Survey On Predictive Analysis Of Drought In India Using AVHRR-NOAA Remote Sensing Data

  • Upload
    ijafrc

  • View
    12

  • Download
    0

Embed Size (px)

DESCRIPTION

This paper encompasses structural model that can be used to analyze the drought that can occur from El Nino and other climatic disorders. With prime emphasis on the parameters used like TCI, VCI and NDVI we can predict the drought with respect to Standard precipitation index (SPI). Various classifications of SPI can be used to predict the severity of drought. Paper also encompasses the use of neural network based model to enhance the predictive power of the drought analyzer. This survey paper establishes a bridge over the AVHRR data to analyze severity and understand the crop required for the land according to the moisture content.

Citation preview

  • International Journal of Advance Foundation and Research in Computer (IJAFRC)

    Volume 1, Issue 8, August2014. ISSN 2348 - 4853

    1 | 2014, IJAFRC All Rights Reserved www.ijafrc.org

    Survey On Predictive Analysis Of Drought In India Using

    AVHRR-NOAA Remote Sensing Data. Mr. Yogesh Gaikwad, Mrs. Rohini Bhosale.

    Department of Computer Engineering, Pillai HOC College of Engineering, Rasayani, Panvel.

    [email protected] *, [email protected]

    A B S T R A C T

    This paper encompasses structural model that can be used to analyze the drought that can occur

    from El Nino and other climatic disorders. With prime emphasis on the parameters used like TCI,

    VCI and NDVI we can predict the drought with respect to Standard precipitation index (SPI).

    Various classifications of SPI can be used to predict the severity of drought. Paper also

    encompasses the use of neural network based model to enhance the predictive power of the

    drought analyzer. This survey paper establishes a bridge over the AVHRR data to analyze severity

    and understand the crop required for the land according to the moisture content. (Abstract)

    Index Terms: TCI, VCI, NDVI, SPI, Neural Network, AVHRR.

    I. INTRODUCTION

    Drought is natural phenomenon which is caused due to the shortage of rainfall. It is affecting the many

    places of the world and can causes natural hazard. The combination of variables such as rainfall,

    temperature humidity, wind, soil moisture can be used to measure the intensity of the drought. The

    monitoring of drought is difficult due to drought spreads over large area, increases slowly. In India for

    the year 2014 drought is going to be an important factor due to the El Nino. El Nino which refers to

    variations in the temperature of the surface of the tropical eastern Pacific Ocean and in air surface

    pressure in the tropical western Pacific. The warm oceanic phase, El Nio, accompanies high air surface

    pressure in the western Pacific. Due to this there may be possibility of Agriculture drought in several

    areas of India. The probability of Drought occurrence can be minimized by modeling the drought and

    managing the water resources.

    There are basically three type of drought is defined: meteorological, agricultural, and hydrological. The

    situation at which the normal precipitation deceases in period of time is called as Meteorological drought.

    Palmer drought severity index [2] and the standardized precipitation index (SPI) [3] are the most

    commonly used method to major the intensity of drought [1].SPI is used to describe the extremely dry or

    wet climate situation. It is standardized measure of precipitation in different regions for different time

    scales. A proper transformation is needed to obtain normally distributed data, since precipitation data do

    not have normal distribution.SPI is the important metrics among others for drought prediction since it

    has variable time scale and uses only rainfall records.

    To monitor the environmental issues most commonly the satellite-based remote sensing data can

    effectively be used. The moisture-related vegetation indices [4] can be extracted by the use of advanced

    very high resolution radiometer (AVHRR) on the National Oceanic and Atmospheric Administration

    (NOAA) satellite. These indices can be used for monitoring vegetation conditions of plant such as

    normalized difference vegetation index (NDVI), vegetation condition index(VCI), and temperature

    condition index(TCI). Among these indices, NDVI have been most effectively used to vegetation and

    drought monitoring.

