HYDROLOGICAL PROCESSES Hydrol. Process. 22, 2235– 2247 (2008) Published online 19 November 2007 in Wiley InterScience (www.interscience.wiley.com) DOI: 10.1002/hyp.6820
Spatial and temporal characteristics of droughts in the western part of Bangladesh Shamsuddin Shahid* Department of Geography, Rhodes University, Grahamstown 6140, South Africa
Abstract: Spatial and temporal characteristics of droughts in the western part of Bangladesh have been analysed. Standardized precipitation index method is used to compute the severity of droughts from the rainfall data recorded in 12 rainfall gauge stations for the period of 1961–1999. An artificial neural network is used to estimate missing rainfall data. Geographic Information System (GIS) is used to map the spatial extent of droughts of different severities in multiple time scales. Critical analysis of rainfall is also carried to find the minimum monsoon and dry months rainfall require in different parts of the study area to avoid rainfall deficit. The study shows that the north and north-western parts of Bangladesh are most vulnerable to droughts. A significant negative relationship between multiple ENSO index and rainfall is observed in some stations. Analysis of seasonal rainfall distribution, rainfall reliability and long-term rainfall trend is also conducted to aid prediction of future droughts in the area. Copyright 2007 John Wiley & Sons, Ltd. KEY WORDS
droughts; rainfall; standardized precipitation index; GIS; Bangladesh
Received 14 June 2006; Accepted 1 May 2007
INTRODUCTION Droughts are recurrent phenomena in the western part of Bangladesh. Since independence in 1971, the country has suffered from nine droughts of major magnitude (Paul, 1998). The impact of droughts was higher in the western part of the country compared to other parts. In recent decades, the hydro-climatic environment of north-western Bangladesh has been aggravated by environmental degradation and cross- country anthropogenic interventions (Banglapedia, 2003). Scientists have become increasingly concerned about the frequent occurrence of drought in western districts of Bangladesh, and this paper reports on studies of drought conditions in the western part of Bangladesh. Although droughts may occur at any time of the year, the impact of droughts during the pre-monsoon period is more severe in Bangladesh. High yield variety Boro rice, which is cultivated in 88% of the potentially available areas of the country, grows during this time. A deficit of rainfall during this period causes huge damage to agriculture and to the economy of the country. As for example, drought in 1995 led to a decrease in rice and wheat production of 3Ð5 ð 106 ton in the country (Rahman and Biswas, 1995). This necessitated the import of huge amount of food grains to offset the shortage in national stocks and meet the national demand on an emergency basis (Paul, 1998). In this paper, pre-monsoon drought as well as droughts due to a deficit of monsoon rainfall have been studied. * Correspondence to: Shamsuddin Shahid, Department of Geography, Rhodes University, Grahamstown 6140, South Africa. E-mail: sshahid
[email protected] Copyright 2007 John Wiley & Sons, Ltd.
Drought is a dynamic phenomenon, which changes over time and space. Therefore, complete analysis of drought requires study of its spatial and temporal extents. Hydrological investigation over a large area requires assimilation of information from many sites, each with a unique geographic location (Shahid et al., 2000). Geographic Information System (GIS) maintains the spatial location of sampling points, and provides tools to relate the sampling data through a relational database. Therefore, it can be used effectively for the analysis of spatially distributed hydro-meteorological data and modelling. In the present paper, GIS is used for the spatial modelling of droughts in western Bangladesh at various time-scales. The common indicators of drought include meteorological variables such as precipitation and evaporation, as well as hydrological variables such as stream flow, groundwater levels, reservoir and lake levels, snow pack, soil moisture, etc. Based on these indicators, numerous indices have been developed to identify the severity of drought conditions (Dracup et al., 1980; Wilhite and Glantz, 1985, 1987). However, most meteorological drought indices are based on precipitation data, e.g. Percentage of Normal Index (Banerji and Chabra, 1964), Precipitation Deciles Index (Gibbs and Maher, 1967), Bhalme–Mooley Drought Index (Bhalme and Mooley, 1980), Standardized Precipitation Index (McKee et al., 1993), Effective Drought Index (Byun and Wilhite, 1999), etc. Among these methods, the Standardized Precipitation Index (SPI) quantifies the precipitation deficit for multiple time steps, and therefore facilitates the temporal analysis of droughts. It has been found that SPI is better able to show how drought in one region compares to drought in another region (Guttman, 1998). It
2236
S. SHAHID
has also been reported that SPI provides a better spatial standardization than the other indices (Lloyd-Hughes and Saunders, 2002). Therefore, SPI is used to study the spatial and temporal characteristics of meteorological drought in western Bangladesh. Critical rainfall analysis, seasonal rainfall distribution, rainfall reliability and longterm rainfall trend are also studied to aid prediction of droughts in the area.
