See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220705259

Development of an Optimal Neural Network for Avalanche Forecast in

Himalayan Region.

Conference Paper · January 2006

Source: DBLP

CITATION READS
1 135

4 authors, including:

Virender Rihani
Shoolini University
12 PUBLICATIONS 61 CITATIONS

SEE PROFILE

All content following this page was uploaded by Virender Rihani on 10 June 2016.

The user has requested enhancement of the downloaded file.

 Development of An Optimal Neural Network For
Avalanche Forecast In Himalayan Region

V.Rihani Atul Parti

Head, Dept. of E&E, PEC Student of ME (CSE & IT), PEC
Sector-12, Chandigarh Sector-12.Chandigarh.
[email protected] [email protected]

Mahesh Bansal Rashpal Kaur

Student of ME, PEC Scientist-“C”
Sector-12.Chandigarh. Snow & Avalanche Study Establishment
[email protected] [email protected]

Abstract - This paper deals with the application temperatures. This data has secondary relevance in
of a well-known data mining technique, multi-layer avalanche formation.
back-propagation neural network, for forecasting of Class III: Meteorological factors, such as the
an avalanche in Himalayan region. Metrological and temperature, winds, precipitation and humidity which
snow data for Himalayan region has been used for bears indirect relevance to avalanche formation.
training the neural network. EasyNN-plus 6.0g, neural Primarily avalanche forecasting [4] is done by an
network software for Microsoft windows, is used for expert, acting on their knowledge and experience with
the development of an optimal neural network- snow stability situations. The physical [5] principles
PUSHPDEV. The system tries to model the decision underlying avalanche formation remain an area of
making process of a pragmatic expert. PUSHPDEV active research as they are not yet fully understood. A
can forecast whether an avalanche will trigger on a forecasting tool can therefore be used as an aid to
particular day from November to April. The network expert in decision-making for prediction of an
accepts eighteen inputs and produces an output whose avalanche. Different types of techniques are used in
value is zero or one, zero for no avalanche and one for avalanche forecast. They can be categorized as non-
avalanche on that day. computer assisted [6] and computer assisted
techniques [4-5][7-17].
Keywords: Avalanche, data mining, data warehouse Avalanches in Himalayan are a serious threat to
both human as well as to property. Avalanches results
1 Introduction in closure of roads which affects the necessary
supplies and services to both army and humans. Fuzzy
An avalanche is a mass of snow or ice that slides rule bases system [14] and nearest neighbour method
down the side of a mountain rapidly. Avalanches [1] [17] are two computer assisted tools, used for
are generated by structural weakness in the snowcover. avalanche forecasting in Himalayan region. The need
Avalanches occur due to the failure of the snowpack to for development of another tool is to use it along with
withstand the stresses placed upon it by its self weight already existing tools so as to have a better avalanche
and other imposed forces. The forecasting factors of warning system for Himalayan region.
avalanche are divided into three classes [2-3]. The Data mining [18] is known as “knowledge
higher the class number less relevance is the data for discovery in databases (KDD)”. It is the process of
avalanche formation. These classes are: extraction of patterns representing knowledge stored
Class I: Stability factors, including current avalanche in either data warehouse or databases. Data mining is a
activity, stability tests (e.g. Rutschblock, Shovel test) multidisciplinary field, drawing work from area
as well as other signs of instability, such as cracking of including database technology, artificial intelligence,
the snowcover. This is most relevant data for neural network, machine learning, statistics,
avalanche formation. knowledge acquisition, high performance computing,
Class II: Snowpack factors, such as the structure of information retrieval, and data visualization. C4.5, C5,
the snowpack, including the existence of weak layers, CART and feedforward backpropagation are different
crystal forms and sizes, densities and snow types of data mining algorithm used. Artificial neural
network (feedforward backpropagation), a form of speed, fresh snow amount, fresh duration and standing
data mining, has the advantage that it can handle snow etc. These parameters are taken twice in a day, at
inconsistent data. Artificial neural network can 8:30 and 17:30 hours.
therefore is a good choice for forecasting problems 2) Avalanche occurrence data, which specify when
like avalanche where data collected over years is an avalanche occurred.
inconsistent. Artificial neural network is an attempt to
mimic the biological decision making process of the 3 Method
human. Artificial neural network [18] development is
a two-step process. In the first step, supervised type of The development of an optimal neural network
learning is used to build a model. In step two, the has three main processes: objective determination,
model is used for the classification and the accuracy is data preparation and neural network modeling. Detail
