19
JOURNAL OF INFORMATION SCIENCE AND ENGINEERING 22, 751-769 (2006) 751 Drought Forecast Model and Framework Using Wireless Sensor Networks * HSU-YANG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN + Department of Management Information Systems + Department of Forestry National Pingtung University of Science and Technology Pingtung, 912 Taiwan E-mail: {kung; m9356014; cct}@mail.npust.edu.tw Taiwan faces a serious challenge with an increasing frequency of drought in recent years. Therefore, it is important to utilize the state-of-the-art sensing and communication technologies to monitor and forecast effectively the drought, and then notify the relevant departments for taking preventive measures against this natural disaster. This paper pro- posed and developed a Drought Forecast and Alert System (DFAS), which is a 4-tier system framework composed of Mobile Users (MUs), Ecology Monitoring Sensors (EMSs), Integrated Service Server (ISS), and Intelligent Drought Decision System (ID 2 S). DFAS combines the wireless sensor networks, embedded multimedia communi- cations and neural network decision technologies to effectively achieve the forecast and alert of the drought. DFAS analyzes the drought level of the coming 7 th day via the pro- posed drought forecast model derived from the Back-Propagation Network algorithm. The drought inference factors are the 30-day accumulated rainfall, daily mean tempera- ture, and the soil moisture to improve the accuracy of forecasting drought. These infer- ence factors are detected, collected and transmitted in real-time via the Mote sensors and mobile networks. Once a region with possible drought hazard is identified, DFAS sends altering messages to users’ appliances. System implementation results reveal that DFAS provide the drought specialists and users with complete environment sensing data and images. DFAS makes it possible for the relevant personnel to take preventive measures, e.g., the adjustment of agricultural water, for a reduced loss. Keywords: wireless sensor networks, decision and support mechanisms, back-propaga- tion network, embedded multimedia communication, drought disaster 1. INTRODUCTION Located nearby the Tropic of Cancer, Taiwan features a subtropical island climate with a mean annual rainfall approx. 2515mm, much higher than worldwide mean annual rainfall of 973mm. However, Taiwan faces a serious situation with an increasing fre- quency of drought in recent years. The major reason for drought lies in extremely non-uniform distribution of rainfall all the year round, especially in south Taiwan where the major source of rainfall is concentrated during plum rains (May to June) and typhoon season (July to October). Thus, the high flow period is from May to October, and low flow period from October to next April, with a ratio of 9:1 [1]. The rainfall in south Tai- Received October 3, 2005; accepted April 11, 2006. Communicated by Yau-Hwang Kuo. * This paper was partially supported by the National Science Council of Taiwan, R.O.C., for financially sup- porting this research under contract No. NSC 94-2218-E-020-003.

Drought Forecast Model and Framework Using Wireless - CiteSeer

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Drought Forecast Model and Framework Using Wireless - CiteSeer

JOURNAL OF INFORMATION SCIENCE AND ENGINEERING 22, 751-769 (2006)

751

Drought Forecast Model and Framework Using Wireless Sensor Networks*

HSU-YANG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN+

Department of Management Information Systems +Department of Forestry

National Pingtung University of Science and Technology Pingtung, 912 Taiwan

E-mail: {kung; m9356014; cct}@mail.npust.edu.tw

Taiwan faces a serious challenge with an increasing frequency of drought in recent

years. Therefore, it is important to utilize the state-of-the-art sensing and communication technologies to monitor and forecast effectively the drought, and then notify the relevant departments for taking preventive measures against this natural disaster. This paper pro-posed and developed a Drought Forecast and Alert System (DFAS), which is a 4-tier system framework composed of Mobile Users (MUs), Ecology Monitoring Sensors (EMSs), Integrated Service Server (ISS), and Intelligent Drought Decision System (ID2S). DFAS combines the wireless sensor networks, embedded multimedia communi-cations and neural network decision technologies to effectively achieve the forecast and alert of the drought. DFAS analyzes the drought level of the coming 7th day via the pro-posed drought forecast model derived from the Back-Propagation Network algorithm. The drought inference factors are the 30-day accumulated rainfall, daily mean tempera-ture, and the soil moisture to improve the accuracy of forecasting drought. These infer-ence factors are detected, collected and transmitted in real-time via the Mote sensors and mobile networks. Once a region with possible drought hazard is identified, DFAS sends altering messages to users’ appliances. System implementation results reveal that DFAS provide the drought specialists and users with complete environment sensing data and images. DFAS makes it possible for the relevant personnel to take preventive measures, e.g., the adjustment of agricultural water, for a reduced loss. Keywords: wireless sensor networks, decision and support mechanisms, back-propaga- tion network, embedded multimedia communication, drought disaster

1. INTRODUCTION

Located nearby the Tropic of Cancer, Taiwan features a subtropical island climate with a mean annual rainfall approx. 2515mm, much higher than worldwide mean annual rainfall of 973mm. However, Taiwan faces a serious situation with an increasing fre-quency of drought in recent years. The major reason for drought lies in extremely non-uniform distribution of rainfall all the year round, especially in south Taiwan where the major source of rainfall is concentrated during plum rains (May to June) and typhoon season (July to October). Thus, the high flow period is from May to October, and low flow period from October to next April, with a ratio of 9:1 [1]. The rainfall in south Tai-

Received October 3, 2005; accepted April 11, 2006. Communicated by Yau-Hwang Kuo. * This paper was partially supported by the National Science Council of Taiwan, R.O.C., for financially sup-porting this research under contract No. NSC 94-2218-E-020-003.