by De Martonne and Thornthwaite methods are 20Ð89 and 64Ð04, respectively, in the central-western and northwestern parts of Bangladesh. As the dryness index values in the region are close to those of a dry zone, the climate of these regions of Bangladesh can be considered very close to ‘dry’. The total annual evapotranspiration in this part of Bangladesh is also lower than or equal to the annual rainfall in some years. The location map of the study area is shown in Figure 3.
HYDRO-CLIMATE OF BANGLADESH Geographically, Bangladesh extends from 20° 340 N to 26° 380 N latitude and from 88° 010 E to 92° 410 E longitude. Climatically, the country belongs to the sub-tropical region where monsoon weather prevails throughout the year in most parts of the country. The average temperature of the country ranges from 17 ° C to 20Ð6 ° C during winter and 26Ð9 ° C to 31Ð1 ° C during summer. The average relative humidity for the whole year ranges from 70Ð5% to 78Ð1%, with a maximum in September and a minimum in March. Three distinct seasons can be recognized in Bangladesh from the climatic point of view: (i) the dry winter season from December to February; (ii) the pre-monsoon hot summer season from March to May; and (iii) the rainy monsoon season, which lasts from June to October (Rashid, 1991). The spatial distribution of rainfall over the country is shown in Figure 1a. The map has been prepared from rainfall data for the 30 years 1970–1999, available at 50 meteorological stations situated in and around the country. The average annual rainfall of the country varies from 1329 mm in the north-west to 4338 mm in the north-east (Shahid et al., 2005). The map shows that the western part of Bangladesh receives much lower rainfall than other parts of the country. The monthly distribution of rainfall over the western part of the country is shown on the graph in Figure 1b. The monthly distribution is calculated from rainfall data for the 39 years 1961–1999 available at 12 stations in the study area. The right vertical axis of the graph represents rainfall in millimetres and the left vertical axis represents the rainfall as a percentage of annual total rainfall. The graph shows that rainfall is very much seasonal in the area, almost 77% of rainfall occurring during the monsoon. In summer, the hottest days experience temperatures of 45 ° C or even hotter. In the winter the temperature falls to 5 ° C in some places (Banglapedia, 2003). Thus, the region experiences two extremities that clearly contrast with the climatic conditions of the rest of the country. A dryness study of Bangladesh, carried out using the De Martonne aridity index (Figure 2a) and the Thornthwaite precipitation effectiveness index (Figure 2b) methods (Essenwanger, 2001) from climatic data for the 30 years 1970–1999 available at 50 meteorological stations situated in and around Bangladesh, shows that western side of Bangladesh can be classified ‘sub-humid’, the central part ‘humid’ and a small part of the northeastern side ‘wet’. The lowest index values obtained Copyright 2007 John Wiley & Sons, Ltd.
