Next Article in Journal
LTSCD-YOLO: A Lightweight Algorithm for Detecting Typical Satellite Components Based on Improved YOLOv8
Next Article in Special Issue
Carbon Storage Simulation and Land Use Optimization for High-Water-Table Resource-Based Cities Based on the Coupled GMOP-PLUS-InVEST Model
Previous Article in Journal
Estimation of Spatial–Temporal Dynamic Evolution of Potential Afforestation Land and Its Carbon Sequestration Capacity in China
Previous Article in Special Issue
Detection of Cliff Top Erosion Drivers through Machine Learning Algorithms between Portonovo and Trave Cliffs (Ancona, Italy)
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 16
Issue 16
10.3390/rs16163100
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Aeolian Desertification Dynamics from 1995 to 2020 in Northern China: Classification Using a Random Forest Machine Learning Algorithm Based on Google Earth Engine

by
Caixia Zhang
1,*,
Ningjing Tan
1,2 and
Jinchang Li
3
1
Key Laboratory of Ecological Safety and Sustainable Development in Arid Lands, Northwest Institute of Eco-Environment and Resources, Chinese Academy of Sciences, No. 320, West Donggang Road, Lanzhou 730000, China
2
University of Chinese Academy of Sciences, Beijing 100049, China
3
Institute of Loess Plateau, Shanxi University, Taiyuan 030006, China
*
Author to whom correspondence should be addressed.
Remote Sens. 2024, 16(16), 3100; https://doi.org/10.3390/rs16163100
Submission received: 11 July 2024 / Revised: 16 August 2024 / Accepted: 20 August 2024 / Published: 22 August 2024
(This article belongs to the Special Issue Remote Sensing Interpretation Systematic Engineering for Natural Resources Monitoring and Management (Second Edition))
Browse Figures
Versions Notes

Abstract

Machine learning methods have improved in recent years and provide increasingly powerful tools for understanding landscape evolution. In this study, we used the random forest method based on Google Earth Engine to evaluate the desertification dynamics in northern China from 1995 to 2020. We selected Landsat series image bands, remote sensing inversion data, climate baseline data, land use data, and soil type data as variables for majority voting in the random forest method. The method’s average classification accuracy was 91.6% ± 5.8 [mean ± SD], and the average kappa coefficient was 0.68 ± 0.09, suggesting good classification results. The random forest classifier results were consistent with the results of visual interpretation for the spatial distribution of different levels of desertification. From 1995 to 2000, the area of aeolian desertification increased at an average rate of 9977 km2 yr−1, and from 2000 to 2005, from 2005 to 2010, from 2010 to 2015, and from 2015 to 2020, the aeolian desertification decreased at an average rate of 2535, 3462, 1487, and 4537 km2 yr−1, respectively.

1. Introduction

Desertification is defined as land degradation in arid, semi-arid, and semi-humid areas and is caused by interactions among climate change and human activities [1]. Desertification can be divided into aeolian desertification, water erosion desertification, salinization desertification, and freeze–thaw desertification. As the largest and most widely distributed desertification type, aeolian desertification is characterized by wind erosion that leads to sand transport and occurs during environmental degradation in arid and semi-arid areas and in some semi-humid areas [2,3,4]. Because aeolian desertification has serious impacts on socioeconomic development and the ecological environment [1,2], it is necessary to monitor the spatial and temporal dynamics of aeolian desertification [5,6,7].
Many researchers have monitored desertification using a range of methods, and this has been an active field of research in China. Visual interpretation is the most reliable method [8], and it has been used to evaluate the desertification dynamics in northern China [9] and in smaller regions [10,11,12]. In addition, computer technology, including both supervised and automatic classification, has been explored. The maximum likelihood method is a commonly used supervised classification method. For example, Gao and Liu [13] analyzed the dynamics of land use change in Tongyu County, Jilin Province, in 1992 and 2002, and the classification accuracy was as high as 75%. Based on the maximum likelihood method, Yang et al. [14] quantified land degradation in the Hunshandake Sandy Land in 1975, 1989, 1992, and 2001. The classification accuracy was 89 to 96%. Moumane et al. [15] extracted information on land use and land cover changes in Morocco’s Ternata oasis from 1991 to 2021 by using maximum likelihood classification, and the overall accuracy ranged from 90.8 to 92.0%.
The k-means dynamic clustering algorithm is often used in unsupervised classification. For example, Mao et al. [16] used an iterative self-organized data analysis (ISODATA) algorithm to analyze desertification in eight townships in Dingbian County, at the southern edge of the Mu Us Sandy Land, from 1991 to 1997, and the classification accuracy was between 78.6 and 82.6%. Wang et al. [17] applied the information decision tree method to classify desertified land and found that the hierarchical extraction method used by the decision tree reduced the probability of pixel mixing, but the standard for image classification in the hierarchical process required further discussion. Ma et al. [18] used a Fisher classifier to extract variables that were closely related to the desertification severity and established a discrimination equation (with an accuracy of 68.9%) to evaluate the desertification severity in the Ordos region. Qiao et al. [19] extracted desertified areas in the northwestern area of Keshiketeng Banner in Inner Mongolia based on a neural network method, and the accuracy of the classification was 84%. Meng et al. [20] used changes in the center of gravity and intensity analysis models to monitor desertification dynamics from 1990 to 2020 in Mongolia.
As research on “big data” has progressed, researchers have developed methods based on the integration of multiple indicators for desertification risk assessment. For example, Abuzaid and Abdelatif [21] identified desertification areas by integrating an erosion quality index with a modified Mediterranean Desertification and Land Use (MEDALUS) method, with good results; the R2 and root mean square error for the difference between the model’s prediction and the value calculated by the MEDALUS model were 0.88 and 0.11, respectively. Duan et al. [22] used the statistical tree classification method to develop an extraction model for aeolian desertification based on multiple indicators; the overall classification accuracy was 96.5%.
Machine learning methods have been improved in recent years [20,23]. One promising approach is the random forest classification method, which has shown high accuracy [24,25]. Sun et al. [26] generated optimal random forest models based on the relationships between species richness and 19 bioclimatic variables and mean monthly temperature and precipitation records. Chan and Paelinckx [27] recommended this method because it can be trained faster and is more stable in the detailed mapping of ecotopes. Wang et al. [28] used a state-of-the-art multi-model random forest ensemble method to integrate the results of 10 models and reproduced the terrestrial vegetation carbon density in China from 1982 to 2010; their multi-model ensemble mean method for estimating terrestrial vegetation carbon density overestimated the actual value by 2%, whereas their multi-model random forest ensemble method underestimated the actual value by only 0.2%. Zhang et al. [29] estimated the area of karst rocky desertification in 2001, 2011, 2016, and 2020 using a random forest model, and achieved an overall accuracy of 94.7%.
Given the ease of training and high accuracy of the random forest approach, we adopted this approach for the current study. Our goal was to evaluate the dynamics of desertification in northern China from 1995 to 2020. The following aspects of our study are novel: to the best of our knowledge, this is the first use of random forest models to quantify desertification in northern China based on multiple variables that are related to desertification, and we integrated existing desertification datasets for multiple indicators. To do so, we trained the random forest classifier using these variables and existing desertification datasets.