  • International Journal of Advance Foundation and Research in Computer (IJAFRC)

    Volume 1, Issue 8, August2014. ISSN 2348 - 4853

    2 | 2014, IJAFRC All Rights Reserved www.ijafrc.org

    II. RELATED WORK

    Most of the researchers studied the relation between the SPI and the vegetation indices like NDVI, TCI,

    and VCI. The tool for early detection of drought in East Asia based on NOAA-AVHRR NDVI data is

    developed. The NDVI time series is correlated with the SPI obtained for US Great Plains and found the

    best correlation value for 3 month SPI. In the proposed system the neural network architectures will be

    used to apply to classification task. Artificial neural network have been previously applied to drought

    prediction. To predict SPI values linear stochastic models and recursive multistep neural networks were

    applied and resulted good performance. The neural network model developed previously was based on

    SPI and did not use any satellite based data to predict the drought severity. The past research is made for

    small area and with small number of stations.

    III. DATA PREPARATION

    A. Region of study

    India is seventh largest country by area and located in South Asia. South of India is bounded by Indian

    ocean , South-west by the Arabian Sea, and South-west by the Bay of Bengal, land borders is shared with

    Pakistan to the west; China, Nepal, and Bhutan to the north-east; and Burma and Bangladesh to the east.

    The Indian climate is strongly influenced by the Himalayas and the Thar Desert, to drive the summer and

    winter monsoons in India both plays an important role. Cold Central Asian katabolic winds from blowing

    in prevented by Himalayas, keeping the bulk of the Indian subcontinent warmer than most locations at

    similar latitudes. The role of Thar Desert in attracting the moisture-laden south-west summer monsoon

    winds that, between June and October, provide the majority of India's rainfall. Four major climatic

    groupings predominate in India: tropical wet, tropical dry, subtropical humid, and mountain [10].

    B. Metrological Data [5] :

    For this study the Metrological data is used. Data can be monthly rainfall and temperature. Data

    reliability needs to be checked before applying it to the model extraction of SPI indices. Those station

    which does not provide sufficient metrological data are removed. Values could be reconstructed for some

    of the stations.

    There are various methods which are used to verify the consistency of the rainfall data. Such methods are

    double mass curve analysis, Von Neumann ratio test, and likelihood ratio test. Those stations showing

    inconsistency are removed from further analysis. Some stations having some missing data are needs to

    be reconstructed. Linear regression method is used to reconstruct missing data. Verification methods

    must be applied to the data before any other processing steps.

    SPI is one of the widely used metrics for drought modeling and forecasting. SPI not only capture

    information about the amount of rainfall but also provide a measure representing rainfall condition

    against a long term mean. It describes how the precipitation is more than normal in a period of time. It is

    invariant it can be calculated at any desired time. To describe the precipitation time series the probability

    density function should be determined. Gamma distribution is desired for precipitation data to makes

    normalized distribution. The cumulative probability of the precipitation is calculated and a normal

    Gaussian function with mean zero is applied to it. This generates the mean SPI for desired location and

    time to be zero. Table I shows the correspondence with SPI values and Drought severity conditions.

  • International Journal of Advance Foundation and Research in Computer (IJAFRC)

    Volume 1, Issue 8, August2014. ISSN 2348 - 4853

    3 | 2014, IJAFRC All Rights Reserved www.ijafrc.org

    To classify the stations into different climate zone the De Mortan coefficient (index)

    I = P/T+10

    Where T is annual mean temperature in C and P is annual mean precipitation in mm. Relation between

    De Morton index and climate zone is shown in table. The above classification is used to model the

    drought.

    C. AVHRR-NOAA Data [6]

    The drought monitoring carried out using indices derived from remote sensing data. The sensors are

    installed in earth observation satellites. Due to less water the capacity to carry out chlorophyll an

    function on the part of vegetation is reduced. In drought condition the vegetation is reduced due to lack

    of water for photosynthesis process. This occurrence given or described or demonstrated by spectral

    response. Because of this characterization, indices are introduced for drought modeling such as NDVI,

    NDVI-DEV, VCI, TCI.

    1. Normalized Difference Vegetation Index

    NDVI is most widely used index. It is used to calculate amount of vegetation cover in the land. It is

    calculated as

    N=[(bNIR-bred)/(bNIR+bred)]

    Where N is the NDVI and bNIR and bred are the red bands respectively.

    2. NDVI-DEV

    The intensity of drought may be calculated be deviation of NDVI from its long term mean. The difference

    between the NDVI for the current time and a long term mean NDVI for that month is named as deviation

    of NDVI.