DATA AND METHODS Rainfall data for the 39 years 1961–1999 from 12 meteorological stations in the western part of Bangladesh was used to study the characteristics of meteorological drought. The main problem encountered during the study of droughts is missing rainfall data. The methods used to estimate the missing rainfall data and to study drought characteristics are discussed below. Estimation of missing rainfall data Numerous methods for estimating missing data have been described in the literature (Creutin and Obled, 1982; Seo et al., 1990; Kuligowshi and Barros, 1998; Schneider, 2001; Teegavarapu and Chandramouli, 2005). In the present study, a feedforward artificial neural network (ANN) approach similar to that proposed by Teegavarapu and Chandramouli (2005) is used for the estimation of missing rainfall data. ANNs are computer models that mimic the structure and functioning of the human brain, and are known for their ability to generalize well on a wide variety of problems and are well suited to prediction applications (Bishop, 1995). Unlike many statistical methods, ANN models do not make dependency assumptions among input variables and can solve multivariate problems with nonlinear relationships among input variables. The efficiency of ANN models does not depend on the density of measuring stations, rather on the number of stations used for the estimation of missing data (Teegavarapu and Chandramouli, 2005). As the density of rain gauges in the study area is low and ANNs are supposed to be suited to any distribution of rainfall stations, the method is used in this paper for the estimation of missing rainfall data. The missing rainfall data is random in most stations, however, continuous missing data for several years is also evident at some stations. The percentage of missing rainfall data varies between 6% and 22% from station to station, except one station (Khepupara), where about 39% of the data is missing. The average level of missing rainfall data in the study area is 14%. Although the performance of ANNs improves with increasing percentage of training data, studies have shown that training with 60% of the total data can reliably estimate unknown data (Teegavarapu and Chandramouli, 2005). Therefore, it can be assumed that the ANN model estimated missing data in the present study with acceptable accuracy. Hydrol. Process. 22, 2235– 2247 (2008) DOI: 10.1002/hyp
DROUGHTS OF BANGLADESH
2237
Figure 1. (a) Spatial distribution of annual mean rainfall over Bangladesh; (b) monthly distribution of rainfall in the western part of Bangladesh
The topology of the ANN used for the estimation of missing rainfall data is 6 : 4 : 1, as shown in Figure 4. The topology was selected using a trial and error procedure. The input neurons use values from six neighbouring stations around the station of interest and the output neuron of the ANN provides the missing value at the station of interest. Neural network training is done using a supervised back-propagation training algorithm (Rumelhart and Mclelland, 1986; Haykin, 1994). The choice of learning rate, momentum factor and activation function for the ANN determines the rate and reliability of the training of the network. In the present case, a learning rate of 0Ð1 and momentum factor of 0Ð4 was used. These factors were obtained by a trial and error method (Haykin, 1994). A gradient descent technique was used to adopt weights in the ANN structure to minimize the mean squared difference between the ANN Copyright 2007 John Wiley & Sons, Ltd.
output and the desired output. In the hidden and output layers, a sigmoidal activation function was used to model the transformation of values across the layers. After computing the missing rainfall data, a geospatial database of rainfall time series is developed within a GIS by following the concept proposed by Goodall et al. (2004). Calculation of standardized precipitation index The standardized precipitation index (SPI, Mckee et al., 1993) is a widely used drought index based on the probability of precipitation for multiple time scales, e.g. 1-, 3-, 6-, 9-, 12-, 18- and 24-month. It provides a comparison of the precipitation over a specific period with the precipitation totals from the same period for all the years included in the historical record. For example, a 3-month SPI at the end of May compares the March-April-May precipitation total in that particular year with the March Hydrol. Process. 22, 2235– 2247 (2008) DOI: 10.1002/hyp
2238
S. SHAHID
Figure 2. Aridity maps obtained by (a) De Martonne aridity index; and (b) Thornthwaite precipitation effectiveness index methods from rainfall data for the 30 years 1970– 1999
to May precipitation totals of all the years. Consequently, it facilitates the temporal analysis of drought phenomena. To compute SPI, historic rainfall data at each station are fitted to a gamma probability distribution function:
ˇD
x ˛
where:
lnx
A D lnx
gx D
1 ˛1 ex/ˇ ˛ ˇ ˛
for
x>0
where ˛ > 0 is a shape parameter, ˇ > 0 is a scale parameter, x > 0 is the amount of precipitation, and ˛ defines the gamma function. The maximum likelihood solutions are used to optimally estimate the gamma distribution parameters, ˛ and ˇ for each station and for each time scale: 1 ˛D 4A
1C
4A 1C 3
Copyright 2007 John Wiley & Sons, Ltd.
