0.268 |
294.301 |
0.022 |
0.165 |
0.222 |
278.151 |
0.009 |
0.302 |
0.428 |
429.969 |
|
|||
|
80m |
0.025 |
0.188 |
0.267 |
294.017 |
0.027 |
0.159 |
0.216 |
276.191 |
0.018 |
0.286 |
0.356 |
417.861 |
|
|||
Model C |
|
|
|
|
|
|
|
|
|
|
|
|
|
||||
|
3m |
0.014 |
0.187 |
0.287 |
294.425 |
0.048 |
0.170 |
0.219 |
277.253 |
0.186 |
0.276 |
0.388 |
416.755 |
|
|||
|
20m |
0.069 |
0.184 |
0.261 |
292.407 |
-0.012 |
0.185 |
0.238 |
282.498 |
0.141 |
0.281 |
0.398 |
420.392 |
|
|||
|
30m |
0.032 |
0.184 |
0.266 |
293.759 |
0.056 |
0.166 |
0.218 |
276.970 |
0.045 |
0.298 |
0.419 |
427.452 |
|
|||
|
80m |
0.011 |
0.189 |
0.269 |
294.507 |
0.118 |
0.163 |
0.206 |
272.967 |
0.036 |
0.286 |
0.353 |
416.574 |
|
|||
Model D |
|
|
|
|
|
|
|
|
|
|
|
|
|
||||
|
3m |
0.017 |
0.185 |
0.268 |
342.304 |
0.046 |
0.171 |
0.219 |
325.347 |
0.069 |
0.311 |
0.415 |
473.784 |
|
|||
|
20m |
0.240 |
0.163 |
0.234 |
333.304 |
-0.009 |
0.189 |
0.237 |
330.410 |
0.064 |
0.299 |
0.416 |
474.087 |
|
|||
|
30m |
0.018 |
0.184 |
0.268 |
342.247 |
0.050 |
0.167 |
0.219 |
325.184 |
0.018 |
0.315 |
0.426 |
477.351 |
|
|||
|
80m |
0.008 |
0.190 |
0.269 |
342.615 |
0.063 |
0.163 |
0.212 |
322.957 |
0.033 |
0.288 |
0.353 |
464.784 |
|
|||
Model E |
|
|
|
|
|
|
|
|
|
|
|
|
|
||||
|
3m |
0.005 |
0.189 |
0.270 |
502.724 |
0.025 |
0.169 |
0.222 |
486.067 |
0.027 |
0.306 |
0.424 |
636.725 |
|
|||
|
20m |
0.021 |
0.188 |
0.268 |
502.147 |
0.025 |
0.183 |
0.233 |
483.589 |
0.027 |
0.309 |
0.424 |
636.733 |
|
|||
|
30m |
0.008 |
0.188 |
0.269 |
502.611 |
0.095 |
0.176 |
0.214 |
483.589 |
0.052 |
0.304 |
0.418 |
634.982 |
|
|||
|
80m |
0.048 |
0.183 |
0.264 |
501.167 |
0.019 |
0.160 |
0.217 |
484.476 |
0.035 |
0.281 |
0.353 |
624.630 |
|
|||
Model F |
|
|
|
|
|
|
|
|
|
|
|
|
|
||||
|
3m |
0.006 |
0.189 |
0.270 |
550.685 |
0.000 |
0.174 |
0.225 |
534.887 |
0.033 |
0.311 |
0.423 |
684.315 |
|
|||
|
20m |
0.222 |
0.173 |
0.239 |
542.136 |
-0.020 |
0.181 |
0.238 |
538.751 |
0.005 |
0.311 |
0.429 |
686.212 |
|
|||
|
30m |
0.009 |
0.189 |
0.269 |
550.584 |
0.033 |
0.174 |
0.221 |
533.790 |
0.025 |
0.304 |
0.424 |
684.829 |
|
|||
|
80m |
0.014 |
0.188 |
0.269 |
550.400 |
0.019 |
0.160 |
0.217 |
532.476 |
0.014 |
0.284 |
0.356 |
674.107 |
|
|||
Notes: Model A, PlanetScope+DEM; Model B, PlanetScope+Sentinel-1+DEM; Model C, Sentinel-2+DEM; Model D, Sentinel-2+Sentinel-1++DEM; Model E, PlanetScope+Sentinel-2+DEM; Model F, PlanetScope+Sentinel-2+Sentinel-1+DEM.