    Ndev=Ni-Ni,mean

    Where Ni=NDVI value for month i and Ni,mean is long term mean for month i over the period. Positive

    values of Ndev indicate the above-Normal Vegetation condition and similarly Negative values of Ndev

    indicates the below-Normal vegetation condition and suggest drought situation. NDVI-DEV index has

    some limitations. The deviation from the mean does not take into account the standard deviation. Hence

    it can be misinterpreted if the variability in vegetation conditions in a region is very high.

    3. Vegetation Condition Index

    Vegetation condition index(VCI) is to calculate how close the NDVI of current month to the minimum

    NDVI of long term record. It is defined as,

  • International Journal of Advance Foundation and Research in Computer (IJAFRC)

    Volume 1, Issue 8, August2014. ISSN 2348 - 4853

    4 | 2014, IJAFRC All Rights Reserved www.ijafrc.org

    Vj= [(Nj-Nmin)/(Nmax-Nmin)]*100

    Where Vj is vegetation condition index for month j and Nmax and Nmin are the maximum and minimum

    value of VCI from long term record for that month. VCI is measured in perfect and it is approximate

    measure of how dry the current month j is.

    4. Temperature condition Index

    TCI includes the deviation of the current value from the maximum NDVI of long term records.

    Tj= [(TBmax-TBj) / (TBmax-TBmin)]*100

    Where Tj is TCI value of month j and TB, TBmax, TBmin are absolute maximum and minimum smoothed

    monthly brightness temperature. Moisture shortage is accompanied by high temperature which causes

    thermal effect. TCI gives an opportunity to identify subtle changes in vegetation health due to thermal

    effect.

    To calculate NDVI, NDVI-DEV, VCI and TCI, AVHRR data with spatial resolution of 1.1km X 1.1km can be

    used High temporal resolution low data volume and low cost compared to other high resolution satellite

    images are the advantages of NOAA-AVHRR data. The rough resolution can be the limitations of NOAA-

    AVHRR images. The data in high-resolution picture transmission (hrpt) format [7]. For radiometric and

    geometric correction on the images can be carried out by applying standard procedures. The number of

    cloud detection methods can be applied to remove cloud effect. Threshold tests are often used for cloud

    masking. Threshold test are based on the principle that is if the measured brightness temperature in one

    of AVHRRs infrared window channels is smaller than a predefined threshold or if the measured

    reflectance in one of AVHRRs visible channel is higher than a predefined threshold then the pixel is

    regarded as cloud contaminated.

    (b4gt 265) and ((b1+b2) lt 70)

    (b1+b2) lt 90

    (b4gt 285)

    Where b1, b2, and b4 are the corresponding AVHRR bands and gt and lt are greater than and less than

    respectively. Second expression is Reflectance test. By the use of Reflectance test, pixel is masked as

    cloudy one if reflectance of a pixel in visible bands is lower than the threshold value. The third expression

    is Brightness temperature test. Pixel is masked as cloudy one if the brightness temperature (band4) of a

    pixel is higher than the threshold value. First expression is Reflectance and brightness test. Because of

    insufficient contrast with the surface radiance, some cloudy types are difficult to detect such as thin

    cirrus, low status at night and small cumulus and such clouds can be detected by first expression.

    With the NDVI data related to each site, the GPS coordinates of the meteorological stations need to co-

    register. For each site two types of time series is constructed, one for SPI and one for satellite data.

    IV. MODELING/FORECASTING BASED ON NEURAL NETWORKS

    Aim of proposed system is to use the satellite images to model and predict the intensity of drought in

    India. The number of features is considered like NDVI, NDVI-DEV, TCI, and VCI. The input to the model is

    satellite images and the drought intensity based on SPI as its output.

    A. Neural Network models

    Artificial Neural networks have been applied on many applications such as modeling classification,

    smoothing, filtering, prediction, function approximation, and optimization. The individual neuron works

    simultaneously to model the complex task. The number of neural network architecture can be used to

    predict the drought condition. Prediction can be made based of features extracted from satellite images.

