Thesis STRUCTURAL ANALYSIS OF ASTRONOMICAL IMAGES: A TWO STEP METHOD BASED ON GAUSSIAN MIXTURE REPRESENTATIONS
Date
2017
Authors
VILLANUEVA AHUMADA, MARTÍN ANDRÉS
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
La identificación y cuantificación automática de componentes y fuentes extendidas, es un desafíoclave para el análisis de datos astronómicos. El aumento de volumen, resolución y sensibilidad delas observaciones, requiere de nuevos algoritmos para la identificación de fuentes. Adicionalmente,la naturaleza jerárquica y acoplada de estructuras complejas, como las nubes moleculares, requierenque estos algoritmos sean capaces de recuperar las relaciones jerárquicas entre las fuentes. Eneste trabajo se propone un nuevo enfoque para la identificación de fuentes en datos astronómicos.En este, primero se representan los datos como una mezcla de Gaussianas, y luego se realizaun análisis automático basado en la reducción de componentes de la mezcla. Se identifican lasfuentes y sus relaciones jerárquicas directamente sobre tal representación. Esto es diferente deotros algoritmos de identificación de fuentes por dos razones: 1) Aquí se realiza la identificaciónsobre una representación continua, en vez de ocupar la representación discreta de pixeles, tomandoventaja de las propiedades de la continuidad. 2) Se construye un árbol jerárquico para obtenerlas relaciones entre los componentes. Esto nos permite seleccionar representaciones con diferentenúmero de componentes identificados, a un bajo costo computacional. Se evalúa la propuestacuantitativamente y cualitativamente usando datos desde el archivo de ALMA Science verification, yademás sobre datos generados sintéticamente. También se comparan los resultados con algoritmos deidentificación de fuentes del estado del arte. Los experimentos y comparaciones realizados muestranque la propuesta es un modo efectivo de inspeccionar y representar la estructura jerárquica de lasfuentes. El método fue capaz de identificar y representar correctamente las emisiones acopladas defuentes sobrepuestas.
The automatic identification and quantification of components in extended sources is a keychallenge for astronomical data analysis. The increasing volume, resolution and sensitivity ofobservations, requires novel algorithms to identify sources. Additionally, the intrinsically hierarchicaland blended nature of complex structures, as molecular clouds, requires these algorithms to be ableto retrieve the hierarchical relationships between sources. We propose a novel approach for theidentification of sources in astronomical data. It first represents the data as a mixture of Gaussiansand then performs automatic analysis based on reduction of the mixture components. We identifythe sources and its hierarchical relations directly over the mixture representation. This is dierentfrom other source identification algorithms for two reasons. 1) We perform the identificationover a continuous representation instead of using discrete pixel representation, taking advantageof the continuity properties. 2) We build a hierarchical tree to obtain the relationships amongthe components. This allows us selecting representations with a dierent numbers of identifiedcomponents with no extra significant computational cost. We asses our proposal quantitativelyand qualitatively using data from ALMA Science verification archive, and also with syntheticgenerated data. We also compare the results with state-of-the-art source-identification algorithms.The experiments and comparisons show that our approach is an eective way to inspect and representthe hierarchical structure of sources. Our method was able to accurately identify and represent theblended emission of overlapped sources.
The automatic identification and quantification of components in extended sources is a keychallenge for astronomical data analysis. The increasing volume, resolution and sensitivity ofobservations, requires novel algorithms to identify sources. Additionally, the intrinsically hierarchicaland blended nature of complex structures, as molecular clouds, requires these algorithms to be ableto retrieve the hierarchical relationships between sources. We propose a novel approach for theidentification of sources in astronomical data. It first represents the data as a mixture of Gaussiansand then performs automatic analysis based on reduction of the mixture components. We identifythe sources and its hierarchical relations directly over the mixture representation. This is dierentfrom other source identification algorithms for two reasons. 1) We perform the identificationover a continuous representation instead of using discrete pixel representation, taking advantageof the continuity properties. 2) We build a hierarchical tree to obtain the relationships amongthe components. This allows us selecting representations with a dierent numbers of identifiedcomponents with no extra significant computational cost. We asses our proposal quantitativelyand qualitatively using data from ALMA Science verification archive, and also with syntheticgenerated data. We also compare the results with state-of-the-art source-identification algorithms.The experiments and comparisons show that our approach is an eective way to inspect and representthe hierarchical structure of sources. Our method was able to accurately identify and represent theblended emission of overlapped sources.
Description
Catalogado desde la version PDF de la tesis.
Keywords
IDENTIFICACION DE FUENTES , REDUCCION DE MEZCLAS GAUSSIANAS , REPRESENTACION GAUISSIANA