Page 2: Drought Forecast Model and Framework Using Wireless - CiteSeer

HSU-YAUNG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN

752

wan is concentrated during high flow period covering plum rains and typhoon season. It results that here isn’t enough accumulated water supply available for the water consump-tion during low flow period [2]. Besides, given the fact of unique terrain, steep moun-tains and torrential rivers in Taiwan, water accumulation in reservoirs is very difficult. Furthermore, coupling with rapid social development and ever-growing water demand, Taiwan meets a challenge of developing new water source distribution to cater for civil and industrial applications.

Among a great number of industries, agriculture depends much upon water re-sources. In the event of shortage of water sources due to drought, the farmers can do nothing to meet the water demand of agricultural plants, especially during the low flow period. As a result of fallowing or giving up cultivation, it will not only waste the pro-duction cost of the farmers, but also affect the food supply. In the case of higher fre-quency and intensity of drought, the government will find it difficult to formulate poli-cies and decide on when or how fallowing is required. Due to expanded loss of the farm-ers, the government has to compensate them annually to the possible extent. As com-pared to typhoon and cloudburst leading to immediate disaster, drought has no any sign indicating the slowly developed disaster. In such case, the opportunity for deci-sion-making will be lost when faced with the threat of drought. For this reason, a drought forecast system shall be set up for early decision-making and warning.

This paper proposes and develops a Drought Forecast and Alert System (DFAS), which is a 4-tier system framework composed of Mobile Users (MUs), Ecology Moni-toring Sensors (EMSs), Integrated Service Server (ISS), and Intelligent Drought Decision System (ID2S). DFAS is used to monitor and collect continuously environmental drought data in combination with MODIS satellite images, wireless sensor networks, embedded multimedia communications and wireless/mobile networks, thereby achieving the objec-tives of timely supervision, warning and notification. DFAS monitors and collects all spatial and temporal ground surface information by using wireless sensor networks and network camera. The collected environmental drought data include the soil and air mois-ture, air temperature, the location of sensor devices, and 24-hour monitored drought im-ages, which are sent and stored back to the rear database through wired/wireless network or third-generation mobile system.

Intelligent Drought Decision System (ID2S) uses the 30-day accumulated rainfall and daily mean temperature as the forecast inference factors, and then analyzes the drought level of the coming 7th day via the proposed drought forecast model derived from backpropagation network. To identify more clearly the disaster, The ID2S improves the accuracy of drought forecast model by combining the forecast drought level with real-time soil moisture, NDVI index (vegetation index) and drought indicator of crops (soil moisture suitable for growth) based on the environmental sensor networks. With the real-time environmental data, this model helps to judge if this region is a dangerous one with potential drought. If the danger indicator is reached, the farmers and relevant de-partments shall be duly informed to take preventive measures and minimize the possible loss according to the drought level.

The rest of this paper is organized as follows. Section 2 describes the research back-ground and related work. Section 3 describes the DFAS system architecture and the drought decision model design. Section 4 describes the system implementation and us-ages. Conclusions are finally drawn in section 5.

Page 3: Drought Forecast Model and Framework Using Wireless - CiteSeer

DROUGHT FORECAST MODEL AND FRAMEWORK USING WIRELESS SENSOR NETWORKS

753

2. RESEARCH BACKGROUND AND RELATED WORK

The research background and relevant technologies required for DFAS include: (1) the definition of drought, (2) wireless sensor network technology, and (3) the neural net-work algorithm.

2.1 Definition of Drought

Drought refers to a lasting water shortage due to water consumption or water de-mand higher than water supply [3]. This leads to unbalanced water cycling, for example, insufficient rainfall and strong vaporization bring about continued loss of soil water for subsequent drought. Drought is most often defined according to rainfall as it’s primarily caused by long-lasting shortage or considerable decline of rainfall. Some scholars in Taiwan defined drought based on the scale of rainfall including (i) “rainless day” with rainfall less than 0.6cm, (ii) “small scale drought” without rainfall lasting 50 days, and (iii) “large scale drought” without rainfall over 100 days [3]. The Ministry of Economic Affairs defined the drought disaster as: “the loss of creature, environment, society, gen-eral public and industries arising from the direct or indirect influence of water shortage involving rainfall, river, groundwater and reservoir. The direct influence is related to jeopardy of creature, decline of food yield, forest, greenbelt, worsening water/air quality, sanitation and higher fire prevention risk. The indirect influence is related to decline of food supply, price upsurge, reduction of income or salary and degradation of living qual-ity” [4]. It means that drought may exert great negative impact upon social economy through progressive steps.

Owing to wide influence of drought, there are a variety of definitions depending upon the research purposes and objects. For example, Dracup et al. [5] classified it into meteorological drought, hydrological drought and agricultural drought due to different applications in these areas. Agricultural drought indicates that crops cannot grow nor-mally due to insufficient soil moisture caused by water shortage within a certain period of time. Meteorological drought refers to the drought caused by abnormal climate and maladjusted rainfall. Hydrological drought indicates that there is no enough water supply for various applications due to water shortage in ground surface, such as lower water level of river or reservoir. Therefore, DFAS takes the rainfall and soil moisture as major variables for drought identification and monitoring covering meteorological and agricul-tural drought.

Soil water is the most fundamental substance for the plants, so soil moisture has great influence upon crop growth, cultivation, planting quality and soil temperature. The soil moisture is also referred to as soil humidity indicating the water content in the soil and water supply for the farm plants [6]. In the case of extremely low soil moisture, the drought will come to impair the normal photosynthesis, leading to lower yield and qual-ity of crops. Moreover, the crops will wither if soil moisture declines below wilting point. Conversely, extremely high soil moisture will block off the breathing and growth of crops’ root system, resulting in the outbreak of insect damages to the crops.