  • International Journal of Advance Foundation and Research in Computer (IJAFRC)

    Volume 1, Issue 8, August2014. ISSN 2348 - 4853

    5 | 2014, IJAFRC All Rights Reserved www.ijafrc.org

    The network such as MLPs and RBFs are universal approximations. They are used to approximate any

    function with a desired function. The Neural networks are correct choice for the prediction if they are

    trained properly. In previous research, the linear model such as regression and adaptive linear model

    were used to calculate the SPI value from the vegetation indices. Least mean square or Delta rules the

    methods can be used to determine the free parameters for these models. The drawback of the linear

    model is that they are incapable of capturing interrelations in highly nonlinear system.

    The MLP networks consist of two layers one is input and another is output layer. The following figure

    shows the general architecture of MLP network. They may consist of computational elements included in

    a number of hidden layer. Each neuron in hidden layer and output layer has a nonlinear function relating

    its output too its input. Let us consider neuron j in the output layer. The output calculated as

    Yj= (wkjyk)

    Where yk is the output of kth neuron. Kth neuron is in the hidden layer just before the output layer and

    wkj is the weight associated to the link between neuron. is the activation function of the neurons. After

    the invention of the back-propagation (BP) learning algorithm, the outputs of the neurons in the hidden

    layers can also be computed based on the outputs of their preceding layers. BP is simple learning

    algorithm based on the steepest descent optimization method. This method consists of two steps. First

    step is that the output of the neurons in different layers is computed and in second step the error signal is

    obtained.

    BP algorithm is the best tool for optimization task in MLP. MLPs with only one hidden layer do not

    produce the accepted results. As we increases the number of hidden layer will increase the accuracy and

    gives good performance. Usually two hidden layers are enough to get accepted performance. The error

    signal for neuron j in the output layer is obtained as

    ej = dj-yj

    Where dj is the desired output for that neuron. Yj is output of jth neuron. These error signals are then

    back propagated to the hidden layers. Then weights are updated using the following equation

    Wkj(new) = Wkj(old) + ykj

    Where Wkj is the weight of the link from neuron k to neuron j, is learning rate, yk is the output of

    neuron k and j is the local gradient of neuron j. where j of neuron in output layer is computed as

    j= -ej`(vj)

    and in the hidden layer is computed as

    k= `(vk) wkjj

    where vk is the input to neuron k, which is obtained by summing all the input neuron k receives from

    those of the previous layer.

    RBFs network is similar to MLPs network. RBFs Network consists of three layer, an input layer, hidden

    layer, and an output layer. The neurons in the output layer and the hidden layer are different as they are

    same in MLPs. The output neurons are linear and hidden neurons are nonlinear. Output of neuron j in

    the output layer is computed as

    yj= wkjexp(||x-ck||)

  • International Journal of Advance Foundation and Research in Computer (IJAFRC)

    Volume 1, Issue 8, August2014. ISSN 2348 - 4853

    6 | 2014, IJAFRC All Rights Reserved www.ijafrc.org

    Where wkj is the weight of link between neuron k in the hidden layer to neuron j in the output layer. x is

    the input vector to the network. ck is the center of the basis function of neuron k in the hidden layer.

    V. PREDICTION MODEL

    The standard autoregressive moving average model (ARMA) [8] needs to be considered to construct the

    model. The output of the model is SPI and features like NDVI, NDVI-DEV, VCI, and TCI are the input. Let

    us consider that the TCI is the input to the model. The general relation between SPI and TCI is

    S (t + p) = f (T (t), T (t-1), T(t-2),..T(t-q))

    Where S indicates the SPI values and t denotes the current time. The values of SPI are predicted p Steps

    ahead. q Indicates how many past samples of input is used in the model. F (.) is the function that is indeed

    determined by MLP and RBF.

    As the q increases, the accuracy of the model also increases with the price of increase in computation

    time. For the current month, to predict the SPI values, past 12 months TCI values are fed into the model

    as input. Next is to obtain the values of NDVI, TCI, VCI and NDVI-DEV for the pixel where the station is

    located. The window of 3x3 pixel size is considered with the station in the central pixels. By making the

    average of nine values, the values for central pixel are obtained. There may be possibility that the station

    is located at without enough vegetation, and surrounding lands have a vegetation cover. Averaging makes

    it possible to have smooth vegetation index [9][10].