2. Methods and Data Sources

2.1. Methods

Random forest is an integrated classifier based on decision trees. It extracts multiple samples from the original samples, then models each sample using a decision tree. When this analysis is complete, it then combines multiple trees for prediction and finally obtains prediction results by voting, which involves choosing the prediction that is provided by the largest number of decision trees [30]. Random forest has the advantages of high prediction accuracy, no requirement for preprocessing samples, a high tolerance for outliers and noise, small prediction errors, and s fast running speed [31,32,33]. In this study, we constructed five categories of decision trees: trees based on Landsat images, remote sensing inversion data, climate data, land use data, and soil type data. The final result was based on the majority voting results (Figure 1). Since the random forest method is sensitive to the sampling design [25], we selected the desertification dataset for northern China in 2000 as the training data from National Cryosphere Desert Data Center (http://www.ncdc.ac.cn, accessed on 1 November 2023). These data were used as the seed to produce data in other years (i.e., sample migration). The dataset was processed using the same criteria for all parts of China, so it is generally considered to be a high-quality dataset and is broadly accepted in China [33]. The dataset contains four types of desertification (wind erosion, water erosion, salinization erosion, and freeze–thaw erosion), each of which is divided into four severity classes: slight, moderate, severe, and extremely severe. These categories were defined according to the classification criteria of Wang et al. [34]. We used Google Earth Engine to implement models and to map the model outputs. The method we used automatically accounted for differences in the resolution of the datasets. We used the desertification dataset for northern China in 2000 as the training sample and selected at least 80,000 random points for training within each region of interest. We used 70% of the random points for training and the remaining 30% to validate the training results.
Given the large spatial scale and regional differences in northern China and limits on the calculation dataset and the processing speed of Google Earth Engine, it is necessary to carry out partition processing in remote sensing image interpretation. We divided the desertification area in northern China into 16 regions of interest based on different geomorphologies (Figure 2). We then established classifiers for each region to use in training and obtained data for the classifiers corresponding to the characteristics of each region. Finally, we used the classifiers to classify the images from 1995 to 2000 in each region of interest.

