A New Approach of Probabilistic Cellular Automata Using Vector Quantization Learning for Predicting Hot Mudflow Spreading Area by ijcsis


More Info
									                                                               (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                               Vol. 9, No. 2, February 2011

  A New Approach of Probabilistic Cellular Automata
  Using Vector Quantization Learning for Predicting
           Hot Mudflow Spreading Area
                        Kohei Arai                                                              Achmad Basuki
            Department of Information Science                               1) Department of Information Science, Saga University
                    Saga University                                            2) Electronic Engineering Polytechnic Institute of
                       Saga, Japan                                                       Surabaya (EEPIS), Indonesia
               Email: arai@is.saga-u.ac.jp                                                Email: basuki@eepis-its.edu

Abstract— In this letter, we propose a Cellular Automata using             The previous approach assumes that hot mudflow has similar
Vector Quantization Learning for predicting hot mudflow                    characteristics to lava flow such as thermal changing, fluid
spreading area. The purpoe of this study is to determine                   mass transport rules and material mixing.
inundated area in the future. Cellular Automata is an easy                 It is difficult to describe some physical phenomena caused by
approach to describe the complex states of hot mudflow disaster            complex human made landscape objects such as levees,
that have some characteristics such as occurring on the urban
area, levees and surface thermal changing. Furthermore, the
                                                                           buildings, and other environmental properties. Avolio et al. [4]
Vector Quantization learning determines mass transport in the              have proposed an alternative Cellular using minimization
surrounding area in accordance with equilibrium state using                differences to simulate lava flow. This approach has
clustering of landslide. Evaluating of prediction result uses              stochastically state changing. The key-point of this approach is
ASTER/DEM and SPOT/HRV imaging. Comparison study shows                     easy to develop. Recently, D’Ambrossio et al. [5] and Del
that this approach obtains better results to show inundated area           Negro et al. [6] have applied the stochastic approach to
in this disaster.                                                          simulate soil erosion. This approach also uses minimization
                                                                           differences based on Cellular Automata for other fluid flow
    Keywords: Probabilistic cellular automata, vector quantization,        phenomena. The idea of the use of the stochastic approach
hot mudflow spreading, prediction, mass transport Introduction             makes the alternative approach describe complex landscape
                                                                           object problems on the hot mudflow disaster [7]. The problem
                       I.    INTRODUCTION                                  of this idea is how to fix probability value of mass transport on
Simulating hot mudflow in the plane and urban area requires                each neighbor-cell.
understanding how the surface changing properties vary with                The aim of this letter is a new approach of cellular automata
time and space. In order to generate complex flow about                    model for predicting hazardous area in the hot mudflow
interactions between natural and human made topography, we                 disaster. This approach uses some ideas such as minimization
need the model of the main mechanical features of hot mud                  difference model and vector quantization to make cluster of
depending on landscape data. Another difficulty is to compute              mass transport possibility depend on altitude, height of mud
the simulation of hot mudflow at acceptable rates. However,                and plant [8]. Because of cluster continuity by vector
they are difficult to apply in general conditions.                         quantization, it looks like the statistical behavior of landscape
Argentini [1] introduced a CA approach to simulate fluid                   object in the urban area. Vector Quantization determines
dynamic with some obstacles and fluid flow parameters. This                cluster of inundated area [9] that makes flow difference in
approach used basic rules in the two-dimensional spaces.                   neighborhood area easy to define in probability values. A
Vicari [2] introduce CA approach to simulate lava flow. This               similar approach has not yet been undertaken for mudflow and
approach used Newtonian fluid dynamic concept.                             lava flow in any other place, which appeared in the landslide
Combination of both approach obtained a discrete approach                  area. However, a simple cellular automata approach is
for predicting hot mudflow [3]. This approach yielded correct              considered there.
location and direction of hazardous area, but the intersection             Simulation results use the landscape map using ASTER DEM,
area between prediction area and real area of hazardous area is            and initial parameters of hot mudflow. This paper shows some
around 36.44%. This approach is a deterministic approach                   simulation result on map view in the varying time and
based on Cellular Automata to estimate the areas potentially               percentage of predicting performances. We also show the
exposed to hot mudflow inundation, concentrate mudflow                     comparison of predicting on inundated area and direction with
characteristics, combine fluid flow and lava flow properties,              the other previous approach.
and neglect difficulty to describe a model of complex human
made landscape data and random behavior of state changing.

