Spatial-temporal variability of snow cover over the Amur River Basin inferred from MODIS daily snow products in recent decades

XiaoLin Lu,WanChang Zhang*,ShuHang Wang,Bo Zhang,QuanFu Niu,JinPing Liu,4,Hao Chen,4,HuiRan Gao,4

1.Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China

2. National Engineering Laboratory for Lake Pollution Control and Ecological Restoration, Chinese Research Academy of Environmental Science,Beijing 100012,China

3.Lanzhou University of Technology,Lanzhou,Gansu 730050,China

4.University of Chinese Academy of Sciences,Beijing 100049,China

ABSTRACT MODIS snow products MOD10A1MYD10A1 provided us a unique chance to investigate snow cover as well as its spa‐tial-temporal variability in response to global changes from regional and global perspectives. By means of MODIS snow products MOD10A1MYD10A1 derived from an extensive area of the Amur River Basin,mainly located in the Northeast part of China, some part in far east area of the former USSR and a minor part in Republic of Mongolia, the reproduced snow datasets after removal of cloud effects covering the whole watershed of the Amur River Basin were generated by us‐ing 6 different cloud-effect-removing algorithms. The accuracy of the reproduced snow products was evaluated with the time series of snow depth data observed from 2002 to 2010 within the Chinese part of the basin,and the results suggested that the accuracies for the reproduced monthly mean snow depth datasets derived from 6 different cloud-effect-removing algorithms varied from 82% to 96%, the snow classification accuracies (the harmonic mean of Recall and Precision) was higher than 80%,close to the accuracy of the original snow product under clear sky conditions when snow cover was sta‐bly accumulated. By using the reproduced snow product dataset with the best validated cloud-effect-removing algorithm newly proposed, spatial-temporal variability of snow coverage fraction (SCF), the date when snow cover started to accu‐mulate (SCS) as well as the date when being melted off (SCM) in the Amur River Basin from 2002 to 2016 were investi‐gated.The results indicated that the SCF characterized the significant spatial heterogeneity tended to be higher towards East and North but lower toward West and South over the Amur River Basin. The inter-annual variations of SCF showed an insignificant increase in general with slight fluctuations in majority part of the basin. Both SCS and SCM tended to be slightly linear varied and the inter-annual differences were obvious. In addition, a clear decreasing trend in snow cover is observed in the region. Trend analysis (at 10% significance level) showed that 71% of areas between 2,000 and 2,380 m a.s.l.experienced a reduction in duration and coverage of annual snow cover.Moreover,a severe snow cover reduction during recent years with sharp fluctuations was investigated. Overall spatial-temporal variability of Both SCS and SCM tended to coincide with that of SCF over the basin in general.

Keywords:MODIS;SCF;SCS;SCM;Amur River Basin;cloud effect removal

1 Introduction

Snow is a major source of streamflow in many re‐gions of the world, and poses important impacts espe‐cially on hydrological regimes and water availability over high altitude regions (Qiuet al., 2017; Tanget al., 2017). Hence, understanding the spatial distribu‐tion and temporal variation of snow is crucial for proper management of water resources in these re‐gions (Sunet al., 2019). As an important element of the cryosphere that plays essential role in the Earth-at‐mosphere interaction, Eco-hydrological regime and ecosystem function of the Earth system (Liuet al.,2018; Sunet al., 2019), pioneer studies to quantita‐tively investigate snow cover variability mainly con‐centrated on site monitoring with snow poles to ob‐serve snow cover data in early 19th century (Wanget al.,2017).With the quick advancement of earth obser‐vations from space in recent decades, satellite remote sensing in association with conventional ground-based observations have become the practical approach in snow cover observations. At present, passive micro‐wave remote sensing, benefited from its strong pene‐trating capability in all weather conditions, such as SMMR (Fosteret al., 1996), SSM/I (Wanget al.,2001),AMSE-R (Chang and Rango, 2000) and MWRI(Jianget al.,2014),therefore superior to optical remote sensing,such as Landsat series AVHRR and MODIS in some aspects (Wanget al., 2017; Wanget al., 2018),has been widely utilized to directly investigate snow in‐formation including snow water equivalent and snow depth,etc.. However, relative lower spatial resolution of snow cover information derived with microwave re‐mote sensing compared with those derived by optical remote sensing, such as Landsat TM series (Rosenthal and Dozier, 1996), AVHRR (Zhao and Fernandes,2009) and MODIS (Hallet al., 2002), yet remained as a shortcoming in its applications(Qiuet al.,2017).Till now MODIS snow products MOD10A1/MYD10A1 and 8-day standard snow cover product MOD10A2/MYD10A2 are still the reliable and most frequently used data source for relevant studies world widely,for which their capabilities in providing a spatially distributed and accurate mapping of snow cover have been well demonstrated (Parajka and Bl?schl,2006; Hall and Riggs, 2007; Wanget al., 2008; Xieet al.,2009;Houet al.,2018).

