You are on page 1of 24

Mitig Adapt Strateg Glob Change

DOI 10.1007/s11027-013-9464-0

Artificial neural networks models for predicting effective
drought index: Factoring effects of rainfall variability
Muthoni Masinde

Received: 9 February 2013 / Accepted: 26 March 2013
# Springer Science+Business Media Dordrecht 2013

Abstract Though most factors that trigger droughts cannot be prevented, accurate, relevant
and timely forecasts can be used to mitigate their impacts. Drought forecasts must define the
droughts severity, onset, cessation, duration and spatial distribution. Given the high probability of droughts occurrence in Kenya, her heavy reliance on rain-fed agriculture and lack of
effective drought mitigation strategies, the country is highly vulnerable to impacts of
droughts. Current drought forecasting approaches used in Kenya are not able to provide
short and long term forecasts and they fall short of providing the severity of the drought. In
this paper, a combination of Artificial Neural Networks and Effective Drought Index is
presented as a potential candidate for addressing these drawbacks. This is demonstrated
using forecasting models that were built using weather data for thirty years for four weather
stations (representing 3 agro-ecological zones) in Kenya. Experiments varying various
input/output combinations were carried out and drought forecasting network models were
implemented in Matrix Laboratory’s (MATLAB) Neural Network Toolbox. The models
incorporate forecasted rainfall values in order to mitigate for unexpected extreme climate
variations. With accuracies as high as 98 %, the solution is a great enhancement to the
solutions currently in use in Kenya.
Keywords Drought Forecasts . Artificial Neural Networks(ANNs) . Effective Drought
Index(EDI) . Available Water Resource Index(AWRI) . Rainfall Variations . Kenya

1 Introduction
Droughts are among the most expensive disasters in the world and their negative impacts
span economic, social and environmental aspects of the affected society. It is for this reason
M. Masinde (*)
Department of Information Technology, Central University of Technology, Private Bag X20539,
Bloemfontein, 9300, South Africa
M. Masinde

Mitig Adapt Strateg Glob Change

that research on drought has attracted a lot of interests from environmentalists, ecologists,
hydrologists, meteorologists, geologists, economists and agricultural scientists (Mishra and
Desai 2006). In Byun and Wilhite (1999), literature on droughts is classified under 4
categories; (1) causes of droughts (atmospheric science); (2) frequency and severity of
droughts in a given region; (3) impacts of droughts; and (4) response, mitigation and
preparedness for droughts; this paper focuses on some aspects of the latter. Natural disasters
triggered by critical climate extremes especially droughts continue to affect millions of
people in Africa. According to the World Disasters Report of 2011, Africa contributed over
50 % of the droughts that occurred in the world between 2001 and 2011 (Armstrong et al.
Kenya has consistently contributed the highest number of people affected by natural
disasters in the Continent (Deely et al. 2010). Droughts are heavily felt in Kenya because of
three factors: (1) regularity – droughts are so regular in Kenya; they were for instance
experienced in 50 % of the years between 1980 and 2008; (2) like most other developing
countries of Africa, Kenya’s rain-fed agriculture is the backbone of the economy. Over 80 %
of Kenyans rely on the agricultural sector that contributes 24 % of the Gross Domestic
Product (GDP) (Mwagore 2002). The sector is highly sensitive to droughts; and (3) inability
of Kenya’s population to prepare and adapt to droughts. Being a developing country with
46 % of the rural population living below the poverty line, the government’s priority list is
filled with items such as providing basic education, implementing democratic constitution,
peace initiatives, providing basic healthcare, and so on. This leaves no room for developing
drought mitigation strategies. Two examples of recent droughts that affected Kenya include:
the 1999–2002 drought in western and central Kenya that affected 23 million people (WHO
Collaborating Centre for Research on the Epidemiology of Disasters (CRED) 2012) and the
2005–2006 in the northern and eastern regions of Kenya that left over 70 % of the livestock
dead (CERF 2009).
Droughts are quantified using indices such as the Effective Drought Index (EDI). EDI is able
to quantify droughts in absolute terms and also provide answers to: (1) the when, (2) the how
long (onset to cessation) as well as (3) the severity of droughts/floods (Byun and Wilhite 1999;
Mishra and Singn 2010 ). EDI quantifies droughts in terms of droughts classes composed of
positive and negative real values; for example, −2.50 indicates extreme drought, +3.28 indicates
extreme floods and 0.98 indicates close to normal wetness. EDI further qualifies droughts by
providing Available Water Resource Index (AWRI) that can for example reliably inform a
farmer of the amount of the available soil moisture at any given day.
Though the well-developed drought indices such as EDI perform very well in mapping
droughts in spatial and timescale dimensions, they only detect the events already happening
(Masinde and Bagula 2011). In this paper, Artificial Neural Networks (ANNs) play the role
of forecasting/predicting future droughts. The numerical approaches that are used in Kenya
produce forecasts with low accuracies; 70 % at best (Mutua 2011). With such forecasts, it
becomes difficult to make sound policy decisions to mitigate for droughts.
The work presented in this paper is part of a larger project whose main objective was to
develop an effective, sustainable and relevant ‘home-grown’ Drought Early Warning System
(DEWS) for the Sub-Saharan Africa by making use of Artificial Intelligence technologies to
integrate: (1) indigenous knowledge on droughts; (2) scientific weather forecasts; (3)
Wireless Sensor Networks (WSNs); and (4) mobile phones (Masinde and Bagula 2012).
The main component of this DEWS is ‘Drought Monitoring and Prediction’. While the
evaluation of the strengths of EDI as a tool for monitoring droughts in terms of their severity,
onset, cessation and probability of their occurrence has been described by Masinde and
Bagula (2012), this paper tackles the Drought Prediction sub-component of DEWS. It