                                                                      32                              http://sites.google.com/site/ijcsis/
                                                                                                      ISSN 1947-5500
                                                                 (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                 Vol. 9, No. 2, February 2011
  II.    OVERVIEW OF FLUID DYNAMIC CELLULAR AUTOMATA                       parameters such as viscosity and surface thermal changing.
    Most numerical approaches to modeling landscape                        This approach is powerful to simulate fluid flow and easy to
evolution simulate the physical flow such as mass transport of             develop.
fluid particles, erosive effects of water discharge, infiltration
and absorption by solving complex differential equations. CA                                 III.    PROPOSED APPROACH
is an alternative approach to simulate fluid flow using a simple
approach. The current implementation is primarily based on                 A. General Characteristic of Hot Mudflow Disaster
D’Ambrossio et al. [5] because it uses "very simple                            On 29 May 2006, the gas exploration operation had caused
approximations intended to describe complex geographical                   cauldron of hot mud in 6.3 km depth spray out hot mud to
effect" and it able to offer "insight into how thermal and                 surrounding areas on Sidoarjo, East Java, Indonesia
viscous fluid parameter affects the evolution of landscapes"               (7.530553°S; 112.709684° E) [13][14]. This disaster located at
despite its simplicity.                                                    the urban area near Sidoarjo (Figure 2-top). Hot mud had
    The CA algorithm simulates first-order processes                       spilled over 5000 m3 per-day. It increased over 170,000 m3
associated with fluvial erosion by iteratively applying a set of           per-day as reported by Cyranoski [15] and over 150,000 m3 as
simplified rules to individual cells of a digital topographic grid         reported by Harsaputra [16].
[10]. The state represents a number of fluid particles in the
topographic grid, and the subsequent movement and behavior
(diffusion, and erosion) of the cell is controlled by the rules and
a few parameters of the current cell and its surrounding
neighbors [11]. The same rules are applied to all grid cells, i.e.,
there is no outside-imposed distinction between slope and
channel; the model forms its own channels [11].
    Figure 1 illustrates how the algorithm works. For example,
fluid particles move to lower elevations, simulating fluid flow
in the landslide grid. There are two varying flows; erosion and
diffusion. The amount of erosion and diffusion each produces
is proportional to the local slope, simulating speedier erosion of
steeper slopes and lesser erosion of hard rock surfaces.

        Figure 1. Schematic diagram showing how CA model works

    Xiaoming Wei [12] introduced the simple CA approach for
highly viscous fluid. Its movement is mainly a result of gravity,
viscosity damping and friction. This approach uses four
variables to indicate the expanding potential of a liquid cell;
there is solid, liquid, amount of material and energy. Setting a                       Figure 2. The location of hot mudflow disaster
certain threshold for this variable enables to control the
expanding behavior of the liquid. For each liquid cell, if its                 Hot mudflow had an immense impact on environment,
energy is higher than a certain threshold, it has the potential to         economic and human resource in the future if no
spread along its horizontal neighboring cells [17]. This                   countermeasure is conducted (Figure 2-bottom) [17]. Within
approach uses four nearest neighbors and four second nearest               the first two years, the mud flow disaster destroy some villages,
neighbors.                                                                 farm lands, factories and public facilities such as schools,
                                                                           markets, roads, water pipes and gas pipes. Over 17,000 people
    Another CA approach to simulate fluid flow uses the
                                                                           had lost their houses and jobs. If facts, approximately mud
minimization difference approach that was introduced by
                                                                           blows out 150,000 m3 per-day with the assumption that
Avolio [4] and D’Ambrossio [5]. This approach is one
                                                                           contains 70% by water. This implies that water come out by
alternative approach to solve fluid dynamic without
                                                                           687,000 barrel a day. This situation is different from some
sophisticated mathematical formulation. It obtains a
                                                                           disaster areas where the previously occurred other locations
satisfactory model to simulate the lava flow with various
                                                                           because it has overmuch mud [18].