2.2. Data Sources

We used remote sensing images of the study area from 1995 to 2020. These included Landsat-5 TM and Landsat-7 ETM bands 1 to 5 and 7 and Landsat-8 OLI bands 2 to 7. We used the median method for image fusion. Remote sensing inversion data included the normalized difference vegetation index (NDVI), land surface albedo, the modified soil-adjusted vegetation index (MSAVI), topsoil grain size index (TGSI), modified normalized difference water index (MNDWI), bare soil index (BSI), and normalized difference barren index (NDBI). Table 1 summarizes how the indexes are calculated. As the seasonal changes in desertified land are very different from those in non-desertified land, studying the annual vegetation change will help to improve the classification accuracy. To support this analysis, we calculated frequency statistics for the NDVI from March to October and calculated multiple percentiles and percentile interval mean values: the 50, 60, 75, 90, 95, and 100% (maximum) percentiles and the 50 to 75%, 75 to 90%, and 90 to 100% interval mean values.
The climate baseline dataset consists of monthly precipitation, average temperature, minimum temperature, and maximum temperature in China, with a spatial resolution of 0.0083333° (about 1 km) from 1991 to 2020. The data can be downloaded from China’s National Earth System Science Data Center (http://www.geodata.cn, accessed on 5 March 2023). According to the recommendations of the organization [35], countries should start to use the dataset from 1981 to 2010 as a baseline. The new benchmark for 1991 to 2020 should be used starting in 2021, and it will be updated every 10 years thereafter.
The land use and cover type data used in this study had a spatial resolution of 1 km and were obtained from 1995 to 2020 at 5-year intervals (http://www.geodata.cn/index.html, accessed on 10 March 2023). We adopted the secondary classification system that included paddy fields, dry land, closed-canopy woodland, shrub communities, open woodland, other woodland, grassland with high, medium, and low coverage (>50%, 20 to 50%, and 5 to 20%, respectively), canals, lakes, reservoirs and ponds, permanent glaciers, tidal flats, beaches, urban land, rural residential areas, other construction land, sandy land, gobis, saline or alkali land, marshland, bare land, and bare rock land.
We obtained the spatial distribution of soil types in China from the Harmonized World Soil Database (https://www.fao.org/soils-portal/data-hub/soil-maps-and-databases/harmonized-world-soil-database-v12/en/, accessed on 3 April 2023). We followed the FAO-90 soil classification system based on the composition of the soil, which is defined based on the gravel, sand, silt, and clay contents.

3. Results

3.1. Desertification Distribution across Northern China in 1995

The area with slight desertification was distributed mainly in the Hulunbuir Sandy Land, the western part of the Mu Us Sandy Land, the southern part of the Tengger Desert, the Qaidam Basin, and the Gonghe Basin (Figure 3). The area with moderate desertification was distributed mainly in the northern part of the Gurbantunggut Desert. The areas with severe desertification were distributed mainly in the northwestern part of the Taklamakan Desert, the southern part of the Gurbantunggut Desert, and the eastern part of the Badain Jaran Desert. The areas with extremely severe desertification were distributed mainly in the northern part of the Gurbantunggut Desert, lower elevations of the Heihe River Basin, the eastern part of the Mu Us Sandy Land, and the southern part of the Horqin Sandy Land.

3.2. Changes in Desertification from 1995 to 2000 in Northern China

From 1995 to 2000, the area with slight desertification increased by 17,618 km2, the area of moderate desertification increased by 14,289 km2, the area of severe desertification increased by 10,059 km2, and the area of extremely severe desertification increased by 7918 km2 (Figure 4). Overall, the desertification area in northern China increased by 49,884 km2. During this period, the area of aeolian desertification increased at an average rate of 9977 km2 yr−1. Slight desertification developed near the Kumutage Desert, and severe desertification developed near the southern part of the Tengger Desert. The severe and extremely severe desertification developed near farming areas and pastoral ecotones, including the Horqin Sandy Land, Otindag Sandy Land, and Mu Us Sandy Land.

3.3. Changes in the Area of Desertification from 2000 to 2005 in Northern China

From 2000 to 2005, the area with slight desertification decreased by 7318 km2, the area with moderate desertification decreased by 14,714 km2, the area with severe desertification decreased by 11,872 km2, and the area with extremely severe desertification decreased by 12,678 km2 (Figure 5). Thus, the area of desertification decreased by 46,582 km2 during this period in northern China. The area of aeolian desertification decreased at an average rate of 2535 km2 yr−1. A widespread trend in desertification reversal occurred across northern China during this period.

3.4. Changes in the Area of Desertification from 2005 to 2010 in Northern China

From 2005 to 2010, the area of slight desertification increased by 3230 km2, the area of moderate desertification decreased by 4096 km2, the area of severe desertification decreased by 3623 km2, and the area of extremely severe desertification decreased by 12,825 km2 (Figure 6). Thus, the area of desertification during this period in northern China decreased by 17,314 km2. The area of aeolian desertification decreased at an average rate of 3462 km2 yr−1. The slight desertification near the Qaidam Basin, Horqin Sandy Land, and Otingdag Sandy Land increased, but other areas showed decreasing desertification.

3.5. Changes in the Desertification Area from 2010 to 2015 in Northern China

From 2010 to 2015, the area of slight desertification decreased by 2398 km2, the area with moderate desertification decreased by 4994 km2, the area with severe desertification decreased by 4210 km2, and the area with extremely severe desertification increased by 4169 km2 (Figure 7). Thus, the area of desertification decreased by 7433 km2 during this period in northern China. The area of aeolian desertification decreased at an average rate of 1487 km2 yr−1. The area of extremely severe desertification increased near the Ulan Buh Desert, the lower elevations in the Heihe River Basin, the southern part of the Tengger Desert, and the northwestern part of the Taklimakan Desert. The desertification in Northern China during this period was reversed.

3.6. Changes in the Area of Desertification from 2015 to 2020 in Northern China

From 2015 to 2020, the area of slight desertification decreased by 12,546 km2, the area of moderate desertification decreased by 5630 km2, the area of severe desertification decreased by 573 km2, and the area of extremely severe desertification decreased by 3935 km2 (Figure 8). Thus, the area of desertification decreased by 22,685 km2 during this period in northern China. The area of aeolian desertification decreased at an average rate of 4537 km2 yr−1. By 2020, the area of desertification in the Mu Us Sandy Land had greatly decreased.

