The dynamic monitoring of aeolian desertification land distribution and its response to climate change in northern China

Aeolian desertification is poorly understood despite its importance for indicating environment change. Here we exploit Gaofen-1(GF-1) and Moderate Resolution Imaging Spectroradiometer (MODIS) data to develop a quick and efficient method for large scale aeolian desertification dynamic monitoring in northern China. This method, which is based on Normalized Difference Desertification Index (NDDI) calculated by band1 & band2 of MODIS reflectance data (MODIS09A1). Then we analyze spatial-temporal change of aeolian desertification area and detect its possible influencing factors, such as precipitation, temperature, wind speed and population by Convergent Cross Mapping (CCM) model. It suggests that aeolian desertification area with population indicates feedback (bi-directional causality) between the two variables (P < 0.05), but forcing of aeolian desertification area by population is weak. Meanwhile, we find aeolian desertification area is significantly affected by temperature, as expected. However, there is no obvious forcing for the aeolian desertification area and precipitation. Aeolian desertification area with wind speed indicates feedback (bi-directional causality) between the two variables with significant signal (P < 0.01). We infer that aeolian desertification is greatly affected by natural factors compared with anthropogenic factors. For the desertification in China, we are greatly convinced that desertification prevention is better than control.

Desertification is a type of land degradation in which a relatively dry land region becomes increasingly arid, typically losing its bodies of water as well as vegetation and wildlife. It is caused by a variety of factors, such as climate change and human activities 1 . Meanwhile aeolian desertification is the most important desertification type in China with serious environmental and socioeconomic problems in arid, semi-arid, and dry sub-humid zones. Continuous aeolian desertification has a serious influence on the biosphere. It is also highly related to issues such as declining productivity, biodiversity loss, land degradation, and declining ecosystem services [2][3][4][5][6] . Many studies showed that desertification was resulting from various processes and reasons including natural and anthropogenic factors [7][8][9][10][11] . Here we used Convergent Cross Mapping (CCM) model to explore the causality of aeolian desertification. The result shows that natural factors are the primary reason for aeolian desertification in northern China during the past 15 years.
In China, desertification area survey was conducted once every five years since 1994. Desertification lands occupy an area about 2.61 million km 2 and spread across 18 provinces accounting for 27.20% of the country's land area by 2014 12 . This survey takes a lot of manpower, material and financial resources with a lack of sequential dynamic monitoring. In this study, a remote sensing method was used to obtain the aeolian desertification land distribution in northern China. A new spectral index called Normalized Difference Desertification Index (NDDI) derived from MODIS surface reflectance data was used to acquire the aeolian desertification land distribution. Results will provide a basis for combating desertification. In northern China, natural vegetation is being transformed into agricultural lands at a faster rate, endangering ecosystem services and increasing soil-loss potential, which may trigger land degradation. This region is sensitive to climate change and human intervention. It becomes an original region of sandstorms. To alleviate the multifaceted environmental degradation, Chinese government has implemented several ecological restoration programs that have deeply affected the structure and function of grassland ecosystems. Understanding desertification processes and causes are important to provide reasonable and effective control measures for preventing desertification. The study area of this paper is located in northern China (31°09′ N-53°23′ N, 73°40′ E-126°04′ E) (Fig. 1). It includes Xinjiang Uyghur Autonomous Region, Qinghai Province, Gansu Province, Ningxia Hui Autonomous Region and Inner Mongolia Autonomous Region with eight famous deserts in China. Generally, a semi-arid or desert climate prevails in Xinjiang. The entire region is marked by great seasonal differences in temperature. This region includes Gurban Tunggut Desert, Taklamakan Desert and Kumtag Desert. Qinghai has quite cold winters, mild summers, and a large diurnal temperature variation. Significant rainfall occurs mainly in summer, while precipitation is very low in winter and spring, and is generally low enough to keep much of the province semi-arid or arid. This region includes Qaidam Basin Desert. Gansu generally has a semi-arid to arid continental climate with warm to hot summers and cold to very cold winters. Most of the limited precipitation is delivered in the summer months. This region includes Badain Jaran Desert, Tengger Desert and Kumtag Desert. Ningxia Hui Autonomous Region has a continental climate with average summer temperatures rising to 17 to 24 °C in July and average winter temperatures dropping to between − 7 to − 15 °C in January. Annual rainfall averages from 190 to 700 millimetres, with more rain falling in the south of the region. This region includes Tengger Desert in Shapotou. Inner Mongolia has a wide variety of regional climates. The winters in Inner Mongolia are very long, cold and dry. The spring is short, mild and arid, with large, dangerous sandstorms, whilst the summer is very warm to hot and relatively humid except in the west where it remains dry. Autumn is brief and sees a steady cooling, with temperatures below 0 °C reached in October in the north and November in the south. It includes Badain Jaran Desert, Tengger Desert, Kubuqi Desert and Ulan Buh Desert.