                                                                      33                                http://sites.google.com/site/ijcsis/
                                                                                                        ISSN 1947-5500
                                                                       (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                       Vol. 9, No. 2, February 2011
    Although one possible solution is spillway to Porong River,                   approach. The algorithm of Minimizatin Differences is as
it does cost and takes a long time and vast human resource.                       follow:
Therefore, strong demands on prediction of mudflow spreading
volume and mudflow disaster area as well as on how to                             (a) A is the set of cell not eliminated. Its initial value is set to
evacuate from the area of which the levee that was constructed                        the number of its neighbors. Each cell on position (i,j)
to prevent mudflow spillover are there for people who are                             has two components such as soil and mud. The height of
living in the disaster areas. If inundated area are predicted                         them are gij and sij. Total height of this cell is: hij = gij +
before the mud comes, the Indonesia government makes                                  sij. There is dynamic soil uij, but it is the small portion of
countermeasures to reducing the impact.                                               soil and we adjust on normal distribution of pm.
                                                                                  (b) The average height is found for the set of A of non-
    This simulation uses map on February 2008 (Figure 3a) as                          eliminated cells:
initial map and map on August 2008 as target map (Figure 3b).
This map is landscape approximation using ASTER/DEM and                                                      hc + ∑ ci .hi
                                                                                                                       i∈ A
the height data on the some observation points. The map size is                                           m=                                              (1)
approximate 3.705km×4.036km. The red area is mud inundated                                                           nA + 1
area. In this simulation, mud blows from the main crater (big                           Where:
hole) that has a diameter around 20m [8], and mud moves to                              hc is height of the center cell.
other locations depend on slope difference and mudflow                                  hi is height of the non-eliminated neighbor cells.
parameters. The key process is mass transport that defines the                          nA is number of non-eliminated neighbor cells.
amount of mud moving.                                                                   c is current mass-transport weighting from the learning
                                                                                  (c) The cells with height larger than average height are
                                                                                        eliminated from A.
                                                                                  (d) Go to step (b) until no cell is to be eliminated.
                                                                                  (e) The flows, which minimize the height differences locally,
                                                                                        are such that the new height of the non-eliminated cell is
                                                                                        the value of the average weighting height.
                                                                                                               ∑ ci .hi
                                                                                                          hi = A                           (2)
                                                                                      When we used probability adjustment depend on height
             (a)                                   (b)                            differences in the previous research, we use Vector
                                                                                  Quantization learning to make cluster space of mass transport
 Figure 3. (a) Initial map on February 2008, (b) target map on August 2008
                                                                                  as a probability adjustment in the neighborhood area. We select
                                                                                  some points in the previous map and the nearest points in the
B. Model Definition                                                               current map as paired point. We use standard competitive
    This model is 2D CA model. It uses two-dimensional grids                      learning to determine height of points around the surrounding
to describe set of cells. The state of cell S is floating point value             area.
that shows the amount of mud and soil particles. In this
research, we define two-type variables of state; the amount of
                                                                                              c new = c old + τ c pair + c old   )                        (3)
mud st(x,y) and the amount of soil ht(x,y). Mud is moving                         Where:
material. It moves from one cell to its neighbors using
probability of move pmov. The other hand, the small part of mud                   c new is a new inundated point in the surrounding area.
also changes into the soil using probability of deposition pvis.                  c old is an inundated point in the previous map.
The model state is as shown in Figure 4.                                          c pair is an inundated point in the current map.
                                                                                  τ is a learning rate.
                st(x,y)                                                               In each point, there are some parameters that influence of
                                      pvis                                        mass transport on simulation process such as altitude (ground
                                                                                  height), mud height and landslide [8]. Because of the
                                                                                  discontinuous distribution of abrupt mass movement hazards
                      Figure 4. Mud and soil states.                              [19], VQ obtains an alternative method to quickly assess the
                                                                                  degree of hazard for each unit. It creates groups without
                                                                                  considering whether or not the units in the same group are
C. Model Definition
                                                                                  continuously distributed. Figure 5 shows the processing
   In this research, we use probability Cellular Automata                         schema of hot mudflow spreading simulation. The learning
based on Minimization Differences [5][7] as the main                              process using vector quantization determines a cluster space
                                                                                  that describes the probability of mass transport. The probability