4. Discussion

4.1. Accuracy Evaluation

To evaluate the classification results, we used the kappa coefficient, which we calculated based on a confusion matrix [36,37]. We selected the 2000 classification data and compared them with the 2000 desertification dataset to calculate the overall classification accuracy and the kappa coefficient. The average accuracy was 91.6% ± 5.8 [mean ± SD], and the average kappa was 0.68 ± 0.09, suggesting good classification results (Table 2). The random forest classifier achieved an accuracy less than 90% only in the Mu Us Sandy Land, Kubuqi Desert, and the Otindag and southern Horqin Sandy Lands. The random forest method had relatively higher classification accuracy in arid areas.

4.2. Validation of the Spatial Results of the Random Forest Classifier

The spatial distribution of the random forest results was strongly similar to the results of the visual analysis (Figure 9a,b). Although the total area of desertification differed between the visual and random forest methods, with continuous classification for the visual method and pixel-based classification for the random forest method, both methods produced very similar results (Figure 9c,d). From the classification results, the prediction results show a discrete distribution in space. However, the results based on remote sensing visual interpretation, due to the operator’s experience and knowledge, show a relatively continuous spatial distribution. Although the method classification results were differing from the visual interpretation at the pixel level, we could still see its advantages in desertification classification assisting. Furthermore, for the random forest method, the development of appropriate post-classification-impact processing methods should become the focus of the future application of machine learning in desertification classification. The Xinjiang region (Figure 9) provides an example. The desertification in the Xinjiang region was mainly distributed in the lower reaches of inland rivers (i.e., rivers that do not reach the ocean), such as the Ertix He River and the Ulurgur He River and at the periphery of oases and the edges of the Gurbantunggut Desert and Taklimakan Desert, which experienced aeolian erosion and sand transport (Figure 9c,d).
Desertification in the central part of northern China was mainly distributed in the lower reaches of inland rivers (i.e., rivers that do not reach the ocean), such as the Shiyanghe, Heihe, and Shulehe River Basins, desert–oasis ecotones, and areas with unsustainable groundwater extraction. The desertified land in the Qaidam Basin is mainly distributed in the wind erosion area in the northwestern piedmont plain. In the source regions of the Yangtze River and the Yellow River, desertification is mainly distributed in river valleys, ancient river beds, and the foothills of the floodplains around lakes. The desertification of grasslands is mainly distributed in areas with overgrazing. Desertification in the Mu Us Sandy Land mainly follows a scattered distribution throughout the region (Figure 10). Under the interaction of climate change and human intervention, the area of desertification in the Mu Us Sandy Land decreased since the year 2000, which accorded with the results of Feng et al. [8] and Wang et al. [38].
Desertification in the northeastern part of northern China mainly included the Songnen Sandy Land, Hulunbuir Sandy Land, Otindag Sandy Land, and Horqin Sandy Land. In the Songnen Sandy Land, the desertification was mainly distributed on terrace plains, and in alluvial fans. In the Horqin Sandy Land, the desertification was mainly distributed in areas with degraded grassland and in agricultural reclamation areas. In the Otindag Sandy Land, the desertification was mainly distributed in the southern grassland and agricultural reclamation areas (Figure 11).

4.3. Changes in the Desertification Area since 1995 in Northern China

From 1995 to 2020, we observed the overall aeolian desertification trends in northern China. Desertification developed rapidly at first and then reversed slowly (Figure 12). From 1995 to 2000, the area of aeolian desertification increased at an average rate of 9977 km2 yr−1, and from 2000 to 2005, from 2005 to 2010, from 2010 to 2015, and from 2015 to 2020, the aeolian desertification decreased at an average rate of 2535, 3462, 1487, and 4537 km2 yr−1, respectively. Desertification reversed after 2000, which is consistent with many previous research results [8,11,39,40].
From 1995 to 2000, the area with slight desertification increased rapidly, and then decreased until 2020, with some fluctuation. The areas of moderate, severe, and extremely severe desertification decreased from 2000 to 2020, with the small exception of extremely severe desertification, which increased from 2010 to 2015. Desertification increased from 1995 to 2000 and then decreased.

5. Conclusions

We used the random forest method based on Google Earth Engine to evaluate the desertification dynamics in northern China from 1995 to 2020. We used Landsat data, seven types of remote sensing inversion data, climate baseline data, land use data, and soil type data to develop random forest models and chose the best model based on majority voting.
(1)
The method’s average classification accuracy was 91.6% ± 5.8, and the average kappa coefficient was 0.68 ± 0.09. The random forest classifier results therefore provided a relatively accurate prediction of the distribution of desertified land and desertification severity.
(2)
From 1995 to 2000, the area of aeolian desertification increased at an average rate of 9977 km2 yr−1, and from 2000 to 2005, from 2005 to 2010, from 2010 to 2015, and from 2015 to 2020, the aeolian desertification decreased at an average rate of 2535, 3462, 1487, and 4537 km2 yr−1, respectively. From 1995 to 2000, the area with slight desertification increased rapidly and then decreased until 2020, with some fluctuation.
(3)
The areas of moderate, severe, and extremely severe desertification decreased from 2000 to 2020, with the small exception of the changing trend in extremely severe desertification during 2010 to 2015. Desertification increased from 1995 to 2000 and then decreased.
The random forest classifier results were highly consistent with the sample data in terms of the predicted severity and its spatial distribution, Therefore, for the random forest method, the development of appropriate post-classification-impact processing methods should be the focus of the future application of machine learning methods in desertification classification. Furthermore, the uncertainty of remote sensing data including NDVI data [40] may affect the research results.