Comments 2: Figure 1 Add basic cartographic elements: compass, latitude and longitude.
Response 2: Compass, latitude and longitude are added in Figure 1 as follows:
Figure 1. (a) Elevation and distribution of soil sample points in the study area; (b) RGB image acquired by the PlanetScope optical satellite sensor (image date: July 27, 2017); and (c) distribution of paddy field and dry land in the study area.
Comments 3: Conclusions section: Add detailed figures for states such as "higher accuracy" (Line 610), "accuracy has shown a decline" (Line 625), "improve the prediction accuracy" (Line 619).
Response 3: Detailed figures are added in the Conclusions section as follows:
1) Different satellite data have their own strengths and weaknesses, and the selection and combination of satellite remote sensing data greatly influence the performance of prediction models. Compared to using a single remote sensing dataset, the synergistic utilization of multi-source remote sensing data can achieve complementarity, resulting in higher accuracy in model predictions. The combination of PlanetScope, Sentinel-2, and topography factors gave satisfactory predictions for dry land (R2 = 0.673, MAE = 0.107%, RMSE = 0.135%). The addition of Sentinel-1 indicators gave the best predictions for paddy field (R2 = 0.699, MAE = 0.114%, RMSE = 0.148%).
2) According to comparative analysis, it has been shown that considering multi-scale modeling contributes to optimizing prediction models and improving the predictive capability of target soil properties. Overall, it is generally observed that under moderate modeling resolutions, more ideal predictions of SOC can be achieved compared to fine and coarse resolutions.
3) The acquisition time of remote sensing images has a certain impact on the interpretation of models. The spectral variables obtained from winter images are the main interpretation variables for SOC.
4) In regions with complex land use types, constructing prediction models for each of them can help improve the prediction accuracy of SOC. Modeling for paddy field and dry land separately resulted in higher prediction accuracies compared to modeling across the entire region, with improvements (in terms of R2) of 36.0% and 33.4%, respectively.
5) The land use types to some extent affect the interpretation of models by satellite data. In models that combine PlanetScope and Sentinel-2 data, the inclusion of radar remote sensing variables has improved the prediction accuracy for paddy field. However, for dry land, the prediction accuracy declines, with R² dropping by10.2%.
Comments 4: 3) should be a paragraph. (Line 615)
Response 4: Done.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe manuscript proposes the use of multisource and multitemporal remote data for SOC estimation. This is a topic that fully falls within the objectives of Remote Sensing and of potential interest to its readers.
The use of data measured with different sensors and in different modalities poses major challenges that should be addressed properly and with appropriate methods that produce scientifically valid results. However, the manuscript has several flaws that the authors should fix to make it suitable for publication in Remote Sensing.
The manuscript has very ambitious objectives and uses very different types of data, which penalises it in terms of clarity. Arguably, this study could be more usefully divided into at least two manuscripts that enhance the work done by the authors and make the communication of the results to the scientific community clearer.
The abstract does not adequately summarise the manuscript and does not give the reader a clear idea of its contents.
Keywords cannot be sentences or duplicate words in the title. Keywords are intended to allow proper indexing of the manuscript and should therefore be chosen from those relevant to the manuscript and commonly used by other authors.
Although the Introduction is very long, it is not sufficiently structured and clear. The issue of the different size of the support to which the data is associated, which in the case of remote sensing imagery coincides with the spatial resolution, should be clearly explained and, above all, it should be explained how it has been solved by others and mention a possible novelty proposed by the authors that would improve existing knowledge.
The use of multi-temporal imaging to mosaic the study area to obtain a composite image of bare ground is the current standard. The authors propose to use the spectral response of the vegetation as proxy information for the estimation of the SOC without using the image mosaicking. This objective seems to me very ambitious and lacking a solid scientific basis.
Objectives should be clearer and more concise. Exploring or investigating are not doing research.
The materials and methods section should be better structured and provide all necessary details so that readers can follow the development of the manuscript. The methods are many and should also be supported by relevant and original bibliographic references. There always remains the question of how data of different types are made homogeneous in terms of spatial resolution and how point data (soil samples) can be correctly combined with other data.
The subsection on statistical analysis of the data should provide details of ANOVA and LSD and how these metrics were used and why. These should also be supported by relevant and original literature references.
The rest of the manuscript should be revised in connection with the new Materials and Methods section.