However, cloud obscuration due to cloud cover‐age in the visible and near infrared wavelengths is the major limitation of optical sensors for snow mapping,especially in mountainous areas where the impact of cloud cover is the biggest obstacle to the use of MO‐DIS snow production in most cases.This problem has been addressed by many studies (Tekeliet al., 2005;Andreadis and Lettenmaier, 2006; Mcguireet al.,2006; Parajka and Bl?schl, 2006; Wanget al., 2009).It was reported that the average annual cloud cover‐age was about 50% in northern Xinjiang, China, and the probability of continuous cloud covered day was as higher as 95% (Xieet al., 2009). And 63% of the areas in average over Austria were found covered by cloud, and the probability of cloud covered day was even greater in winter (Parajka and Bl?schl, 2006).Therefore, different methods have been developed to address this issue, nevertheless, space yet remained for further improvement (Parajka and Bl?schl, 2006;Gafurov and Bárdossy, 2009; Wang and Xie, 2009;Xieet al., 2009; Paudel and Andersen, 2011; Huanget al.,2012;Dietzet al.,2013;Liuet al.,2017;Qiuet al.,2017;Houet al.,2018).

Different indexes are used to investigate the snow cover spatiotemporal changes in different studies(Tanget al., 2013; Tahiret al., 2014; Gascoinet al.,2015; Marchaneet al., 2015). In this study, we uti‐lized the most popular indexes including snow cover‐age fraction (SCF), the date when snow cover started to accumulate (SCS) as well as the date when being melted off (SCM). In addition, we developed some other suitable indexes for further investigation of spa‐tial and temporal variability of snow cover areas in the Amur River Basin from 2002 to 2016 with the best val‐idated cloud-effect-removing algorithm newly pro‐posed. The aim of this paper was to investigate spatio‐temporal snow cover changes using an improved cloudfree approach and different snow cover indexes.By us‐ing the cloud-removed snow cover product; the trend of snow cover spatiotemporal changes was analyzed statistically via different snow cover indicators. This paper was structured as follows: in Section 2 the study area and the datasets are described. Section 3 displays the methodology used in this paper. Results and discussions were presented in Section 4 and final‐ly,in Section 5 the study findings are concluded.

2 Study area and datasets

2.1 Study area

The Amur River Basin, one of the largest water‐sheds in Northeast Asia, is a transnational watershed situated on China, Russia and Mongolia extending to about 1.68×104km in length and covering an area of approximately 2.09×106km2(Huanget al., 2017). As shown in Figure 1,the topography of the basin charac‐terizes the high mountainous terrain over the west and gentle hills and lower plains in east with an extensive rangeland distributed in between with elevations rang‐ing from 0 m to 2,585 m a.s.l..Owning to its geo-loca‐tion on the eastern edge of the Eurasian continent ad‐jacent to the Pacific Ocean,the Amur River Basin fea‐tured complicated climatic conditions that are affect‐ed jointly by monsoon climate, ocean currents and to‐pography. Over the basin, the largest wetlands, exten‐sive permafrost and perennial snow covers extensively developed in this middle-high latitude region,which fa‐vor the land of the region rich of fertilizer (Chenet al.,2018; Gaoet al., 2020). Spatial-temporal variations of snow cover under the acerated global warming attract great concerns on the local ecological environment,hy‐drological regimes and sustainable development of the region(Simonov and Dahmer,2008;Yuet al.,2014).

Figure 1 Geo-location map showing the overview of Amur River Basin,including topography,major rivers,lakes,geographic locations of 78 meteorological stations and MODIS tile grid information in the Amur River Basin

2.2 Data

The data used in present study are mainly down‐loaded from MODIS daily snow products (NCDC,http://nsidc.org), DEM (http://srtm.csi.cgiar.org), and in situ snow depth observation data for verification provided by hydro-meteorological stations within the basin.

