ARTICLE IN PRESS

Contents lists available at ScienceDirect

Journal of Arid Environments

Journal of Arid Environments 72 (2008) 2251–2263

0140-19

doi:10.1

� Cor

3173-25

E-m

journal homepage: www.elsevier.com/locate/jaridenv

Assessing Mongolian snow disaster risk using livestockand satellite data

K. Tachiiri a,c,�, M. Shinoda b,c, B. Klinkenberg a, Y. Morinaga c,d

a Department of Geography, The University of British Columbia, 1984 West Mall, Vancouver, BC, Canada V6T 1Z2b Arid Land Research Center, Tottori University, Hamasaka 1390, Tottori 680-0001, Japanc Institute of Arid Asian Studies, Meiji University, 1-9-1 Eifuku, Suginami-ku, Tokyo 168-8555, Japand School of Commerce, Meiji University, 1-9-1 Eifuku, Suginami-ku, Tokyo 168-8555, Japan

a r t i c l e i n f o

Article history:

Received 3 December 2007

Received in revised form

8 April 2008

Accepted 23 June 2008Available online 27 August 2008

Keywords:

Drought

Dzud

Mongolia

NDVI

Regression tree

Snow Water Equivalent

63/$ - see front matter & 2008 Elsevier Ltd. A

016/j.jaridenv.2008.06.015

responding author. Present address: Frontier

Showamachi, Kanazawa-ku, Yokohama 236-

ail address: tachiiri@jamstec.go.jp (K. Tachiiri

a b s t r a c t

In Mongolia, several record-breaking disastrous dzuds (mass livestock loss directly

induced by harsh winter conditions but often influenced by drought in the previous

summer) occurred from 1999 to 2003. To understand the mechanism of this climatic

disaster, we conducted a tree regression analysis. The predictor variables included two

indices developed from remote sensing data—the Normalized Difference Vegetation

Index (NDVI) and the Snow Water Equivalent (SWE)—as well as the previous year’s

livestock numbers and mortality rates. According to the model, serious livestock

mortality was associated with low NDVI values (i.e., poor vegetation) in August of the

previous year, high SWE values (i.e., significant snow accumulation) in December of the

previous year, a high previous year’s mortality, and high previous year’s livestock

population. This result suggests that for dzud risk assessment, we need to monitor

snowfall in winter, the vegetation condition in the previous summer, and the density

and health condition of the livestock. The tree-based model developed in this study is

effective only for a white dzud (deep snow), the most common type of dzud. The large

cross-validation error indicates that more data are needed before using the model in

order to make predictions.

& 2008 Elsevier Ltd. All rights reserved.

1. Introduction

The most widely felt climatic disaster in arid regions is drought; it has been identified as the world’s most hazardousnatural disaster (e.g., Bryant, 1991; Obasi, 1994; Wilhite, 2000). In particular, rural populations dependent on agriculture orlivestock farming are significantly impacted by droughts, and since droughts inevitably occur, rural populations are alwaysexposed to the risk of this type of disaster. However, in dry and cold (and in many cases, high latitude and inland) regions,such as Mongolia, people living in rural areas are subjected not only to drought in the summer but also to another naturaldisaster in the winter. Harsh winter conditions can prevent livestock from accessing pastures and can result in a largenumber of livestock deaths. Even when winter conditions are comparatively moderate, if the pasture conditions areinadequate due to poor conditions (e.g., drought) in the previous growing season, livestock may not survive the winter.

In Mongolia, this type of disaster—a mass livestock loss directly induced by a harsh winter climate but often influencedby drought in the previous summer—is called a dzud (sometimes spelled zud). Dzud is experienced throughout central

ll rights reserved.

Research Center for Global Change, Japan Agency for Marine–Earth Science and Technology,

0001, Japan. Tel.: +8145 778 5698; fax: +8145 778 5707.

).

ARTICLE IN PRESS

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–22632252

Asia. For example, winter conditions, not drought, have historically controlled livestock numbers in Kazakhstan and winterdisasters, including heavy snow and the refreezing of melted snow, are collectively called dzhuut there (Robinson andMilner-Gulland, 2003). The United Nations and Government of Mongolia (2001) described a dzud as a winter disaster thatinvolves the mass debilitation, starvation, and death of livestock that seriously damages the livelihoods of the herderhouseholds who depend upon them.

Previous studies of this climatic disaster have been limited to simple analyses of the historic data and classification ofthe conditions. Begzsuren et al. (2004), in their study in southern Mongolia, concluded that dzud (noting that they use theterm dzud to mean severe winter weather in general) resulted in higher livestock mortality than drought. They also foundthat the average livestock mortality resulting from drought combined with dzud is 18.8%, the mortality rate associated withdzud by itself is 13.3%, drought only is 11%, and the mortality rate attributable to neither drought nor dzud is 8.8%. Thus, adisastrous situation for the livestock sector in Mongolia is likely to occur when a summer drought is followed by severewinter conditions. In the United Nations Development Programme’s (UNDP) report on natural disasters in Mongolia, dzudwas found to dominate in terms of the severity of damage (as measured by livestock loss), although wildfire was the mostcommonly occurring natural disaster in Mongolia (UNDPa). According to Templer et al. (1993), there were 24 dzuds inMongolia between 1944 and 1993, with the most disastrous one occurring in 1944 (a nationwide dzud that caused 37%livestock mortality).

Begzsuren et al. (2004) reported that Mongolian researchers classified dzuds into five types: white dzud (caused bydeep snow), black dzud (no accumulated snow), combined dzud (deep snow and sudden temperature drop), storm dzud(increased wind speed and heavy snow), and iron dzud (impenetrable ice cover over pastureland). Begzsuren et al. (2004)concluded that, among all types of dzud, white dzud combined with drought (unfortunately, the timing of the drought isnot clearly stated in their paper) resulted in the highest rates of livestock mortality (32% and 24.1%, including naturalmortality, in 1962 and 1983, respectively). Morinaga and Shinoda (2005) refined the concept of the combined dzud,calling it a cold dzud (sudden temperature drop), and they added a sixth category, hoof dzud (overgrazing). Morinaga andShinoda (2005) observed that all six dzud types can be explained using three factors: snow or ice cover, stormy weather,and lack of pasture.