Author Response
Comments and Suggestions for Authors
The manuscript proposes the use of multisource and multitemporal remote data for SOC estimation. This is a topic that fully falls within the objectives of Remote Sensing and of potential interest to its readers.The use of data measured with different sensors and in different modalities poses major challenges that should be addressed properly and with appropriate methods that produce scientifically valid results. However, the manuscript has several flaws that the authors should fix to make it suitable for publication in Remote Sensing.
Response: We appreciate you very much for your work.
We have carefully revised the manuscript according to your suggestions and comments. These changes have greatly improved the quality of the manuscript, and we believe it now meets the requirement for publication in Remote Sensing.
The point-to-point responses to the comments were given below.
Comments 1: The manuscript has very ambitious objectives and uses very different types of data, which penalises it in terms of clarity. Arguably, this study could be more usefully divided into at least two manuscripts that enhance the work done by the authors and make the communication of the results to the scientific community clearer.
Response 1: The objective of this study is to evaluate the effectiveness of multi-source and multi-temporal remote sensing data for predicting SOC across arable land under vegetation cover condition in the Wuling Mountain region of Southwest China. To do this, we analyzed (i) the utility of synergistic application of multisource remote sensing data in predicting SOC; (ii) the feasibility of using multi-temporal and multiscale remote sensing imagery for SOC content mapping; and (iii) whether constructing SOC prediction models under individual land use types can improve prediction accuracy.
The results show that the synergy of multi-source remote sensing data significantly outperforms single remote sensing data. Modeling under four different spatial resolutions (3m, 20m, 30m, and 80m) reveals that a medium scale (20m) modeling resolution generally achieves the best performance. Within multi-temporal remote sensing data, winter optical imagery plays a dominant role in SOC prediction, suggesting that data from specific time periods have a crucial impact on model performance. Constructing SOC prediction models exclusively under specific land use types indeed helps to improve prediction accuracy. For instance, establishing separate models for paddy field or dry land can significantly enhance their precision. This study demonstrates the great potential of multi-source and multi-temporal remote sensing data in predicting SOC in areas with complex vegetation cover and emphasizes the importance of medium modeling resolution and data from specific time periods. The potential readers can gain a comprehensive understanding of how the combination of remote sensing data, modeling resolution, temporal phases, and land use influence prediction outcomes. This provides valuable reference material for researchers in related fields.
In future work, we will expand our research scope to include other agriculturally significant regions and incorporate data spanning different temporal scales to enhance the robustness of our conclusions. Furthermore, we plan to broaden our experimental framework to encompass a wider range of land use types, with the objective of evaluating the adaptability of our methods across diverse environmental settings and land management practices. Additionally, we intend to integrate multi-source data, including climate, soil conditions, crop growth patterns, and fertilization practices, to further refine the predictive accuracy of our models. This approach is expected to improve both the precision and generalizability of our methods, enabling us to provide more region-specific agricultural management recommendations. We anticipate synthesizing these insights into a comprehensive research paper that validates the reproducibility of our results and elucidates the applicability of these methodologies across varied agricultural landscapes.
Comments 2: The abstract does not adequately summarise the manuscript and does not give the reader a clear idea of its contents.
Response 2: The abstract is modified to summarise the work as follows,
Accurate prediction of soil organic carbon (SOC) is important for agriculture and land management. Methods using remote sensing data are helpful for estimating SOC in bare soils. To overcome the challenge of predicting SOC under vegetation cover, this study extracted spectral, radar, and topographic variables from multitemporal optical satellite images (high-resolution PlanetScope and medium-resolution Sentinel-2), synthetic aperture radar satellite images (Sentinel-1), and digital elevation model, respectively, to estimate SOC content in arable soils in the Wuling Mountain region of Southwest China. These variables were modeled at four different spatial resolutions (3m, 20m, 30m, and 80m) using the eXtreme Gradient Boosting algorithm. The results showed that modeling resolution, the combination of multisource remote sensing data, and temporal phases all influenced SOC prediction performance. The models generally yielded better results at a medium (20m) modeling resolution than at fine (3m) and coarse (80m) resolutions. The combination of PlanetScope, Sentinel-2, and topography factors gave satisfactory predictions for dry land (R2 = 0.673, MAE = 0.107%, RMSE = 0.135%). The addition of Sentinel-1 indicators gave the best predictions for paddy field (R2 = 0.699, MAE = 0.114%, RMSE = 0.148%). The values of R2 of the optimal models for paddy field and dry land improved by 36.0% and 33.4%, respectively, compared to that for the entire study area. The optical imageries in winter played a dominant role in prediction of SOC for both paddy field and dry land. This study offers valuable insights into effectively modeling soil properties under vegetation cover at various scales using multisource and multitemporal remote sensing data.