estimated. If estimated accuracy of the model is descriptions of these processes are
acceptable then the model is used for prediction.
Artificial neural networks have been developed in 3.1 Objective Determination
south central Colorado of United States [4]. They uses The basic aim of this process is to obtain clear
fifteen metrological and snow parameters, and output understanding of the model to be developed. Objective
is an index whose value ranged from zero to one, with determination leads to the actual specification to be
higher values corresponding to increased likelihood's used for the development of the model. The output of
of avalanche activity. These models have the accuracy this phase is the problem definition and expected
of predicting avalanche activity up to 78 to 91 percent result. Problem definition for our work is “To develop
at different learning rate. an optimal neural network model to predict whether
EasyNN-plus 6.0g, is neural network software for an avalanche will trigger on a particular day or not in
Microsoft windows, is used for the development of an Himalayan region.” The expected output of the model
optimal neural network. All the component parts of is a value, which is zero or one, zero for no avalanche
EasyNN-plus are implemented as C++ reusable and one for avalanche on that day.
classes. The inspiration to use EasyNN-plus was based
on the fact that it has a graphical user interface and
Data Pre-Analysis
backpropagation algorithm. EasyNN-plus uses double
linked lists for implementation to store the examples,
the nodes and the connections. The lists can be
processed quickly in both directions simultaneously, Data-Preprocessing
and it can also be extended and contracted
dynamically.
Table 1 Data used in PUSHPDEV Data Post-Analysis
1999 2001 2002 2003 2004 2005
Figure 1 Data preparation scheme
Jan √ √ √ √ √ *
Feb √ √ √ √ √ * 3.2 Data Preparation
Mar √ √ √ √ √ * Data preparation is an important and critical step
April √ √ √ √ √ * in neural network modeling. Data preparation
Nov √ √ √ √ * includes all operation that is carried out on data for
Dec √ √ √ √ * development of the model. Data preparation [19]
√ : Data used for training * : Data used for testing includes data pre-analysis, data preprocessing and data
post-analysis. Data preparation scheme is shown in
Figure 1.
2 Data
The present study lies in Lower Himalayan Zone 3.2.1 Data Pre-Analysis
and data for a particular site, is provided by SASE Data pre-analysis phase includes collection of
(Snow & Avalanche Study Establishment). It includes metrological and snow data for last few years. It also
metrological and snow data, which falls in Class III, identifies main data variables from the collected data,
least relevant data for avalanche forecast. Data which was used later for the development of the model.
available for different years is shown in Table 1. The selections of variables were based on their
Following is the description of data. importance to problem and quantity of data available.
1) Daily weather data, which records meteorological The parameters identified and used as input to
and snow parameter like minimum temperature, PUSHPDEV are shown in Table 2.
maximum temperature, dry temperature, average wind
 Table 2 Eighteen input to the PUSHPDEV 3.2.3 Post Analysis
Sr.
No.
Input Last step in data preparation is post analysis,
1 Max temperature which include dividing data into two parts, one for
2 Min temperature training and other for testing to verify the network
3 Dry temperature performance. Data divided for training (including 99
4 Dry temperature difference records randomly selected for validation) and testing
5 Average maximum(Average of last three reading) (for an entire season) is shown in Table 1.
6 Average minimum(Average of last three reading)
7 Difference of average maximum(Difference of current
and last average maximum temperature) 3.3 Neural Network Modeling
8 Difference of average minimum(Difference of current Neural network modeling is an iterative process.
and last average minimum temperature)
9 Pressure difference
It includes selecting an initial configuration. Then a
10 Average speed number of experiments with each configuration are
11 Difference of average speed performed and the best network is retained. On each
12 Fresh snow continuous (sum of current and last fresh experiment, if under-learning occurs, then new
snow amount) network is tried by adding more neurons to the hidden
13 Fresh snow duration layer/layers or by adding an extra hidden layer. If
14 Fresh snow water equivalent (log (1+x)/log181, where over-learning occurs, then new network is tried by
x is the value of Fresh snow water equivalent).
removing hidden units. After training the network
15 Standing Snow
16 Difference in snow temperature
under predefined average error, network performance
17 Snow Penetration below Surface(Is the depth of is verified against data kept for testing. If network
snowpack up to which the ram penetrometer penetrates performance against test data is satisfied, then the
with its own weight) network is selected.
18 Characteristic of weather during proceeding one hour Table 3 Specification of PUSHPDEV
(new code developed by Code conversion method Number of input neurons 18
describes later.)
Number of output neuron 1