In December, 1999, NASA successfully launched first Terra satellite onboard MODIS, also referred to as MODerate-resolution Imaging Spectroradiometer. The satel-lite images present excellent spectrum and spatial resolution, thus providing dynamic and abundant ground surface radiation images [7]. Due to different water content and chlo-

Page 4: Drought Forecast Model and Framework Using Wireless - CiteSeer

HSU-YAUNG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN

754

rophyll content of leaves under the same wavelength, the spectrum reflection strength shows a considerable difference [8]. In the case of drought, physiological change of plants, such as reduction of chlorophyll occurs, leading to yellowing and withering of leaves. Also, this difference can be monitored and analyzed using spectrum reflection feature of MODIS satellite images. In DFAS, images NDVI indexes via satellites and environmental drought data via sensor networks are considered as parameters for identi-fying drought disaster. 2.2 Wireless Sensor Networks Applied on Environmental Monitoring

Wireless sensor networks (WSN) combining the mobile computing, telecommuni-cation and sensing equipments, can operate automatically with least power consumption. As depicted in Fig. 1, a standard WSN device comprises the sensing units, processing units, transceiver units and power units [9]. The functionality of the units is described as follows.

Location finding system Mobilizer

Sensing Unit Unit

Processor

Storage

Power unit

Processing

Sensors ADC Transceiver

Power generator

Fig. 1. The system architecture a standard WSN device.

(1) Sensing Unit. A sensing unit comprises the sensor and the analog-to-digital converter.

The sensor is responsible for detecting and collecting the environmental data, which represent with the analog signals. The analog-to-digital converter converts the analog signals into the digital data and sends the data to the processing unit.

(2) Processing Unit. Processing unit comprises the processor and the storage unit. Stor-age unit stores the collected environmental data. Processor processes the data ac-cording to the pre-defined program codes.

(3) Transceiver Unit. Transceiver unit is responsible for the communications between the sensor devices.

(4) Power Unit. Power unit provides the electric power and is the most important unit of a WSN device.

The communication topology of WSNs is based on mobile ad-hoc protocols. As de-

picted in Fig. 2, the communication topology of WSNs comprises the sensor nodes, the sink nodes and the manager nodes [10]. The functionality of each node is described as follows.

Page 5: Drought Forecast Model and Framework Using Wireless - CiteSeer

DROUGHT FORECAST MODEL AND FRAMEWORK USING WIRELESS SENSOR NETWORKS

755

Sensor Sensor Region Nodes

Sink Nodes

MANET

D

C B A E

Manager Nodes

Internet/Mobile Networks

Fig. 2. The communication topology of WSN.

(1) Sensor node is capable of detecting and collecting the environmental data. Sensor

node processes the collected data and transmits them to the sink node. (2) Sink node is a gateway node, which is responsible for receiving the sensor data and

re-transmitting these data to the manager node via Internet or mobile networks. (3) Manager node is responsible for processing and displaying the sensor data.

The WSN applications include the monitoring of wild animals, environment moni-toring/forecast and health monitoring. This paper intends to use Motes wireless sensor networks, which enable data collection covering soil and air humidity, air temperature, light and GPS. All data from every sensor can be transmitted via ZigBee network trans-mission protocol, thus forming a mobile ad hoc network [11]. The environmental drought data are sent back to the rear database via wireless network and mobile telecom network, thereby contributing to real-time and continuous monitoring of drought. The drought model decision system automatically analyses the environmental drought data and the results are sent to the end users in real-time. This reduces human error for a more accu-rate drought forecast.

2.3 Neural Network Algorithm

Neural network primarily aims to simulate human neural system, which has perfect performance with respect to language, hearing, imaging, vision, and other field applica-tions. Neural network is composed of many non-linear operating units, i.e., neurons, and links within them. All these operating units are operated in a parallel and distributed manner to process many data at the same time. The neural network is the combination of neurons. The connection between neurons has two types, which are the inhibitory con-nection and excitatory connection [12]. Fig. 3 depicts the neuron model of the neural network and describes as follows. (1) Xi is the input of a neuron i. (2) Wi is the synaptic weight to represent the linkages strength of neuron i. The high

synaptic weight highly effects the operation of the neural network. On the other hand, the low synaptic weight has lightly effects on the neural network. The low synaptic weights usually are removed to reduce the computing time of the neural network. The operation of the neural network is to adjust the synaptic weights of neuron link-ages and determine the suitable synaptic weights to have accuracy results.

Page 6: Drought Forecast Model and Framework Using Wireless - CiteSeer

HSU-YAUNG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN

756

X1 W1

W2 X2

Wn

Xn

Σ ϕ( ) Yj

Outputs : Decision Results

Inputs : Decision Initialization

Neuron

Output Layer

HiddenLayer

Input Layer

Fig. 3. The neuron model of neural network. Fig. 4. The back-propagation network model.

(3) Σ is responsible for the summation of each neuron’s synaptic weight. (4) φ( ) is the activation function, which is the non-linear type to convert the summation

value into the output. (5) Yj is the output of the neural network.

This paper utilizes the most prevailing Back-Propagation Network (BPN) algorithm to analyze the potential drought degrees. The BPN algorithm is a typical Supervised Learning Network [13], which is to learn the internal reflection and regulations between inputs and outputs. The regulations are the synaptic weights of network neurons. For analyzing any new cases, the input values or independent variables are inputted into the neural network and get the inferential related output values quickly.

