Aproksimacija višinske referenčne ploskve z umetnimi nevronskimi mrežami

delo je pripravljeno v skladu s Pravilnikom o podeljevanju Prešernovih nagrad študentom, pod mentorstvom doc. dr. Tomaža Ambrožiča in somentorstvom doc. dr. Mirana Kuharja

Igor Derenda (Author), Tomaž Ambrožič (Mentor), Miran Kuhar (Co-mentor)

Abstract

Keywords

geodezija;raziskovalna naloga;fakultetna Prešernova nagrada;višinska referenčna ploskev;aproksimacija;umetna nevronska mreža;eksperiment;analiza;primerjava;

Data

Language:	Slovenian
Year of publishing:	2008
Source:	Ljubljana
Typology:	2.25 - Other Monographs and Other Completed Works
Organization:	UL FGG - Faculty of Civil and Geodetic Engineering
Publisher:	[I. Derenda]
UDC:	004.7:528.21
COBISS:	4386145
Views:	1727
Downloads:	587
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Approximation of height reference surface by artificial neural networks
Secondary abstract:	With some of geodesy’s tasks, it is crucial to have precise knowledge of height reference surfaces and of local geoid respectively. It is impossible to define geoid’s surface by using simple mathemtical functions due to the following reasons: geoid is defined through gravity potential, which is not a directly measurable quantity itself and furthermore, geoid’s curve is subjected to a constant change, depending on the factors of relief change and density of Earth’s interior. Geoid’s surface can be represented by a multitude of discrete points on the one hand and with transformation of these points into a function or a math series on the other. Approximation with the use of artificial neuron networks is one of the recent methods, which is also demonstrated in this task. Theoretical knowledge of relations between geoid heights is not necessary for the purpose of using artificial neuron networks, since the latter absorb these relations on the basis of sufficient number of incoming (geographic latitude and longitude) and outgoing information (geoid height); moreover, networks in question can also predict accurate output values for incoming information, which were not involved in the process of learning. Three different artificial neuron networks are used for the purpose of approximating height reference surface: Kohonen’s counter-propagation artificial neural network, Levenberg-Marquardt’s artificial neural network and the radial basis artificial neural network. Computer programs have been designed for every artificial neural network. The results of the artificial neural networks have been compared and analyzed in respect to four different samples combinations of study and test points (25/100, 50/75, 75/50, 99/26) on the territory of the land Baden-Württemberg. The assignment was brought to conclusion by comparing and analyzing the obtained results with the results of previous research on this field.
Secondary keywords:	geodesy;research paper;faculty Prešeren award;height reference surface;approximation;artificial neural;network;experomentation;analysis;comparison;
File type:	application/zip
Type (COBISS):	Research project (high school)
Thesis comment:	Univ. v Ljubljani, Fak. za gradbeništvo in geodezijo
Pages:	XIII, 154 str., pril.
Type (ePrints):	thesis
Title (ePrints):	Approximation of height reference surface by artificial neural networks
Keywords (ePrints):	višinska referenčna ploskev;aproksimacija;umetna nevronska mreža;eksperiment;analiza;primerjava
Keywords (ePrints, secondary language):	height reference surface;approximation;artificial neural network;experomentation;analysis;comparison
Abstract (ePrints):	Pri nekaterih nalogah geodezije je nujno natančno poznavanje višinske referenčne ploskve oziroma lokalnega geoida. Ker je geoid definiran s pomočjo težnostnega potenciala, ki ni neposredno izmerljiva količina in zaradi neprestanega spreminjanja ukrivljenosti geoida s spremembo reliefa in gostote zemeljske notranjosti, je podajanje ploskve geoida nemogoče podati z enostavnimi matematičnimi izrazi. Ploskev geoida je lahko upodobljena z množico diskretnih točk ali pa s pretvorbo teh točk v funkcijo oziroma matematično vrsto. Ena novejših metod je aproksimacija z umetnimi nevronskimi mrežami, kar prikazujemo tudi v tej nalogi. Za uporabo umetnih nevronskih mrež ne potrebujemo teoretičnega poznavanja odnosov med geoidnimi višinami, ampak se umetna nevronska mreža nauči teh relacij iz dovolj velikega števila vhodnih (geografska dolžina in širina) in izhodnih podatkov (geoidna višina) in lahko napove pravilne izhodne vrednosti tudi za vhodne podatke, ki niso sodelovali v procesu učenja. Za aproksimacijo višinske referenčne ploskve so uporabljene tri različne umetne nevronske mreže: Kohonenova protitočna, Levenberg-Marquardtova in radialno bazična umetna nevronska mreža. Za vsako umetno nevronsko mrežo so bili sestavljeni računalniški programi. Rezultati umetnih nevronskih mrež so primerjani in analizirani na štirih različnih kombinacijah vzorcev učnih in testnih točk (25/100, 50/75, 75/50 in 99/26) na območju nemške zvezne dežele Baden-Württemberg. Na koncu naloge pa so dobljeni rezultati primerjani in analizirani z rezultati dosedanjih raziskav na tem področju. IV
Abstract (ePrints, secondary language):	With some of geodesy’s tasks, it is crucial to have precise knowledge of height reference surfaces and of local geoid respectively. It is impossible to define geoid’s surface by using simple mathemtical functions due to the following reasons: geoid is defined through gravity potential, which is not a directly measurable quantity itself and furthermore, geoid’s curve is subjected to a constant change, depending on the factors of relief change and density of Earth’s interior. Geoid’s surface can be represented by a multitude of discrete points on the one hand and with transformation of these points into a function or a math series on the other. Approximation with the use of artificial neuron networks is one of the recent methods, which is also demonstrated in this task. Theoretical knowledge of relations between geoid heights is not necessary for the purpose of using artificial neuron networks, since the latter absorb these relations on the basis of sufficient number of incoming (geographic latitude and longitude) and outgoing information (geoid height); moreover, networks in question can also predict accurate output values for incoming information, which were not involved in the process of learning. Three different artificial neuron networks are used for the purpose of approximating height reference surface: Kohonen’s counter-propagation artificial neural network, Levenberg-Marquardt’s artificial neural network and the radial basis artificial neural network. Computer programs have been designed for every artificial neural network. The results of the artificial neural networks have been compared and analyzed in respect to four different samples combinations of study and test points (25/100, 50/75, 75/50, 99/26) on the territory of the land Baden-Württemberg. The assignment was brought to conclusion by comparing and analyzing the obtained results with the results of previous research on this field.
Keywords (ePrints, secondary language):	height reference surface;approximation;artificial neural network;experomentation;analysis;comparison
ID:	8311098