the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 01 Sep 2025
Monitoring planted forest expansion from 1990–2020 in China
Abstract. China has undertaken extensive afforestation efforts in recent decades. However, the effectiveness of these plantings varies with different environmental conditions. Whether China's forest expansion is primarily due to intentional planting or natural reforestation remains uncertain. Thus, assessing the growth of planted forests (PF) is crucial for monitoring forest quality and supporting China’s commitment to carbon neutrality. In this study, using 30 m Landsat time-series, we proposed a Continuous Change Detection and Classification (CCDC)-based PF expansion monitoring (C-PFM) method. Based on the C-PFM, 30 m annual maps for PF and natural forests (NF) across China from 1990 and 2020 were produced. The resulting PF map in 2020 achieved a F1-score of 79.2 % for PF and an overall accuracy of 90.8 % when validated against visually interpreted reference data. The PF maps for the years 1998, 2003, 2008, 2013, and 2018 were evaluated using data from the 5th, 6th, 7th, 8th, and 9th National Forest Inventory (NFI) data across 34 provinces and autonomous regions of China. The results demonstrated that all Pearson’s product-moment correlations were larger than 0.86. According to the C-PFM results, we found 8.06 million ha (Mha) of net forest gains across China from 1990 to 2020, with 16.15 Mha net gains of PF and 8.09 Mha net loss of NF. In eight forestry ecological engineering areas, we observed that the upper and middle reaches of Yangtze river Shelterbelt Program and Pearl River Shelterbelt Program experienced the most significant PF expansion. The resulting dataset can serve as valuable scientific data for policymakers, researchers, and forest managers, guiding appropriate planting, environment enhancement, and carbon sequestration efforts. The produced 30 m annual maps for PF and NF in China are publicly available at https://doi.org/10.5281/zenodo.15559086 (Xiao, 2025).
This preprint has been withdrawn.
-
Withdrawal notice
This preprint has been withdrawn.
-
Preprint
(5519 KB)
Interactive discussion
Status: closed
- CC1: 'Comment on ：Monitoring planted forest expansion from 1990-2020 in China', Sanglin Zhao, 14 Sep 2025
-
RC1: 'Comment on essd-2025-489', Anonymous Referee #1, 30 Sep 2025
1 It is recommended that the author revise the title. The primary focus of this study is the production of a continuous annual time-series dataset of plantations. As for the change analyses based on this data, they do not appear to be particularly innovative.
2 The challenges in producing annual data products include obtaining time-series samples and acquiring high-quality remote sensing imagery each year. Regarding the samples, they do not seem to be obtained through field surveys but rather through visual interpretation or change detection from remote sensing images. How does the author ensure the reliability of these samples? To my knowledge, distinguishing plantations from natural forests in remote sensing imagery is quite difficult, especially in southern China, which does not seem easily achievable. Additionally, acquiring annual cloud-free remote sensing images covering the entire Chinese region was likely challenging in the early years. How did the author address this issue? Approximately how many remote sensing images were used each year? These methodological details need to be supplemented by the author.
3 Sections 4.2 and 4.3, which discuss the impact of different sampling strategies on classification accuracy and model parameters, do not seem necessary in the Results section and are better suited for the Discussion section.
4 The author has employed best practices for accuracy evaluation, which is rigorous. However, the distribution of classification uncertainties, which is crucial for users of the produced data, has not been provided. It is hoped that the author will supplement this content and share it along with the data.
5 The figures in the paper, particularly the statistical charts, require careful optimization by the author. Many figures could be merged, and the font sizes in the figures are inconsistent.
Citation: https://doi.org/10.5194/essd-2025-489-RC1 -
RC2: 'Comment on essd-2025-489', Anonymous Referee #2, 09 Oct 2025
This work by Xiao et al. focuses on mapping PF and NF in China for long time periods. Although the topic is important for ecosystem monitoring and the carbon budget, the methods used are neither new or proper for this topic. Please see my comments below.
Major:
- I can not follow how the training data was generated. Specifically, in lines 191-192, the authors claim "PF exhibits higher disturbance frequencies than NF. The disturbance frequencies were calculated using the CCDC algorithm." How exactly did you distinguish PF from NF according to the CCDC time series detection? Did you use the CCDC to find the number of break points (representing disturbance frequency) of the time series, or did you compare the slope of each segment of PF with NF? This is completely missing.
- Following this statement, I also question if PF really exhibits higher disturbance frequencies than NF. How many PF pixels show clearly disturbed patterns in the time series in Fig. 1? Maybe some PFs are indeed more vulnerable to climate change but we are not sure, right? You will only know this information with ground-based records for the whole time period. However, the description of NFI ground data is also very vague. It seems the data is only used for validation. Given all of these, using the trait of disturbance frequency for all PFs could risk generality and the results are therefore not convincing.