Comments 3: Keywords cannot be sentences or duplicate words in the title. Keywords are intended to allow proper indexing of the manuscript and should therefore be chosen from those relevant to the manuscript and commonly used by other authors.
Response 3: The keywords are modified as follows,
Multispectral imaging; Synthetic aperture radar data; Machine learning; Soil property prediction; Land use. (P1 L19-20)
Comments 4: Although the Introduction is very long, it is not sufficiently structured and clear. The issue of the different size of the support to which the data is associated, which in the case of remote sensing imagery coincides with the spatial resolution, should be clearly explained and, above all, it should be explained how it has been solved by others and mention a possible novelty proposed by the authors that would improve existing knowledge.
Response 4: The objective of this study is to evaluate the effectiveness of multi-source and multi-temporal remote sensing data for predicting SOC across arable land under vegetation cover condition in the Wuling Mountain region of Southwest China.
The novelty of this study lies in the synergistic utilization of optical and radar remote sensing data to comprehensively capture surface information, along with the dynamic monitoring of vegetation growth cycles using multi-temporal remote sensing data. This approach aims to optimize prediction accuracy in complex environmental conditions. Additionally, by employing multi-scale modeling to refine resolution and constructing independent models tailored for different land use types, the method effectively enhances the accuracy of existing SOC prediction methods.
Please refer to Page 1 Line 21 to Page 3 Line 136 for modifications of the introduction.
Comments 5: The use of multi-temporal imaging to mosaic the study area to obtain a composite image of bare ground is the current standard. The authors propose to use the spectral response of the vegetation as proxy information for the estimation of the SOC without using the image mosaicking. This objective seems to me very ambitious and lacking a solid scientific basis.
Response 5: Currently, the application of multi-temporal remote sensing data in SOC mapping is primarily concentrated in areas with exposed soil (e.g., Guo, 2021; Stevens et al., 2010; Íala et al., 2019; Wang et al., 2018; Tayeb et al., 2021). These studies are typically limited to fallow lands, agricultural seedbeds, and arid and semi-arid regions with sparse vegetation, because, in such areas, exposed soil is more easily observed within specific time windows, allowing multi-temporal remote sensing data to effectively capture the reflective characteristics of the soil surface (Wang et al., 2018). However, in the area under vegetation cover, the signal within the pixels of satellite imagery might be generated jointly by both vegetation and soil, making it challenging to isolate information that is purely attributable to the soil.
Numerous studies have demonstrated a significant correlation between vegetation growth conditions and soil properties (Wei et al., 2022; Wang et al., 2018; Chen et al., 2023). For instance, the health status of plants often directly reflects the nutrient content in the soil around the root zones, which is closely related to Soil Organic Carbon (SOC) (Chen et al., 2023). Satellite imagery such as Sentinel-2, MODIS, and Sentinel-1 can capture the spectral characteristics of vegetation, which can serve as indirect proxy variables to reflect soil properties (Wang et al., 2021; Paul et al., 2020; Hengl et al., 2017; Dahhani et al., 2024; Maynard and Levi, 2017; Pellerin et al., 2019). By analyzing these spectral data, we can infer vegetation characteristics associated with SOC, thereby achieving more accurate predictions of SOC.
Multi-source remote sensing data provide a more comprehensive and detailed description of surface features, while multi-temporal remote sensing data allow for dynamic tracking of information over different periods, thereby reducing biases introduced by short-term climatic conditions or incidental events affecting single observations (Luo et al., 2022). Although some studies have applied multi-source and multi-temporal remote sensing data for SOC prediction and achieved positive results (Wang et al., 2021; Grabska et al., 2020), the potential of high-resolution satellite imagery for estimating SOC in densely vegetated areas is still limited.
Therefore, we try to utilize multi-source (high- and medium-resolution, and radar) and multi-temporal remote sensing data to estimate SOC in this vegetation cover area.
