Multi-Year Cropland Mapping Based on Remote Sensing Data: A Case Study for the Khabarovsk Territory, Russia

Round 1

Reviewer 1 Report

The manuscript uses cubic polynomial, Fourier series, and double sinusoidal function to fit the time series of the NDVI data, and selects DF for data recovery and construction. Meanwhile, the SVM, RF, and GB classifiers are evaluated comparatively, among which the GB performs best. The DF and GB are then combined for cropland mapping. It is useful in Crop identification based on remote sensing data. Some specific comments on the manuscript are as follows.

1. For the study, have the authors tested that the optimal number of terms in the Fourier series expansion is two, as in line 268?

2. Could the authors provide some more discussion as to why the PA(grasses) of GB is less than that of RF (in Figure 9 and 10)? And it can be seen that the number of misclassified pixels in the other 4 classes is all larger in GB than those in RF.

3. There is a redundant “has not” in line 121.

4. In the second line of the Figure 2, it would be better to change the commas to dots.

There is no major problem with the English language.

Author Response

Thank you for your valuable comments to improve the quality of our research paper.

Comment 1. For the study, have the authors tested that the optimal number of terms in the Fourier series expansion is two, as in line 268?

Response 1: Yes, indeed, this is important and we have verified it. We calculated MAPE for the averaged time series of all classes in our study for 2021 and 2022, fitted by different numbers of terms in the Fourier series: one, two, three. MAPE were 11.8%, 6.2%, 4.6%, respectively. Using the ANOVA analysis, we revealed the significance of the differences (p<0.05). By pairwise comparison using Scheffe's criterion it was found, that the mean error for 2 terms and for 3 terms did not differ significantly (p>0.05), while the mean for 1 term and 2 terms and for 1 term and 3 terms differed significantly (p<0.05). An increase in the number of terms of Fourier series leads to an increase in the number of model parameters, which, firstly, requires more observations in time series (that can cause limitations in case for lack of uncloudy data), and secondly, increases the complexity of calculations and interpretation of the model.

Comment 2: Could the authors provide some more discussion as to why the PA (grasses) of GB is less than that of RF (in Figure 9 and 10)? And it can be seen that the number of misclassified pixels in the other 4 classes is all larger in GB than those in RF.

Response 2: You are right, F1 (grasses) for GB (0,62) is less that of RF (0,69). The issue of grasses recognition on fields where grass was not mowed or mowed partially is complex task. NDVI on this field was similar to fallow. Both algorithms are ensembles of trees, which implies some which implies share of indeterminism, especially for individual fields of individual classes. One of the algorithms has drawn the boundary for similar samples a little more clearly, but it is problematic to determine the cause of this unambiguously. In general, the RF and GB classifiers achieved a similar classification accuracy (Tukey test, p>0.05), the GB algorithm could significantly reduce the time cost (Tukey test, p<0.01). F1 for other 4 classes are similar: GB has unsignificant advantage in soybean and buckwheat recognition, RF – in fallow and oat recognition. For us, as well as for many other researchers, computational performance is important for numerous calculations (cross-validation, parameter selection). In such a situation, an order of magnitude acceleration of computational speed when applying the GB method is a decisive argument.

Comment 3: There is a redundant “has not” in line 121.

Response 3: Checked.

Comment 4: In the second line of the Figure 2, it would be better to change the commas to dots.

Response 4. Commas have been replaced by dots.

Author Response File: Author Response.pdf

Reviewer 2 Report

Approximation by nonlinear functions was proposed. Time series of weekly NDVI composites were plotted using multispectral Sentinel-2 (Level-2A) images at a resolution of 10 m for sites in Khabarovsk District from April to October 17 in the years 2021 and 2022. The overall accuracy (OA) for site pixels during cross-validation (Fourier series restored) was 67.3-85.9% for the SVM, RF, and GB classifiers, respectively. In general, this paper is written clearly and can be published if the authors addressed the following questions:

1.     The paper is quite technique with very implications and discussion on the results. How is you mapping compared with other datasets? The whole study also lacks of the sensitivity analysis and uncertainty analysis, which should be discussed in more detail.