Fig. 4 depicts the Back-Propagation Network model, which have three system layers and described as follows.

(1) Input Layer comprises the inputs of the BPN and represents the initial values of de-cision.

(2) Hidden Layer comprises the neurons, which are responsible for adjusting the synap-tic weights of neuron linkages and determining the suitable synaptic weights. To have accuracy results, the hidden layer is composed of several sub-layers to learn the internal reflection and regulations between inputs and outputs.

(3) Output Layer comprises the outputs of the BPN and represents the final decision results at this training operation.

The control procedure of the BPN algorithm divides into the following operation steps.

(1) Set up the network parameters. (2) Set up weighted matrixes, i.e., W_xh and W_hy, and the initial values of bias vector,

i.e., θ_h and θ_y, by uniformly random numbers. (3) Calculate the output quantity of the Hidden Layer. (4) Set up the tolerant difference quantity between the Output Layer and the Hidden

Layer.

Page 7: Drought Forecast Model and Framework Using Wireless - CiteSeer

DROUGHT FORECAST MODEL AND FRAMEWORK USING WIRELESS SENSOR NETWORKS

757

(5) Calculate the difference quantity, i.e., δ, between the Output Layer and the Hidden Layer.

(6) Determine whether the difference quantity between the Output Layer and the Hidden Layer is greater than the tolerant difference quantity Φ. If the difference quantity δ is smaller than the tolerant difference quantity Φ, the optimal regression model is ob-tained.

(7) If the difference quantity δ is greater than the tolerant difference quantity Φ, the weighted matrixes W_xh and W_hy, and the corrections of bias values θ_h and θ_y in the Output Layer and the Hidden Layer have to be computed.

(8) Revise the weighted matrixes and the bias values in the Output Layer and the Hidden Layer, and repeat steps (3) to (7) until the difference quantity lies within the range of the tolerant difference quantity.

(9) Finally, compare the correlation of sensitivity correction to find out the optimal re-gression model.

3. DFAS SYSTEM FRAMEWORK AND DESIGN

Fig. 5 depicts the system and network architecture of DFAS, which is a 4-tiers framework including Mobile Users (MUs), Ecology Monitoring Sensors (EMSs), Inte-grated Services Server (ISS) and Intelligent Drought Decision System (ID2S). Each de-signed component is described as follows.

Mobile Users Ecology Monitoring Sensors Integrated Service Server Intelligent Drought Decision System

User Message Interface Interface

IE Module

Windows Operation System

Communication Interface

Intelligent Agents

Information Agent

Rainfall Agent

Multimedia Agent

Push Agent

Predication and Alert Agent

Location Aware Agent

Satellite Index and Area Agent

Integrated Services Interface

WE

B S

ervi

ce/X

ML

PDA Table PC Laptop

GPS

Wireless Sensor

Networks

ZigBee

RS-232

Emergency Action

Knowledge Base

Knowledge Acquisition

Module

Emergency Action Inference Engine

Drought Disaster

Inference Engine

Model Operation

Agent

Environment Database

GPRS/3G GPRS/3G

Ecology Monitoring Engine Application Server Intelligent Drought Decision System Server

Router Router

GPRS/3G Gateway

Core Network TCP/IP Network Backbone

GPRS/3G Cellphone

Fig. 5. Drought forecast and alert system and network architecture.

Page 8: Drought Forecast Model and Framework Using Wireless - CiteSeer

HSU-YAUNG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN

758

3.1 Mobile Users (MUs) Component

The mobile appliances include desktop computer, PC and handheld devices for mo-bile users. DFAS is used by farmer users and the staff members of the drought monitor-ing and warning center. The MUs component on the mobile appliances is designed and integrated in such a manner to enable the mobile users to access the information by link-ing DFAS to any device, mobile network and Internet. MUs component is designed with: (1) real-time multimedia receiver module, (2) customized information service module and (3) real-time information receiver module as described below. (1) Real-time Multimedia Receiver Module. This module is primarily used to receive

video images of drought information from Integrated Services Server (ISS) to pro-vide personnel with drought forecast and monitoring multimedia information.

(2) Customized Information Service Module. This module delivers customized services to the mobile users, such as optimum linking, different interfaces based on the mo-bile appliances, the personal functions and information.

(3) Real-Time Information Receiver Module. This module provides mobile users to automatically download the drought data and images from ISS. This module also transmits the latest alert message to the users in the event of upcoming drought.

3.2 Ecology Monitoring Sensors (EMSs) Component

The EMSs located at the observation stations is used to collect environmental drought data and images. The main control modules of the EMSs component are the Ecology Monitoring Engine (EME) and the Motes wireless sensor networks. Fig. 6 de-picts the monitoring and sensor devices at a sensing point. Every sensing point comprises two sensors and a network camera. Among two sensors, i.e. MDA300 and MTS420, MDA300 sensor is used to sense soil moisture, air temperature and humidity. MTS420 sensor is used to sense air pressure and GPS coordinate. EME is responsible for analyz-ing the collected sensing data. The EME control modules are: (1) Environment Data Monitoring Agent, (2) Connection Detection Agent and (3) Stand-by Database as de-scribed below.

Fig. 6. Monitoring and sensor devices at a sensing point.

3.2.1 Environment data monitoring (EDM) agent

EDM agent aims at collecting environmental drought data and images sensed and

Page 9: Drought Forecast Model and Framework Using Wireless - CiteSeer

DROUGHT FORECAST MODEL AND FRAMEWORK USING WIRELESS SENSOR NETWORKS

759

shot by Motes and network camera, and converting original data into standard digital form for engineering applications. Then, these data and images are immediately sent back to rear drought database via mobile and wireless networks.