The recent historic dzud in 1999–2003, however, was beyond our conventional understanding of this type of climaticdisaster. Severinghaus (2001) identified the 1999–2000 dzud as the most severe one in the past 50 years. The InternationalFederation of Red Cross and Red Crescent Societies (2004) determined that about 8.5 million or 25% of Mongolia’s herdperished from 1999 to 2003. The International Monetary Fund (2003) reported that droughts and dzuds between 1999 and2002 killed 8.2 million livestock and that 3.0 million female livestock miscarried. It estimated that, if there were nodroughts or dzuds during this period, Mongolia’s annual economic growth rate would have reached about 8%, which isabout 4–7% higher than the actual economic growth rate of 1–4%. Additionally, Natsuagdorj (2003) observed that, inthe 3 years from 2000 to 2002, more than 100 counties (soums) were affected by drought for the first time since 1989, andfor the first time in the same period, more than 100 counties were also affected by dzud. The dzuds that occurred in1999–2002 were described as ‘‘multiple dzuds’’ by the United Nations and Government of Mongolia (2001), a combinationof multiple types of dzuds over the years; we classified the dzuds as a combined dzud using the classification of Begzsurenet al. (2004).

Lotsch et al. (2005), using a vegetation index derived from remote sensing data, observed that between 1999 and 2002 alarge and geographically extensive decrease in vegetation activity that coincided with below normal rainfall in theNorthern Hemisphere, and they associated that decrease with synchronous patterns of ocean circulation anomalies in thePacific, Atlantic, and Indo-Pacific oceans. The possibility exists that the abnormal climate in Mongolia during this periodwas part of this larger scale phenomenon.

The significant impact of the dzuds of 1999–2003 prompted a need to better understand this type of climatic disaster,knowledge that would assist in the development of an early warning system (EWS) for Mongolia. In addition tosummarizing historical records and classifying the nature of the dzud, a major objective of this study was to further ourunderstanding of dzud, particularly when it starts and how it temporally progresses, and how drought and/or other effectscontribute to this process.

The results of this study were considered in a drought/dzud EWS developed as part of an international co-operation project that was launched to strengthen the systems used to track droughts and dzuds in Mongolia (Shinoda andMorinaga, 2005).

In the following sections as this study focuses on the social impact of this climatic disaster, the term dzud refers toabnormally large livestock mortality in winter–spring that occurs as a result of drought and a harsh winter–spring climate(e.g., significant snowfall). Our primary focus was on the white dzud, the most serious type.

2. Study area and data used

2.1. Study area

The study area encompasses all of Mongolia, which has 21 provinces (aimags), three of which are significantly smallerthan the others, and the capital (Ulaanbaatar), which forms an independent jurisdiction (Fig. 1a). Templer et al. (1993)

ARTICLE IN PRESS

31

2

4

65

78

9

10

11

12

1314

1516

17

181920

2122

2. Sukhbaatar 16. Khuvsgul 3. Khentii 17. Zavkhan 4. Dornogovi5. Govi-Sumber6. Ulaanbaatar7. Erdenet8. Selenge

1. DornodMongolia

Average NDVIin late August (1981-2003)

0.01 - 0.10.1 - 0.20.2 - 0.30.3 - 0.40.4 - 0.50.5 - 0.60.6 - 0.31

Decimal Degree5

Average SWE in january(1988-2003) (mm)

N.A.0 - 1010 - 2525 - 5050 - 7575 - 100100 - 196

Elevation (m)

580 - 750750 - 1,0001,000 - 1,2501,250 - 1,5001,500 - 2,0002,000- 3,523

9. Tuv10.Dundgovi 11.Umnugovi 18. Bayankhongor 12. Uvurkhangai 19. Govi-Altai 13. Arkhangai 20. Khovd 14. Bulgan 21. Uvs 15. Darkhan-Uul 22. Bayan-Ulgii

Fig. 1. Study area (Mongolia): (a) Aimag boundaries and location of Mongolia, (b) elevation (Source: GTOPO30 [http://edc.usgs.gov/products/elevation/

gtopo30/gtopo30.html]), (c) average NDVI values in late August from 1981 to 2003 (GIMMS [http://glcf.umiacs.umd.edu/data/gimms/]), and (d) average

SWE values in January from 1988 to 2003 (NSIDC [http://nsidc.org/]).

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–2263 2253

identified the dzud risk for each aimag by summarizing the historical frequency of dzuds. Provinces in the eastern regionhave the lowest risk (no more than one major dzud in 14 years), whereas provinces in the other regions have medium andhigh levels of risk.

The majority of Mongolia is considered to be highlands, particularly the western part (Fig. 1b). The northern region has ahigher average annual rainfall (about 400 mm) than the southern region (about 100 mm), part of which lies within the GobiDesert. As such, the northern provinces have, on average, better vegetation cover, whereas the southern provinces have, onaverage, poorer vegetation cover (Fig. 1c). With respect to the snow pack, the mountainous areas have, on average, a heaviersnow cover in winter (Fig. 1d), whereas the low-lying eastern region has lower snow cover. Using remote sensing data,Suzuki et al. (2003) identified the peak level of vegetation cover as occurring around the beginning of August; observationdata from the National Agency of Meteorology, Hydrology and Environmental Management (NAMHEM) identified it asoccurring in late August. The heaviest snow is typically observed in January (Morinaga et al., 2003).

Agriculture, which is dominated by the livestock sector, combined with hunting and forestry accounted for 43.8% of thenational GDP in 1996; this contribution has since decreased to 21.7% in 2005 (National Statistical Office of Mongolia, 2004, 2006).

2.2. Data used

2.2.1. Overview

We used two datasets derived from satellite imagery in our analyses. The first relates to vegetation condition, and it isestimated using the Normalized Difference Vegetation Index (NDVI). The second relates to snow conditions, which areestimated using the Snow Water Equivalent (SWE), an index that can be calculated by multiplying snow depth by itsdensity. Monitoring vegetation cover (NDVI) is useful for detecting the effects of drought, while SWE is useful formonitoring snow conditions. In addition, we used the livestock mortality data published by the Mongolian government asthe indicator of dzud damage. The aimag boundary map, the base map for all the data in this study, was provided by the GISSection of the Information and Computer Center, NAMHEM, Mongolia. The following sections contain more detailedexplanations of the NDVI, SWE, and livestock data used in the study.