The latter are mainly issued two times a year a few weeks before each of the two main rain seasons. (3) Seven-day forecast. Kenya contributed a whopping 32. 2 Drought forecasting in Kenya Droughts are the most common disasters in Kenya. (2) March-April-May (MAM). Apart from the normal meteorological observations.go. Currently. the amount of rain and moisture index in each of the above season is used to classify Kenya under 14 climate and agro-ecological zones (http:// www. The data used in this research was obtained from this Department. no research known to the author has attempted to model the micro (daily) level of details like it has been done in this paper.meteo. and (5) Seasonal forecast. All this data is stored in semi-automated formats at the Department’s Head Quarters in Nairobi. They run 3 main types of stations that are currently managed by the Climatological Section of the Department (http:// www. (2) 62 temperature stations and (3) 27 synoptic stations. (2) Four-day forecast. When they occur. droughts affect more that 25 % of Kenya’s population not mentioning the ripple effects such as inadequate hydro-electric power supply. The data is available to interested stakeholders and on request. Such disasters cannot be avoided but can be managed through effective early warning systems. they result from variations in weather/ climate. increased commodity prices and loss of jobs just to mention a few. In the period 1999 to (1) 700 rainfall stations. MAM is the main monitoring of climatic/weather variations in Kenya is the mandate of the Kenya Meteorological Department (KMD). The Meteorological Department uses the data collected to provide five main types of forecasts: (1) Daily forecast for main cities/ towns in Kenya. Kenya was among the countries at the Horn of Africa that experienced a devastating drought that was described as “the worst drought in 60 years” (http:// www. The models have been tested using historical weather data from 4 weather stations in Kenya. February 2010 and February 2011. Among other factors (not part of this research work). The Department is handicapped in the sense that all they have are sparsely distributed weather stations. Forecasted rainfall values are used as input into the ANNs models to take care of unexpected extreme rainfall variations patterns. Sometimes the rainfall may occur in the period June-July-August (JJA).ke).co. . February 2009. The Country has two main rainy seasons: (1) October-NovemberDecember (OND).go. other observations by the Agrometeorological Section include: soil temperature. the correlation among EDI values of similar calendar months was utilised. given EDI values for February 2008.meteo. data is remitted from these stations every 10 accessed on 10 October 2012). but there are occasional floods too.Mitig Adapt Strateg Glob Change describes drought-forecasting component that combines the use of Artificial Neural Networks (ANNs) and Effective Drought Index (EDI) to forecast drought for periods ranging from one day to four years with accuracies ranging from 98 % to 75 %. In designing the long-term category of network models.85 % of people affected by natural disasters in the Africa (Collins et al. radiation. one can forecast EDI value for February 2013. For instance. This is a phenomenal improvement on the accuracy (most existing solutions’ accuracy is about 70 %) of the forecasts. (4) Monthly forecast. In August 2011. 2009). The Agrometeorological Section on the other hand manages 13 stations related to agriculture. sunshine duration. pan evaporation and potential evapotranspiration. Though the integration of EDI and ANNs has been attempted elsewhere in the world especially in South-East Asia.

Like the PDSI. hence the most critical. To the economist. As a matter of fact. however its major undoing (among others) is the predication granularity. which provided a solution to the inappropriately higher frequency of extreme drought (Wells et al 2004).or below-normal rainfall. the severity of a drought is a function of the drought duration and probability distribution of the drought variable and its autocorrelation structure. these are then compared against the forecasts from neighbouring countries. There are three main categories of drought: meteorological. To the hydrologist. it is not as accurate as required. This way. which is used for water supply monitoring. and (5) AWRI that expresses the actual amount of available water (Wilhite and Glantz 1985). Conversely. meteorological drought is precursor for the other categories of droughts. it means a shortage which affects the established economy” (Palmer 1965). 1993). Palmer came to this conclusion as early as 1965 when he stated “Drought means various things to various people depending on their specific interest. historically. Its weaknesses however range from its complexity to poor applicability (underlying computation is based on the climate of the Midwestern United States). (2) the how long (onset to termination) as well as (3) the severity of droughts/floods. As a disaster. causing the disruption of the water supply to the natural and agricultural ecosystems. PDSI is the best-known index. lakes. (3) self-calibrated PDSI. (2) Palmer Hydrological Drought Index (PHDI). the severity is defined in the form of indices such as the Palmer Drought Severity Index (PDSI) (Palmer 1965). as well as against the forecasts from several global dynamical models. Since then. hydrological and agricultural (Wilhite and Glantz 1985.Mitig Adapt Strateg Glob Change KMD uses statistical and dynamical models to generate seasonal rainfall forecasts for each region in Kenya. issued by Meteorological agencies in Europe and the United States. it suggests below average water levels in the streams. There is no one universally accepted definition of drought yet. 1980). Combining EDI with Artificial Neural Networks(ANNs) leads to a more effective drought forecasting solution described in this paper. which happens to be monthly scale. (4) Normalized Difference Vegetation Index (NDVI) on the other hand takes advantage of the reflective and absorptive characteristics of plants in the red and nearinfrared portions of the electromagnetic spectrum. reservoirs. In meteorological drought. drought corresponds to the failure of the precipitation regime. 3 Drought severity indices and effective drought index Drought is an insidious hazard of nature which according to Elsa et al. According to Panu and Sharma (2002). This makes it difficult for key stakeholders to develop solid strategic plans. attempts have been made to define the term drought. near. and the like. Effective Drought Index (EDI) is able to quantify droughts in absolute terms and also provide answers to: (1) the when. it has been the most commonly implemented. it does not . SPI has its own share of popularity. Meteorological drought is defined as a period of abnormally dry weather sufficiently prolonged for the lack of water/rainfall to cause serious hydrologic imbalance in the affected area (Huschke 1970). Dracup et al. This has led to various variations of the Index such as: (1) the Standardized Precipitation Index (SPI) that measures the deviation of the actual precipitation from the average conditions in a given area (Mckee et al. (2008) qualifies as a hazard because it is a natural accident of unpredictable occurrence but of recognizable recurrence. To the farmer drought means a shortage of moisture in the root zone of his crops. These are then compiled to give probabilistic forecast of rainfall under three categories: above-. One of the shortcomings of KMD’s approach is the fact that their forecasts provide conceptual indications of droughts/floods without giving operational indicators.

For decades.Mitig Adapt Strateg Glob Change reflect daily/weekly patterns. retrieval by content and process control. This function can be altered by varying the values of the connections (weights) between these elements. a series of indices can be calculated to highlight different characteristics of a station’s water resources. clustering. Calculation is therefore made with consideration of the fact that the quantity of rainfall that can be used as a water resource drops gradually over time after the rain has fallen. prediction. These are: Mean Effective Precipitation (MEP). (2) It considers drought continuity. Some desirable features of EDI are: (1) More accurately calculates current level of available water resources. Effective Precipitations (EPs) are used to compute deficiency or surplus of water resources for a particular date and place. ANNs are composed of simple elements (inspired by biological nervous systems) that mimic the brain neurons. Deviation of EP (DEP) and standardised value of DEP (EDI). not just for a limited period. summation for a year which is the most dominant precipitation cycle worldwide. The network function is determined by the connections between elements. (2010) described and compared AI techniques’ performance in modelling environmental system. EP here refers to the summed value of daily precipitation with a time-dependant reduction function. EPi ¼ Xi n¼1  Pn m¼1 Pm n  ð1Þ where Pm is the precipitation of m days before and the index i represents the duration of summation (DS) in days. it makes use of Eq. Chang-Shian et al. it was developed to address weaknesses identified in the existing (at the time) drought indices. The 365 can then be a representative value of the total water resources available or stored for a long time. function approximation. Once the daily EP is computed. Researchers such as Byun and Wilhite (1999) developed the EDI to address SPI shortcomings. pattern recognition techniques. EP2 becomes [P1+P1+P2)/2]. For instance. EDI produced satisfactory results in measuring drought severity and was adopted (and adapted) to analyse 200-year drought climatology of Seoul. EDI is different from the rest of the indices in a number of ways. this is referred to as training of the network. 1 below. ANNs are highly interconnected systems that operate by mimicking the operation of the human brain. and (4) It considers daily water accumulation with a weighting function for time passage. ANNs provide one of the . learning and reasoning to solve problems. (3) It is computed using precipitation alone. it can therefore diagnose prolonged droughts that continue for several years. Intelligent Agents and Artificial Neural Networks (ANNs). 2009) and to analyse droughts in Kenya (Masinde and Bagula 2011) 4 Artificial neural networks’ in forecasting droughts Artificial Intelligence (AI) refers to the computing paradigm that aims to develop solutions that mimic human perception. regression models and autoregressive moving average (ARMA) and autoregressive integrated moving average (ARIMA) models for statistical time series have been used for drought forecasting. that is. optimisation. then m varies from 1 to 2. if i=2. they singled out ANNs approach as being excellent in solving data-intensive and multivariable problems with unclear mapping rules. they are designed to operate in parallel to achieve a goal. Korea (Kim et al. Here i=365 is used. Two AI techniques stand out when it comes to modelling environmental systems such as droughts. ANNs perform seven categories of tasks: pattern classification.

