Análise de atributos espectrais e de textura na classificação de espécies de vegetação aquática emersa por técnica de aprendizado de máquina
Carregando...
Arquivos
Data
2023-04-20
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A construção de reservatórios ocasiona várias alterações ambientais, que favorecem o crescimento descontrolado de vegetação aquática, podendo comprometer os usos múltiplos da água. Assim, identificar essas plantas, bem como conhecer suas principais características e distribuição espacial, são informações importantes para fazer um plano de manejo adequado. Neste sentido, com o advento dos veículos aéreos remotamente pilotados no contexto do sensoriamento remoto, é possível obter imagens de alta resolução espacial que podem dar suporte no mapeamento de vegetação aquática. O alto nível de detalhamento dessas imagens permite extrair atributos de textura, uma característica importante da imagem que é utilizada para auxiliar na identificação de objetos ou regiões de interesse da superfície terrestre. Assim, este estudo teve por objetivo avaliar a capacidade de discriminação de diferentes espécies de vegetação aquática emersa (VAE) por meio da integração de atributos espectrais e de textura, pelo método de análise orientada a objeto, segmentação com Superpixel seguida de classificação dos segmentos com o algoritmo Random Forest. Uma das formas de extrair textura foi pela técnica Scale Invariant Feature Transform (SIFT), um algoritmo robusto da área de visão computacional. Os resultados mostraram que os atributos de textura são importantes, independente da procedência, mas que o SIFT teve um desempenho inferior em relação às texturas usualmente empregadas, como Grey-Level Co-occurrence Matrix. Classificar somente com atributos de textura levou à obtenção acurácias menores, mas ainda assim expressivas, indicando que utilizar apenas a informação de textura já explica parte das características intrínsecas de cada espécie. Os atributos de textura foram combinados com outros atributos e, ao final, a combinação de atributos espectrais e de textura não advinda do SIFT, índices de vegetação e densidade de pontos (textura SIFT) obteve a maior acurácia global e coeficiente Kappa, 95% e 0,92, respectivamente. Logo, a hipótese baseada na possibilidade de descrever as informações de organização espacial, estrutura da folha, particularidades de sombra etc. pelos atributos de textura foi confirmada, e quando combinados com atributos espectrais o ganho não foi tão considerável, mas foi comprovada sua importância na discriminação das espécies de VAE.
The construction of reservoirs causes several environmental changes, which may lead to uncontrolled growth of aquatic vegetation, which in turn, might compromise the multiple uses of water. Therefore identifying these plants, as well as knowing their main characteristics and spatial distribution, is important information for making an appropriate management plan. With the advent of unmanned aerial vehicles in the context of remote sensing, it is possible to obtain high spatial resolution images that can support the mapping of aquatic vegetation. The high level of detail of these images allows the extraction of texture features, an important characteristic of the image that is used to help identify objects or regions of interest on the Earth's surface. Thus, this study aimed to evaluate the ability to discriminate between different species of emergent aquatic vegetation (EAV) by integrating spectral and texture features, using object-based image analysis, Superpixel segmentation, followed by Random Forest classification. The Scale Invariant Feature Transform (SIFT) technique was one of the ways used to extract texture, a robust computer vision algorithm. The results showed that texture features are important, regardless of origin, but that SIFT was not as consistent as common textures, such as the Grey-Level Co-occurrence Matrix. Classifying with texture features alone resulted in lower, but still significant, accuracies, indicating that by using texture information alone, part of the intrinsic characteristics of each EAV species can already be explained. The texture features were combined with other features and, in the end, the combination of spectral and texture features not derived from SIFT, vegetation indices and point density (texture feature from SIFT) obtained the highest overall accuracy and Kappa coefficient, 95% and 0.92, respectively. Therefore, the hypothesis based on the possibility of describing information on spatial organization, leaf structure, shade particularities etc. by texture features was confirmed, and when combined with spectral features the gain was not as considerable, but its importance in discriminating EVA species was proven.
The construction of reservoirs causes several environmental changes, which may lead to uncontrolled growth of aquatic vegetation, which in turn, might compromise the multiple uses of water. Therefore identifying these plants, as well as knowing their main characteristics and spatial distribution, is important information for making an appropriate management plan. With the advent of unmanned aerial vehicles in the context of remote sensing, it is possible to obtain high spatial resolution images that can support the mapping of aquatic vegetation. The high level of detail of these images allows the extraction of texture features, an important characteristic of the image that is used to help identify objects or regions of interest on the Earth's surface. Thus, this study aimed to evaluate the ability to discriminate between different species of emergent aquatic vegetation (EAV) by integrating spectral and texture features, using object-based image analysis, Superpixel segmentation, followed by Random Forest classification. The Scale Invariant Feature Transform (SIFT) technique was one of the ways used to extract texture, a robust computer vision algorithm. The results showed that texture features are important, regardless of origin, but that SIFT was not as consistent as common textures, such as the Grey-Level Co-occurrence Matrix. Classifying with texture features alone resulted in lower, but still significant, accuracies, indicating that by using texture information alone, part of the intrinsic characteristics of each EAV species can already be explained. The texture features were combined with other features and, in the end, the combination of spectral and texture features not derived from SIFT, vegetation indices and point density (texture feature from SIFT) obtained the highest overall accuracy and Kappa coefficient, 95% and 0.92, respectively. Therefore, the hypothesis based on the possibility of describing information on spatial organization, leaf structure, shade particularities etc. by texture features was confirmed, and when combined with spectral features the gain was not as considerable, but its importance in discriminating EVA species was proven.
Descrição
Palavras-chave
Vegetação aquática emersa, Atributos de textura, Object-based image analysis (OBIA), Scale invariant feature transform (SIFT), Emergent aquatic vegetation, Texture features
Como citar
MATSUDA, Letícia Mayumi. Análise de atributos espectrais e de textura na classificação de espécies de vegetação aquática emersa por técnica de aprendizado de máquina. Orientador: Nilton Nobuhiro Imai. 2024. 101 f. Dissertação (Mestrado em Ciências Cartográficas) - Faculdade de Ciências e Tecnologia, Universidade Estadual Paulista, Presidente Prudente, 2023.