Considering the differences in crop growth across different seasons in vegetated areas, utilizing multi-temporal data not only effectively reduces uncertainties that may arise from single-time-point data but also comprehensively considers factors such as annual climatic conditions and vegetation growth cycles’ impact on SOC (Luo et al., 2022). The vegetation indices used in our research have been widely applied for indirectly measuring soil fertility and organic matter content (Huete et al., 2002; Wang et al., 2021; Veloso et al., 2017; Trudel et al., 2012; Yu et al., 2020; He et al., 2022; Haj-Amor et al., 2022). These indices capture the corresponding plant responses during different growth stages and seasonal variations, providing reliable information for SOC estimation. Many researchers commonly employ image fusion techniques to integrate data when using multi-temporal remote sensing data for SOC prediction (Meng et al., 2022; Li et al., 2023; Žížala et al., 2019). While this approach can reduce input variables and lower model complexity, it may lead to information blurring during the fusion process, making it difficult to identify which temporal data contribute more significantly to the prediction results during feature variable importance analysis (Zhou et al., 2022; Zhu et al., 2012; Forkuor et al., 2017). In our study, spectral variables from each temporal phase are extracted separately as proxies for SOC prediction. This approach avoids unnecessary information mixing and ensures that each variable provides unique and valuable insights, clarifying which season’s remote sensing data is more crucial for SOC prediction and enhancing the interpretability of the results.
In summary, by integrating multi-source and multi-temporal remote sensing data, we not only mitigate issues caused by plant spectral reflectance on soil property observations but also capture unique information from different periods, thereby enhancing the accuracy of SOC predictions. This approach presents a novel and effective method for predicting SOC in densely vegetated areas and introduces new research perspectives to this field.
References
Guo, J., Zhu, Q., Zhao, X., Gou, X., Han, Y., Xu, J, 2020. Hyper-spectral inversion of soil organic carbon content under different land use types. Chinese Journal of Applied Ecology, 31(3), 863-871. (In Chinese with English abstract)
Stevens, A., Udelhoven, T., Denis, A., Tychon, B., Lioy, R., Hoffmann, L., Van Wesemael, B, 2010. Measuring soil organic carbon in croplands at regional scale using airborne imaging spectroscopy. Geoderma: An International Journal of Soil Science, 158, [specific pages not provided].
Íala, D., Minaík, R., Zádorová, T, 2019. Soil Organic Carbon Mapping Using Multispectral Remote Sensing Data: Prediction Ability of Data with Different Spatial and Spectral Resolutions. Remote Sensing, 11(24), 2947.
Wang, B., Waters, C., Orgill, S., Gray, J., Cowie, A., Clark, A., Liu, D. L, 2018. High-resolution mapping of soil organic carbon stocks using remote sensing variables in the semi-arid rangelands of eastern Australia. Science of the Total Environment, 630(July 15), 367-378.
Tayebi, M., Rosas, J. T. F., Mendes, W. D. S., Poppiel, R. R., Demattê, J. A. M, 2021. Drivers of organic carbon stocks in different LULC history and along soil depth for a 30 years image time series. Remote Sensing, 13(11), 1-32.
Chen L, Jiahui S, Taogetao B. Changes in soil organic carbon and nitrogen stocks following revegetation in a semi-arid grassland of North China[J]. Journal of Environmental Management, 2023, 346: 118995.
Yajuan W, Meiying L, Ji W, et al. The effects of vegetation communities on soil organic carbon stock in an enclosed desert-steppe region of northern China[J]. Soil Science and Plant Nutrition, 2022, 68(2): 284-294.
Wang X, Li Y, Chen Y, et al. Spatial pattern of soil organic carbon and total nitrogen, and analysis of related factors in an agro-pastoral zone in Northern China[J]. PLoS One, 2018, 13(5): e0197451.
Wang, H., Zhang, X., Wu, W., Liu, H, 2021. Prediction of soil organic carbon under different land use types using Sentinel-1/-2 data in a small watershed. Remote Sensing, 13(7), 1229.
Paul, S. S., Coops, N. C., Johnson, M. S., Krzic, M., Chandna, A., Smukler, S. M, 2020. Mapping soil organic carbon and clay using remote sensing to predict soil workability for enhanced climate change adaptation. Geoderma, 363, 114177.
Hengl, T., Leenaars, J. G., Shepherd, K. D., Walsh, M. G., Heuvelink, G. B., Mamo, T., Tilahun, H., Berkhout, E., Cooper, M., Fegraus, E, 2017. Soil nutrient maps of Sub-Saharan Africa: assessment of soil nutrient content at 250 m spatial resolution using machine learning. Nutrient Cycling in Agroecosystems, 109, 77-102.