Results
Automatic monitoring of aeolian desertification land. Taking a part of GF-1 data as the experimental data ( Fig. 2), three fusion algorithms as commonly used in the image fusion experiments were applied. Mean and standard deviation values of fusion image using multiplicative algorithm are far from raw image (Fig. 3a,b). Entropy and correlation coefficient values of fusion image using PCA algorithm are greater than Brovey Transform (Fig. 3c,d). It shows that Principal Component Analysis (PCA) image fusion algorithm is the best choice for GF-1 data fusion (Table 1). Then PCA image fusion algorithm was finally used on entire GF-1 data.
GF-1 remote sensing image was classified by Support Vector Machine (SVM) classification algorithm. The main land cover type obtained by this algorithm is shown in Fig. 4. Results show that the SVM classification algorithm can meet the precision requirements based on visual interpretation of GF-1.
It is a problem to achieve the combination of high and low resolution remote sensing data. In this study, the first step was to establish a 1 km × 1 km grid frame (consistent with the resolution of re-sampled MODIS) with vector format and each grid was identified with a unique identity (Fig. 4). The second step was to use this frame to respectively perform statistical analysis for SVM classification result by GF-1 data in each grid. Finally, calculate the proportion of each land use type in each grid frame. Taking the proportion of aeolian desertification area was greater than 70% as the pure pixel for aeolian desertification land, the proportion of vegetation (others) area was greater than 90% as the pure pixel for vegetation (others). Then changes of reflectance values for different land use types were acquired. Total correlation index(r) value of all types between band1 and band2 was the lowest. Band1 and band2 exhibited a large disparity in their spectral responses of different land covers. So these two bands were used to derive Normalized Difference Desertification Index (NDDI) in this study.
Aeolian desertification area was extracted through above-mentioned method by using GF-1data. Mean value of MODIS-NDDI time series curves for different land use types are shown in Fig. 5. Filtered MODIS-NDDI time series curve of aeolian desertification land is shown in Fig. 6.
Mean Absolute Distance (MAD) 13 was used to compare the MODIS-NDDI time series image of aeolian desertification land with the MODIS-NDDI image of the study area for each pixel. A lower image value illustrated a closer MAD, which indicated a greater possibility of aeolian desertification. A threshold was set on the MAD map based on the prior knowledge, by considering the official data of aeolian desertification area. The difference between the estimated aeolian desertification area and the official data was smallest when the threshold value is 0.051. In this study, a p-tile algorithm was adopted for threshold selection 14 . Aeolian desertification distribution of northern China in 2001 to 2015 is shown in Fig. 7. The eolian desertification land area estimated from MODIS and investigation data are shown Table 2.
Factor analysis. Generally, desertification means the ratio of annual precipitation to potential evapotranspiration falls within the range from 0.05 to 0.65; and evapotranspiration is highly related to temperature. Wind is the power of desertification. Meanwhile, population is one of the most important anthropogenic factors of desertification. So temperature, precipitation and wind speed as the natural factors and population as the anthropogenic factor in combination of CCM model were used to analyze cause-and-effect relationship in this study (Fig. 8).
Based solely on the relationship between library length and Pearson correlation coefficient, results for this CCM test suggest that aeolian desertification area with population indicates feedback (bi-directional causality) between the two variables (P < 0.05; Fig. 9a), but forcing of aeolian desertification area by population is weak. Based on the same diagnostic tests as we used above, we find aeolian desertification area is significantly affected by temperature, as expected (Fig. 9b). However, there is no obvious forcing for the aeolian desertification area and precipitation (Fig. 9c). Aeolian desertification area with wind speed indicates feedback (bi-directional causality) between the two variables with significant signal (P < 0.01; Fig. 9d).