    As mentioned in previous section India had different climate zones. The relation between the SPI and

    vegetation index depends upon the specific climate zones where the station is placed. According to the

    climate zones, separate model is needed for each zone leads to 4 models. To train the models the data

    from different stations can be used. The mean absolute error and accuracy are the metrics which can be

    used to achieve effectiveness of models. As shown in table I classification is the task in which

    classification of drought is done with respective of SPI values. Accuracy of the classification task is

    Accuracy= Noc/ No

    Where No are the total number of stations and Noc are number of stations those correctly predicted the

    drought conditions. Accuracy indicates number of stations for which the model is successful for

    prediction of drought.

    The process of data acquisition for proposed system is shown in figure 1. Block diagram comprise of

    various steps to calculate SPI. The steps are as follows:

    Step 1: Captured satellite image is processed by using the Matlab software.

    Step 2: The required features such as NDVI, TCI, VCI extracted from the satellite image and given as input

    to the model.

    Step 3: Model is used to analyze the given data to get the SPI.

    Step 4: Model is trained with existing SPI data to get better accuracy.

  • International Journal of Advance Foundation and Research in Computer (IJAFRC)

    Volume 1, Issue 8, August2014. ISSN 2348 - 4853

    7 | 2014, IJAFRC All Rights Reserved www.ijafrc.org

    Figure 1. Data acquisition block diagram

    VI. CONCLUSION

    Finally we conclude that the many researchers have worked to predict the severity of drought condition.

    The proposed system will decide which crop is needed to be taken in a particular drought condition

    strategy.

    VII. ACKNOWLEDGMENT

    I personally thank Prof. Hemant Palivela from Department of Information Technology, Mukesh Patel

    School of Technology Management and Engineering, NMIMS University for helping me understand the

    related topics of the field.

    VIII. REFERENCES

    [1] Keyantash, John, and John A. Dracup. "The quantification of drought: an evaluation of drought

    indices." Bulletin of the American Meteorological Society 83.8 (2002): 1167-

    1180.J.ClerkMaxwell,ATreatiseonElectricityandMagnetism,3rded.,vol.2.Oxford:Clarendon,1892,p

    p.6873.

    [2] Alley, William M. "The Palmer drought severity index: limitations and assumptions." Journal of

    climate and applied meteorology 23.7 (1984): 1100-1109.

    [3] Khan, S., H. F. Gabriel, and T. Rana. "Standard precipitation index to track drought and assess

    impact of rainfall on water tables in irrigation areas." Irrigation and Drainage Systems 22.2

    (2008): 159-177.

    [4] Ji, Lei, and Albert J. Peters. "Assessing vegetation response to drought in the northern Great Plains

    using vegetation and drought indices." Remote Sensing of Environment 87.1 (2003): 85-98.

    [5] Barlow, Mathew, Heidi Cullen, and Bradfield Lyon. "Drought in central and southwest Asia: La

    Nia, the warm pool, and Indian Ocean precipitation." Journal of Climate 15.7 (2002): 697-700.

    [6] Sobrino, J. A., N. Raissouni, and Zhao-Liang Li. "A comparative study of land surface emissivity

    retrieval from NOAA data." Remote Sensing of Environment 75.2 (2001): 256-266.

  • International Journal of Advance Foundation and Research in Computer (IJAFRC)

    Volume 1, Issue 8, August2014. ISSN 2348 - 4853

    8 | 2014, IJAFRC All Rights Reserved www.ijafrc.org

    [7] Miller, Peter, et al. "Panorama: a semi-automated AVHRR, and CZCS system for observation of

    coastal and ocean processes." Proceedings of the Remote Sensing Society, Reading. 1997.

    [8] Akaike, Htrotugu. "Maximum likelihood identification of Gaussian autoregressive moving average

    models." Biometrika 60.2 (1973): 255-265.

    [9] Bhuiyan, C., R. P. Singh, and F. N. Kogan. "Monitoring drought dynamics in the Aravalli region

    (India) using different indices based on ground and remote sensing data." International Journal of

    Applied Earth Observation and Geoinformation 8.4 (2006): 289-302.

    [10] Jain, Sanjay K., et al. "Application of meteorological and vegetation indices for evaluation of

    drought impact: a case study for Rajasthan, India." Natural hazards 54.3 (2010): 643-656.

    [11] Chang, Jen-Hu. "The Indian summer monsoon." Geographical review (1967): 373-396.