2.2.1 MODIS Daily Snow Products

The daily snow cover products MOD10A1/MYD10A1 from the National Snow and Ice Data Cen‐ter (NCDC, http://nsidc.org), currently the most wide‐ly used long-term sequence of snow cover data, was the main data source for the spatial-temporal variabili‐ty study of snow cover in the Amur River Basin.MOD10A1 and MYD10A1 data were derived from the Terra and Aqua satellites respectively. The Terra was launched in December 1999 and the data was available in February 2000. The Aqua began to pro‐vide data in July 2002, and the double-star transit time difference was 3 h,which means that two images are available in most parts of the globe on the same day.Some observations were missing during study pe‐riod, but MOD10A1 and MYD10A1 complement one another because the missing data happened on differ‐ent dates and from different sensors. The global im‐age had been divided into 36×18 tiles,the Amur River Basin was covered by 6 tiles (h24v03, h25v03,h25v04, h26v03, h26v04, h27v04). The daily snow cover products MOD10A1/MYD10A1 from Septem‐ber 1, 2002 to August 31, 2016 were utilized for anal‐yses.The MODIS V5 Snow Cover Daily Global 500 m product contains snow cover map, snow cover frac‐tion, snow albedo and QA in the data product file(Hall and Riggs, 2001), and in this study, only snow cover map was used.

The snow cover map was generated by snow map‐ping algorithm that was especially developed for land cover classification of total 11 categories with MODIS imaginaries (Hallet al., 2002). In order to facilitate the implementation of the cloud-effects removing algo‐rithm, the land cover classification was recoded, and the recoding rules were given as shown in Table 1.

Table 1 The code and significance of MODIS snow cover products as well as rules of recoding

2.2.2 DEM

The DEM data with spatial resolution of 90 m was collected from the SRTM digital elevation data re‐leased by the USGS EROS data center, and was downloaded free of charge on the CGIAR-CSI web‐site(http://srtm.csi.cgiar.org).

2.2.3 Snow depth observation data

In order to verify the accuracy of the cloud-free snow cover data sets, the in-situ snow depth observa‐tion dataset obtained from 2002 to 2010 in 83 meteo‐rological stations over Northeast China were collect‐ed, among which only 78 snow depth observation da‐ta were used for verification because of observing date were not consistent or in poor quality for some stations. Due to the spatial differences between site data and image data, the snow depth data was divided into three categories referring to Wanget al.(2017):(1) SD=0 cm, which implied that the corresponding pixels of MODIS image were non-snow covered; (2)1 cm≤SD ≤3 cm,which assumed that the correspond‐ing pixels of MODIS image were partly snow cov‐ered; (3) SD ≥4 cm, which indicated the correspond‐ing pixels of MODIS image were completely snow covered.

3 Methodology

3.1 Cloud Effects Removing Algorithm

On the basis of the continuous six-step cloud ef‐fects removing algorithm proposed by Gafurov and Bárdossy (2009), in order to combine relevant spatiotemporal information of snow cover for removing the influence of the cloud much effectively, the improved algorithm as illustrated in Figure 2 was proposed in present study. As the flow chat presented, the input of each step is just the output of the previous one to ensure enough accuracy of the classified snow pixels while re‐moving the effects of cloud effectively. Andreaset al.(2013)and Liuet al.(2017)used similar approach to re‐move cloud effects, and the satisfactory results were obtained (the overall accuracy was above 90%). For the detailed explanation of the algorithm, please refer to www.hydrol-earth-syst-sci.net/13/1361/2009/.