Discussion
Land degradation and desertification has been ranked as a major environmental and social issue in the coming decades in China. It has received a great attention, especially the northern China. Desertification area in China has increased since the 1950s and reached its maximum during the 1970s and early 1980s. Since then desertification area has decreased continuously to the present [15][16][17] . It suggests that reforestation/afforestation policy has played a significant role in controlling the desertification in China. Small field experiments prove that vegetation in desertified/degraded land could recover if isolated from human activities. Since 1998, natural recovery has become one powerful national force to prevent land desertification and recover natural vegetation 17 , so anthropogenic factors are not the main factor exacerbating the desertification distribution during 2001 to 2015 in northern China. Numerous scientists have claimed that land desertification in China is primarily due to human impacts. Wang et al. 15 suggested that desertification in China has been primarily caused by climate change 15 . We also believed that desertification in northern China is mainly controlled by natural factors during the past 15 years. In addition, although overall land desertification area decreases year by year, desertification situation in some regions shows the worsened tendency, such as regions around rivers and lakes. We must pay more attention to environment deterioration of rivers, lakes and the nearby areas in the future. In this study, the proposed NDDI is able to obtain the aeolian desertification land distribution on a large scale and assess the aeolian desertification area variability simply and effectively. The results show that aeolian desertification area with population indicates  feedback (bi-directional causality) between the two variables (P < 0.05). Forcing of aeolian desertification area by population is weak (P = 0.046). However, aeolian desertification area forcing population has more significant signal (P = 0.02). In fact, land desertification in northern China is affected by anthropogenic factors with small fluctuation. Once the land desertification happens, it will bring population migration with population growth pressure around desertification regions. Based on the same diagnostic tests as we used above, we find aeolian desertification area is significantly affected by temperature, as expected. However, there is no obvious forcing for the aeolian desertification area and precipitation. We supposed several situations about this: (i) there is a small fluctuation for the precipitation in northern China during the past 15 years, so we derived no obvious causality; (ii) precipitation is not the root cause of land desertification. Aeolian desertification area with wind speed indicates feedback (bi-directional causality) between the two variables with significant signal (P < 0.01). In conclusion, we infer that aeolian desertification is greatly affected by natural factors compared with anthropogenic factors. From the result, aeolian desertification area covers a large area in northern China threatening human life. If the desertification land area continues to grow, it will reduce the habitable zone and lead to a large number of population migration and growth 18 . It is also a big dust origin in northern China. For the desertification in China, we are greatly convinced that desertification prevention is better than control. We should pay more attention to the impact of climate change on the desertification distribution.   limited impact factors including precipitation, temperature, wind speed and population were used to analyze the cause-and-effect relationship. More field survey data in combination of more high spatial resolution images should also be adopted in the further research.   county from 2001 to 2015 were used in this study (Fig. 10). And 33 fieldwork investigation points as interpretation key were used to perform SVM classification for GF-1 data.
SVM classification. SVM is a range of classification and regression algorithm that has been formulated from the principles of statistical learning theory developed by Vapnik 19 . This type of classification method for remote sensing images has many advantages. The most direct advantages are that the internal structure is uniform, the boundaries between different categories are more obvious, and the classification accuracy is improved [20][21][22] . In this study, the kernel type is polynomial and the degree of kernel polynomial is two for the SVM algorithm.
Normalized Difference Desertification Index (NDDI) calculation. In order to effectively monitor aeolian desertification distribution, band1 and band2 of MODIS were selected to calculate NDDI using the following equation 23 . Where i is the line number of the pixel, j is the column number of the pixel, d ij is the MAD, m k is the number of pixels in the k serie, x ik and x jk are the image values in the k serie, and n is the total number of MODIS-NDDI time series (n = 46).

Convergent cross mapping (CCM) model. Convergent cross mapping (CCM) model is a statistical test
for a cause-and-effect relationship between two time series variables that, like the Granger causality test, seeks to resolve the problem that correlation does not imply causation. While Granger causality is best suited for purely stochastic systems where the influences of the causal variables are separable (independent of each other), CCM is based on the theory of dynamical systems and can be applied to systems where causal variables have synergistic effects. The test was developed in 2012 by the lab of George Sugihara of the Scripps Institution of Oceanography, La Jolla, California, USA 24 .