Dahhani, S., Raji, M., Bouslihim, Y, 2024. Synergistic use of multi-temporal radar and optical remote sensing for soil organic carbon prediction. Remote Sensing, 16(11), 1871.
Maynard, J. J., Levi, M. R, 2017. Hyper-temporal remote sensing for digital soil mapping: Characterizing soil-vegetation response to climatic variability. Geoderma.
Pellerin, S., Bamière, L., Constantin, J., Launay, C., Richard, G, 2019. A model-based assessment of the soil C storage potential at the national scale: A case study from France. Science of the Total Environment, 579, 1094-110.
Peng L, Wu X, Feng C, et al. Assessing the potential of multi-source remote sensing data for cropland soil organic matter mapping in hilly and mountainous areas[J]. CATENA, 2024, 245: 108312.
Xie B, Ding J, Ge X, et al. Estimation of soil organic carbon content in the Ebinur Lake wetland, Xinjiang, China, based on multisource remote sensing data and ensemble learning algorithms[J]. Sensors, 2022, 22(7): 2685.
Luo, C., Zhang, X., Meng, X., Zhu, H., Ni, C., Chen, M., Liu, H, 2022. Regional mapping of soil organic matter content using multitemporal synthetic Landsat 8 images in Google Earth Engine. Catena, 209, 105842.
Grabska, E., Frantz, D., Ostapowicz, K, 2020. Evaluation of machine learning algorithms for forest stand species mapping using Sentinel-2 imagery and environmental data in the Polish Carpathians. Remote Sensing of Environment, 251, 112103.
Huete A, Didan K, Miura T, et al. Overview of the radiometric and biophysical performance of the MODIS vegetation indices[J]. Remote sensing of environment, 2002, 83(1-2): 195-213.
Veloso, A., Mermoz, S., Bouvet, A., Le Toan, T., Planells, M., Dejoux, J., Ceschia, E, 2017. Understanding the temporal behavior of crops using Sentinel-1 and Sentinel-2-like data for agricultural applications. Remote Sensing of Environment, 199, 415-426.
Trudel, M., Charbonneau, F., Leconte, R, 2012. Using RADARSAT-2 polarimetric and ENVISAT-ASAR dual-polarization data for estimating soil moisture over agricultural fields. Canadian Journal of Remote Sensing, 38(4), 514-527.
He, W., Wang, H., Ye, W., Tian, Y., Hu, G., Lou, Y., Pan, H., Yang, Q., Zhuge, Y, 2022. Distinct stabilization characteristics of organic carbon in coastal salt-affected soils with different salinity under straw return management. Land Degradation & Development, 33(13), 2246-2257.
Haj-Amor, Z., Araya, T., Kim, D., Bouri, S., Lee, J., Ghiloufi, W., Yang, Y., Kang, H., Jhariya, M. K., Banerjee, A, 2022. Soil salinity and its associated effects on soil microorganisms, greenhouse gas emissions, crop yield, biodiversity and desertification: A review. Science of the Total Environment, 843, 156946.
Meng X, Bao Y, Wang Y, et al. An advanced soil organic carbon content prediction model via fused temporal-spatial-spectral (TSS) information based on machine learning and deep learning algorithms[J]. Remote Sensing of Environment, 2022, 280: 113166.
Li X, Li Z, Qiu H, et al. Soil carbon content prediction using multi-source data feature fusion of deep learning based on spectral and hyperspectral images[J]. Chemosphere, 2023, 336: 139161.
Žížala D, Minařík R, Zádorová T. Soil organic carbon mapping using multispectral remote sensing data: Prediction ability of data with different spatial and spectral resolutions[J]. Remote Sensing, 2019, 11(24): 2947.
Zhou, X., Dandan, L., Huiming, Y., Honggen, C., Leping, S., Guojing, Y., Qingbiao, H., Brown, L., Malone, J. B, 2002. Use of landsat TM satellite surveillance data to measure the impact of the 1998 flood on snail intermediate host dispersal in the lower Yangtze River Basin. Acta Tropica, 82(2), 199-205.
Zhu Z, Woodcock C E, Olofsson P. Continuous monitoring of forest disturbance using all available Landsat imagery[J]. Remote sensing of environment, 2012, 122: 75-91.
Forkuor G, Hounkpatin O K L, Welp G, et al. High resolution mapping of soil properties using remote sensing variables in south-western Burkina Faso: a comparison of machine learning and multiple linear regression models[J]. PloS one, 2017, 12(1): e0170478.
