Comparative study of image texture analysis and machine learning methods for classification of Phragmites australis using true-color high resolution images

DSpace Repository

A- A A+

Comparative study of image texture analysis and machine learning methods for classification of Phragmites australis using true-color high resolution images

Show full item record

Title: Comparative study of image texture analysis and machine learning methods for classification of Phragmites australis using true-color high resolution images
Author: da Silva Casagrande, Luan Carlos
Abstract: Phragmites australis (common reed) comumente encontrada em zonas úmidas costeiras pode alterar rapidamente a ecologia por competir e superar as plantas nativas por espaço e pelos recursos. Além disso, este tipo de vegetação representa um perigo de navegação para embarcações menores, prejudicando a visibilidade ao longo do litoral e em torno de curvas e canais de rios. Os esforços de gerencialmento direcionados a plantas não nativas de Phragmites dependem fortemente de um mapeamento preciso das áreas invadidads. No entanto, o mapeamento de Phragmites representa um desafio único por diferentes razões. Identificar e mapear Phragmites pode ajudar os gerentes de recurso a restaurar zonas húmidas afetadas. Neste trabalho, quatro técnicas de extração de características foram testadas: gabor filters, grey level co-occurrence matrix, segmentation-based fractal texture analysis e wavelet texture analysis. Estes algoritmos foram combinados com três estruturas de rede neural artificial: multilayer perceptron, probabilistic neural network e radial basis function network. Além disso, objetivando reduzir o tempo computacional, uma implementação na Graphics Processing Unit do melhor método identificado foi realizada. O estudo de avaliação foi realizado com imagens adquiridas no delta de Pearl River localizado no sudeste da Louisiana e no sudoeste do Mississippi, Estados Unidos da América. Em comparação com os resultados apresentados no estado da arte, wavelet texture analysis com probabilistic neural network e segmentation-based fractal texture analysis com probabilistic neural network apresentaram melhorias em várias variáveis estatísticas como acurácia geral e o kappa. Além disso, o nível de Phragmites agreement aumentou considerávelmente. Nos mostramos que os erros de omissão e comissão restantes geralmente estão localizados ao longo dos limites das áreas identificadas como Phragmites, o que reduz os esforços desnecessários para os gerentes de recursos na busca de áreas inexistentes.Phragmites australis (common reed) commonly found in the coastal wetlands can rapidly alter the ecology by outcompeting with natives for space and resources. In addition, this type of vegetation presents a navigation hazard to smaller boats by impairing visibility along shorelines and around bends of canals and rivers. Management efforts targeting non-native Phragmites rely heavily on accurately mapping invaded areas. However, mapping Phragmites represents a unique challenge for different reasons. Identifying and mapping Phragmites can help resource managers to restore affected wetlands. In this work, four feature extraction methods were tested: gabor filters, grey level co-occurrence matrix, segmentation-based fractal texture analysis, and wavelet texture analysis. These algorithms were combined with three artificial neural network architectures: multilayer perceptron, probabilistic neural network, and radial basis function network. In addition, aiming to reduce the computational cost, a graphics processing unit implementation of the best result was performed. Evaluation study was conducted with imagery acquired in the delta of Pearl River located in southeastern Louisiana and southwestern Mississippi, United States of America. In comparison to state-of-art results, wavelet texture analysis with probabilistic neural network and segmentation-based fractal texture analysis with probabilistic neural network presented presented improvements in several statistical variables such as overall accuracy and kappa value. Furthermore, the Phragmites agreement increased considerably. We show that the remaining omission and commission errors are generally located along boundaries of patches with Phragmites, which reduces unnecessary efforts for resource managers while searching for nonexistent patches.
Description: TCC(graduação) - Universidade Federal de Santa Catarina. Campus Araranguá. Engenharia da Computação.
URI: https://repositorio.ufsc.br/xmlui/handle/123456789/182274
Date: 2017-12-14


Files in this item

Files Size Format View
TCCF.pdf 46.59Mb PDF View/Open

This item appears in the following Collection(s)

Show full item record

Search DSpace


Browse

My Account

Statistics

Compartilhar