Other projects in which ANNs have been used for predicting droughts/floods include drought prediction in north-east Brazil(Freitas and Billib 1997) and summer flood forecasting in Moravia (Bodri and Cermak 2001). This is the weakness that is addressed by Decision Group BPN (Chang-Shian et al. Though invented in 1943 by Mcculloch and Pitts (1943). 2010). Antonić et al. Besides it being difficult to determine the representativeness of learning data. using BPN in hydraulic modelling is faced with the problem that BPN is a non-linear black box that does not consider/explain the underlying physical process of watershed (Chang-Shian et al. Using a data (precipitation) set for 36 years (1965 to 2001). and Chang-Shian et al. Kung et al. (b) Data set division criteria. results are produced at the output layer(s) (Morid et al. 2001.Mitig Adapt Strateg Glob Change nonlinear non-stationary alternative models for drought forecasting as explained in (Mishra and Desai 2006). it is the coming-on-stage of the Back-Propagation training algorithm for feed-forward ANNs in 1986 that boosted the application of ANNs in several real-life application areas such as drought forecasting (Panu and Sharma 2002). water quality and precipitation estimation (American Society of Civil Engineers (ASCE) Task Committee on Application of Artificial Neural Networks in Hydrology 2000). Most neural network tools such as . Applying both the Recursive Multi-step Neural Network (RMSNN) and the Direct Multi-step Neural Network (DMSNN) approaches. The greatest strength of ANNs. they can be used to solve problems with nonlinear/unknown multivariate and less controlled environments (Bodri and Cermak 2001). Mishra and Desai (2006) observed that most ANNs’ implementations adopted the multi-layer feed-forward neural network and were based on BPN. In most hydrological applications. 1986. validation and testing phases of ANNs. (2001) used seven climatic variables obtained from 127 weather stations in Croatia to develop empirical drought models using BPN. 2010). Maier and Dandy (2000) described five key considerations to successful implementation of ANNs as: (a) ANNs’ performance evaluation criteria: how will the performance of the model be judged? The relevant criterion for this study was ‘accuracy of predicted results’. BPN has three general layers: input. Freitas and Billib 1997. ANNs are also flexible and less assumption-dependent and this has seen them find applications in modelling extremely complex hydrological domains such as rainfallrunoff. 2009). BPN has been applied in modelling animal/plant population growth to predict hydrological/meteorological phenomena such as droughts and floods. and has two processing types: learning/training and recalling. 2006. This also happens to be the most commonly used criterion (Demuth et al. stream flows. deals with how to share the available dataset among the training. At the input layer(s) data are fed to the system while the processing takes place within the hidden layer(s) and finally. 2010). Shin and Salas (2000) used ANNs to quantify the spatial and temporal patterns of meteorological droughts for the south-western region of Colorado. multi-layer perceptron (MLP) neural network model is adopted with feed-forward back propagation (BPN) as the training algorithm (Rumelhart et al. is its ability to weave together various mathematical components capable of tackling very complex physical systems such as droughts. Bodri and Cermak 2001. Mishra and Desai (2006) compared the performance of BPN and ARIMA for different Standard Precipitation Index (SPI) lead-times. Antonić et al. though. 2007). they found that RMSNN performed better than both DMSNN and ARIMA for one-month lead-time and that DMSNN performed better than the other two for a lead-time of four months. hidden and output. ANNs have several advantages to this end.

and how many hidden layers and how many nodes in each of the hidden layers. the mapminmax function implemented in the MATLAB Neural Network Toolkit was applied. (d) Determining model input. Data composed of daily readings for temperature (highest and lowest).Mitig Adapt Strateg Glob Change MATLAB Neural Network Toolkit adopt 70:15:15 percentages for training. 2009). validation and testing respectively (Demuth et al. detailed Confirmatory Single Case Study using data from the fourth weather station was used. Further. These are: Dagoretti. The latter puts all the input values into [−1:1] range. In the this paper. this is purely based on a priori knowledge of causal variables in conjunction with inspections of time series plots of potential inputs and outputs. Embu. From the rainfall values. wind (direction and speed). (c) Data pre-processing with the aim of ensuring some form of standardisation. In Kenya. (e) Selecting a suitable network architecture which is made up of two aspects: The type of connection (are there loops?) among the nodes. only . The number at the input and output layers is generally determined by the problem domain. Four weather stations were selected to represent four (excluding the rift valley) these regions. central rift valley and the highlands and in a narrow strip along the Indian Ocean coast. relative humidity (at 06Z and at 12Z). 5 Materials and methods 5. In the last phase of the case study design. This in turn ensured that all variables involved had equal representation during the training. this is referred to as the geometry of the network. these areas are found in western. This data set division criterion was adopted in this paper. First it was used to determine if EDI could be used to qualify/quantify droughts and two.1 Case study with experimental design Yin (1993) defines case study approach as an empirical enquiry that ‘investigates a contemporary phenomenon within its real life context. when the boundaries between phenomenon and context are not clearly evident and in which multiple sources of evidence are used’. The design of the case studies followed a three-phase experiment made up of pilot. Each of the case studies started off with a PilotSingle Case Study of one weather station followed by recursive Exploratory Multiple Case Studies of two weather stations. atmospheric pressure and rainfall was sourced from KMD. In this paper. This led to the variation of Case Study with Experimental Design as shown in the Fig. The case study design was used to analyse historical daily weather data for over 30 years for four weather stations in Kenya. rainfall. to investigate the suitability of ANNs in forecast droughts. In this paper an array of weather parameters such as daily readings of temperature. exploratory and confirmatory experiments. all rainfall data was first subjected to EDI computation and hence. 5. 1.2 Data sample Droughts impacts are more pronounced in highly populated areas whose main economic activity is agricultural production. both daily and monthly EDI and AWRI were computed and extensively used as inputs/outputs of the neural networks. relative humidity and wind direction were considered. it was already standardised/homogenisation. Kakamega and Makindu.

250 mm A Java program was written to covert the original format to formats acceptable by the EDI Fortan program: A Fortan program developed by Byun and Wilhite (1999) and available at (http://atmos. this translated into 4×365×31 records (Table 1). WMO (World Meteorological Organisation) The values of the annual average rainfall in this table were calculated (by the author) from daily rainfall values for each station for years 1979 to 2009 in Kenya.133 m 1. 1 The 3-steps of the case study with experimental design daily precipitation data for years 1979 to 2009 from these stations was used.493 m 2.913 mm 593 mm Ecological-zone Geographical region 11 Highland (Nairobi) 11 Central 13 Western 10 Coast ICAO (International Civil Aviation Organisation). 1 Table 1 Geo-Data of the weather stations studied Name Dagoretti Embu Kakamega Makindu WMO number 63741 63720 63687 63766 ICAO HKNC HKEM HKKG HKMU Year opened 1954 1975 1957 1904 Latitude 01 18S 00 30S 00 17 N 2 17S Longitude 36 45E 37 27E 34 47E 37 50E Elevation 1.798 m 1. 2. (2) Data 5. . were carried out as per the flow chart shown in Fig.Mitig Adapt Strateg Glob Change ANNs Design Experiments Pilot Single-CaseStudy Eploratory MultipleCase-Studies Confirmatory SingleCase-Study • Data for 1 weather Station (Embu) • Formulate theory • Data for 2 weather stations (Dagoretti and Makindu) • Test theory • Data for 1 weather station (Kakamega) • Confirm and extend theory Fig.023.000 m Annual average rainfall 1.php (lastly accessed on 10 October 2012) was then used to compute EDI values.3 EDI calculation Two Phases: (1) Data cleaning. This computer program uses Eq.pknu.