2.      In the introduction I am missing about the impact of accurate land use map for estimating cropping nitrogen and carbon emissions and their impact on ecosystems, which I think it's very important part. Food-driven N emissions pose a significant part of human N input to ecosystems and oceans. See refs: Liu et al., Modeling global oceanic nitrogen deposition from food systems and its mitigation potential by reducing overuse of fertilizers, https://doi.org/10.1073/pnas.2221459120;  

3.     Are there any social-economic drivers on changes or mapping in cropland mapping and emissions? For instance, urbanization can change the land use substantially. I hope to see some additional discussions on changes in cropland and their social-economic drivers both from your study and others. How about the urbanization and ageing affect cropland mapping and emission pollution?  Deng et al. 2024 Nature communications, https://doi.org/10.1038/s41467-023-44685-y).

4.     I would like to see some discussion on how climate change/extreme weather affect cropland mapping or future predictions of yields and emissions, which I think it's an essential part of future efforts on maintaining food production. 

English should be improved by native speakers

Author Response

Thank you for your important comments. We added additional Shapley values analysis for GB to evaluate model sensitivity and uncertainty. Discussion were References [51-55] were added to discuss problems of N emissions and climate change. The Discussion has been updated.

Comment 1.     The paper is quite technique with very implications and discussion on the results. How is you mapping compared with other datasets? The whole study also lacks of the sensitivity analysis and uncertainty analysis, which should be discussed in more detail.

Response 1: Cross-validation was performed to minimize the uncertainty for accuracy estimation. At each iteration, a small independent dataset with rows of fields not included in the training set was used as a test set. Additionally, we have calculated Shapley values for all weeks to define contributions of every week (feature) for a class prediction. To assess the sensitivity of the developed model to NDVI variations throughout the vegetation season, the contribution of each model parameter (NDVI composite for a particular week) on the prediction result of the GB method was assessed. Calculation of the Shapley values suggested that the contribution of individual elements of the series for most classes is fairly uniform. The most important weeks for identification also vary from crop to crop. This is quite natural, given that different crops are harvested at different times, have different NDVI peaks, and sowing dates. Thus, NDVI s throughout the season are important for multi-class classification, and shortening the time series will lead to a decrease in model accuracy.

Comment 2.     In the introduction I am missing about the impact of accurate land use map for estimating cropping nitrogen and carbon emissions and their impact on ecosystems, which I think it's very important part. Food-driven N emissions pose a significant part of human N input to ecosystems and oceans. See refs: Liu et al., Modeling global oceanic nitrogen deposition from food systems and its mitigation potential by reducing overuse of fertilizers, https://doi.org/10.1073/pnas.2221459120.

Response 2: Accurate cropland mapping and vegetation analysis can also help assess the rational use of fertilizers, whose contribution to total nitrogen emissions is large enough that it can lead to significant changes in ecosystems.

Comment 3.     Are there any social-economic drivers on changes or mapping in cropland mapping and emissions? For instance, urbanization can change the land use substantially. I hope to see some additional discussions on changes in cropland and their social-economic drivers both from your study and others. How about the urbanization and ageing affect cropland mapping and emission pollution?  Deng et al. 2024 Nature communications, https://doi.org/10.1038/s41467-023-44685-y).

Response 3: Effective management in agriculture, coupled with well-managed urbanization, will allow for faster adoption of modern agricultural methods, which in turn will reduce nitrogen emissions, potentially improving air and water quality.

Comment 4.     I would like to see some discussion on how climate change/extreme weather affect cropland mapping or future predictions of yields and emissions, which I think it's an essential part of future efforts on maintaining food production.

Response 4:  The climate in Russia is warming about 2.5 times more intensely than the global average. The softening of winters in Russia could be an important factor in the future development of agriculture. An increase in precipitation and climate warming leads to an increase in soybean yields. In general, the change in agro-climatic conditions in the Middle Amur River Region is relatively favorable for soybean (which is the major crop in this region) cultivation.

Author Response File: Author Response.pdf