3.2.2 Data Preprocessing Number of neuron in first hidden layer 11

Number of neuron in second hidden
Data pre-analysis step was followed by data layer 8
preprocessing. Data preprocessing is used to extract
key features from data and transforming data in the Learning rate 0.35
form that neural network can accept. Data Momentum 0.35
preprocessing improves the quality of data, which Average training error 2%
influence the neural network models to be built. High Average validation error recorded 5.20%
quality data improves performance and learning time
for the network. Different types of operations were Validating (99 random example) 96.00%
performed on data to yield good results. These Number of cycles 3121
operations include substituting data for missing values, Accuracy of model for a season 87.38%
taking average, difference of values/average and log of
values for different parameters. A new method called
“Code-Conversion” is proposed, for the use of non- 4 Main Result
numeric data in artificial neural network. Non-numeric Models were developed by using multi-layer
data used in our case is “characteristic of weather (feedforward) backpropagation algorithm, with
during proceeding one hour”, is the weather code neurons on each layer fully connected to the neurons
assigned for different weather condition. Their impact on the following layers. Various architectures were
on avalanche is not known, but they play an important considered and selected model’s specifications are
role in the avalanche to trigger. The proposed method shown in Table 3. Accuracy of PUSHPDEV is defined
calculates the number of times each code exists and as the ratio of correct predictions against total number
the number of times avalanche has triggered for that of predictions made. It shows 87.38% overall accuracy
code, and then their ratio is calculated. This ratio is for an entire season i.e. 1st Nov, 2004 to 19th April,
called “Conversion Ratio”. New codes are assigned 2005.
corresponding to the conversion ratio that is higher
conversion ratio has higher new code. The idea of 5 Conclusions
introducing new code is to eliminate longer learning
time and improve performance of trained network. The present study shows encouraging results for
using artificial neural network to assist an expert in
 prediction of an avalanche. The performance of neural Avalanche Review, Vol 7, No. 6, 1989.
network depends upon data used, which in our case [7] Barbolini, M. and Keylck, C.J, “A new method
falls in Class-III, is the least relevant data for for avalanche hazard mapping using a
avalanche forecast. Further artificial neural network combination of statistical and deterministic
can be developed by using stratigraphy data of the models”, Natural Hazards and Earth System
snowcover, which falls in Class II. However, it is Sciences (2002) 2: 239-245, 2002.
important to realize that one particular model is not [8] Lehning, M., Bartelt, P., Brown B., Fierz, C., and
suited to all areas on the globe. This is especially true Satyawali, P., “A physical SNOWPACK model
when attempting to model avalanche forecasts in the for the Swiss avalanche warning Part II. Snow
Himalayan region. The model developed is for a microstructure”, Cold Regions Science and
particular site in Himalayan region and this model Technology 35, 147– 167, 2002.
may not be a good choice for forecasting of an [9] Leprettre, B., Navarre, J.P., Taillefer, A.,
avalanche in all parts of Himalayan region. For Danielou, Y., Panel, J.M and Touvier, F.,”
