Categorization of Snow Profile Data into Predefined Class Templates

Christian Schaiter

Categorization of Snow

Profile Data into

Predefined Class

Templates

Academic advisors: Univ.-Prof. Dr. Günther Specht

DI Robert Binna

MASTER THESIS

Leopold-Franzens-University Innsbruck

Institute of Computer Science

Categorization of Snow Profile Data into Predefined Class Templates 2

Content

Introduction and motivation 1

What you may expect in the next 15 minutes...

Overview of the developed system 2

Alignment of hardness profiles 3

Conclusion 4


Overview


Current chapter: Introduction


Conclusion


4


Introduction and motivation

Master thesis was done in cooperation with the Tyrolean

avalanche warning service (TAWS)

Goal: Find a way to

determine the potential

magnitude of an

avalanche

Aim of the thesis



TAWS frequently takes snow profiles from different areas

Most relevant property of a snow profile for this thesis:

Hardness profile

Snow profiles, hardness profiles, template types

Sn

ow

heig

ht

[cm

]

Snow hardness

Hardness

profile



Some example hardness profiles (collected from the TAWS):




Idea: Introduce a group of predefined hardness profile types

(template types, class templates)

These template types describe the overall composition of the

snowpack

They may be used to

estimate the magnitude

of avalanches

10 types defined so far


Profile types (class templates)




Hardness profile

Determine type

Goal: Automatically find the associated profile type for

each collected hardness profile

Reject hardness profile if no reasonable profile type is found

Profile types (class templates)

Snow hardness

Sn

ow

heig

ht

[cm

]


Overview

Overview of the developed system


Current chapter: Overview of the developed system

2


Conclusion 4


Overview of the developed system

Schematic view of the proposed classification system


Overview


Current chapter: Alignment of hardness profiles

Conclusion 4




Alignment of hardness profiles

Hardness profiles may be viewed as

time series (clockwise rotated by 90°)

Height corresponds to time axis

Hardness values map to the amplitude

dimension

Desired properties of an internal data

format:

Height independence

Shape preservation

Local height warping

Representation format of hardness profiles

90°



Hardness profiles may be viewed as

time series (clockwise rotated by 90°)

Height corresponds to time axis

Hardness values map to the amplitude

dimension

Desired properties of an internal data

format:

Height independence

Shape preservation

Local height warping

Representation format of hardness profiles

matching



Idea: Use symbolic representation (weighted strings)

A weighted string is composed of weighted characters

Direction character:

D (down), U (up)

Height (as a percentage)

Hardness difference

(among adjacent layers)

Compact form uses only

direction characters

Hardness profiles as weighted strings

6 5 4 3 2 1

3050 ,.,D

110 ,.,D

50350 .,., D

610 ,.,U U

D

D

D

U

D

D

D

53050 .,.,U

51150 .,., D

1150 ,.,D

51050 .,., D



Hardness profiles are aligned based on their compact strings

Weights are required to calculate a penalty score

Goal: Find a global alignment with a maximum number of

matching characters (optimal alignment)

Edit transcript describes operations (match M, deletion D,

insertion I) required to transform the first string into the second

Principle of hardness profile alignment

The compact string UUDUD and DUDUU may be aligned as

– U U D – U D U U D U D – –

D – U D U U – or – – D U D U U , etc I D M M I M D D D M M M I I

String alignment examples



Principle of hardness profile alignment

Well-matching profile Badly-matching profile

U D D – U U D D D U D D U U D D D

U D D D – U D D D U – D – – – – –

M M M I D M M M M M D M D D D D D



Calculation of optimal alignments

Algorithms for calculating optimal string alignments:






After all optimal alignments have been found: Compute penalty

scores

2 types of penalties:

Match-penalty

for differences in:

- Height

- Hardness

Mismatch-penalty

for mismatching blocks

- Much more severe

Computation of penalty scores

match-penalty

Intrusion Detection Systems for SOA 18

Overview


Current chapter: Conclusion


Conclusion 4



Conclusion

Template types are used to

estimate magnitude of avalanches

Classification is based on string

alignment techniques

Well matching strings similar

hardness profiles

Achieved success rate of > 90% for

example set (if not rejected)

Approach may be used for any time

series data

What you have heard in the last 15 minutes

Any Question?

Thank you for your attention!


Backup slides

Distance measures: Cosine of angle, Euclidean Distance

Similarity measure: Cosine of angle

Dissimilarity measure: Euclidean Distance

Euclidean Distance:

Take the smallest distance

between points

in the vector space

Cosine of angle:

Take the smallest angle

between vectors


Backup slides

Problems with the Euclidean Distance

Consider the hardness profile

(in blue) and its dedicated

template (in green)

In principle they are

well-matching

Nevertheless, a large error

occurs (in red) when applying

the Euclidean Distance

measure

ED does not handle “height

warps”


Backup slides


Compute a distance table (with dynamic programming)

Example: Compact strings UUDUD and DUDUU


Backup slides


Based on the distance table, compute the edit graph and

perform a traceback

U U D U D – –

– – D U D U U

D D M M M I I

– U U D – U D

D – U D U U –

I D M M I M D


Backup slides

Template type duplication

Template type versioning:

Create different versions with multiple gradation steps


Backup slides

Template type duplication

Variations of template type versions:

Scale up one layer at the cost of the other layers


Backup slides

Profile matching

Alignment of hardness layers: Limitations

Example should be of type 4, but perfectly matches type 3


Backup slides

Profile matching


Problem of deceptive local extremes

Causing false negatives

badly matching

perfect match


Backup slides

Profile matching


Problem of abusing misalignments

Causing false positives

Apparently wrong profiles may achieve too good penalty scores

(see examples)


Backup slides

Genetic Algorithm principles

Mechanics of a simple Genetic Algorithm: 5 Steps

Step I: Encoding of the search domain with a small alphabet

Step II: Creation of an initial string population

Step III: Reproduction of strings

Step IV: Crossover of strings

Step V: Mutation of strings

Technology

Categorization of Snow Profile Data into Predefined Class Templates