Figure 2 Flow chart of could-effects removing algorithm cited and improved from Andreas et al.(2013)and Liu et al.(2017)

The first step took advantage of the different im‐age acquisition time of Terra and Aqua to obtain the maximum snow coverage boundary (Gafurov and Bárdossy, 2009;Wang and Xie, 2009; Paudel and An‐dersen, 2011; Dietzet al., 2013; Liuet al., 2017) with the accuracy claimed to be about 92% (Dietzet al.,2013).The second step aimed to combine the datasets derived from three consecutive days to determine if a cloud pixel exists, pixel values in the previous day and in the next day were assigned to take place of the cloud information.This procedure has been utilized to analyze the snow cover of the Tianshan Mountains on the borders of China, Kazakhstan and Kyrgyzstan,and overall agreement with in situ measurements was reported with accuracy about 90% (Xieet al., 2009).The third step combined the four direct "side-border‐ing" neighboring pixels of the cloud pixels by judging as if there were three snow pixels or three land pixels of its "side-bordering" pixels, the cloud pixels were redefined as snow or land, the overall accuracy of this approach was found about 92.38% (Gafurov and Bárdossy, 2009). The fourth step was based on the classified elevation features of snow-covered pixels with common knowledge that snow accumulated earli‐er at high altitude and melted out later than at low alti‐tude (Paudel and Andersen, 2011). With this step, the threshold altitude,i.e., the lowest and the highest snow line where the snow cover distributed,can be detected(Gafurov and Bárdossy, 2009). The cloud pixels be‐low the lowest snow line were regarded as land, and above the highest snow line were classified as snow,the accuracy of this approach was acclaimed to ap‐proximately 93.15%(Gafurov and Bárdossy,2009).

However,due to extensive area featured with com‐plex terrain,high landscape heterogeneity and the lim‐ited altitude range mainly concentrated within 1,000 m altitude range of the Amur River Basin, the regional applicability of the cloud-effects removing algorithm presented in Figure 2 should be improved, especially for step 4.The Amur River Basin was thus divided in‐to 321 sub-watersheds as shown in Figure 3 with Arc-Hydro module from Arc-info software according to the river network extraction by setting sub-watershed threshold at least 25% cloudless covered (Paudel and Andersen, 2011). Otherwise, the original classifica‐tion was retained and the next step was directly initiat‐ed. For decreasing probability of cloud pixels being misclassified as snow, the Equations(1)and(2)as listed in the below were adopted according to Paudel and Andersen(2011):

wherexandyare the number of row and column, re‐spectively;tis the date for the pixelsS,H(x,y)is the ele‐vation of cloud pixels;Hmean(t) is the average eleva‐tion of all snow pixels;Hmin(t) is the average eleva‐tion of all land pixels;Czrepresents the percentage of cloud pixels in the study area.

Figure 3 The sub-watersheds delineated with ArcHydro module for the Amur River Basin

The fifth step was to design 3×3 spatially moving window centered on each of cloud pixels.If all neigh‐boring pixels were viewed as cloud free, the center pixel needs to be re-examined. The accuracy of this step was claimed to 94% according to Gafurov and Bárdossy(2009).

The step 6 was based on systematic analyses on a time series of pixels in a hydrological year. In this study, the period from September 1 to August 31 of next year was defined as the hydrological year, which means that the classification type of each pixel in the entire hydrological year began mainly with the types of either land or clouds until snow cover type ap‐peared for some period and gradually disappeared un‐til land surface reappeared again with the melted away of snow covers. Using this objective informa‐tion,the averaged snow cover appearance date as well as the disappearing date of the basin were investigat‐ed as thresholds to reclassify the cloud pixels (Gafu‐rov and Bárdossy,2009).

From the beginning of snow accumulation to the melted away of the snow cover, the contingency of snowfall and snowmelt events often made the correct determination of thresholds very difficult. Therefore,the progressively increased two direction method was adopted to extract the first date snow started to accu‐mulate as well as the last date the snow cover was melted away. A time window combined with three consecutive days was used to detect positively from the pixel of snow cover on the 1st of September of the year until the pixel of snow cover was encountered for three consecutive days. Similarly, from the pixels where snow cover was melted away on August 31 of the following year, the similar procedure as described previously was conducted versa a visa until the snow cover was encountered again within three consecutive days. Using the time window detection method to ex‐tract the threshold for snow cover duration can effec‐tively reduce the influence of sporadic snow pixels.

3.2 Criterion for Evaluating Snow Cover Classifications

The best way to validate the performance of the improved cloud effect removal algorithm is to use in situ observations as ground truth for comparison.However,for such extensive area like Amur River Ba‐sin (Refer to Figure 1), continuous snow cover obser‐vation records available only from 2002 to 2010 for the Chinese part can be obtained. Regardless scale ef‐fect and snow partly covered pixels,the pixel value of the image was extracted based on the location of hy‐dro-metrological station, then the pixel value was compared with the site observed data for evaluating the accuracy of the improved cloud effect removal algorithm.