2.2.2. NDVI

For the vegetation cover data, we used the Global Inventory Modeling and Mapping Studies dataset (GIMMS, http://glcf.umiacs.umd.edu/data/gimms/) provided by the Global Land Cover Facility, University of Maryland (College Park, MD,

ARTICLE IN PRESS

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–22632254

USA). The Global Mosaic of GIMMS provides semimonthly NDVI data derived from National Oceanic and AtmosphericAdministration/Advanced Very High Resolution Radiometer (NOAA/AVHRR) data from 1981 to 2003. The original cell sizeof this data is 0.0731 by 0.0731.

2.2.3. SWE

For the SWE data, we used the Global Monthly EASE-Grid Snow Water Equivalent Climatology dataset (version 1)provided by the National Snow and Ice Data Center (NSIDC) of the United States (http://nsidc.org/). This dataset presentsglobal monthly SWE data from November 1978 through June 2003. These data, available in Northern and Southern 25-kmEqual-Area Scalable Earth Grids (EASE-Grids), are derived from Scanning Multichannel Microwave Radiometer (SMMR)data from 1978 to 1987 and from selected Special Sensor Microwave/Imagers (SSM/I) data from 1987 to 2003. SSM/I iscarried onboard the Defense Meteorological Satellite Program (DMSP) series of polar orbiting satellites. The original datawere converted from a Lambert Azimuthal Equal Area Projection (25-km grid) to geographical coordinates to match theNDVI data and resampled into 0.251 grids.

2.2.4. Livestock data

In this study, we assumed that the intensity of the dzud is reflected in the livestock loss rate, which was obtained bydividing livestock deaths (as of December) by the previous year’s total livestock numbers. Livestock mortality data, inwhich deaths resulting from drought/dzud are included, are collected by the National Statistical Office of Mongolia. TheMinistry of Livestock Husbandry collects the mortality data each December, and the National Statistical Office of Mongolia(2004, 2006) has total livestock numbers for each aimag (also as of December). Combining these data sources, we produceda livestock loss rate dataset for 1991–2002 for each aimag. For the dzud of 1999–2003, supplemental data were alsoobtained from reports produced by international organizations (e.g., UNDPb).

3. Methods

3.1. Data preprocessing

Although the results are not presented, models initially were developed using simple differences but they resulted inpoorer performance than models using anomalous ratios (normalized by the long-term average).

Two NDVI values were used for each month from April to September. We used these semimonthly values to calculatethe ratio of each NDVI value to the 1981–2003 (1982–2003 for April, May, and June) average for each aimag.

To maximize the amount of NDVI data available for our analysis, the following process was carried out on theindependent variables. If the average value for a given period was less than 0.1, a shift was carried out to prevent ‘‘divisionby zero’’ as well as to preclude negative values from being derived. Thus, the ratio, NDVIr, was calculated as follows(note that in case of NDVIao0.1, [0.1�NDVIa] was added to both the denominator and the numerator so that thedenominator is never less than 0.1):

NDVIr ¼ðNDVIiÞ

ðNDVIaÞðif NDVIaX0:1Þ,

NDVIr ¼ðNDVIi þ 0:1�NDVIaÞ

0:1ðif NDVIao0:1Þ, (1)

where NDVIi is the NDVI value in the period of concern and NDVIa is the average value (1981 or 1982–2003). NDVIr wascalculated for every half-month from April to September; thus, there are 12 ratios for each year.

The final step was to take the average within each aimag of the adjusted ratio of NDVI values obtained by the above-mentioned processes and to subtract 1 so that the average values were shifted from 1 to 0.

Monthly SWE values from October to April were used as the snow-related variables. In calculating the averagevalues, only grid cells containing data for at least eight of the 16 years the dataset covers (i.e., from October 1987 toApril 2003) were considered. If either the average value or the monthly value for a grid cell had no data (i.e., the SWEvalues within a grid cell were not measured by the NSIDC), the pixels were not used in the analysis. The SWE ratiowas determined for every grid cell (for which data were available) for every month from October through April. To preventthe division-by-zero problem, a value of 1 mm was added to both the overall average and to the monthly SWEvalue. Similarly to NDVI, the data were averaged within each aimag and final SWE values were obtained by subtracting 1from the ratio.

Small aimags were integrated with an adjacent larger one in order to produce more stable rates: the data for the aimags

of Erdenet, Darkhan-Uul, and Govi-Sumber were combined with the data for Selenge, Bulgan, and Dornogovi, respectively(see Fig. 1a). Similarly, Ulaanbaatar’s data were combined with Tuv’s, the aimag that surrounds the capital. Consequently,given that there were 18 combined aimags and 11 years of livestock mortality data, there are 198 observations available.Because data from the previous year is used as an independent factor, our analysis only covers livestock losses from 1992to 2002.

ARTICLE IN PRESS

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–2263 2255

3.2. Methods of analysis

In addition to using the NDVI and SWE values as predictors, we also included the previous year’s livestock numbers(as a ratio to the total livestock within that aimag in 1990, just before the study period) and livestock loss rates (Table 1).Including the previous year’s livestock numbers enabled us to investigate the effect of density on livestock mortality(i.e., the impact of carrying capacity on mortality), and including the livestock loss rate enabled us to consider thecumulative impact of previous disasters on mortality.

Table 1Variables used in the study

NDVI

Variable name Half Month Year

Apr1_0 1st April This year

Apr2_0 2nd

May1_0 1st May

May2_0 2nd

Jun1_0 1st June

Jun2_0 2nd

Jul1_0 1st July

Jul2_0 2nd

Aug1_0 1st August

Aug2_0 2nd

Sep1_0 1st September

Sep2_0 2nd

Apr1_1 1st April Previous year

Apr2_1 2nd

May1_1 1st May

May2_1 2nd

Jun1_1 1st June

Jun2_1 2nd

Jul1_1 1st July

Jul2_1 2nd

Aug1_1 1st August

Aug2_1 2nd

Sep1_1 1st September

Sep2_1 2nd

SWE

Variable name Month Year

Jan2 January Previous year

Feb2 February

Mar2 March

Apr2 April

Oct1 October

Nov1 November

Dec1 December

Jan1 January This year

Feb1 February

Mar1 March

Apr1 April

Oct0 October

Nov0 November

Dec0 December

Others

Pop1 Previous year’s livestock population

Loss1 Previous year’s livestock loss rate

The last digit in a variable name indicates the year (i.e., 0 refers to the current season and 1 to the previous season), while for the NDVI data the second last

digit represents the 1st or 2nd half of the month. For SWE, in order to respect seasonal continuity, the same last digit is used throughout one winter from

October to the next April.