3.2.2 Connection detection (CD) agent

The functionality of CD agent is as follows. (i) Off-line detection. This function helps to judge if the network is in an off-line state. In such case, CD agent closes data transmission function between the original EDM agent and environmental drought data-base. (ii) Storage control. CD agent stores the off-line environmental drought data col-lected by EDM agent into a stand-by database of EME until the on-line communication is resumed. Then, CD agent informs EDM agent to transmit and store data into the rear drought database.

3.2.3 Stand-by database

Stand-by database is responsible for storing the environmental drought data under off-line state. In such situation, environmental drought data collected by EDM agent are temporarily stored into the stand-by database until the network communication is re-sumed. Then, the data are sent back to the rear environmental drought database.

3.3 Integrated Services Server (ISS) System

ISS is installed at the drought forecast and alert center. ISS system is designed to allow multiple intelligent agents for data collection, searching, classifying, processing and notification. The major purpose of ISS system is to separate numerous ecological information and multimedia data, thereby making the users to access the rear database in a time-saving manner. The control agents of ISS system are (1) Location Aware Agent, (2) Predication and Alert Agent, (3) Multimedia Agent, (4) Push Agent, (5) Satellite Im-age Index and Area Agent, (6) Information Agent and (7) Rainfall Agent as described below. (1) Location Aware (LA) Agent. LA agent aims at finding out rapidly accurate drought

regions. LA agent calculates and stores the accurate positions of sensors by combin-ing GPS and GIS. During the preliminary configuration or replacement of sensor, GPS value is used to display or update the location of sensor on GIS map. Addition-ally, LA agent periodically senses the GPS value so as to monitor if sensor is pas-sively relocated due to external factors.

(2) Predication and Alert (PA) Agent. PA agent notifies the users of the determined drought level and hazard identification results obtained from the drought forecast model.

(3) Multimedia Agent. In combination with real-time multimedia receiver module of MUs component, the multimedia agent transmits the images to mobile appliances of users to answer visualization requirements for mobile applications. The major pur-pose is to provide the mobile user a multimedia environment, which covers historical images and real-time images for retrieval and monitoring. When the mobile user re-

Page 10: Drought Forecast Model and Framework Using Wireless - CiteSeer

HSU-YAUNG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN

760

quests to receive historical images, the multimedia agent acquires the historical im-ages into database and sends to the user’s mobile appliance. In addition, the mobile user may also select real-time images taken by the network camera. The multimedia agent sends the real-time images to the user’s mobile appliance. These real-time im-ages are also stored into the rear database via the multimedia agent.

(4) Push Agent. Push agent adopts push technology to enable automatic downloading of latest information. Firstly, the push agent observes regularly if the database is avail-able with latest drought data, alert information and relevant publications for the drought monitoring and warning center. If latest information is found, the push agent automatically informs the MUs component for drought forecast and alert. If the mo-bile user desires to obtain the latest information, the push agent acquires the latest information from the database and sends to mobile appliance of the mobile user.

(5) Satellite Image Index and Area (SI2A) agent. To guarantee detailed and accurate haz-ard identification and the drought level, the drought forecast model has to consider the real-time soil moisture and NDVI satellite image index. To avoid the large iden-tification errors due to the difference between satellite image index and geographical locations associated with environmental drought data, SI2A agent stores the appro-priate satellite image indexes, and allows ID2S to obtain these indexes and drought data from rear database and calculate by the drought forecast model.

(6) Information Agent. Information agent provides the MUs component to look-up and store real-time drought information for drought forecast and warning. Information agent not only stores the real-time data input from the MUs component into the rear drought database, but also receives and determines the appropriate searching condi-tions and sends them to MUs component.

(7) Rainfall Agent. Rainfall agent acquires daily rainfall data of south Taiwan from FTP stations of Central Weather Bureau on a day-by-day basis. Then, rainfall agent stores the rainfall data into the rear database. Furthermore, rainfall agent also calculates previous 30-day accumulated rainfall and monthly accumulated rainfall for operation of the drought forecast model.

3.4 Intelligent Drought Decision System (ID2S)

The main propose of ID2S is to forecast the drought level, infer current hazard and provide emergency action messages. The inference factors include the 30-day accumu-lated rainfall, daily mean temperature, soil moisture, NDVI indicators and crop drought indicators. Soil moisture and temperature obtained from sensor networks belong to real-time and temporal dimension. Rainfall obtained from the rainfall server appertains to non-real-time and temporal dimension. NDVI index obtained by manual input pertains to non-real-time and spatial dimension. Using the inference factors, the drought forecast model adopts the back- propagation network to infer the drought level and disaster iden-tification. Based on drought level and disaster identification results, ID2S can infer and then provide MUs some emergency action messages in response to different drought conditions. The emergency action messages may provide a reference for real-time deci-sion-making and facilitate rapid and correct emergency actions.

The control agents of ID2S include (1) Model Operation Agent, (2) Drought Disas-

Page 11: Drought Forecast Model and Framework Using Wireless - CiteSeer

DROUGHT FORECAST MODEL AND FRAMEWORK USING WIRELESS SENSOR NETWORKS

761

ter Inference Engine, (3) Emergency Action Inference Engine, (4) Knowledge Acquisi-tion Module, (5) Emergency Action Knowledge Base and (6) Environment Database as described below.

3.4.1 Model operation (MO) agent

The MO agent uses 30-day accumulated rainfall and daily mean temperature as in-ference factors to analyze and infer 7th-day drought level through the drought forecast model, which adopts the Back-Propagation Network (BPN) algorithm. The output target of training examples of BPN algorithm has to determine the required drought level. As such, a drought level classification principle has to be formulated to obtain target output vector of training examples.