                                                                             34                                    http://sites.google.com/site/ijcsis/
                                                                                                                   ISSN 1947-5500
                                                                          (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                          Vol. 9, No. 2, February 2011
values add some weighting under flow process in minimization                         resolution; minimization differences algorithm (48.15%-
differences approach.                                                                65.67%) in our previous research, Avolio’s approach (45.75%-
                                                                                     63.34%) and Vicari’s approach (43.25%-60.25%). Comparison
                                                                                     of these methods is shown in figure 8.

       Figure 5. The schematic of hot mudflow spreading simulation

                      IV.       SIMULATION RESULTS                                                 (a)                                    (b)
   In this simulation, we use the current resolution of
ASTER/DEM (30m×30m). The mud blow volume is around
150.000 m3 per day using Gaussian random number around this
volume. The mixing particle is 70% water and 30% solid

A. Simulation Results
    The simulation result is shown as Figure 6. In this figure,
we show the total inundated area (Figure 6a) and the new
inundated area (Figure 6b). The red area is the real inundated
area, the blue area is the predicted area, and the pink area is
intersection between real area and predicted area. In Figure 7a,
the intersection area is above 95% that show this approach                                         (c)                                    (d)
yield a good result of prediction. It is not fair because the                         Figure 7. Comparison of (a) Vicari’s approach, (b) Avolio’s approach, (c)
prediction accuracy is only for new inundated area. Therefore,                           CA using Minimum Difference approach, (d) CA using VQ approach
we compare the predicted area and the real area in new
inundated area only. Figure 7b shows that the intersection area
in new inundated area is 71.85%. This result is better that the
previous result that uses minimization difference approach
(56.44%) [7]. Figure 7 shows the comparison between this
approach and other approach.

                                                                                                  Figure 8. Comparison with the other approaches

                                                                                     B. Resolution Influences
                                                                                         This simulation runs in some resolution. In normal size, we
                                                                                     use ASTER/DEM map that has resolution 30m and image size
                                                                                     300x300 pixels. The minimum resolution is 200 pixels (map
                          (a)                    (b)                                 resolution is 45m). The maximum resolution is 700 pixels (map
 Figure 6. The simulation result: (a) total inundated area, (b) new inundated        resolution is 12.9m). The prediction performance increases by
                         area using this approach                                    increasing resolution and become stable on higher resolution as
                                                                                     shown in Fig. 9. This figure shows there are two peak points of
    Figure 8 shows combination of CA approach and online                             intersection area; in resolution 30m and in resolution 20m.
clustering using vector quantization obtain better performance                       They occur because the resolution of our ASTER/DEM data is
to predict new inundated area (54.13-69.13%) than previous                           30m, and we use another data (height data on critical points)
methods in 3x3 Von-Newmann neighborhood system in all                                that have resolution 20m.