ARTICLE IN PRESS

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–22632256

After reviewing and checking the data, we first considered a linear regression approach using the variables describedabove (after removing some of the variables as a result of high intercorrelations amongst some of them), but the results(Fig. 2) demonstrated that a linear model could not adequately represent the strongly nonlinear data. After considering thealternatives, and in light of the need to use the results of our analysis in the development of an EWS, a tree-based modelapproach was selected in order to predict dzud damage. All analyses, including data assessment and tree-based modeling,were carried out using R (Venables et al., 2006).

Tree-based models are a relatively new method of analysis that provides an alternative to linear and additive models forregression problems (regression trees) and to linear logistic and additive logistic models for classification trees. Tree-basedmodels employ binary recursive partitioning, whereby a dataset is successively split into increasingly homogeneoussubsets until it is infeasible to continue (Clark and Pregibon, 1992). The following explanation is an overview of thismethodology summarized from Hastie et al. (2001).

The model can be described as a binary tree. The full dataset sits at the top of the tree and observations satisfying thestated condition at each junction are assigned to the left branch; all others are assigned to the right branch. The terminalnodes, or leaves, of the tree correspond to the regions having a single predicted value. For each leaf, the average (which canbe derived from the minimization of the sum of squares) is normally chosen as the representative value (if the absolutevalue is selected as the error function, the predicted value of each leaf will be the median). For every step or branch ofthe tree, a splitting variable and a splitting point are determined so that the result minimizes the sum of the squared error(or an alternative error function).

In order to limit the tree size (i.e., the number of branches) and to prevent over-fitting (i.e., having too many parameters,which may decrease the training error but increase the validation error), additional parameters are used to limit the tree’ssize (i.e., we could reduce the total error of the model to zero by using as many nodes as there are observations). A preferredmethod is, first, to grow a large tree T0, stopping the splitting process only when some a priori minimum node size (given inadvance considering the application of the model output) is reached, and then to prune the tree by using a complexityparameter. The complexity criterion is

CaðTÞ ¼XjTj

m¼1

NmQmðTÞ þ ajTj, (2)

where a, |T|, Nm, and Qm(T) are the complex parameter, number of leaves, number of data points in leaf m, and the errorfunction for leaf m, respectively. The error function is expressed here as

QmðTÞ ¼1

Nm

X

xi2Rm

ðxi � cmÞ2, (3)

where cm, Rm, and xi are the representative value of the leaf, the space covered by leaf m, and the observed data included inleaf m, respectively. Starting from T0 (i.e., a ¼ 0), we increase a continuously and successively delete the internal node sothat Ca(T) is minimized. This process ends when we reach a one-leaf (no split) tree. As a is increased, a smaller tree isselected as the best model. Thus, we obtain a set of candidate trees for varied a (the weight for the simplicity).

Y = XR2 = 0.46

-0.1

0

0.1

0.2

0.3

0.4

0.5

-0.1

Obs

erve

d lo

ss r

ate

Model output

0 0.1 0.2 0.3

Fig. 2. Scatterplot of the output of the linear regression model and observed loss rate.

ARTICLE IN PRESS

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–2263 2257

From a set of candidate sub-trees, the best model is chosen through five- or 10-fold cross validation (CV, Breiman et al.,1984). This process is called pruning. We carried out five- and 10-fold CVs 10 times each and used the average error valuesto help select the best model. We used the minimum validation error method to select the best model because thedifference in error between each model was too small to consider applying the alternative 1 S.E. method (although 1 S.D.would seem to be a more appropriate name for the method, we decided to follow the conventional naming in theliterature) that is, taking the smallest tree that has an error within 1 S.D. of the minimum (Venables and Ripley, 2002).

An n-fold CV error for the tree-based model was assessed for 1/n of the data through comparison with the tree derivedfrom (n�1)/n of the data (Breiman et al., 1984). The resultant CV error (relative deviance) for tree regression is typicallyreported to be about 0.6–0.7 (Deconinck et al., 2005; Laaha and Bloschl, 2006; Negron, 1998) or greater (Nerini et al., 2000).Nirel and Dayan’s (2001) model had a training error of about 0.7, which indicates that the (unreported) CV error was larger.

In this study, the rpart (which stands for Recursive PARTitioning) package of R was used to develop the tree-basedmodel. Upon developing the initial tree, the minimum number of observations for each terminal node (m1) and theminimum number of observations a parent node must possess prior to being split (m2; obviously m2X2m1) weredetermined as follows: considering the dataset size, we concluded that an m1 of about 5 is suitable, and several values ofsuch an m1 (3, 4, 5, 6, 10) and typical m2 (2m1 and 3m1) were tested to select the best combination. The test resultsdemonstrated that, in general, (1) smaller m1 values result in a larger CV error, although they also produce smaller trainingerrors (i.e., over-fitting); (2) setting m2 ¼ 3m1 results in a smaller cross-validation error as compared with m2 ¼ 2m1 for ourdata; and (3) if m145, the leaves become too heterogeneous (i.e., the model is over constrained). After considering theresults, a setting of (m1, m2) ¼ (5, 15) was identified as producing the best results.

In order to strengthen the robustness of the tree-based model (e.g., Schroder, 2006), the dependent data were firstconverted to ordinal (ordered categorical) data. By this transformation, the nonlinear relation between the livestockmortality and potential explaining variables is significantly moderated, and it was confirmed through observation that theresultant leaves become more homogenous. This transformation is based on the concept that a loss rate less than a certain

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

0.01

0.

02

0.03

0.