3.4.1.1 Drought level classification principle

Owing to a slow and progressive process, drought always develops from normal conditions to slight, moderate and serious disaster. According to the previous research [14], drought level is classified into five levels including (i) non-drought, (ii) slight drought, (iii) moderate drought, (iv) serious drought and (v) extremely serious drought. Based on the previous research [15], meteorological drought in Taiwan is defined as “30-day continuous accumulated rainfall is lower than 2nd decile of the same period and accumulated drought severity reaches over 130mm.” Therefore, the value of drought severity is calculated by Eq. (1). Considering the comprehensive influence of accumu-lated drought severity and temperature upon drought, drought severity and temperature are as operating parameters to determine the drought level and establish the drought clas-sification list.

( )Drought severity ( 1)

t dt tp

tx x

n dnd

+−∑

= + − (1)

xt is the n-day accumulated rainfall. xtp is the truncation level. d is the drought delay.

The truncation level required for drought severity is based on monthly truncation level of southern counties/cities, which is calculated from 30-day continuous accumu-lated rainfall of various rainfall stations in 1992 to 2003 [7]. This is a typical truncation level of recent drought as it is obtained from accumulated rainfall distribution within recent 10 years. Besides, the rainfall and temperature from 27 southern rainfall stations in 2002 to 2003, where drought occurred in recent years, are taken as the research samples of calculating drought severity.

After calculating daily drought severity, the determination of the “maximum accu-mulated drought severity,” “maximum air temperature” and “minimum air temperature” is derived from the daily accumulated drought severity and maximum/minimum air tem-perature. The value between 130mm and maximum accumulated drought severity, and the value between the minimum air temperature and maximum air temperature are uni-formly classified into 5 intervals. We use symbols S1, S2, S3, S4 and S5 respectively rep-resent the mean range of five respective accumulated drought severity levels. Symbols A1,

Page 12: Drought Forecast Model and Framework Using Wireless - CiteSeer

HSU-YAUNG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN

762

A2, A3, A4 and A5 respectively represent the mean range of five respective temperature levels.

Eq. (2) is used to determine the drought level indicator with the accumulated drought severity and temperature. For example, S1 and A5 represents D = 1 and T = 5, respectively. Since the accumulated drought severity has a stronger influence upon drought than temperature, the accumulated drought severity level is squared and then calculated with temperature level.

Drought level indicator = DL = log5(D2 × T) (2)

D is the accumulated drought severity level and T is the temperature level.

Table 1 depicts the definition and values of drought levels derived from Eq. (2). Ta-ble 2 depicts the completed drought level classification, which is the foundation for fu-ture drought classification comprising drought severity and temperature.

Table 1. Definition and values of drought level indicators.

Drought level Non-drought Slight drought Moderate drought

Serious drought

Extremely serious drought

Drought level indicator range

0 ≤ DL ≤ 0.6 0.6 ≤ DL ≤ 1.2 1.2 ≤ DL ≤ 1.8 1.8 ≤ DL ≤ 2.4 2.4 ≤ DL ≤ 3

Table 2. Drought level classification.

Low High Acc.

Temp Drought (Ai) (Si)

Below 130mm

(S1)

130mm~ 1286mm

(S2)

12864mm~ 25598mm

(S3)

25598mm~ 38332mm

(S4)

Above 38332mm

(S5) Below

13.72°C (A1)

Non-drought Slight

drought Moderate drought

Moderate drought

Serious drought

13.72°C~ 19.16°C

(A2) Non-drought

Moderate drought

Moderate drought

Serious drought

Extremely serious drought

19.16°C~ 24.61°C

(A3)

Slight drought

Moderate drought

Serious drought

Serious drought

Extremely serious drought

24.61°C~ 30.05°C

(A4)

Slight drought

Moderate drought

Serious drought

Extremely serious drought

Extremely serious drought

Low

High

Above 30.05°C

(A5)

Slight drought

Serious drought

Serious drought

Extremely serious drought

Extremely serious drought

3.4.1.2 Drought forecast model design

To forecast the future occurrence probability of drought disaster, the drought fore-cast model is designed according to the Back-Propagation Network (BPN) method. The

Page 13: Drought Forecast Model and Framework Using Wireless - CiteSeer

DROUGHT FORECAST MODEL AND FRAMEWORK USING WIRELESS SENSOR NETWORKS

763

drought classification serves as target output values required for the BPN training opera-tions. The training samples of BPN algorithm include output and input variables. The major function is to learn the principle of inherent correspondence of input and output variables.

Given the progressive process of drought disaster, prediction on alternative days makes no sense. To the contrary, a long-lasting prediction time may yield inaccurate re-sults due to changing climate. In general, the forecast cycle of weather bureau is 7 days. Therefore, the drought level is forecasted on 7th day according to the general time inter-val of weather forecast. The time factors are significant to the prediction results of the drought forecast model. If the drought forecast is performed on tth day, the model re-quires the previous n-day data. This paper uses the previous 30-day accumulated rainfall and daily mean temperature as the input variables of tth day and the drought level classi-fication as output variable. Therefore, upcoming m-day is set as 7 days, and the drought level of the coming 7th day is determined to be output variable of t-day through drought level classification principle, based on previous 30-day accumulated rainfall and daily mean temperature on the coming 7th day.