- The feature selection for RF classifier is careless. The authors just threw all spectral reflectance or vegetation indices into the classifier, which could generate large redundancy and bring bias, especially given that the classifier (RF) is a machine learning-based method that cannot actually handle high-order complex and non-linearities. A deep learning algorithm may be more applicable for PF and NF classification.
Minor:
- Lines 165-169: What are these shapefiles used for?
- Lines 96-105: These are methods and should not be within the introduction part.
- Lines 10-11: "proposed". You cannot use the word "propose" here as this work is just a CCDC-based procedure used for land cover classification, not a new framework or model at all.
- The authors need to largely improve the English expression as awkward and redundant expressions are in many parts now and we can not actually follow in many parts what the authors are talking about. For instance, lines 96-105, line 203 "tile", which tile? This is a generally remote sensing term, e.g., the MODIS tile, but apparently the authors are not talking about remote sensing.
Citation: https://doi.org/10.5194/essd-2025-489-RC2
Interactive discussion
Status: closed
-
CC1: 'Comment on ：Monitoring planted forest expansion from 1990-2020 in China', Sanglin Zhao, 14 Sep 2025
Comment on ：Monitoring planted forest expansion from 1990-2020 in China
- 1.Core advantages
Innovative and practical method: Proposed the C-PFM method (CCDC+RF), which automatically generates 1.34 million training samples, integrates multiple sources of high confidence regions and disturbance frequency discrimination samples, and reduces the limitations of traditional manual sampling; Generate annual PF/NF maps with a resolution of 30 meters from 1990 to 2020, breaking through the limitation of 5-year intervals in similar research and accurately capturing dynamic processes.
Complete verification system: In 2020, the map OA reached 90.8%, the F1 score of PF was 79.2%, and the Pearson correlation coefficient with NFI data was high, showing better performance compared to four existing products; Conduct spatial stratification analysis based on climate zones and forest types to avoid regional bias.
Outstanding application value: Quantify the net increase of 16.15 Mha in national PF and the net decrease of 8.09 Mha in NF, clarify the expansion focus of PF in the engineering area, and provide a basis for carbon neutrality; The data is fully publicly available through Zenodo, in line with the trend of open science.
- 2.Main shortcomings
Sample and classification uncertainty: There are insufficient samples in ecologically fragile areas (such as Xinjiang and Qinghai), with 27.2% of PF pixels only recognized by C-PFM; Without introducing dynamic features to distinguish semi natural forests, the boundaries are fuzzy and prone to misclassification.
Weak model generalization: The local RF model has a validation accuracy of only 62.3% -73.8% in new regions; The robustness of CCDC core parameters has not been analyzed.
Shallow analysis of driving and ecology: Without combining policy nodes to quantify the contribution of policies to PF expansion; Only qualitative mention of PF ecological impact, lacking quantitative analysis.
Insufficient early validation: From 1990 to 2000, the sample size was small (89) and could not reflect the accuracy of early classification.
- 3.Suggestion
Optimize samples and classification: supplement fragile area samples and use active learning to optimize semi natural forest labeling; Add dynamic features (such as disturbance frequency change rate) to distinguish semi natural forests.
Improve model generalization: Build a "global local joint training" model to supplement climate and terrain features; Conduct sensitivity experiments on CCDC parameters.
Deepening analysis: Consider using the DID model to quantify policy effects; Quantitatively evaluate the ecological benefits of PF based on GEDI and GRACE data.
Improve early validation: Integrate Landsat 5 imagery and provincial forestry data to supplement early samples.
-
RC1: 'Comment on essd-2025-489', Anonymous Referee #1, 30 Sep 2025
1 It is recommended that the author revise the title. The primary focus of this study is the production of a continuous annual time-series dataset of plantations. As for the change analyses based on this data, they do not appear to be particularly innovative.
2 The challenges in producing annual data products include obtaining time-series samples and acquiring high-quality remote sensing imagery each year. Regarding the samples, they do not seem to be obtained through field surveys but rather through visual interpretation or change detection from remote sensing images. How does the author ensure the reliability of these samples? To my knowledge, distinguishing plantations from natural forests in remote sensing imagery is quite difficult, especially in southern China, which does not seem easily achievable. Additionally, acquiring annual cloud-free remote sensing images covering the entire Chinese region was likely challenging in the early years. How did the author address this issue? Approximately how many remote sensing images were used each year? These methodological details need to be supplemented by the author.
3 Sections 4.2 and 4.3, which discuss the impact of different sampling strategies on classification accuracy and model parameters, do not seem necessary in the Results section and are better suited for the Discussion section.
4 The author has employed best practices for accuracy evaluation, which is rigorous. However, the distribution of classification uncertainties, which is crucial for users of the produced data, has not been provided. It is hoped that the author will supplement this content and share it along with the data.
5 The figures in the paper, particularly the statistical charts, require careful optimization by the author. Many figures could be merged, and the font sizes in the figures are inconsistent.