04

0.05

0.

06

0.07

0.

08

0.09

0.

10

0.11

0.

12

0.13

0.

14

0.15

0.

16

0.17

0.

18

0.19

0.

20

0.21

0.

22

0.23

0.

24

0.25

0.

26

0.27

0.

28

0.29

0.

30

0.31

0.

32

0.33

0.

34

0.35

0.

36

0.37

0.

38

0.39

0.

40

0.41

0.

42

0.43

0.

44

Mor

e

Cum

ulat

ive

(%)

Freq

uenc

y

Loss rate

FrequencyCumulative (%)

Fig. 3. Histogram of livestock loss rate.

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

1991

Liv

esto

ck lo

ss r

ate

Year

1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002

Fig. 4. Distribution of livestock loss rate per year.

ARTICLE IN PRESS

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–22632258

value can be considered to be normal, whereas loss rates beyond a certain threshold can be thought to be out of theordinary (eventually being classified as ‘‘emergency’’ rates). The following classification was determined on the basis ofthe distribution of the livestock loss rates (L, presented in Figs. 3 and 4): 1 is for Lo0.05 (131 cases), 2 is for 0.05pLo0.11(42 cases), 3 is for 0.11pLo0.17 (12 cases), 4 is for 0.17pLo0.23 (7 cases), and 5 is for 0.23pL (6 cases).

Although the transformed data are ordered categorically, the classification was specified such that it is roughly possibleto consider the distance between category 1 and category 3 as twice that between category 1 and category 2; as such, theclasses to some degree can be considered to be on an interval scale (i.e., the concept of root mean square errors (RMSEs)works as an error evaluation). Kramer et al. (2001) concluded, although noting future work is needed, that using theregression tree methodology with ordinal data results in reasonable performance in RMSE and the Spearman’s rankcorrelation coefficient.

4. Results

4.1. Overview of the 2000–2003 dzud

The distribution of cases with a loss rate of less than 10% is close to a normal distribution. However, when the caseswith loss rates above 10% are included, there is an obvious non-normal distribution (Fig. 3). Furthermore, it is apparentthat almost all of the high loss rate cases were observed in the period from 2000 to 2002 (Fig. 4; note that data for 2003are unavailable).

The spatial distribution of dzud damage for 2000 through 2002 is presented in Fig. 5. In 2000, the most serious damagewas observed in the center of the country. All aimags recorded some dzud damage in 2001, with the northern aimags

exhibiting the highest livestock loss rates. In 2002, the dzud occurred principally in the southwestern region and thenortheast region was not impacted. On average, the aimag of Bayankhongor experienced the most serious dzud damage,and, basically, the further away from this aimag, the more moderate the dzud damage.

4.2. Tree-based model

The final tree regression model (Fig. 6) has an estimated RMSE of 0.441. The first node uses Aug1_1 as the branchingcondition (i.e., NDVI in early August in the previous year). Thus, if the NDVI in this period is greater than normal(X�0.0065) (note that in this section, references to NDVI and SWE values refer to the transformed values used in theanalysis), the cases branch to the left. This branch contains comparatively moderate livestock loss rates. On the other hand,if the NDVI is below this threshold, the cases are assigned to the right branch that contains the higher livestock loss rates. Asimilar methodology was used for the remaining branches: if the data satisfies the condition given as an inequality, thecases are assigned to the left branch, and if the data does not match the condition, the cases are assigned to the rightbranch. In this tree, the most severe losses are assigned to the leaf that contains cases in which NDVI for Aug1_1o�0.0065,Loss1 (the previous year’s loss rate) X0.0595, SWE for Dec1 X�0.6315, and NDVI for Aug2_1o�0.0635. There are two

Fig. 5. Dzud damage, as indicated by livestock mortality, for the years (a) 2000, (b) 2001, and (c) 2002. The categories of livestock loss rate are the same as

those used in the tree-based model in the study. The average presented in (d) is a simple average for the 3 years.

ARTICLE IN PRESS

|Aug1_1>=-0.0065

Feb1< 0.3005

Sep2_1>=-0.0845

Nov1< -0.052

Jun2_0>=-0.233

Jul1_1>=0.02

Loss1< 0.0595

Pop1< 0.397

Mar1< -0.158

Nov0>=-0.485

Loss1< 0.031

Jun2_1>=-0.1155

Jun1_1< -0.008

Dec1< -0.6315

Aug2_1>=-0.0635

1.561

1.216

1.116

1.078

1.0131 1.2 1.462

1.8

1.8131.444 2.286

2

1.645

1.526

1.233

1.1251 1.5 1.667

1.852

1.636

1 2

2.8

3

2.88

1.125 3.706

3 4.714

(1) (2) (3)

(4)(5) (6)

(7) (8) (9)

(10) (11)

(12)

(14)

(15) (16)

(13)

Fig. 6. The final, pruned tree, model. The numbers are the average of each node, and the numbers in parentheses indicate the leaf numbers, which

correspond with those in Appendix 1 (see Appendix A).

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–2263 2259

leaves that contain the next most severe set of cases: when the same conditions as above are met except forAug2_1X�0.0635, and when NDVI Aug1_1o�0.0065, Loss1o0.0595, and the previous year’s livestock population X0.397.That is, if NDVI in the first-half of August in the previous year is equal to or above �0.0065 (i.e., 99.35% of the average), theresultant livestock loss rate is generally low, at worst being in category 3 (i.e., a livestock loss rate between 11% and 17%).

It should be noted that in the model the main predictors, except for Loss1 and Pop1, are the previous summer’s NDVIand the previous winter’s SWE. Only two predictors (Jun2_0 and Nov0) that reflect conditions after Apr1 (April of thecurrent year) are included, neither of which is associated with significant impact to the values finally predicted.

There is only one case wherein smaller SWE values result in higher livestock mortality; that is, if Nov0o�0.485(i.e., approximately smaller than half of the average SWE value for that month), leaf (9) branches from leaves (7) and (8).This case may be related to the occurrence of a black dzud. Approximately, equal numbers of NDVI (7 out of 24 in all) andSWE (5/14) predictors are used as conditions in the final model (Fig. 6).