2 Flow chart showing data cleaning and analysis phases for the Kenya’s case study Phase I: Data Cleaning Weather Data in text files. reformat the data to Day. Source: KMD Extract rainfall data for 4 stations Using a Java Program. Month.Mitig Adapt Strateg Glob Change Fig. an output file containing the EDI and AWRI values was generated. Some of these can be found in (Masinde and Bagula 2011). Year and Station Identify and remove data gaps Formatted Data EDI Program Computed Monthly & Daily EDI Excel Charts Generator Phase II: Data Analysis Weather EDI Web-Based Decision Support Database System (MySQL) (presented in Section 3) to compute daily/monthly EDIs and outputs them into text files. . The raw data that was acquired from the Kenya KMD was first cleaned and formatted to suit the pre-defined format (generating an input file to the EDI program) for each of the 4 weather stations. For each input file. These were then exported and stored in a My Structured Query Language (MySQL) database from where various data analysis reports were generated.

EDI and AWRI.3 Data division criteria and network training For each data set. validation and testing respectively was used.6 146.2 Network architecture One hidden layer with 20 neurons was selected. precipitation (used to derive EDI and AWRI) is in itself a function of these parameters. an attempt to include temperature (and other parameters) led to large errors because among other reasons. Splitting of the data during training was done randomly using dividerand() function. precipitation is itself a function of so many other linear and nonlinear functions (Weichert and Burger 1998).66 … 1981-01-31 0. The ANNs described in this work utilised only EDI. This was selected because it is the fastest backpropagation algorithm in the MATLAN ANNs toolbox that was used in this reasearch (Demuth et al. the number of hidden layers was varied between 10 and 40 at an interval of 5. 5. Further. default data preprocessing function.1 0.4. for each station.7 −0. To normalize the input values.1 Overview The ANNs network models were developed using the output from the EDI computation described above. During the network building.0 208.20 .96 … … … … 2009-12-31 26.Mitig Adapt Strateg Glob Change 5.4.0 … 117. For example. hence the decision to use the default value 20. during the training of the ANNs for predicting EDI value for Dagoretti for 7-day lead-time. it emerged that the network with 20 hidden layers had the best performance. 5.4.4 Artificial neural networks’ models development 5. The Mean Square Error (MSE) and Regression(R) values were used to rank them. the ratio of 70 %:15 %:15 % for training. 2009) Table 2 Format of daily EDI output Date Total Precipitation AWRI EDI 1980-01-01 … 0. The format of the daily EDI/AWRI file is as shown in Table 2 while the monthly one had the format: [Month/Year|Precipitation|AWRI|EDI]. The network models were created by combing various Inputs/Targets. Training was carried out using Levenberg-Marquardt backpropagation (trainlm) algorithm. AWRI and precipitation values for 30 years (1980 to 2009).0 … −0. some experiments of varying the number of neurons were carried out but the value 20 outperformed them. mapminmax that maps the range of input values to the range [−1 1] was used. the data set contained daily or monthly readings for precipitation.

Pn−3. forecast the EDI value for the next day. Heterogeneous inputs. En−1. En−1. (i) Homogeneous inputs. En−2. En−2. En−4. it is the measure of the correlation between the inputs and the outputs. forecast the EDI value for the next day. 5. forecast the EDI and AWRI values for the next day.76521) for the same station. En−2. En−1. the higher the value. En−3. homogeneous output example: & (iii) Network 14: En+1 =f((Pn.4. In the case of MSE: the smaller the value the better the network. Wn−2. Wn−3. En−5)):Given Precipitation. En−5) (Wn. the outputs/targets for Networks 15 to 21 are those of AWRI whose mean for Embu weather station is 194. In order to get the actual implications of the networks. RMSE and the percentage of the RMSE were computed. Pn−2. Wn−1. Pn−3. the higher the rank. En−1. Pn−2. Pn−5)):Given EDI. heterogeneous output example: & (iv) Network 2: En+1.58 while for Networks 7 to 14 are for EDI values that ranges between −2. Pn−4. Wn−3. forecast the EDI and AWRI values the next day.4 Network geometry The networks were first designed to predict future EDI and/or AWRI values for a given combination of EDI (E). Wn−5) ((Pn.Wn+1 =f((En.Wn+1 =f(En.Mitig Adapt Strateg Glob Change 5. The networks were categorized as follows: & & & 7 Networks with 2(EDI and AWRI) outputs 7 Networks with EDI as the output 7 Networks with AWRI as the output A simple Java program was then created to convert input files in the format on Table 2 to the following inputs-targets mapping. precipitation (P) and AWRI (W) for 6 previous days/months. Wn−2. En−5):Given EDI values for the last 6 days.5 ANNs performance analysis Both R and MSE values were used to select the networks models with the best performance. homogeneous output example: & (ii) Network 8: En+1 =f(En. En−2.32 (an absolute mean of ±0. AWRI and Precipitation values for the last 6 days. En−3. Wn−4. En−4. EDI and AWRI values for the last 6 days. For Regression Analysis. Pn−4. En−4. Pn−5 ) (En. Pn−1. However. Homogeneous inputs. Pn−1. Wn−1. . Wn−4. Root Mean Square Error (RMSE) and Percentage RMSE were then used to determine the implications of the errors on the resulting forecasts. it refers to the average squared difference between outputs and targets. the MSE value is bound to be directly proportional to the size of the outputs values.28 to 4. En−3. Heterogeneous inputs. heterogeneous output example: & Network7: En+1.4. En−5) (Wn. En−5):Given EDI values for the last 6 days. En−3. For example. En−4.