Four binary indexes were used to identify a MO‐DIS pixel covered snow or not: true positive (a), false negative (b), false positive (c) and true negative (d).The accuracy of the algorithm was evaluated by the formula as described below (Rittgeret al., 2013; Qiuet al.,2017):

whereAccuracyindicates the probability that a pixel is correctly classified, but may be misleading when there was only few snow pixels in the image;Recallpresents the probability of accurate detection of a snow-covered pixel;Precisionis the ratio of the num‐ber of snow-covered pixels correctly identified to the number of snow-covered pixels in the image; TheFwas defined as the harmonic mean ofRecallandPre‐cision,which can be used to better reflect the misjudg‐ment of snow classification(Rittgeret al.,2013).

3.3 Snow cover indexes

A set of snow cover indexes were used to investi‐gate snow cover changes, such as the snow cover fre‐quency (SCF) was defined by Equation(7)(Chu,2016; Liuet al., 2017; Yanget al., 2017; Labaduoma and Cizhen,2018):

where,CsandCtare the number of the snow pixel and the total number of images for cloud-free snow cover dataset in one hydrological year,respectively.

The date when snow cover started to accumulate(SCS) and the date when snow cover was melted away (SCM) in a hydrological year are also vital in‐dexes for climate and Phenology studies (Liuet al.,2018). These two indexes were mapped and tracked over time to refer their sustainable changes as indica‐tors for climatic and downstream phenological investi‐gations, which can be calculated according to Equa‐tions(8)and(9)(Dietzet al.,2013),respectively:

whereFdis the fixed date, representing the date of the maximum snow cover between 2002 and 2014, it was fixed on March 5th in this study,i.e., the 64th (65th)day of the hydrological year; theSCDbFdandSCDaFdrepresented number of days snow accumulated before and after the fixed date,respectively.It was worth not‐ing that, it doesn't directly reflect the actual date of the SCS as well as the SCM in this way, but it is the best way to show the trend of them(Liuet al.,2018).

4 Results and discussions

4.1 Performance assessment of cloud effect removing algorithm

Due to the limited number of in-situ snow depth observed data, the image data used in each step be‐tween 2002 and 2010 was selected for cloud effect-re‐moving efficiency and accuracy analysis. Monthly mean percentage of cloud cover pixels after imple‐mentation were shown in Figure 4. It was observed that the cloud cover ratio of the original MOD10A1 and MYD10A1 images accounted for about 53% and 60% between 2002 and 2010, respectively. Moreover,the highest cloud cover ratio period for a specific year was usually during November to March in the hydro‐logical year. In the entire six-steps of cloud effect re‐moving algorithm described in Section 3.1,the contri‐bution of each step to the cloud effect removal was es‐timated different from 2002 to 2016 about 11%, 9%,1%, 7%, 2%, 26% in average, respectively. The im‐proved snow-line algorithm didn't work well in cloud effect removal in the Amur River Basin,which was at‐tributed mainly to that the sub-watersheds snow line method is suitable for cloud effect removal to those cloud pixels at very high or very low elevations.It was difficult to get better cloud removal effects meanwhile to keep high accuracy of snow classification because of effects of topographical features and higher coverage of cloud pixels over the Amur River Basin. The most efficient step in cloud effects removal of the entire al‐gorithm was seasonal filtering that considerably de‐pends on the accurate extraction of the first date when snow started to accumulate and the last date when the snow cover was melted away, in which consecutive three-day time window played an essential role.

Figure 4 Monthly mean percentage of cloud covers of MOD10A1,MYD10A1 and the monthly mean Accuracy of the images after implementation of each step of the cloud effect removal algorithm from 2002 to 2010

The monthly mean Accuracy of image after imple‐mentation of each steps of the algorithm was ranged from 83% to 98%, and a slight accuracy loss was of‐ten found between each two adjacent date, except for March (Figure 4). However, false positive result caused by large number of none-snow pixels might be existed, special attention should be paid to the snow classification statistic result such asF, Precision and Recall. TheFvalue is a harmonic index of Precision and Recall, which can be used to well reflect the mis‐judgment of snow-covered pixels.As shown in Figure 5, monthly mean accuracy varied seasonally, and re‐duced to 0 between May and September as none-snow pixels existed during this period, which was attributed to the data scale effect and the inherent shortcomings of snow detection by satellites. For example, in the early stage of snow accumulation, the snow-covered duration was relatively short, the meteorological sta‐tions may record the depth of snow accumulation but remotely sensed images probably failed to record it.Even though the snow depth less than 4 cm was ex‐cluded when performing true-value sample statistics,this situation yet existed. Moreover, the snow depth recorded by meteorological stations were the point ob‐served data, well the estimated remotely sensed im‐age data was area one, which were also the resource to introduce errors to decrease the performance of the algorithm.