The correspondence between the actual and the predicted category (by the model in Fig. 6) is presented in Appendix 1(see Appendix A) (electronic version only). The condition column in Appendix 1 (see Appendix A) was determined asfollows: (1) leaf nodes containing a majority of category 1 cases and no cases 3 or higher are called normal; (2) leaf nodeswith only one case of 3 (or higher), or having a majority of cases higher than category 1, are labeled cautious; (3) leaf nodeswith no category 1 cases and an average of less than 3 are called moderate dzud; (4) leaf nodes with an average of 3are called intense dzud; and (5) leaf nodes with an average above 4, in which all cases are 4 or 5, are identified asemergency dzud.

The plot of total deviance (training error) for all the candidate trees obtained when we increase the complex parameter(a) shows that later splits make smaller improvements (Fig. 7). All the candidate sub-trees are presented in Appendix 2(see Appendix A) (electronic version only).

The average of 10 times the 5- and 10-fold CV error is 0.96.

5. Discussion

The significant contribution of Pop1 and Loss1 in the tree regression model (Fig. 6) should be noted. The positivecontribution of Pop1 in the model indicates that there is an inherent carrying capacity of livestock numbers and, if it isexceeded, more livestock are prone to abnormal deaths and dzuds control the livestock number. On the other hand, thepositive contribution of Loss1 suggests that a dzud’s effect sometimes extends into the following year; that is, livestockweakened in the first year fail to survive through the following year if stressful conditions are experienced.

ARTICLE IN PRESS

Dev

ianc

e

200

180

160

140

120

100

80

60

40

20

00 5 10 15

Number of splits

Candidate sub-treesTrees never selected ascandidate trees for any α

Fig. 7. Reduction in deviance of the tree-based model as a function of the number of splits.

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–22632260

Although the final regression tree model (Fig. 6) provides a convincing view of dzud (e.g., the environmental conditionsincluded in the model are consistent with the seasonal livestock mortality trend; Batjargal et al., 2000), the large total CVerror (a CV error of 0.96 means there is only a 4% improvement over a mean [constant] model) and the inapplicability of the1 S.E. method to choose the best tree indicates that the results presented may not represent the best ‘‘global’’ solution. Themain reason for the large CV error is our data’s strongly biased distribution (Figs. 2 and 3). Because the average of our datais 1.56, taking the average value as the ‘‘best’’ predictor results in a small error for categories 1 and 2 (the majority of thecases), but it results in a large error for categories 3–5. These characteristics result in our model producing a large relativedeviance for categories 1 and 2, and a relatively small deviance for categories 3–5 (the significant categories in disastermanagement). For 10-fold CV, for example, the average relative deviance for high mortality categories (3–5) is improved to0.65, in line with the results of previous studies. Additional data should be obtained that would improve the reliability ofthe tree-based model.

The split associated with Nov0, wherein a relatively small amount of snow resulted in a relatively larger livestock loss,is the only case in the model that cannot be explained by the white dzud mechanism. This split may be related to ablack dzud.

Although the white dzud was the main target of our study, other abrupt abnormal climatic events occur in winter andspring. These include serious snowstorms in late March or April (i.e., after sheering), which sometimes result in largelivestock losses, as reported in 1993 (15 March–15 May, Templer et al., 1993), 2001 (5–9 April, UNDPc), and 2002 (18–22March, UNDPd, and 5–9 April, UNDPe). In particular, a snowstorm on 5–9 April 2001 (UNDPc) resulted in damage in at least12 aimags; 20 people and 150,000 livestock were killed. Storm dzuds, which occur rapidly and are difficult to predict inadvance, are beyond the scope of this study, but they should also be considered in the future, given the significant impactthey can have on people and livestock. Although some of these effects are considered in our model (i.e., SWE in March andApril), the compounding effect of wind, in particular, has not been considered.

In the future, a model using absolute physical values, as opposed to relative ones, should be attempted. It shouldconsider the relation between snow depth, snow density, and the availability of pasture (e.g., Tuvaansuren, 1986).Incorporating such relations in conjunction with an animal biology model (e.g., Hobbs, 1989) integrated with a plantgrowth model and a climatic model may be possible in the future, but the development of such a complex model is stillseveral years away, so our tree-based model is appropriate for now.

Another important consideration revealed in Tuvaansuren’s (1986) study was the different impacts that snowconditions have with respect to each type of livestock. In Mongolia, associated with the regional vegetation and climaticconditions, each region has a unique composition of livestock types. For example, in the southern region, goats are morecommon than sheep, whereas sheep are more common in the other regions. The regional differences in livestockcomposition and the morphological differences associated with each livestock type in the different regions result indifferent resistances to snow and drought. According to data from the National Statistical Office of Mongolia, a higher levelof camel mortality is found in the northern region, whereas other livestock mortality is higher in the southern region.Significant levels of goat mortality are evident in Zavkhan, Bayankhongor, and Govi-Altai (which also have the highestmortality in total livestock as well). This factor is implicitly considered in our analysis since livestock holders have adaptedover time to the average regional climate and abnormalities from the average will impact the livestock regardless of thetype of livestock common to that region.

ARTICLE IN PRESS

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–2263 2261

Some error is expected in estimating SWE and vegetation conditions from satellite imagery. A comparison of observedSWE (the product of snow depth and its density observed by NAMHEM) and estimated SWE values indicates that there is aspatial variation in the ratio of remotely sensed and observed SWE values. Existing studies (e.g., Tedesco et al., 2004)concluded that the inversion of the emission model can include significant bias dependent upon snow grain size,stratification, and snow pack; thus, even the empirical SWE detection method, such as that applied in the dataset used inthis study, results in some bias. However, using ratios of the data (to the average), as we did in our analyses, shouldmoderate any effect that uncertainty in the values has on our results. In general, SWE indices derived from remote sensingdata are insensitive to SWE conditions above about 100 mm (e.g., Tedesco et al., 2004). Fortunately, according toTuvaansuren (1986), this threshold is high enough to avoid serious problems when evaluating livestock’s accessibility tosnow-covered pasture.

The timing of the collection of livestock numbers by the Mongolian government should also be considered. A dzud endsin spring, so the current timing of the data collection (in December) is not best for dzud assessment. The government pointsout that it is easier to count livestock when the livestock population is stable. It would be best if the livestock numberscould be determined after a dzud has ended, if the numbers are to be used in monitoring the impacts of dzud, although thismay not be technically feasible.