Dagoretti and Makindu weather stations. 20 and 21.Experiments results From the results in Table 3 and the analysis in above. data analysis point towards more frequent and pronounced rainfall variations as well as a rise in temperatures in Kenya. 9. and 18 not perform well. 19. Though Network14 had the best performances in Category 2. Secondly. the one-input Network 17 had the best performance. however. What about creating two different networks that use all the three inputs (P. These are Networks 7.Mitig Adapt Strateg Glob Change 6 Results and discussion 6. Forecasted (by KMD) rainfall values were incorporated into the ANNs models to address such situations. The networks with average (Rank 1 to 3) performance and that have Precipitation as an input were selected for further investigation. 14. 20 and 21) were identified for further analysis during the Exploratory Experiments.2 Pilot phase . 8 and 15) had the worst performance. E and W) to predict both EDI and AWRI respectively? This is achieved in Network 14 and 21.99029).1 Rainfall and temperature variability in Kenya As shown in Figs. 8 and 15 that have precipitation alone as input had the worst performance. 10 and 16 give poor performance. This made Networks 2. 7. For Embu. Network 21 was the third after Network 19 and 17 in Category 3. both Networks 9 and 11 had equal values (0. 6. E and W) and gives both EDI and AWRI as outputs is ideal because it ensures that only one network is needed to solve the current problem. a table similar to Table 4 was generated. 11. This scenario is represented in Networks 7 which had the best performance in Category 1 of the networks. the two best performing networks in each of these categories were selected. . This is an indication that there is no any form of relationship between the two. 14. the rainfall variability being witnessed in Kenya and elsewhere in the world would lead to larger errors in the predicted drought values. For each weather station. combining EDI and Precipitation to predict AWRI or AWRI with Precipitation to predict EDI did not work either. 11. Network 9 had the lowest value (0. these (1. Networks 1. For Dagoretti. There is evidence of extreme within-season variability especially of the rains on-set leading to unseasonal rainfall patterns. 12 and 14) as output. Finally. the 10 Neural Networks were trained and their performances analysed. by selecting a network that uses all the 3 inputs (P. An attempt to use either EDI to compute AWRI and vice versa did not yield desirable results.01887) of MSE in the category with EDI (Networks 9. 6. 3 and 4. 10 Neural Networks models (5. in order to exhaustively investigate the best networks for building bi-network (two network. For AWRI category. 3. Using data for Embu. data for one (Embu) weather station was used. 13. for EDI and AWRI respectively). 11. An ideal Network would be the one that predicts EDI and/or AWRI given precipitation only. results of the runs are tabulated in Table 3. 12. in the category of Networks for computing EDI(Networks 9. Considering that the design and operation of ANNs is heavily dependent on learning historical patterns and using these to predict the future. For the R value.3 Exploratory phase I . Finally.Experiments results In order to select the best network to use. This made Networks 4. 17. 11. 6.

Mitig Adapt Strateg Glob Change Fig. 3 Temperature variability in selected weather stations in Kenya .

4 Rainfall variability in selected weather stations in Kenya .Mitig Adapt Strateg Glob Change Fig.

02 0.78 0.01 861.703. the value of V1 is preferred because it is the measure of the network’s ability to carry out the future forecasting (classify dataset it has not encountered before).94 0.25 3 0.38 6 0.85 0.13 59.32 7 Network9 N Y N Y N 0.63 5 0.00 32.67 78.358. Network7 has the lowest error in the Category 1 while Network 14 and Network 19 have the lowest error in Category 2 and 3 respectively) Though the ranking of the networks was based on the average of the performance (MSE and R) of the three data sets (Tr.33 45.99 0.447.74 42.82 0.38 0.78 0.60 72.69 5 0. Network 17 was selected to represent this category.99 0.85 0.45 35.75 6 1 0.99412). these networks had weaker Regression values for all the .62 7 Network2 N Y N Y Y 1.74 0.05 7 0. While for Makindu.97 0.99361(0.82 0. Network 17 had the lowest MSE(64.99307 to 0.24 40.50 0.4371 to 47.17 79.85 7 0.99 0. Similarly. E Effective Drought Index.669.44 97.99 1 Network18 Y Y N N Y 1.365.99068).76 0.15 2 0. Y yes (included in the input/targets).99 1 Network13 Network14 Y Y N Y Y Y Y Y N N 0.87 118.57 77.287.33 0.33 4.566.48 2 0.85 5 Network19 N Y Y N Y 84. Vl and Ts).68 0.05 %)) were too low considering that extra input(s) values are needed.Mitig Adapt Strateg Glob Change Table 3 Performance of the pilot phase neural networks Network Inputs (I) Targets (T) Mean Square Error (MSE) P E Tr Vl Ts E W W Regression (R) Rn Tr Vl Ts Rn Category 1: Two outputs (EDI and AWRI) Network1 Y N N Y Y 2.00 0.17 80.98 0.99 0.02 0.50 4 0.76 0.82 0.02 5 1 0.02 0.36 0.32 0.743.99 1 Network12 N Y Y Y N 0.02 0.02 0.02 2 0.0179) as well as the strongest R(0.97 4 1 Category 2: Output EDI only Network8 Y N N Y N 0.62 885.39 4 3 0. Ts Testing.05 1.7839 (7 %) for MSE and 0. Network 9 performed better among the networks with EDI as output but unlike the case of Embu and Dagoretti.58 0.75 3 Network4 Y Y N Y Y 795.99 1.99 0.97 0.78 0.09 4.00 0.28 2.77 6 Network17 N N Y N Y 87.834.672.99 0.57 91. W Available Water Resource Index (AWRI).77 0.96 0.27 50.84 0.96 0.31 0.99 0.655.97 1 Network6 Network7 Y Y N Y Y Y Y Y Y Y 39.31 0.02 4 0.46 7 Network16 N Y N N Y 2.99 0.99 0.90 2. 12 and 14).03 0. Network 9 had the lowest MSE (0.99 1 Network20 Network21 Y Y N Y Y Y N N Y Y 77.78 6 Network11 Y Y N Y N 0.19 47.847.82 0.77 0.9521) and highest R (0.02 0.66 51. Network 21(3 inputs) had the least MSE and Network 19(has EDI as the second input) had the highest value of R.340.81 80.99 0. Vl Validation.169.99 1 1 P precipitation.99 0. in the category of networks that compute AWRI.76 0. for example.96 5 Network5 N Y Y Y Y 43.99 1.19 2.99 1 Network10 N N Y Y N 0.02 3 0.99 0.02 0.35 0.63 7 0. Networks 5 and 7 have both EDI and AWRI as output and therefore difficult to calculate the implications of MSE.95 6 Network3 N N Y Y Y 41. Rn Rank (the Rank of the Network by performance (MSE and R).01 1.17 1.63 6 0.56 2.50 0.358.05 1 0.97 0. N No (not included in the inputs/targets). Besides.10 1.99 5 1 Category 3: Output AWRI only Network15 Y N N N Y 4.99 0.94 0. Tr Training. Since the improvements (51.