n and n D number of precipitation observations. This allows the rainfall distribution at the station to be effectively represented by a mathematical cumulative probability function given by: x x 1 gxdx D ˛ x ˛1 ex/ˇ dx Gx D ˇ ˛ 0 0 Since the gamma function is undefined for x D 0 and a precipitation distribution may contain zeros, the cumulative probability becomes: Hx D q C 1 qGx Hydrol. Process. 22, 2235– 2247 (2008) DOI: 10.1002/hyp
2239
DROUGHTS OF BANGLADESH
almost once per decade. Details of the SPI algorithm can be found in Guttman (1998; 1999), McKee et al. (1993; 1995) and Hayes et al. (1999). Rainfall reliability To compute rainfall reliability, the coefficient of rainfall variation (CV) in percentage is used, CV D 100 ð υ/R where υ D standard deviation, and R D arithmetic mean of rainfall (mm). Spatial interpolation
Figure 3. Study area and the location of meteorological stations
For the mapping of the spatial extent of rainfall and droughts from point data, a Kriging interpolation method is used. Geostatistical analysis tool of ArcMap 9Ð1 (ESRI, 2004) is used for this purpose. Kriging is a stochastic interpolation method (Journel and Huijbregts, 1981; Isaaks and Srivastava, 1989), which is widely recognized as a standard approach for surface interpolation based on scalar measurements at different points. Studies showed that Kriging gives better global predictions than other methods (van Beers and Kleijnen, 2004). However, Kriging is an optimal surface interpolation method based on spatially dependent variance, which is generally expressed as a semi-variogram. Surface interpolation using Kriging depends on the selected semi-variogram model, and the semi-variogram must be fitted with a mathematical function or model. Depending on the shape of semi-variograms, different models are used in the present study for their fitting.
RESULTS AND DISCUSSION
Figure 4. Topology of artificial neural network used for the estimation of missing rainfall data
where, q is the probability of a zero. The cumulative probability Hx is then transformed to the standard normal distribution to yield the SPI (McKee et al., 1993). As the precipitation rate is fitted to a gamma distribution for different time scales for each month of the year, the resulting function represents the cumulative probability of a rainfall event for a station for a given month of the dataset and at different time scales of interest. This allows one to establish classification values for SPI. McKee et al. (1993) classified drought severity according to SPI values as given in Table I. An SPI of 2 or more represents a very severe drought, and happens about 2Ð3% of the time or about once in every fifty years. An SPI between 1Ð5 and 1Ð99 represents a severe drought, and happens about 4Ð4% of the time or once in every 25 years. An SPI between 1Ð0 and 1Ð49 represents a moderate drought, and happens about 9Ð2% of the time or Copyright 2007 John Wiley & Sons, Ltd.
The occurrence of droughts in the study area is identified from SPI time series of multiple-time steps. In the present study, SPI for 3- and 6-months time steps are computed to study the characteristics of drought in short and medium time periods. The 3-month SPI is used to describe the pre-monsoon drought, while the 6-month SPI is used to characterize seasonal droughts that occur due to rainfall deficit in monsoon and non-monsoon months. Temporal and spatial distribution of drought The regional SPI time series are calculated by a Thiessen polygon method for 3 and 6-months time steps, and are shown in Figure 5a and b, respectively. Major Table I. Drought categories defined for SPI values SPI value 0 to 0Ð99 1Ð00 to 1Ð49 1Ð50 to 1Ð99 2Ð00 and less
Drought category
Probability of occurrence (%)
Near normal or mild drought Moderate drought Severe drought Extreme drought
34Ð1 9Ð2 4Ð4 2Ð3
Hydrol. Process. 22, 2235– 2247 (2008) DOI: 10.1002/hyp
2240
S. SHAHID
Figure 5. Arial values of SPI for (a) 3-month and (b) 6-month time steps
droughts identified from the areal SPI time series are in the years of 1963, 1966, 1968, 1973, 1977, 1979, 1982, 1989, 1992 and 1994–1995. Spatial extension of the 3-month SPI at the end of May and 6-month SPI at the end of November for the four worst drought years in Bangladesh after independence in 1971 is shown in Figures 6 and 7, respectively. The 3-month SPI calculated for May uses the precipitation total for March, April and May while the 6-month SPI calculated for November uses the precipitation total for June to November. The 3-month SPI shows a pre-monsoon drought and the 6-month SPI at the end of November shows a seasonal monsoon drought. Figure 6 shows that in 1982, 62% of the study area was affected by drought, among which, 12% was affected by severe droughts and 9% was by very severe pre-monsoon drought. In 1989 and 1992, the whole study area was affected by drought. About 78% of the area in 1989 and 26% of the area in 1992 was affected by severe drought. In 1995, almost 95% of the area was affected by drought, with 49% of the area experiencing severe drought and 22% experiencing very severe pre-monsoon drought. The spatial extent of the 6-month SPI (Figure 7) shows that in 1982, almost 44% of the study area was affected by drought, with almost 21% of the area affected by severe drought (SPI