The monthly precision of each snow data was slightly varied except for October and April in a hy‐drological year. And the monthly Recall of each snow data was rather stable except for the last one,and Recall statistic largely drives theF(Figure 5),which implied false negatives were more severe than false positives. As for the result of the cloud effect removal algorithm, it was investigated that the de‐crease of accuracy was mainly found in the sixth step of the algorithm, which accounted for 26% of whole cloud pixels that were cloud effect removed based on the characteristics of the snow cover cycle.The thresholds for extracting snow cover cycle caused false negative as result of removal of large number of accidental snowfall and occasional snow‐melt events. Additionally, the extensive area of the watershed also bring big challenge to snow cover and cloud identification as the threshold for extrac‐tion is usually inaccurate. Although there is still un‐certainty in the improved cloud effect removal algo‐rithm, the average monthlyFvalue obtained from the experimental study in the Amur River Basin still reaches about 80%?96% in the snow season. More ground coverage information was provided after the spatial compensation for the continuity of the snow data products.

Figure 5 Monthly mean of three binary indexes?Precision(a),Recall(b),and F(c)for all steps in the algorithm

4.2 Spatial-temporal variability of SCF

The annual mean SCF from 2002 to 2016 was used to reflect the spatial distribution of snow cover,as shown in Figure 6a that snow cover over the Amur River Basin was extremely unevenly distributions.Generally, the east and west side of the watershed were thicker than the central part regarding to snow cover depth, and the snow cover depth over northern part of the watershed was significantly thicker than that over the southern part of the watershed, so as the SCF. The mean SCF in the watershed for 14 years from 2002 to 2016 was about 33.94%,the higher SCF(>50%) was mainly distributed in the eastern margin and the northern part of the watershed, the western margin of the watershed, and the Inner Mongolia Pla‐teau of the central watershed. The common feature of these areas was the high mountainous rangeland(above 1,000 m a.s.l.) covered by thick forests and woody savanna. The area with SCF lower than 20%was mainly distributed in the central south region of the basin, where grasslands extensively developed over the low altitude plains. In summary, the snow cover over the Amur River Basin was very much un‐evenly distributed,which was mainly governed topog‐raphy,landcover and other meteorological reasons.

The standard deviation of SCF from 2002 to 2016 was shown in Figure 6b, and change rate of SCF over the years from 2002 to 2016 was presented in Figure 6c. It was noted that the standard deviation of SCF in the basin spatially varied in an increasing trend from east to west. The most dramatic change of SCF was found closing the border of Russia extending to the western part of the Russian territory till Mongolia.More severe changes of SCF were taken place in the southern part of the basin (in China side), and the rest areas were less changed. The change rate of SCF as shown in Figure 6c suggested that the SCF tended to be decreased mainly in the northwest of the basin, as well as in the eastern margin of the basin.About 45.2%of the basin tended to be decreased in SCF,and mainly located in the Mongolian and in the central western parts of Russia,and partly in Northwest China.

Figure 6 (a)The mean SCF from 2002 to 2016,(b)The standard deviation of SCF from 2002 to 2016,and(c)The change rate of SCF 2002 to 2016 over the Amur River Basin

4.3 Spatial and temporal distributed and variations of SCS and SCM

The date when snow cover started to accumulate(SCS) and the date when snow cover was melted away(SCM)in a hydrological year are also vital indexes for cli‐mate and Phenology studies.These two indexes were cal‐culated for whole hydrological years from 2002 to 2016 according to Formulae(8)and(9), and the mean SCS and SCM over the watershed were illustrated in Figure 7.