Finally, the effects of changes in social conditions after 1990 and how those changes affect the impact that disasters haveon society should be considered. Mearns (2004) pointed out that herd sizes had increased until the last dzud, due in part tothe shift from socialism to a market economy that was triggered by the collapse of the Soviet Union in 1991. Siurua andSwift (2002) stated that one reason for the most recent dzud’s significant toll on livestock was emergence of the marketsystem, which resulted in a large number of unemployed people becoming inexperienced nomads. They reported that theself-security system in the local Mongolian community (i.e., normal winter preparations, such as preparing frozen meatand flour, and a traditional mutual assistance system) endured until 2001 (resulting in almost no human deaths), but thatsystem started to decline in 2002. Bedunah and Schmidt (2004) concluded that governmental policy to protect some areasfrom overgrazing had inadvertently produced negative side effects. Furthermore, dzuds also affect the structure of society.For example, in 2001 after 2 consecutive years of dzuds, the Mongolian prime minister publicly stated that Mongoliansshould stop being nomads and suggested intensive urbanization of Mongolia (Finch, 2002).

6. Conclusion

In this study, motivated by the serious consequences of a recent historic dzud, and the desire to moderate future dzuddamage in Mongolia, we used datasets readily derived from remote sensing data as well as census data of livestocknumbers and mortality rates in a regression tree analysis in order to develop an understanding of the conditions that leadto dzud. With the development of an effective EWS in mind, we particularly focused on the need to identify which factors,selecting from amongst the (semi)monthly vegetation and snow conditions observed in the preceding period, and the

Livestock are weakened by the drought in the previous year (e.g., Aug1_1, Aug2_1)

Large livestock mortality wasobserved by the government

ASOND

MAMJJ

JF

ASOND

MAMJJ

JF

Too many livestock createcapacity problem

Large loss rate in previousyear associated with manyweakened livestock

Small effect by the drought inthe this year’s summer (Jun2_0)

Little snowfall in November(Nov0) results in slightlylarger loss (black dzud?)

Thi

s ye

arPr

ev y

ear

Livestock are killed by deep snow at the end of the previous winter (e.g., Feb1, Mar1)

Fig. 8. Conceptual diagram of the results of this study.

ARTICLE IN PRESS

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–22632262

recent livestock population/mortality rates, have the greatest explanatory potential with respect to predicting a potentialwinter disaster (i.e., a dzud).

The regression tree analysis was used after we confirmed that linear regression models failed to adequately predictlivestock mortality because the relationship between the predictor and the predicted variables exhibited significantnonlinearity. The regression tree method demonstrated that it is robust enough to deal with problematic (i.e., nonlinearand non-normal) data and produce results that are easy to interpret.

The regression tree model results are summarized in Fig. 8. In the model, only two predictors (Jun2_0 and Nov0) after Apr1(April of the current year) are included, neither of which is associated with significant impacts. The minimal contribution bythe current season’s NDVI indicates that drought weakens but does not kill the livestock in many cases. The large contributionof NDVI values from the previous summer and SWE values from the previous winter means that livestock weakened by adrought in the previous year are likely killed by deep snow in the previous winter. That is, as the animals typically die in thelatter half of winter—January or later—they are not included in the December count for the winter in which they died. Theirdeaths are thereby counted by the government almost a full year after the events that led to that mortality.

The model provides satisfactory results in understanding the dzud mechanism, although the large CV error indicatesthat we should be cautious when applying the model to actual monitoring. The tree-based model developed in this study iseffective only for a white dzud. Although the model includes one split which may be related to a black dzud, detailedconsideration of other types of dzud remains as a future work.

Spatial variation in the parameters should also be considered in extending our analyses, although significantly moredata will be needed in order to use spatial regression techniques. Ultimately, recent progress in modeling may enable us tocreate advanced models that connect numerical predictions of climate and physical processes to animal and plantecological models. However, until then, the model presented in this study should be of significant help to governmentofficials trying to prepare for dzuds.

Appendix A. Supplementary materials

Supplementary data associated with this article can be found in the online version at doi:10.1016/j.jaridenv.2008.06.015.

References

Batjargal, Z., Oyun, R., Togtokh, N., Sangidansranjav, S. (Eds.), 2000. Lessons learnt from the dzud 1999–2000. An English summary of 1999–2000 dzudlessons. MON/00/302 Study Project by UNDP.

Bedunah, D.J., Schmidt, S.M., 2004. Pastoralism and protected area management in Mongolia’s Gobi Gurvansaikhan National Park. Development andChange 35 (1), 167–191.

Begzsuren, S., Ellis, J.E., Ojima, D.S., Coughenour, M.B., Chuluun, T., 2004. Livestock responses to drought and severe winter weather in the Gobi ThreeBeauty National Park, Mongolia. Journal of Arid Environments 59, 786–796.

Breiman, L., Friedman, J.H., Olshen, R.A., Stone, C.J., 1984. Classification and Regression Trees. Wadsworth International Group, Belmont, CA.Bryant, E.A., 1991. Natural Hazards. Cambridge University Press, Cambridge.Clark, L.A., Pregibon, D., 1992. Tree-based models. In: Chambers, J.M., Hastie, T.J. (Eds.), Statistical Models in S. Wadsworth & Brooks/Cole Advanced Books