6 past values are input into the neural network.6 ±8. Network 9 generated values that are within acceptable error-range.1539 35. the daily EDI/AWRI values for years 1980 to 2009 were used.99311 0.4754 27.6619 81. in order to forecast future values.0240 0. 2.6930 0.99021 Network17 Network19 68.8056 75. E19 . E17 . In case of AWRI. E is replaced with W.99359 weather stations and were therefore not considered in building the final prediction models. 5. With MSE values ranging from 3 % to 7 % on EDI.99334 0. 6. neural networks were created to forecast future values of EDI (Network 9) and AWRI (Network 17) for the short-term.98751 0.0199 0.3021 34.99211 0.Dagoretti Network Mean Square Error Regression Tr Vl Ts Average Tr Vl Ts Average Network5 34.32 (Extreme Floods) with an error of 3 % will yield 4.2966 68.6352 72. En−4. as such. an EDI value 4. For example to forecast the value of EDI 5-days Lead-Time from 21st January (26th January). only the two single-input Networks (Network 9 for EDI and Network 17 for AWRI) were considered in the next phases.7377 75.99442 0.98408 0.99385 0. For each of the 3 weather stations.6 would yield 194.99068 0.99503 0. input files for each of the neural networks (representing forecast durations above) were created.99457 0.99126 0. Similarly.1 Short-Term forecasts (a) D-Days-Lead-Time Forecast The lead-time (in days) considered included: 1. En−5) where En is the EDI value for day n.8559 33.99816 0.99382 0. the following expression applies: Forecast½E21þ5  ¼ f ðE21 .99029 Network11 0.1 6.99271 0.98864 0.0165 0.5308 68.99373 0.32±0.8462 0.4.0221 0.99347 Network21 68.99382 Network20 64.0188 0.98865 0.98892 Network9 0.1657 78.0178 0. 4. In all the networks except 12b and 14b.01899 0. the following expression was used: Forecast[En+d]=f(En.0151 0. including extra inputs to calculate either EDI or AWRI values did not have significant improvements on the Networks performance.99018 0.98391 Network7 35. En−2.99237 0.98955 0. That is.97997 0. 12 and 14.1539 33.0174 0.Mitig Adapt Strateg Glob Change Table 4 Performance of the exploratory phase’s neural networks .0188 0. the latter ranges from 1 to 6. Further. 7.3141 60.Forecast models results Using the results above. For example.0166 0. to forecast EDI or AWRI value for d-days from n.13. 6 number of values (EDI or AWRI) were included as inputs to the .0195 0.99110 0.99378 0. En−1.1388 0.0254 0.4 Exploratory phase II .7853 53. Given a specific day n.99138 0. E20 .99332 0.99001 0. E18 .98315 0.98986 Network14 0.5809 66.0200 0.4546 59.5366 0. 3.0149 0.9927 67.98793 0.7476 38. E16 Þ For the purpose of training and testing the neural networks.99159 0. En−3.99029 Network12 0. Network 17 has MSE values between 4 % and 8 % and an AWRI value of 194.7444 0.01887 0.99399 0.1977 72. medium-term and long-term timescales as follows: 6.98450 0.

Increasing the number of inputs (previous days’ values) did not have a significant improvement in network performance. the DDays-Lead-Time Forecasting neural networks described in (a) above were modified to include forecasted precipitation values for the lead-time considered. the networks in the EDI category (Network 9) displayed excellent performance. the networks had a Fig. with accuracies of 94. Conventionally. Embu and Makindu respectively. increasing number inputs from 6 to 14 resulted in an increase from 13.03 %. E19 .Mitig Adapt Strateg Glob Change neural network models. E20 . Example. For example.25−6. 93. 5 MSE graph for D-Days-Lead-Time for selected weather stations in Kenya .09 % for Dagoretti.52 % for Dagoretti. 5 and 6.13 to 97. to forecast the drought (EDI and AWRI) values for the 5-days-Lead-Time counting from 21st January (26th January).57 to 97.62 % and 96. ANNs-based drought forecast solution would result in poor forecast skill. Embu and Makindu respectively.25 to 99. Further. the anticipated daily precipitation (as forecasted by professional weather forecasting institutions such as the Kenya Meteorological Department) for the 5 days are included as inputs to the networks: Forecast½E21þ5  ¼ f ððE21 . This means that faced with extreme climate variations. the results indicate that the performance of the network declines as the number of days being forecast increases. P25 . for instance.29 % and 9. E17 . P23 . In an attempt to find out if increasing the number of inputs would significantly improve the network performance. (b) D-Days-Lead-Time Forecast with Precipitation In the era of climate change. 15. the number of inputs were increased to 12 and 14 in 12b and 14b respectively.21 to 15. E18 . E16 Þ. ANNs-based forecasting solutions learn from past events/patterns and use this knowledge to forecast future trends. P24 .61 to 13. The results of this network model are shown in Figs. ðP22 . P26 ÞÞ Where Pi represent the forecasted (approximated) precipitation values As shown in Figs. In order to take care of such events. forecasting 3 days ahead gives a better performance that forecasting 14 days head. extreme climate variations may trigger precipitation either during periods that it does not normally rain or lead to precipitation amounts that are below or above the expected normal for the area/region. 7 and 8.55 %.

4. 7 MSE graph for D-Days-Lead-Time (with precipitation) for selected weather stations in Kenya . 6 Regression graph for D-Days-Lead-Time for selected weather stations in Kenya correlation coefficient values between 0. These values have accuracies as low as 70 % and therefore this will slightly affect the neural networks performance. This is the data that was used for the following long-term and short-term forecasts. It must be noted that the impressive performance of the D-Day-Lead-Time Forecast were achieved because the actual historical precipitation values were used as input.Mitig Adapt Strateg Glob Change Fig.2 Medium and long term forecasts Similar to the daily EDI. monthly EDI/AWRI values are computed using monthly precipitation totals.99. 6. The real networks are expected to use values forecasted by meteorological institutions. Fig.98 and 0.

EMonthM. Year nþ1 ¼ f ððWMonthM.7 and 12 months. For example to forecast the value of EDI 5-months Lead-Time from January 2012 (that is forecast for July 2012). the following expression applies:   Forecast½EJun2012  ¼ f EJan2012 . En−5) where En is the EDI value for month n. Year n1 . WMonthM.Year y3 . PYear Yyþ1 Þ Network 17 : WMonthM. ESep2011 . Forecast[En+m]=f(En. WMonthM. EDec2011 . The lead-time units considered were 2. 82. PYear y5 .6. 3.Mitig Adapt Strateg Glob Change Fig. It also requires the annual precipitation totals for these years as well as an approximate annual precipitation value for year 1986 . Year n4 . PYear n1 . WMonthM. the performance of the neural networks decreases as the length of the forecast duration increases. WMonthM.Year n5 . 8 Regression graph for D-Days-Lead-Time(with precipitation) for selected weather stations in Kenya (a) M-Months-Lead-Time Forecast Six inputs (previous months) were selected as the standard input in training neural networks to forecast monthly EDI/AWRI values.4. PYear n3 . Year n . Network 9 : EMonthM.Year yþ1 ¼ f ððEMonthM. PYear n .Year y . En−3. EMonthM. forecast for Y(1. PYear n2 . En−4. En−1. EMonthM. PYear n5 . EMonthM.Year y2 . PYear nþ1 Þ Illustration: to forecast EDI/AWRI values for August 1986. En−2. EAug2011 As shown in Table 5.2. PYear y1 . PYear n4 . ENov2011 .Year y5 . 84 and 85. PYear y3 . The error rates for majority of the networks are above 30 % resulting in forecasts that with below 70 % accuracy. PYear y2 . the latter ranges from 1 to 6. EMonthM. the neural network requires the EDI/AWRI values for August 80. Year n2 .Year y1 . WMonthM. 81.Year y4 . EOct2011 . 5. 83. PYear y4 . Year n3 . (a) Y–Years-Lead-Time Forecasting Using the expression below. PYear y .3 and 4) years were developed and tested for the 3 weather stations.