                                                                                35                                   http://sites.google.com/site/ijcsis/
                                                                                                                     ISSN 1947-5500
                                                                          (IJCSIS) International Journal of Computer Science and Information Security,
                                                                                                                          Vol. 9, No. 2, February 2011
                                                                                             Spatial Information Science, Volume XXXVIII, Part 8, pp. 237-242,
                                                                                             Kyoto Japan 2010.
                                                                                      [8]    H. A. Nefeslioglu, E. Sezer, C. Gokceoglu, A. S. Bozkir, and T. Y.
                                                                                             Duman, Assessment of Landslide Susceptibility by Decision Trees in the
                                                                                             Metropolitan Area of Istanbul, Turkey, Mathematical Problems in
                                                                                             Engineering Volume 2010, Article ID 901095, 2001.
                                                                                      [9]    Li-Chiu Chang, Hung-Yu Shen, Yi-Fung Wang, Jing-Yu Huang, Yen-
                                                                                             Tso Lin, Clustering-based hybrid inundation model for forecasting flood
                                                                                             inundation depths, Journal of Hydrology 385 (2010) 257–268.
                                                                                      [10]   Wei Luo, Kirk L. Duffin, Edit Peronja, Jay A. Stravers, and George M.
                                                                                             Henry, 2003, A Web-based Interactive Landform Simulation Model
                                                                                             (WILSIM), Computers and Geosciences, accepted Nov., 2003.
                                                                                      [11]   Chase, CG., 1992. Fluvial land sculpting and the fractal dimension of
                                                                                             topography. Geomorphology 5, 39-57. Department Riello Group,
            Figure 9. Prediction performance for each resolution                             Legnago (Verona), Italy, February 2003.
                                                                                      [12]   Xiaoming Wei, Wei Li and Arie Kaufman, Interactive Flowing of
                                                                                             Highly Viscous Volumes in Virtual Environments, Proceedings of the
                      V.     CONCLUSION REMARKS                                              IEEE Virtual Reality 2003 (VR’03).
   Through the simulation study with the proposed model                               [13]   Mazzini, A., Svensen, H., Akhmanov, G.G., Aloisi, G., Planke, S.,
based on Cellular Automata, we may conclude the following,                                   Malthe-Sφrenssen, A., Istadi, B., 2008, Triggering and dynamic
                                                                                             evolution of the LUSI mud volcano, Indonesia, Eart and Planetary
(1) The using vector quantization learning in CA approach                                    Science Letters, Vol. 261, No. 375-388.
    obtain much better performance to predict new inundated                           [14]   Manfred P Hochstein, Sayogi Sudarman, Monitoring of LUSI Mud-
                                                                                             Volcano - a Geo-Pressured System, Java, Indonesia, Proceedings World
    area in hot mudflow disaster.                                                            Geothermal Congress 2010.
(2) The prediction performances depend on resolution.                                 [15]   Cyranoski, D., 2007, Muddy Waters: Hot did a mud volcano come to
    Increasing resolution will increase the prediction                                       destroy an Indonesian Town?, Nature, Vol. 445, 22 February 2007.
    performance and become stable in the higher resolution.                           [16]   Harsaputra, 2007, I., Govt. weight option for battling the sludge, The
(3) The dangerous levee location for spillover can be found                                  Jakarta Post, 29 may 2007.
    with the proposed method.                                                         [17]   Sjahroezah, A.: Environmental Impact of the hot mud flow in Sidoarjo,
                                                                                             East Java. The SPE Luncheon Talk, 19 April 2007.
(4) Cell size effect is clarified. By considering the resolution
                                                                                      [18]   Pramadihanto, D., Basuki A., Barakbah A.R., 2007, “Global Disaster
    of data sources, the resolution of ASTER derived DEM                                     Managemnet System: A Local Disaster Management Model and
    (Digital Elevation Model) is 30m, the most appropriate                                   Knowledge Connecntion between NiCT – EEPIS Inherent Network Case
    number of cells of CA is determined with these                                           Study: Sidoarjo Mud Volcano”, The First International Symposium on
    resolutions.                                                                             Universal Communication (ISUC), Kyoto, 14-15 June 2007.
                                                                                      [19]   J.R. Ni, R.Z. Liu, Onyx W.H. Wai, Alistair G.L. Borthwick, X.D. Ge,
                               REFERENCES                                                    Rapid zonation of abrupt mass movement hazard: Part I. General
                                                                                             principles, Geomorphology 80, pp. 214–225, 2006.

