Ferramentas e estudos de caso para a caracterização de séries temporais de sensoriamento remoto

Author(s)

Nathália Menini Cardoso dos Santos

• Document type: Master's Dissertation
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Instituto de Computação
• Defense date: 2019-04-26

Examining board members

Ricardo da Silva Torres; Leonor Patricia Cerdeira Morellato; Liana Oighenstein Anderson

• Advisor: Ricardo da Silva Torres; Marina Hirota

Abstract

It is of fundamental importance to precisely detect changes in the Earth's surface, which may lead to, for instance, better decision making. In order to detect those changes in the Earth's surface, remote sensing images have been broadly used recently. As for change detection, the vegetation characterization of ecossystems is a topic of great importance to understand how complex natural systems are structured. In a more specific context, natural and modified forest ecossystem, such as the Amazon, may have their characteristics changed significantly due to climate change impacts, such as floods and droughts. In order to investigate such changes, several tools have already been proposed. However, their current usage is still compromised by the necessity of alternating several times across them and by the incompatibility of systems to the variety of user profiles, data types, and formats. Given the research problems listed above, in this thesis we propose to advance on three different and linked research venues: (i) to conceive a novel methodology that uses a structural break detection algorithm in order to obtain direct vegetation recovery measures regarding extreme climatic events, (ii) to develop of a toolbox that integrates several modules and that permits an end-to-end data analysis and learning process of remote sensing data, and (iii) to elaborate a soft computing framework that combines a recurrence plot time series representation, dissimilarity measures, and genetic programming in order to improve classification results in problems involving remote sensing data. In (i), we present a framework based on the correlation of temporal changes detected in precipitation and vegetation time series which leads us to better understand the resilience of the Amazon basin's vegetation related to extreme climatic events. In (ii), we introduce the Tucumã toolbox, developed as a toolkit with several modules that support the acquisition, exploration, analysis, clustering, and classification of remote sensing data. Lastly, in (iii), we propose a novel approach to classify regions of remote sensing images based on their time series properties encoded with recurrence plot representations in combination with genetic programming. Our approach led to more accurate results than several baselines considered, suggesting that it is suitable for classification problems involving remote sensing data with well-defined temporal profiles (AU)

• FAPESP's process: 16/26170-8 - Structural breaks detection in time series and its use in the definition of stability measures
• Grantee: Nathália Menini Cardoso dos Santos
• Support Opportunities: Scholarships in Brazil - Master