different location/site new model should be developed Reliable Estimation of Avalanche Activity Using
based on their historical data. The performance of Seismic Methods”, ISSW Proceeding, 1996.
these models will entirely depends upon the quality [10] Zeidler, A. and Jamieson, B. ”Estimating the
and quantity of data collected over the period of time. strength of faceted snow layers for an avalanche
As each model has its own advantages, so hybrid forecasting model”, Proceeding of the
models can be developed to incorporate advantages of International Snow workshop in Penticton, British
more than one model. Columbia, 29-sept to 4 oct, 2002.
[11] Davis, R. E., K. Elder, D. Howlett and E.
6 Acknowledgement Bouzaglou, “Storm and weather factors related to
dry slab avalanche activity”, Proceedings -
We thank Dr. R.N.Sarwade, Director, Scientist- International Snow Science Workshop 1998,
‘G’, Snow & Avalanche Study Establishment, Him pp.25-34, 1998.
Parisar, Plot No. 1, Sec-37-A, Chandigarh. We also [12] Davis, R. E., and Elder, K. “Application of
like to thanks A.Ganju, joint Director, Scientist-‘F’, Classification and Regression Trees: Selection of
Snow & Avalanche Study Establishment. Last but not Avalanche Activity Indices at Mammoth
the least all the technical staff of Snow & Avalanche Mountain”. Proceeding International Snow
Study Establishment, whose help and support at all Science Workshop, ISSW’94, P.O. Box 49,
time resulted in the development of the project. Snowbird, Utah, USA, 285-294, 1994.
[13] Schweizer, J., Fohn, P.M.B,” Two expert systems
7 References to forecast the avalanche hazard for a given
[1] LaChapelle, E.R., “Avalanche forecasting- A region”. International Snow Science Workshop,
modern synthesis”. Int Association Science Snowbird, Utah, USA, 1994.
Hydrology, Pub. 69, pp. 75-84, 1965. [14] Pant, L.M., and Ganju, A., “Fuzzy rule-based
[2] McClung, D.M, Schaerer, P, 1993: The system for prediction of direct action avalanches,
Avalanche Handbook: The Mountaineers, Seattle, Current Science, Vol. 87, No.1, pp 99-104, 2004.
USA. [15] Gassner, M., and Brabec, B.,” Nearest neighbour
[3] Zeilder, A. and Jamieson, B., “Computer Assisted models for local and regional avalanche
Avalanche Forecasting: Skier-Triggered forecasting”, Natural Hazards and Earth System
Avalanche”, Paper presented Western Snow Sciences, 2: pp 247-253, 2002.
Conference, Vancouver, 2004. [16] Singh, D., Ganju, A. and Singh, A., “Weather
[4] Stephens,J., Adams,E., Huo, X., Hicks, J., and prediction using nearest-neighbour model”,
McCarty,D., “Use of neural networks in Current Science, Vol. 88, No. 8, pp 1283-1289,
avalanche hazard forecasting” ,Proceedings - 2005.
International Snow Science Workshop,30 [17] Singh A, and Ganju A., “A supplement to nearest-
October-3 November, Snowbird, Utah, pp. 327- neighbour method of avalanche forecasting”, Cold
340, 1994. Region Science and Technology (39), pp.105-113,
[5] Joseph, S.W., Lakeman, G.P.,” An expert system 2004.
to support snow avalanche forecasting”, IEEE [18] Han, J, Kamber, M., “Data Mining: Concepts and
International Conference on System, Man and Techniques”, Morgan Kaufmann Publisher, 2001.
Cybernetics (3), pp. 22-25, 1995. [19] Yu, L., Wang, S., Lai, K.K., “An Integrated Data
[6] Fohn, P.M.B., “The ‘Rutschblock’ as a practical Preparation Scheme for Neural Network Data
tool for slope stability evaluation”. The Analysis”, IEEE Transactions on Knowledge and
Data Engineering, Vol. 18, No. 2, February 2006.

View publication stats