[1]   Argentini G, 2003, A first approach for a possible cellular automaton
      model of fluids dynamic. Computer Science - Computational                                                   AUTHORS PROFILE
      Complexity, arXiv:cs/0303003v1.
[2]   Vicari A, Alexis H, Del Negro C, Coltelli M, Marsella M, and Proietti C,        Kohei Arai
      2007, “Modeling of the 2001 Lava Flow at Etna Volcano by a Cellular             He received BS, MS and PhD degrees in 1972,74 and 82, respectively.
      Automata Approach”, Environmental Modelling & Software 22,                      He was with The Institute for Industrial Science and Technology of the
      pp.1465-1471.                                                                   University of Tokyo from April 1974 to December 1978 and also was with
                                                                                      National Space Development Agency of Japan from January 1979 to March
[3]   Kohei Arai, and Achmad Basuki, 2010, A Cellular Automata Based
                                                                                      1990.During from 1985 to 1987, he was with Canada Centre for Remote
      Approach for Prediction of Hot Mudflow Disaster Area, Computational
                                                                                      Sensing as a Post Doctral Fellow of National Science and Engineering
      Science and Its Applications – ICCSA 2010, Part II, Lecture Notes in
                                                                                      Research Council of Canada.He moved to Saga University as a professor in
      Computer Science 6017, Springer-Verlag Berlin Heidelberg, pp. 119-
                                                                                      Department of Information Science in April 1990.He was councilar for the
                                                                                      Aeronoutics and space related technology committee of the Ministry of
[4]   Avolio MV, Di Gregorio S., Mantovani F., Pasuto A., Rongo R., Silvano           Science and Technology during from 1998 to 2000. He was councilar of the
      S., and Spataro W. (2000), Simulation of the 1992 Tessina Landslide by          Saga University for 2002 and 2003. Also he was executive councilar for the
      a Cellular Automata Model and Future Hazard Scenarios, International            Remote Sensing Sciety of Japan for 2003 to 2005. He is now Adjunct Prof. of
      Journal of Applied Earth Observation and Geoinformation, Volume 2,              the University of Arizona, USA since 1998. He also is Vice Chiarman of the
      Issue 1, pp.41-50.                                                              Commission A of ICSU/COSPAR sice 2008. He wrote 26 books and
[5]   D’Ambrosio D., Di Gregorio S., Gabriele S. and Claudio R. (2001), A             published 227 journal papers.
      Cellular Automata Model for Soil Erosion by Water, Physic and
      Chemistry of The Earth, EGS, B 26 1 2001, pp.33-39.
                                                                                      Achmad Basuki
[6]   Ciro Del Negro, Luigi Fortuna, Alexis Herault, Annamaria Vicari                 He received BS and MS degrees in 1992 and 2002 respectively.
      (2008), Simulations of the 2004 lava flow at Etna volcano using the             He was with Electronic Engineering Polytechnic Institute of Surabaya from
      magflow cellular automata model, Bulletin of Volcanology, Volume 70,            April 1994. Now he studies at Department of Information Science, Saga
      Number 7/May, 2008, pp. 805-812, Springer Berlin/Heidelberg, 2008               University for PhD Degree from April 2009. His field is Disaster Spreading
[7]   Kohei Arai, Achmad Basuki, Simulation Of Hot Mudflow Disaster With              Modeling. He wrote 6 books in Indonesian language and published 20
      Cell Automaton And Verification With Satellite Imagery Data,                    publication papers for conferences and journals.
      International Archives of the Photogrammetry, Remote Sensing and

                                                                                 36                                     http://sites.google.com/site/ijcsis/
                                                                                                                        ISSN 1947-5500

To top