Author Contributions

Conceptualization, C.Z. and J.L.; methodology, C.Z.; software, C.Z.; validation, C.Z., J.L. and N.T.; writing—original draft preparation, C.Z.; writing—review and editing, C.Z.; visualization, N.T.; project administration, C.Z.; funding acquisition, C.Z. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the National Key Research and Development Program of China, grant number 2020YFA0608404, the National Nature Science Foundation of China, grant number 41101006, and the Natural Science Foundation of Shanxi Province, grant number 20210302124032.

Data Availability Statement

The original contributions presented in the study are included in the article, further inquiries can be directed to the corresponding author.

Acknowledgments

The authors especially thank Zhenting Wang for his constructive comments. Also, we particularly thank Wenyong Ma for his assistance in code modification.

Conflicts of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References

  1. UNCCD. United Nations: Convention to Combat Desertification in Those Countries Experiencing Serious Drought and/or Desertification, Particularly in Africa. In International Legal Materials; Cambridge University Press: Cambridge, UK, 1994; pp. 1328–1382. [Google Scholar]
  2. Zhu, Z.D.; Liu, S. The concept of desertification and the differentiation of its development. J. Desert Res. 1984, 4, 2–8. [Google Scholar]
  3. Zhu, Z.D. Concept, cause and control of desertification in China. Quat. Sci. 1998, 18, 145–155. [Google Scholar]
  4. Wang, T.; Zhu, Z.D. Study on sandy desertification in China—1. Definition of sandy desertification and its connotation. J. Desert Res. 2003, 23, 209–214. [Google Scholar]
  5. Guo, B.; Yang, F.; Fan, Y.; Zang, W. The dominant driving factors of rocky desertification and their variations in typical mountainous karst areas of southwest China in the context of global change. Catena 2023, 220, 106674. [Google Scholar] [CrossRef]
  6. Shao, W.; Wang, Q.; Guan, Q.; Zhang, J.; Yang, X.; Liu, Z. Environmental sensitivity assessment of land desertification in the Hexi Corridor, China. Catena 2023, 220, 106728. [Google Scholar] [CrossRef]
  7. Elnashar, A.; Zeng, H.; Wu, B.; Gebremicael, T.G.; Marie, K. Assessment of environmentally sensitive areas to desertification in the Blue Nile Basin driven by the MEDALUS-GEE framework. Sci. Total Environ. 2022, 815, 152925. [Google Scholar] [CrossRef]
  8. Feng, K.; Wang, T.; Liu, S.; Yan, C.; Kang, W.; Chen, X.; Guo, Z. Path analysis model to identify and analyse the causes of aeolian desertification in Mu Us Sandy Land, China. Ecol. Indic. 2021, 124, 107386. [Google Scholar] [CrossRef]
  9. Wang, T.; Song, X.; Yan, C.Z.; Li, S.; Xie, J.L. Remote sensing analysis on aeolian desertification trends in northern China during 1975–2010. J. Desert Res. 2011, 31, 1351–1356. [Google Scholar]
  10. Song, X.; Yang, G.; Yan, C.; Duan, H.; Liu, G.; Zhu, Y. Driving forces behind land use and cover change in the Qinghai-Tibetan Plateau: A case study of the source region of the Yellow River, Qinghai Province, China. Environ. Earth Sci. 2009, 59, 793–801. [Google Scholar] [CrossRef]
  11. Yan, C.; Song, X.; Zhou, Y.; Duan, H.; Li, S. Assessment of aeolian desertification trends from 1975’s to 2005’s in the watershed of the Longyangxia Reservoir in the upper reaches of China’s Yellow River. Geomorphology 2009, 112, 205–211. [Google Scholar] [CrossRef]
  12. Song, X.; Wang, T.; Xue, X.; Yan, C.; Li, S. Monitoring and analysis of aeolian desertification dynamics from 1975 to 2010 in the Heihe River Basin, northwestern China. Environ. Earth Sci. 2015, 74, 3123–3133. [Google Scholar] [CrossRef]
  13. Gao, J.; Liu, Y.S. Determination of land degradation causes in Tongyu County, Northeast China via land cover change detection. Int. J. Appl. Earth Obs. Geoinf. 2010, 12, 9–16. [Google Scholar] [CrossRef]
  14. Yang, X.; Ding, Z.; Fan, X.; Zhou, Z.; Ma, N. Processes and mechanisms of desertification in northern China during the last 30 years, with a special reference to the Hunshandake Sandy Land, eastern Inner Mongolia. Catena 2007, 71, 2–12. [Google Scholar] [CrossRef]
  15. Moumane, A.; Al Karkouri, J.; Benmansour, A.; El Ghazali, F.E.; Fico, J.; Karmaoui, A.; Batchi, M. Monitoring long-term land use, land cover change, and desertification in the Ternata oasis, Middle Draa Valley, Morocco. Remote Sens. Appl. Soc. Environ. 2022, 26, 100745. [Google Scholar] [CrossRef]