Figure 7 Spatial distribution of annual mean SCS(a)and SCM(b)from 2002 to 2016 over the Amur River Basin

Obviously, annual mean of SCS varied spatially,and its spatial pattern coincided with the annual mean of SCF as shown in Figure 6a considerably. Snow ac‐cumulation began in the basin as early as September,and until about early January at the latest. Snow be‐gins in September in the higher elevations to the north of the Amur Mountains such as the Kent Mountains,the Sikhote-Alin Mountains and the Bureinsky Moun‐tains, and in November in the mountains of eastern China. The southern part of the basin is situated on the main agricultural production area of China,anthro‐pogenic activities are significantly active here, where snow cover began to occur often around December of a year, which was obviously later than the eastern and western part of the basin. Snow accumulation in the central part of the basin began around January of the next year, which may be due to vegetation cover and human activities in the area, resulting in late accumu‐lation of snow.

While the annual mean of SCM in the basin var‐ied slightly over the basin,and snow cover in most ar‐eas of the basin were completely melted away around April, and the snow cover in the rest few areas over the high elevations was melted away until the end of May. However, for some high-altitude areas the snow cover could be survived till June and July for some specific years.

The monthly mean SCS and SCM over the basin were useful to identify anomalies when analyzing snow cover variations. The variation trends of the monthly mean SCS and SCM over the basin for 14 years from 2002 to 2016 was shown in Figure 8, re‐spectively. The inter-annual difference for these two indexes were obvious and none linear trends were in‐vestigated. In general, SCS and SCM varied corre‐spondingly. Postpone of SCS of specific year over the basin usually accompanied by the advance of SCM for the same year over the basin. Changes of SCS and SCM usually provide important indication to the causes of extreme climate and natural disasters, for ex‐ample, the drought in 2008 (Yanet al., 2019) and the flood in 2013(Han,2014)in the Amur River Basin.

The existing studies shown that the snow cover in Amur River Basin will be reduced due to the climate warming up, which was consistent with the changing trend of SCS and SCM (Figure 8) in the present study(Wanget al., 2017). However, as a typical seasonal snow cover area, the snow cover of the Amur River Basin shows a slight increase trend under the back‐ground of global warming. This is inconsistent with our conventional cognition, but it is true and consis‐tent with the existing research results.There are many reasons for this result, and which is need further re‐gional exploration.

Figure 8 The annual mean SCS(a)and SCM(b)from 2002 to 2016 over the Amur River Basin

5 Conclusions

Statistical analysis on spatial and temporal varia‐tions of SCF, SCS and SCM provided by the estima‐tions inferred from MODIS daily snow products in re‐cent decades revealed the spatial-temporal variations of snow cover over Amur River Basin in the past 14 years. The most dramatic change of SCF was found closing the border of Russia extending to the western part of the Russian territory till Mongolia. More se‐vere changes of SCF were taken place in the south‐ern part of the basin (in China side), and the rest ar‐eas were less changed. The SCF tended to increase mainly over the middle eastern part of the basin, but tended to decrease preliminary centered over the northwestern part of the basin, and the overall in‐creased area for SCF was slightly larger than the de‐creased over the Amur River Basin. Annual mean of SCS varied spatially, and its spatial pattern coincided with the annual mean of SCF.In the higher altitude ar‐ea over the northern part of the Amur River Basin snow cover began to occur in September. Most parts of Russia, Mongolia and the Inner Mongolia Autono‐mous Region of China,as well as the mountainous re‐gions of eastern China snow cover began to occur around November. The southern part of the basin lo‐cates the main agricultural production area of China,anthropogenic activities are significantly active here,where snow cover began to occur often around De‐cember of a year, which was obviously later than the eastern and western part of the basin.The inter-annual difference for SCS and SCM were obvious and none linear trends were investigated. In general, SCS and SCM varied correspondingly.Postpone of SCS of spe‐cific year over the basin usually accompanied by the advance of SCM for the same year over the basin.Changes of SCS and SCM usually provide important indication to the causes of extreme climate and natu‐ral disasters, for example, the drought in 2008 and the flood in 2013 in the Amur River Basin.

Acknowledgments:

This research was funded by the National Key Re‐search and Development Program of China (Grant No. 2016YFA0602302). The MODIS images are pro‐vided by NASA(http://nsidc.org), and the in-situ snow depth observation dataset were provided by Professor XianWei Wang from Sun Yat-Sen University.