& Software, Pacific Grove, CA, pp. 377–419.Deconinck, E., Hancock, T., Coomans, D., Massart, D.L., Heyden, Y.V., 2005. Classification of drugs in absorption classes using the classification and

regression trees (CART) methodology. Journal of Pharmaceutical and Biomedical Analysis 39 (1–2), 91–103.Finch, C.M., 2002. Mongolia in 2001: political consolidation and continued economic reform. Asian Survey 42 (1), 39–45.Hastie, T., Tibshirani, R., Friedman, J., 2001. Additive models, trees and related methods. In: The Elements of Statistical Learning, Springer, New York,

pp. 257–298.Hobbs, N.T., 1989. Linking energy balance to survival in mule deer: development and test of a simulation model. Wildlife Monograph 101, 1–39.International Federation of Red Cross and Red Crescent Societies, 2004. Final Report Mongolia: Dzud 23 November 2004, 18p. Available from: /http://

www.reliefweb.int/library/documents/2004/ifrc/ifrc-mong-23nov.pdfS (accessed October 2006).International Monetary Fund, 2003. Mongolia: Poverty Reduction Strategy Paper, IMF Country Report 03/277. Available from: /http://www.imf.org/

external/pubs/ft/scr/2003/cr03277.pdfS (accessed October 2006).Kramer, A., Widmer, G., Pfahringeer, B., de Groeve, M., 2001. Prediction of ordinal classes using regression trees. Fundamenta Informaticae XXI, 1001–1013.Laaha, G., Bloschl, G., 2006. A comparison of low flow regionalisation methods—catchment grouping. Journal of Hydrology 323 (1–4), 193–214.Lotsch, A., Friedl, M.A., Anderson, B.T., Tucker, C.J., 2005. Response of terrestrial ecosystems to recent Northern Hemispheric drought. Geophysical Research

Letters 32 (6), L06705.Mearns, R., 2004. Sustaining livelihoods on Mongolia’s pastoral commons: insights from a participatory poverty assessment. Development and Change 35

(1), 107–139.Morinaga, Y., Shinoda, M., 2005. Natural disasters in Mongolia—drought and dzud monitoring system at the Institute of Meteorology and Hydrology.

In: Environment Handbook. National Museum of Ethnology, Osaka (in Japanese).Morinaga, Y., Tian, S.-F., Shinoda, M., 2003. Winter snow anomaly and atmospheric circulation in Mongolia. International Journal of Climatology 23 (13),

1627–1636.National Statistical Office of Mongolia, 2004. Mongolia in a market system. Statistical Yearbook, Ulaanbaatar.National Statistical Office of Mongolia, 2006. Statistical Yearbook 2005. Ulaanbaatar.Natsuagdorj, L., 2003. Pasture and animal husbandry. In: Batima, P. (Ed.), Climate Change. Institute of Meteorology and Hydrology, Ulaanbaatar

(in Mongolian).Negron, J.F., 1998. Probability of infestation and extent of mortality associated with the Douglas-fir beetle in the Colorado Front Range. Forest Ecology and

Management 107 (1–3), 71–85.Nerini, D., Durbec, J.P., Mante, C., 2000. Analysis of oxygen rate time series in a strongly polluted lagoon using a regression tree method. Ecological

Modelling 133 (1–2), 95–105.Nirel, R., Dayan, U., 2001. On the ratio of sulfur dioxide to nitrogen oxides as an indicator of air pollution sources. Journal of Applied Meteorology 40 (7),

1209–1222.

ARTICLE IN PRESS

K. Tachiiri et al. / Journal of Arid Environments 72 (2008) 2251–2263 2263

Obasi, G.O.P., 1994. WMO’s role in the international decade for natural disaster reduction. Bulletin of the American Meteorological Society 75, 1655–1661.Robinson, S., Milner-Gulland, E.J., 2003. Political change and factors limiting numbers of wild and domestic ungulates in Kazakhstan. Human Ecology 31

(1), 87–110.Schroder, W., 2006. GIS, geostatistics, metadata banking, and tree-based models for data analysis and mapping in environmental monitoring and

epidemiology. International Journal of Medical Microbiology 296 (Suppl. 1), 23–36.Severinghaus, S.R., 2001. Mongolia in 2000: the pendulum swings again. Asian Survey 41 (1), 61–70.Shinoda, M., Morinaga, Y., 2005. Developing a combined drought–dzud early warning system in Mongolia. Geographical Review of Japan 78 (13), 928–950

(in Japanese with English summary).Siurua, H., Swift, J., 2002. Drought and zud but no famine (yet) in the Mongolian herding economy. IDS Bulletin 33 (4), 88–97.Suzuki, R., Nomaki, T., Yasunari, T., 2003. West–east contrast of phenology and climate in northern Asia revealed using a remotely sensed vegetation index.

International Journal of Biometeorology 47 (3), 126–138.Tedesco, M., Pulliainen, J., Takala, M., Hallikainen, M., Pampaloni, P., 2004. Artificial neural network-based techniques for the retrieval of SWE and snow

depth from SSM/I data. Remote Sensing of Environment 90 (1), 76–85.Templer, G., Swift, J., Payne, P., 1993. The changing significance of risk in the Mongolian pastoral economy. Nomadic Peoples 33, 105–122.Tuvaansuren, G., 1986. Results of zoo-meteorological observation in Mongolia. Papers in Meteorology and Hydrology 11, 132–145 (in Russian).UNDPa. Disasters in 1996–2000. Available from: /http://mirror.undp.org/mongolia/archives/disaster/disasterevents/disasterevents.PDFS (accessed

October 2006).UNDPb. Availble from: /http://mirror.undp.org/mongolia/archives/disaster/#S (accessed October 2006).UNDPc. Available from: /http://mirror.undp.org/mongolia/archives/disaster/dzud2000/statistics.html#003S (accessed October 2006).UNDPd. Available from: /http://mirror.undp.org/mongolia/archives/disaster/aimag%27s%20report/storm%20information/18-22march.htmS (accessed

October 2006).UNDPe. Available from: /http://mirror.undp.org/mongolia/archives/disaster/aimag%27s%20report/storm%20information/Snowstorm%20of%20April.PDFS

(accessed October 2006).United Nations and Government of Mongolia, 2001. Mongolia Winter Disaster Dzud: Appeal for International Assistance (February–May, 2001).Venables, W.N., Ripley, B.D., 2002. Tree based methods. In: Modern Applied Statistics with S, fourth ed. Springer, New York, pp. 251–270.Venables, W.N., Smith, D.M., the R Development Core Team, 2006. An Introduction to R-Notes on R: A Programming Environment for Data Analysis and

Graphics, Version 2.4.0, 93p. Available from: /http://www.r-project.org/S (accessed December 2006).Wilhite, D.A., 2000. Drought as a natural hazard: concepts and definitions. In: Drought, A Global Assessment, vol. 1. Routledge, London/New York, pp. 3–8.