  16. Mao, X.; Zhao, P.; Wang, D. Nonsupervising digital classification and GIS in the dynamic monitoring of soil desertification application. J. Northwest For. Univ. 2005, 20, 6–9. [Google Scholar]
  17. Wang, J.; Dong, G.; Li, W.; Wang, L.; Tang, H. Primary study on the multi-layer remote sensing information extraction of desertification land types by using decision tree technology. J. Desert Res. 2000, 20, 12–16. [Google Scholar]
  18. Ma, J.H.; Cai, D.H.; Xiao, T. Qualified evaluating on the remote sensing of desertification—A case study of Erdos region. J. Lanzhou Univ. (Nat. Sci.) 2007, 43, 1–6. [Google Scholar]
  19. Qiao, P.L.; Zhang, J.X.; Lin, Z.J. An artificial neural network method for the information of desertification extraction. Acta Geod. Cartogr. Sin. 2004, 33, 58–62. [Google Scholar]
  20. Meng, X.; Gao, X.; Li, S.; Li, S.; Lei, J. Monitoring desertification in Mongolia based on Landsat images and Google Earth Engine from 1990 to 2020. Ecol. Indic. 2021, 129, 107908. [Google Scholar] [CrossRef]
  21. Abuzaid, A.S.; Abdelatif, A.D. Assessment of desertification using modified MEDALUS model in the north Nile Delta, Egypt. Geoderma 2022, 405, 115400. [Google Scholar] [CrossRef]
  22. Duan, H.; Wang, T.; Xue, X.; Yan, C. Dynamic monitoring of aeolian desertification based on multiple indicators in Horqin Sandy Land, China. Sci. Total Environ. 2019, 650, 2374–2388. [Google Scholar] [CrossRef]
  23. Bai, Z.; Han, L.; Jiang, X.; Liu, M.; Li, L.; Liu, H.; Lu, J. Spatiotemporal evolution of desertification based on integrated remote sensing indices in Duolun County, Inner Mongolia. Geoecol. Inform. 2022, 70, 101750. [Google Scholar]
  24. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  25. Belgiu, M.; Drăguţ, L. Random forest in remote sensing: A review of applications and future directions. ISPRS J. Photogramm. Remote Sens. 2016, 114, 24–31. [Google Scholar] [CrossRef]
  26. Sun, Y.; Sun, Y.; Yao, S.; Akram, M.A.; Hu, W.; Dong, L.; Li, H.; Wei, M.; Gong, H.; Xie, S.; et al. Impact of climate change on plant species richness across drylands in China: From past to present and into the future. Ecol. Indic. 2021, 132, 108288. [Google Scholar] [CrossRef]
  27. Chan, J.C.-W.; Paelinckx, D. Evaluation of Random Forest and Adaboost tree-based ensemble classification and spectral band selection for ecotope mapping using airborne hyperspectral imagery. Remote Sens. Environ. 2008, 112, 2999–3011. [Google Scholar] [CrossRef]
  28. Wang, Z.; Gong, H.; Huang, M.; Gu, F.; Wei, J.; Guo, Q.; Song, W. A multimodel random forest ensemble method for an improved assessment of Chinese terrestrial vegetation carbon density. Methods Ecol. Evol. 2021, 14, 117–132. [Google Scholar] [CrossRef]
  29. Zhang, Y.; Tian, Y.; Li, Y.; Wang, D.; Tao, J.; Yang, Y.; Lin, J.; Zhang, Q.; Wu, L. Machine learning algorithm for estimating karst rocky desertification in a peak-cluster depression basin in southwest Guangxi, China. Sci. Rep. 2022, 12, 19121. [Google Scholar] [CrossRef]
  30. Ayala-Izurieta, J.E.; Márquez, C.O.; García, V.J.; Recalde-Moreno, C.G.; Rodríguez-Llerena, M.V.; Damián-Carrión, D.A. Land cover classification in an ecuadorian mountain geosystem using a random forest classifier, spectral vegetation indices, and ancillary geographic data. Geosciences 2017, 7, 34. [Google Scholar] [CrossRef]
  31. Kelsey, E.N.; Gunn, G.E.; Shiklomanov, N.I.; Engstrom, R.N.; Streletskiy, D.A. Land cover change in the lower Yenisei River using dense stacking of landsat imagery in Google Earth Engine. Remote Sens. 2018, 10, 1226. [Google Scholar] [CrossRef]
  32. Zeng, H.; Wu, B.; Wang, S.; Musakwa, W.; Tian, F.; Mashimbye, Z.E.; Poona, N.; Syndey, M. A synthesizing land-cover classification method based on Google Earth Engine: A case study in Nzhelele and Levhuvu Catchments, South Africa. Chin. Geogr. Sci. 2020, 30, 397–409. [Google Scholar] [CrossRef]
  33. Wang, Z.; Liu, K.; Peng, L.; Cao, J.; Sun, Y.; Qian, Y.; Shi, S. Analysis of mangrove annual changes in Guangdong Province during 1986–2018 based on Google Earth Engine. Trop. Geogr. 2020, 40, 881–892. [Google Scholar]
  34. Wang, T.; Wu, W.; Xue, X. Spatial-temporal changes of sandy desertified land during last 5 decades in Northern China. Acta Geogr. Sin. 2004, 59, 203–212. [Google Scholar]
  35. World Meteorological Organization. WMO Guidelines on the Calculation of Climate Normals; World Meteorological Organization: Geneva, Switzerland, 2017. [Google Scholar]