Citation: https://doi.org/10.5194/essd-2025-489-RC1 -
RC2: 'Comment on essd-2025-489', Anonymous Referee #2, 09 Oct 2025
This work by Xiao et al. focuses on mapping PF and NF in China for long time periods. Although the topic is important for ecosystem monitoring and the carbon budget, the methods used are neither new or proper for this topic. Please see my comments below.
Major:
- I can not follow how the training data was generated. Specifically, in lines 191-192, the authors claim "PF exhibits higher disturbance frequencies than NF. The disturbance frequencies were calculated using the CCDC algorithm." How exactly did you distinguish PF from NF according to the CCDC time series detection? Did you use the CCDC to find the number of break points (representing disturbance frequency) of the time series, or did you compare the slope of each segment of PF with NF? This is completely missing.
- Following this statement, I also question if PF really exhibits higher disturbance frequencies than NF. How many PF pixels show clearly disturbed patterns in the time series in Fig. 1? Maybe some PFs are indeed more vulnerable to climate change but we are not sure, right? You will only know this information with ground-based records for the whole time period. However, the description of NFI ground data is also very vague. It seems the data is only used for validation. Given all of these, using the trait of disturbance frequency for all PFs could risk generality and the results are therefore not convincing.
- The feature selection for RF classifier is careless. The authors just threw all spectral reflectance or vegetation indices into the classifier, which could generate large redundancy and bring bias, especially given that the classifier (RF) is a machine learning-based method that cannot actually handle high-order complex and non-linearities. A deep learning algorithm may be more applicable for PF and NF classification.
Minor:
- Lines 165-169: What are these shapefiles used for?
- Lines 96-105: These are methods and should not be within the introduction part.
- Lines 10-11: "proposed". You cannot use the word "propose" here as this work is just a CCDC-based procedure used for land cover classification, not a new framework or model at all.
- The authors need to largely improve the English expression as awkward and redundant expressions are in many parts now and we can not actually follow in many parts what the authors are talking about. For instance, lines 96-105, line 203 "tile", which tile? This is a generally remote sensing term, e.g., the MODIS tile, but apparently the authors are not talking about remote sensing.
Citation: https://doi.org/10.5194/essd-2025-489-RC2
Data sets
Monitoring planted forest expansion from 1990-2020 in China Yuelong Xiao https://doi.org/10.5281/zenodo.15559086
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,148 | 84 | 26 | 1,258 | 21 | 20 |
- HTML: 1,148
- PDF: 84
- XML: 26
- Total: 1,258
- BibTeX: 21
- EndNote: 20
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Comment on ：Monitoring planted forest expansion from 1990-2020 in China
Innovative and practical method: Proposed the C-PFM method (CCDC+RF), which automatically generates 1.34 million training samples, integrates multiple sources of high confidence regions and disturbance frequency discrimination samples, and reduces the limitations of traditional manual sampling; Generate annual PF/NF maps with a resolution of 30 meters from 1990 to 2020, breaking through the limitation of 5-year intervals in similar research and accurately capturing dynamic processes.
Complete verification system: In 2020, the map OA reached 90.8%, the F1 score of PF was 79.2%, and the Pearson correlation coefficient with NFI data was high, showing better performance compared to four existing products; Conduct spatial stratification analysis based on climate zones and forest types to avoid regional bias.
Outstanding application value: Quantify the net increase of 16.15 Mha in national PF and the net decrease of 8.09 Mha in NF, clarify the expansion focus of PF in the engineering area, and provide a basis for carbon neutrality; The data is fully publicly available through Zenodo, in line with the trend of open science.
Sample and classification uncertainty: There are insufficient samples in ecologically fragile areas (such as Xinjiang and Qinghai), with 27.2% of PF pixels only recognized by C-PFM; Without introducing dynamic features to distinguish semi natural forests, the boundaries are fuzzy and prone to misclassification.
Weak model generalization: The local RF model has a validation accuracy of only 62.3% -73.8% in new regions; The robustness of CCDC core parameters has not been analyzed.
Shallow analysis of driving and ecology: Without combining policy nodes to quantify the contribution of policies to PF expansion; Only qualitative mention of PF ecological impact, lacking quantitative analysis.
Insufficient early validation: From 1990 to 2000, the sample size was small (89) and could not reflect the accuracy of early classification.
Optimize samples and classification: supplement fragile area samples and use active learning to optimize semi natural forest labeling; Add dynamic features (such as disturbance frequency change rate) to distinguish semi natural forests.
Improve model generalization: Build a "global local joint training" model to supplement climate and terrain features; Conduct sensitivity experiments on CCDC parameters.
Deepening analysis: Consider using the DID model to quantify policy effects; Quantitatively evaluate the ecological benefits of PF based on GEDI and GRACE data.
Improve early validation: Integrate Landsat 5 imagery and provincial forestry data to supplement early samples.