the D-Days Lead-Time Forecast with precipitation had much higher accuracies of up to 98 %. . D-Days-Lead-Time Forecasting with Precipitation and Y-YearsLead-Time Forecast.999 0.3970 0. E Embu station. EJanuary.750 0.01 24. For instance. 3.27 0. P1998 Þ . August.6713 0.809 1. D Dagoretti Station. M Makindu station.19 37. The ANNs models developed accepts as input a combination of EDI/AWRI and precipitation values and generate a set of EDI/AWRI values representing predicted drought. entry 5 implies that the forecast is for the 5th month given the previous 6 months.950 0.3551 0.5045 0.86 28.730 0.1996. . EJanuary. a 2-Years Lead-Time (2 years in advance) Forecasting model could be used to forecast the 1998 floods a follows   Forecast EJanuary.1993 .945 0.15 34.995 0.4204 0.903 0. For example.59 0.845 0. RMSE Root Mean Square Error – computed by calculating the square root of the MSE.60 0.3907 0. MSE Mean Square Error. For example.997 0.17 33.2593 0.625 0.087 1. the network forecasts the value for May 2012 6.009 0.90 33.2477 12 0.1837 0.717 26.919 34.900 1.74 0.5483 0. For each forecast category.552 0.862 0.6868 0. October and November 2011.5255 7 0. Just like it was the case in Exploratory Phase. EJanuary. given the values for June.1994 .3359 0.973 0.514 0. .84 37.7432 3 0.1992 .15 0. %RMSE Percentage Root Mean Square Error – calculated by computing the percentage of the actual mean values (EDI or AWRI) the RMSE represents Values 2. the following networks were identified for this confirmation: DDays-Lead-Time Forecasting.947 0.895 35. P1996 .855 0. The fact that the ANNs models gave these impressive results on data not encountered before helped to reaffirm that the ANNs can indeed be used to enhance the accuracy of drought forecasts. the already trained networks were retrieved and used to evaluate the respective networks’ input-output combinations created using the Kakamega data set.93 36.791 31.37 0. P1997 .043 1.998 0.556 0.86 36.55 31. ….5 Confirmatory phase – Results of evaluating forecast models From the exploratory phase.746 0.3288 0.932 33. September.658 0.6583 0.846 34.811 0. P1993 . Data set for Kakamega weather station with same dimensions as the data for Dagoretti.758 0.928 0.743 0.4733 0.6686 6 0. P1994 . EJanuary. July. P1995 . For example.12 represent the lead-time of the forecast in months.871 0.1995 .715 0. the results showed consistency in performance.97 0.4995 LT(D) Lead-Time in Days.866 0.801 0.004 27.893 0.1 Evaluation Artificial neural network as implemented in MATLAB’s Neural Network Toolbox was used to develop drought prediction models discussed in this paper.24 35.Mitig Adapt Strateg Glob Change Table 5 Network performance for the M-Months-Lead-Time Forecast – Network 9 LT (D) MSE RMSE %RMSE R D E M D E M D E M D E M 2 0. In all the three categories of forecasts.00 31.868 0. EJanuary. . the D-Days Lead-Time Forecast had accuracies of 88 % to 94 % for EDI and 88 % to 95 % for AWRI. 7 Evaluation and conclusion 7. . Embu and Makindu was used to evaluate the performance of the ANNs models. P1992 .6317 4 0.4862 5 0.1991 . .1998 ¼ f ðEJanuary.75 29. P1991 .42 38.

The models were then tested using weather data for over 30 years for 4 weather stations in Kenya. That is. 9 Predicted versus actual values – sample graph . This was evaluated by developing forecasting models for time-scales ranging from one day to four years. the best performance (MSE=0. 0.9 and above were considered acceptable. 92. The accuracy of this value (Fi) was determined by its closeness to the actual drought value (Ai) for this period (1998). Values for other calendar months are computed in a similar version. MSE was then used to determine the accuracy of the forecasts. Figure 9 is a sample plot of such relationships as generated by the ANNs model for 2-years lead-time for the period October 1997 to September 1999 using data for Makindu weather station. It also exploited the correlation among EDI values of calendar months to forecast annual droughts.267 and 0. Networks with correlation values of 0. the solution is a great enhancement to Fig. given 6 EDI and precipitation values for 6 years and predicted precipitation values for the lead period (1997 and 1998). It also requires the annual precipitation totals for these years as well as an approximate annual precipitation value for year 1997 and 1998. the data for Makindu for Y-Years Lead Time category (see Table 6) had MSE values of 0.037 to 4. This analysis was performed for all models and networks with accuracies of over 70 % were considered acceptable. 0. epoch here refers to the presentation of the entire training set to the neural network. 10.205. The AWRI models for Makindu for the same network category gave values ranging from 3. validation error was used o evaluate the performance of our network models. 95 and 96. Throughout the experiments. In the case of 2-Years lead-time EDI forecast model for Makindu. A value of 2. The ‘best’ performance was retrieved from the MATLAB Artificial Neural Network Toolkit as illustrated in Fig. In order to put the MSE values into perspective. further he Percentage RMSE which was easier to understand than the MSE were computerd.661.270 for 1. In addressing this drawback.2 Conclusion Most of the drought indices applied in forecasting droughts in Kenya fall short of providing the severity of the drought. our solution took care of the effects of unprecedented climate variations.Mitig Adapt Strateg Glob Change To forecast EDI values for January 1998. 3 and 4 years lead times respectively.89 % implied that the network had an accuracy of 97. 2. the ANNs model for 2-years lead-time outputs the EDI values (Fi) for 1998. 94. 7. For instance. By including forecasted precipitation values as inputs to the ANNs models. a combination of ANNs and EDI was adopted. With accuracies raging from 75 % to 98 %. the neural network requires the EDI values for January 91.233.37097) was attained after 16 epochs.11 %. 93.

516 19.8521 0. Embu and Makindu) for 30 (1979– 2009) years was used to further evaluate the 10 networks selected above.03 18. and (7) Y-Years-Lead-Time Forecast.285 0.467 0.224 0. Overall.453 17.552 0.47 13.8527 .10 14.533 0. (3) D-Days-Lead-Time With Precipitation Forecast.294 0.233 0.83 14.02 0.Data for three weather stations (Dagoretti.8682 4 0.Mitig Adapt Strateg Glob Change Fig.520 19. EDI and AWRI values.60 0. From these.205 0. three phases were followed: Phase I: . (2) D-Days-Lead-Time Forecast. (5) M-MonthsLead-Time Forecast (6) Next-Year Forecast.7953 0.542 0. After thorough analysis.8618 0.218 0.8761 0.315 0.87 21. monthly forecasts could still be achieved using Table 6 Network Performance for the Y-Years-Lead-Time Forecast –Network 9 MSE RMSE %RMSE R D E M D E M D E M D E M 1 0.46 0. Phase II: .561 0.473 0.8504 0.305 0.294 0. 2 networks (for forecasting EDI and AWRI respectively) were identified and used to run the following forecasts: (1) Next-Day-Forecast. 10 ANNs evaluation – ‘Best’ MSE sample graph the solutions currently in use in Kenya.8167 0.267 0. In developing the ANNs models.8389 0.17 11.9107 3 0.542 0. Dagoretti and Embu had accuracies of between 76–87 % and 72 %–80 % for EDI and AWRI network categories respectively. (4) Next-Month Forecast.482 19. 10 best performing networks were selected for further investigation during the Phase II.51 21.305 0. the D-Days-Lead-Time With Precipitation Forecast (for up to 7 days Lead-Time) had accuracies levels of over 94 % while the Y-Years-Lead-Time Forecast (for up to 4 years Lead-Time) had over 75 % accuracies.81 0.8516 2 0.270 0. However.552 0.using 30-years (1979–2009) data for Embu Station.8487 0. The Next-Month Forecast had poorer performance. 21 neural networks were created using different combinations of Rainfall (Precipitation).20 21.