  36. Foody, G.M. Status of land cover classification accuracy assessment. Remote Sens. Environ. 2002, 80, 185–201. [Google Scholar] [CrossRef]
  37. Zhang, H.K.K.; Roy, D.P. Using the 500m MODIS land cover product to derive a consistent continental scale 30m Landsat land cover classification. Remote Sens. Environ. 2017, 197, 15–34. [Google Scholar] [CrossRef]
  38. Wang, X.; Song, J.; Xiao, Z.; Wang, J.; Hu, F. Desertification in the Mu Us Sandy Land in China: Response to climate change and human activity from 2000 to 2020. Geogr. Sustain. 2023, 3, 177–189. [Google Scholar] [CrossRef]
  39. Decuyper, M.; Chávez, R.O.; Čufar, K.; Estay, S.A.; Clevers, J.G.; Prislan, P.; Gričar, J.; Črepinšek, Z.; Merela, M.; de Luis, M.; et al. Spatio-temporal assessment of beech growth in relation to climate extremes in Slovenia-An integrated approach using remote sensing and tree-ring data. Agric. For. Meteorol. 2020, 287, 107925. [Google Scholar] [CrossRef]
  40. Shen, X.; Liu, B.; Henderson, M.; Wang, L.; Jiang, M.; Lu, X. Vegetation greening, extended growing seasons, and temperature feedbacks in warming temperate grasslands of China. J. Clim. 2022, 35, 5103–5117. [Google Scholar] [CrossRef]
Remotesensing 16 03100 g001
Figure 1. The random forest system variables for the desertification classification. Image bands L5, L7, and L8 represent Landsat-5 TM and Landsat-7 ETM bands 1 to 5 and 7 and Landsat-8 OLI bands 2 to 7. Abbreviations: BSI, bare soil index; MNDWI, modified normalized difference water index; MSAVI, modified soil-adjusted vegetation index; NDBI, normalized difference built-up index; NDVI, normalized difference vegetation index; TGSI, topsoil grain size index.
Figure 1. The random forest system variables for the desertification classification. Image bands L5, L7, and L8 represent Landsat-5 TM and Landsat-7 ETM bands 1 to 5 and 7 and Landsat-8 OLI bands 2 to 7. Abbreviations: BSI, bare soil index; MNDWI, modified normalized difference water index; MSAVI, modified soil-adjusted vegetation index; NDBI, normalized difference built-up index; NDVI, normalized difference vegetation index; TGSI, topsoil grain size index.
Remotesensing 16 03100 g001
Remotesensing 16 03100 g002
Figure 2. The distribution of the desert types in the 16 regions of interest in the random forest study of northern China.
Figure 2. The distribution of the desert types in the 16 regions of interest in the random forest study of northern China.
Remotesensing 16 03100 g002
Remotesensing 16 03100 g003
Figure 3. The spatial distribution of aeolian desertification in 1995.
Figure 3. The spatial distribution of aeolian desertification in 1995.
Remotesensing 16 03100 g003
Remotesensing 16 03100 g004
Figure 4. The spatial distribution of aeolian desertification in 2000.
Figure 4. The spatial distribution of aeolian desertification in 2000.
Remotesensing 16 03100 g004
Remotesensing 16 03100 g005
Figure 5. The spatial distribution of aeolian desertification in 2005.
Figure 5. The spatial distribution of aeolian desertification in 2005.
Remotesensing 16 03100 g005
Remotesensing 16 03100 g006
Figure 6. The spatial distribution of aeolian desertification in 2010.
Figure 6. The spatial distribution of aeolian desertification in 2010.
Remotesensing 16 03100 g006
Remotesensing 16 03100 g007
Figure 7. The spatial distribution of aeolian desertification in 2015.
Figure 7. The spatial distribution of aeolian desertification in 2015.
Remotesensing 16 03100 g007
Remotesensing 16 03100 g008
Figure 8. The spatial distribution of aeolian desertification in 2020.
Figure 8. The spatial distribution of aeolian desertification in 2020.
Remotesensing 16 03100 g008
Remotesensing 16 03100 g009
Figure 9. Spatial distribution of desertification in the Xinjiang region based on (a) visual interpretation and (b) predictions by the random forest model and (c) the results of visual classification interpretation are continuous and centralized in space and (d) the classification results of the random forest model are pixel-based, and its image spot is relatively small.
Figure 9. Spatial distribution of desertification in the Xinjiang region based on (a) visual interpretation and (b) predictions by the random forest model and (c) the results of visual classification interpretation are continuous and centralized in space and (d) the classification results of the random forest model are pixel-based, and its image spot is relatively small.