Input and output variables significantly affect the drought forecast results during the training phase of BPN algorithm. Therefore, the historical data at various levels have to be the training samples, thus allowing learning various drought levels and developing a practicable drought forecast model. 3.4.2 Drought disaster inference engine (D2IE)

The forecasted 7th-day drought level indicates the meteorological drought. However, the agricultural drought may not occur if soil moisture still meets the water demand of crops. Due to varied soil moisture required for different crops, slight meteorological drought may not lead to agricultural drought. Therefore, drought disaster inference en-gine is responsible for integrating the forecasted drought level with real-time soil mois-ture, NDVI indicator (vegetation index) and crop drought indicator, i.e., the suitable soil moisture for growth to determine the agricultural drought disaster in a more accurate manner. Fig. 7 depicts the processing procedure of D2IE for drought level forecast and disaster identification.

Daily average temperature of the

current day

Accumulated

rainfall of the first 30 day

Input variable

of forecast Back-Propagation Network

Drought Forecast

model to infer and predict the

drought level

Forecast result of drought

level

Crop drought indicator

drought level of 7-th day

NDVI indicator

Real-time soil humidity

Input inference

factors

Drought Disaster

Inference Engine

Drought level of 7-th day

Disaster determination

result

Output inference

result

Output forecast result

Fig. 7. The processing procedure of the drought disaster inference engine.

Page 14: Drought Forecast Model and Framework Using Wireless - CiteSeer

HSU-YAUNG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN

764

3.4.3 Emergency action inference engine (EAIE)

The emergency action inference engine (EAIE) offers emergency action measures and knowledge in response to present status. EAIE firstly receives the emergency action measures from the model operation agent, and then check if current environmental con-ditions and drought level correspond to drought emergency action measures in the knowledge base. The predication and alert agent transmits the proper emergency action measures to the users, thereby providing a basis for emergency decision-making.

3.4.4 Knowledge acquisition module (KAM)

The knowledge acquisition module processes the expert solutions and drought con-trol knowledge/experience acquired by the knowledge engineers, and then converts into feasible computing solutions. KAM provides the simple knowledge editing functions and the user interface to turn expert knowledge/experience into the knowledge base.

3.4.5 Emergency action knowledge base

The emergency action knowledge base stores the expert solutions and knowledge for different environments and drought levels, which comprise facts, successful experi-ence and emergency measures/identification principles. All these are modularized and represented in a systematic way, e.g., principle-based knowledge representation.

3.4.6 Environment database

Environment database stores the drought data and images sensed and shot by Motes and network cameras. In addition to offering Drought Forecast Model for DFAS, envi-ronment database also provides the user with environmental drought data and images for effective drought monitoring. Environment database also stores relevant user information, historical drought alert messages and real-time information previously published.

4. SYSTEM IMPLEMENTATION AND USAGES

The implementation environment of DFAS includes JAVA, Java 2 Platform, Micro Edition (J2ME), ArcGIS 8.0 and PostgreSQL. The used hardware includes wireless sen-sor devices, which are one MIB510, three MPR400, one MDA300 and one MTS420, Compaq Pads and network cameras. The experimental location is at Neipu, where is in south Taiwan and usually suffers the situation of low rainfall through the whole year.

Fig. 8 depicts the implementation of wireless sensor networks for environmental information. Fig. 8 (a) shows the sensing area and data inquiry, wherein green square represents the location of the sensor. Fig. 8 (b) indicates the sensing environmental in-formation including sensor’s ID, sensor’s location, sensing time, temperature, humidity and soil moisture after selecting one sensing point.

Fig. 9 depicts the environment monitoring function for multimedia images. Fig. 9 (a) is a display of real-time images by selecting the dedicated location. Fig. 9 (b) displays the historical images to look up. Fig. 10 depicts the implementation results of wireless sensor

Page 15: Drought Forecast Model and Framework Using Wireless - CiteSeer

DROUGHT FORECAST MODEL AND FRAMEWORK USING WIRELESS SENSOR NETWORKS

765

Fig. 10. Display of wireless sensor networks, real-time images and instant messages using on a PC.

(a) (b)

Fig. 8. (a) The sensing area and data inquiry function; (b) Display of the sensing environmental information for the selected sensing location.

(a) (b)

Fig. 9. The environment monitoring functions for (a) real-time images and (b) historical images.

Page 16: Drought Forecast Model and Framework Using Wireless - CiteSeer

HSU-YAUNG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN

766

Fig. 12. Displays of (a) real-time environment sensing information and (b) historical rainfall infor-mation via PDA web browser.

networks and real-time information displayed on a PC. Since PC has the higher comput-ing capability than the mobile appliance has, there are several environmental information displayed simultaneously on the user interface. As depicted in Fig. 10, the rainfall and soil moisture, which are the major variables for drought identification, are displayed with the compound manner to clearly identify the relations between these two variables. Real-time images and instant messages are also provided for the mobile users on the drought area to communicate with the specialists on the drought monitoring center.

DFAS also provides users to look up the real-time and historical drought environ-ment information via the web browser. Fig. 11 depicts the user can look up the real-time environment sensing information using the PC web browser. Fig. 12 depicts the user can use a PDA to look up the real-time and historical drought environment information via the web browser. Fig. 12 (a) is a display of real-time environment sensing information with PDA web browser. Fig. 12 (b) is a display of historical rainfall information with PDA web browser.

Fig. 11. The real-time environment sensing information using the PC web browser.