Comments 6: Objectives should be clearer and more concise. Exploring or investigating are not doing research.
Response 6: A more clearer and concise objective is addressed in the revision as follows,
Hence, the objective of this study is to evaluate the effectiveness of multi-source and multi-temporal remote sensing data for predicting SOC across arable land under vegetation cover condition in the Wuling Mountain region of Southwest China. (P3 L124 to 126)
Comments 7: The materials and methods section should be better structured and provide all necessary details so that readers can follow the development of the manuscript. The methods are many and should also be supported by relevant and original bibliographic references. There always remains the question of how data of different types are made homogeneous in terms of spatial resolution and how point data (soil samples) can be correctly combined with other data.
Response 7: According to your suggestion, we have revised the Materials and Methods section. A flowchart of methodology is added to help readers follow the environmental data process (Figure 2 in Section 2.3). Meanwhile, the relevant references are also added in the revision.
To address the issue of homogenizing spatial resolution across different types of data, we provided a detailed explanation in Section 2.3 and illustrated the specific procedures for processing three types of data in Figure 2. For the two optical remote sensing datasets, identical preprocessing steps were applied, including radiometric calibration and atmospheric correction using the FLAASH model. Subsequently, to standardize pixel resolutions across optical, radar, and topographic data, we employed bilinear resampling within ArcGIS 10.8 to resample all images to resolutions of 3 meters, 20 meters, 30 meters, and 80 meters respectively, thereby achieving consistent spatial resolution among these datasets.
Regarding the accurate integration of point data (soil samples) with other datasets, we provided a comprehensive explanation in Section 2.1. At each designated sampling point, we established 4-6 sub-sampling points centered around the GPS-located point and extended radially by a distance of 30-50 meters for sample collection. The corresponding latitude and longitude coordinates for each sub-sampling point were recorded during this process. Subsequently, soil sampling data were matched with environmental data using the “Extract Multi Values to Points” tool in ArcGIS 10.8 under a common coordinate system to achieve precise data integration.
Figure 2. Environmental data preprocessing processes for modeling.
Comments 8: The subsection on statistical analysis of the data should provide details of ANOVA and LSD and how these metrics were used and why. These should also be supported by relevant and original literature references.
Response 8: One-way analysis of variance (ANOVA) was used to analyse the differences in SOC content between paddy field and dry land. The relevant reference is added in the revision. The sentence is modified as follows:
To analyse the differences in SOC content between paddy field and dry land, one-way analysis of variance (ANOVA) with a confidence interval of 0.05 was used because it effectively determines if there are significant differences between groups (Field A, 2024). These statistical analyses were performed using SPSS V.25 software package.
References
Field A. Discovering statistics using IBM SPSS statistics[M]. Sage publications limited, 2024.
Comments 9: The rest of the manuscript should be revised in connection with the new Materials and Methods section.
Response 9: The manuscript is thoroughly revised according to your comments. These modifications have enhanced the rigor of our research and improved the quality of our manuscript. We believe that the revised version meets the high standards of Remote Sensing and we hope these improvements will be satisfactory to you. Once again, we greatly appreciate your valuable time and professional guidance.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsI am satisfied with the revision of the authors. I only have one recommendation before publication:
I recommend the authors to append their comparisons among XGBoost, Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Support Vector Regression (SVR) in the Appendix in the manuscript.
Author Response
Comments 1: I recommend the authors to append their comparisons among XGBoost, Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Support Vector Regression (SVR) in the Appendix in the manuscript.
Response 1:We are extremely grateful for your thorough review and constructive feedback on our manuscript. Based on your suggestions, we have made the corresponding revisions. Specifically, we have added the predictive results of different model under three additional machine learning models (Random Forest (RF), Gradient Boosting Decision Trees (GBDT), and Support Vector Regression (SVR)) in Appendix B. Additionally, we have supplemented and clarified the content in Section 3.1 of the manuscript (P14 L393-396).
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors sufficiently addressed all comments and revised the manuscript appropriately. Some limitations of the manuscript remain from my point of view, but I believe the manuscript deserves to be published in Remote Sensing so as to encourage more comparison of the approach used in the scientific community potentially interested in it.
Author Response
We would like to extend our sincere gratitude for your thorough and professional review of our manuscript. Your valuable feedback and suggestions are greatly appreciated and will be immensely helpful in improving our work.