Wilhite DA (1999) Objective quantification of drought severity and duration. northeast Brazil.. J Hydrol Eng 5:115–123 Bodri L.H. and Martin. KS L. International Federation of Red Cross and Red Crescent Societies. Ki-Seon C (2009) Evaluation. H. International Federation of Red Cross and Red Crescent Societies.S. modification. Josip K. 1997. Chen F. This way. 1st edn. and Billib. 105. MATLAB Neural Network Toolbox – User’s Guide. Action. Nick M. Drought prediction and characteristic analysis in semiarid Cearâ. Journal of Environmental Informatics 45:155–167 Byun H. KS L. H. M. Soc. B. Geneva Demuth. Phase III –: this is the validation of the results achieved above. Cassidy J (2010) World Disasters Report – Focus on Urban Risk. Acknowledgement Special thanks to the Kenya Meteorological Department for providing the data set used in this research. International Federation of Red Cross and Red Crescent Societies. Chao-Chung Y (2010) Development and application of a decision group Back-Propagation Neural Network for flood forecasting. Meteor. References Antonić O. Addis Ababa. Moreira C. Geneva Deely S. Antun M. J Hydrol 354:116–130 Freitas. 1970: Glossary of Meteorology. Dragan B (2001) Ecol Model 138:255–263 Armstrong S. Dan M (2011) World Disasters Report – Focus on Hunger and Malnutrition. Cermak V (2001) Prediction of extreme precipitation using a neural network: application to summer flood occurrence in Moravia. Hua J. J Clim 12:2747–2756 CERF (Central Emergency Response Fund) (2009) CERF funds jump start emergency aid operations in Kenya. Randolph K. Chen C (2006) Drought forecast model and framework using wireless sensor networks. Hi-Ryong B. Dandy GC (2000) Neural networks for the prediction and forecasting of water resources variables: a review of modelling issues and applications. data set from Kakamega weather station (not used before in the previous two phases) was utilised. Huschke. The resulting accuracies were similar to those achieved in Phase II. E. Korea.A. the ANNs forecasting models have addressed this by incorporating forecasted precipitation values from issued at professional meteorological institutions such as the KMD. phenomena such as climate change may lead to erroneous forecasts due to unexpected increase/decrease of precipitation. Coelho A. Geneva ASCE Task Committee On Application Of Artificial Neural Networks In Hydrology (2000) Artificial Neural Networks in Hydrology: II. Environ Model Softw 15(1):101–124 . The work presented in this paper is a part of an on-going research and one of the immediate tasks is to put the ANNs models described here on to a web-based interface to allow end users (such the meteorologists and staff at meteorological departments) to access and use the system. Anne M. Chou N. United Nations Economic Commission for Africa. International Federation of Red Cross and Red Crescent Societies. Anonymous IAHS. Ethiopia Chang-Shian C. 2009. No. Maarten VA (2009) World Disasters Report – Focus on Early Warning. J Hydrol 385:173–182 Collins A.. 638 pp. ANNs are known to rely on historical trends to forecast future values but in the case of droughts.A. . 240. R. it will be possible to use the ANNs models together with the seasonal climate forecasts carried out in Kenya and beyond. J Hydrol 378:1–12 Kung H. Michele M. Amer. Paulson EG Jr (1980) On the definition of droughts. Jorgelina H. Kim D. Hydrologic Applications. 1st edn. J Inf Sci Eng 22:751–769 Maier HR. M. 6 Dracup JA. Mark C. Sustainability of Water Resources under Increasing Uncertainty. Mark. Boris P.Mitig Adapt Strateg Glob Change the Y-Years-Lead-Time Forecast. The fact that the latter forecasts can have accuracies of as low as 70 % has been factored in to the design and was found that it does not significantly affect accuracies of the ANNs’ outputs. Water Resour Res 16:289–296 Elsa E.. Paulo AA (2008) SPI-Based Drought Category Predication Using Loglinear Models. In this paper. and application of the Effective Drought Index to 200-Year drought climatology of Seoul. David D.

pp.. Salas JD (2000) Regional drought analysis based on neural networks. pp 17–23 Mishra AK. DC. IEEE Xplore. 2012]. M. S. 2012-last update. J Hydrol 391:202–216 Morid S. 1. January. T. Nairobi Palmer. 1–8 Masinde M. Int J Climatol 27:2103–2111 Mutua.. Research Paper 45. Singn VP (2010) A Review of Drought [May/11. and Sharma. Vol. W. J Hydrol Eng 5(2):145–155 Weichert A. pp 318–362 Shin HS.emdat. Bull Math Biophys 5:115–133 Mckee TB. 1965.C. Desai VR (2006) Drought Forecasting Using Fee-Forward Recursive Neural Network.” Parallel distributed processing. The Role of ICTs in Quantifying the Severity and Duration of Climatic Variations – Kenya’s Case. Ecol Model 198:128–138 Mishra AK. Hinton GE. pp. Hayes MJ (2004) A self-calibrating palmer drought severity index.C. US Department of Commerce. 2011. and Bagula.Mitig Adapt Strateg Glob Change Masinde. Kleist J (1993) The relationship of drought frequency and duration to time scales. J Clim 17:2335–2351 WHO Collaborating Centre For Research On The Epidemiology of Disasters (CRED). Available: http://www. Washington. Challenges in drought research: some perspectives and future directions. Pitts W (1943) A logical calculus of the ideas imminent in nervous activity. 2011. Anonymous American Meteorological Society. A. Emergency Events Database (EM-DAT) [Homepage of Government of Belgium. Cambridge. Clim Res 10:83–93 Wells N. Proceedings of ITU Kaleidoscope 2011: The Fully Networked Human? Innovations for Future Networks and Services (K-2011). Bagula A (2012) ITIKI: bridge between African indigenous knowledge and modern science of drought prediction. Meteorological Drought. U. MIT Press.. Printfast Kenya Ltd. Sage. 1–58.. Hydrological Sciences-Journal . Goddard S. Newbury Park . Wilhite DA. Burger G (1998) Linear versus nonlinear techniques in downscaling. Boston. Rumelhart DE. 12–14 December 2011. Smakhtin V. Knowl Manag Dev J 7(03):274–290 Mcculloch WS. Bagherzadeh K (2007) Drought forecasting using artificial neural networks and time series of drought indices. [Online]. for WHO]. In: 8th Conference on Applied Climatology. Massachusetts. Panu. Williams RJ (1986) “Learning internal representations by error propagation. Strengthening Drought Early Warning At The Community and District Levels: Analysis Of Traditional Community Warning Systems In Wajir & Turkana Mwagore D (2002) Land use in Kenya: The case for a national land use policy. Weather Bureau. Doesken NJ. Glantz MH (1985) Understanding the drought phenomenon: the role of definitions.S. 2002. Water International 10:111–120 Yin RK (1993) Applications of Case Study Research.