(a) (b)

Page 17: Drought Forecast Model and Framework Using Wireless - CiteSeer

DROUGHT FORECAST MODEL AND FRAMEWORK USING WIRELESS SENSOR NETWORKS

767

5. CONCLUSIONS

This paper designs and develops Drought Forecast and Alert System (DFAS), which is a 4-tier system framework composed of Mobile Users (MUs), Ecology Monitoring Sensors (EMSs), Integrated Service Server (ISS), and Intelligent Drought Decision Sys-tem (ID2S). DFAS establishes an effective drought forecast model and alert system by combining WSN technology for ecology monitoring, embedded multimedia software technology for pervasive communications, and agro-ecology knowledge for drought dis-aster. This paper is expected to realize the following contributions. (1) Firstly propose a feasible system framework, which uses the state-of-the-art wireless

sensor network and embedded multimedia communication technologies to automati-cally collect environmental drought information. The designed DFAS aims to save manpower and prevent incomplete data due to human error. Also, DFAS allows for effective and real-time drought forecast and alert due to continuous data collection around the clock.

(2) Firstly propose the drought forecast model based on the Back-Propagation Network algorithm to effectively predict the occurrence of drought in south Taiwan. This pa-per defines the inference formula to drive the drought level to achieve more accurate and clearer disaster identification.

(3) The DFAS users can receive the environmental drought information, inquiry infor-mation or drought images, and feed back to rear database via mobile/wireless net-work at any time and anywhere.

(4) This paper successfully applies the wireless sensor networks on agro-ecology fields by investigating environmental situations. The completed real-time and historical environment information is expected to help the agro-ecological specialists achieve efficient management and utilization of agro-ecological resources.

REFERENCES

1. Sinotech Engineering Consultants, Ltd., “Planning decision support system for drought defense (2/3),” Water Resources Agency, Ministry of Economic Affairs, 2002.

2. G. S. Yu and M. D. Chuang, “Studies on the characteristics of drought in Taiwan,” Taiwan Water Conservancy, Vol. 40, 1992, pp. 20-33.

3. R. Y. Wang and C. D. Chao, “Simulation of regional drought and its application to the agricultural water-resources planning in tsengwen river basin,” Taiwan Water Conservancy, Vol. 38, 1990, pp. 15-35.

4. Ministry of Economic Affairs, http://www.wra.gov.tw/sp.asp?xdURL=Rules/rules_ con.asp&no=15&sec_no=4&comefrom=full.

5. J. A. Dracup, K. S. Lee, and E. G. Paulson, “On the definition of droughts,” Water Resources Research, Vol. 16, 1980, pp. 297-302.

6. S. S. Wan, Fundamentals of Soil Physics, National Institute for Compilation and Translation, 2001, pp. 136-145.

7. C. N. Hsi, “Study on integrating GIS techniques and MODIS image into drought

Page 18: Drought Forecast Model and Framework Using Wireless - CiteSeer

HSU-YAUNG KUNG, JING-SHIUAN HUA AND CHAUR-TZUHN CHEN

768

monitoring,” Master dissertations, Department of Forestry, National Pingtung Uni-versity of Science and Technology, Taiwan, 2004, pp. 1-17.

8. C. T. Chen, C. Y. Lee, and G. Yang, “A feasibility study on using NOAA satellite image into drought forecast,” Photogrammetry and Remote Sensing, Vol. 7, 2002, pp. 75-86.

9. I. F. Akyildiz, S. Weilian, Y. Sankarasubramaniam, and E. Cayirci, “A survey on sen-sor networks.” IEEE Communications Magazine, Vol. 40, 2002, pp. 102-114.

10. L. B. Ruiz, J. M. Loureiro, and A. A. F. Loureiro, “MANNA: a management archi-tecture for wireless sensor networks,” IEEE Wireless Communication Magazine, Vol. 41, 2003, pp. 116-125.

11. C. Evans-Pughe, “Bzzzz zzz : ZigBee wireless standard,” IEE Review, Vol. 49, 2003, pp. 28-31.

12. J. T. Tsau, Management Information System, Tingmao Publish Company, 2003. 13. Y. C. Ye, Application and Implementation on Neural Network Models, Scholars

Books Co., Ltd., 2004. 14. R. S. Wu, W. R. Su, P. W. Wen, and C. T. Lin, “A study for drought model at local

area,” in Proceedings of 11th Hydraulic Engineering Conference, 2000, pp. O115-O121.

15. J. T. Shiau, “Defining meteorological droughts by successively cumulative precipita-tion,” Taiwan Water Conservancy, Vol. 49, 2001, pp. 52-64.

Hsu-Yang Kung (龔旭陽) received his B.S. degree from Tatung University, M.S. degree from National Tsing Hwa Uni-versity, Ph.D. degree from National Cheng Kung University, Taiwan, all in Computer Science and Information Engineering. He is currently an Associate Professor at the Department of Man-agement Information Systems, National Pingtung University of Science and Technology, Taiwan. His research interests include distributed multimedia systems, wireless sensor network, and the embedded multimedia applications.

Jing-Shiuan Hua (華瀞萱) is currently a Master student in the National Pingtung University of Science and Technology majoring in Management Information Systems in Taiwan. Her current research is focused on mobile communications, mobility management and wireless sensor networks.

Page 19: Drought Forecast Model and Framework Using Wireless - CiteSeer

DROUGHT FORECAST MODEL AND FRAMEWORK USING WIRELESS SENSOR NETWORKS

769

Chaur-Tzuhn Chen (陳朝圳) is a Dean, College of Agri-cultural and a Professor at the Department of Forestry, National Pingtung University of Science and Technology, Taiwan. He is received his M.S. and Ph.D. degree from Department of Forestry, National Chung Hsing University, Taiwan. His main research interests are agricultural drought theory, geographic information system, timber management, biometry, and remote sensing.