Remotesensing 16 03100 g009
Remotesensing 16 03100 g010
Figure 10. Spatial distribution of desertification severity in the central part of northern China in 2000 based on (a) visual interpretation and (b) predictions by the random forest model.
Figure 10. Spatial distribution of desertification severity in the central part of northern China in 2000 based on (a) visual interpretation and (b) predictions by the random forest model.
Remotesensing 16 03100 g010
Remotesensing 16 03100 g011
Figure 11. Spatial distribution of desertification in the northeastern part of northern China in 2000 based on (a) visual interpretation and (b) predictions by the random forest model.
Figure 11. Spatial distribution of desertification in the northeastern part of northern China in 2000 based on (a) visual interpretation and (b) predictions by the random forest model.
Remotesensing 16 03100 g011
Remotesensing 16 03100 g012
Figure 12. Changes in the area of desertification 1995 to 2020 in northern China.
Figure 12. Changes in the area of desertification 1995 to 2020 in northern China.
Remotesensing 16 03100 g012
Table 1. The calculation equations for the retrieval parameters. Abbreviations: B3 to B6, Landsat band numbers; BSI, bare soil index; MNDWI, modified normalized difference water index; MSAVI, modified soil-adjusted vegetation index; NDBI, normalized difference built-up index; NDVI, normalized difference vegetation index; TGSI, topsoil grain size index.
Table 1. The calculation equations for the retrieval parameters. Abbreviations: B3 to B6, Landsat band numbers; BSI, bare soil index; MNDWI, modified normalized difference water index; MSAVI, modified soil-adjusted vegetation index; NDBI, normalized difference built-up index; NDVI, normalized difference vegetation index; TGSI, topsoil grain size index.
NameFormula
NDVILandsat 8: NDVI = (B5 − B4)/(B5 + B4)
Landsat 5, 7: NDVI = (B4 − B3)/(B4 + B3)
AlbedoLandsat 8: Albedo = 0.356 B2 + 0.130 B4 + 0.373 B5 + 0.085 B6 + 0.072 B7 – 0.0018
MSAVILandsat 8: MSAVI = {2 B5 + 1 − sqrt [(2 B5 + 1)2 − 8 (B5 − B4)]}/2
Landsat 5, 7: MSAVI = {2 B4 + 1 − sqrt [(2 B4 + 1)2 − 8 (B4 − B3)]}/2
TGSILandsat 8: TGSI = (B4 − B2)/(B4 + B3 + B2)
Landsat 5, 7: TGSI = (B3 − B1)/(B3 + B2 + B1)
MNDWILandsat 8: MNDWI = (B3 − B6)/(B3 + B6)
Landsat 5, 7: MNDWI = (B2 − B5)/(B2 + B5)
BSILandsat 8: BSI = 100 [(B6 + B4) − (B5 + B2)]/[(B6 + B4) + (B5 + B2)] +100
Landsat 5, 7: BSI = 100 [(B5 + B3) − (B4 + B1)]/[(B5 + B3) + (B4 + B1)] +100
NDBILandsat 8: NDBI = (B6 − B5)/(B6 + B5)
Landsat 5, 7: NDBI = (B5 − B4)/(B5 + B4)
Table 2. The overall accuracy and the kappa coefficient.
Table 2. The overall accuracy and the kappa coefficient.
RegionAccuracy (%)KappaRegionAccuracy (%)Kappa
Mu Us Sandy Land74.780.64Qaidam Basin97.330.71
Kubuqi Desert87.350.63Source region of the Yellow River94.110.75
Shiyanghe Basin93.390.75Kumutage Desert96.590.69
Heihe River Basin98.110.83Middle of the Otindag Sandy Land87.990.69
Hulunbuir Desert92.160.80Other parts of the Otindag Sandy Land89.780.59
Junggar Basinand Tianshan Mountains90.870.67Southern part of the Horqin Sandy Land85.570.72
Turpan–Hami Basin95.400.73Northern part of the Horqin Sandy Land92.410.54
Southern Xinjiang 94.040.70Songnen Sandy Land95.210.50
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Zhang, C.; Tan, N.; Li, J. Aeolian Desertification Dynamics from 1995 to 2020 in Northern China: Classification Using a Random Forest Machine Learning Algorithm Based on Google Earth Engine. Remote Sens. 2024, 16, 3100. https://doi.org/10.3390/rs16163100

AMA Style

Zhang C, Tan N, Li J. Aeolian Desertification Dynamics from 1995 to 2020 in Northern China: Classification Using a Random Forest Machine Learning Algorithm Based on Google Earth Engine. Remote Sensing. 2024; 16(16):3100. https://doi.org/10.3390/rs16163100

Chicago/Turabian Style

Zhang, Caixia, Ningjing Tan, and Jinchang Li. 2024. "Aeolian Desertification Dynamics from 1995 to 2020 in Northern China: Classification Using a Random Forest Machine Learning Algorithm Based on Google Earth Engine" Remote Sensing 16, no. 16: 3100. https://doi.org/10.3390/rs16163100

APA Style

Zhang, C., Tan, N., & Li, J. (2024). Aeolian Desertification Dynamics from 1995 to 2020 in Northern China: Classification Using a Random Forest Machine Learning Algorithm Based on Google Earth Engine. Remote Sensing, 16(16), 3100. https://doi.org/10.3390/rs16163100

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop