Detecção e mapeamento de nematoides na cultura cafeeira por meio de imagens multiespectrais do MSI/Sentinel-2 e classificação baseada em aprendizado de máquina
Carregando...
Arquivos
Data
2020-03-04
Autores
Marra, Aline Barroca [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
O Brasil é responsável por grande parte da produção mundial de café. Assim como em outras culturas agrícolas, a infecção por parasitas é comum em cafeeiros. Nematoides são parasitas do solo e do sistema radicular da planta que afetam a absorção de água e de nutrientes, podendo causar estresse hídrico e comprometer o desenvolvimento da parte aérea. As plantas infectadas por nematoides tem seu comportamento espectral alterado em relação às plantas sadias. Especificamente quanto à região espectral do infravermelho de ondas curtas (SWIR - short wave infrared), região sensível a variações no conteúdo de água da planta, considera-se que a ocorrência de nematoides no cafeeiro possa ser detectada com maior eficiência. Nesse sentido, o objetivo desta pesquisa foi estudar e avaliar o potencial das imagens multiespectrais do sistema orbital Sentinel-2 na detecção de áreas parasitadas por nematoides em plantação cafeeira, bem como desenvolver um método para mapeamento de áreas infectadas pelo parasita, baseado em uma abordagem de aprendizado de máquina. O estudo foi desenvolvido em uma área cafeeira do Cerrado Mineiro na qual foi constatada e registrada a ocorrência de talhões infectados por nematoides. A fim de avaliar o potencial das bandas espectrais originais e de diferentes tipos de índices de vegetação na detecção do parasita, áreas sadias e infectadas por nematoides foram identificadas e utilizadas como referência para a realização de dois experimentos. Esses experimentos auxiliaram no processo de seleção dos atributos mais significativos para uma comparação do desempenho dos classificadores Florestas Aleatórias (FA) e Redes Neurais Artificiais (RNAs). Os resultados indicaram que o FA apresentou melhores indicadores de acurácia e menor tempo de processamento. Quanto aos mapas gerados após a classificação, os dois classificadores apresentaram produtos semelhantes, porém, no que se refere à definição das áreas onde a atribuição de um pixel a uma classe foi mais incerta, a RNA teve desempenho superior ao FA. Assim, o estudo desenvolvido mostrou que a utilização de bandas na região espectral do (SWIR) do Sentinel-2, ofereceu eficiência na detecção de nematoides em plantação de café, por ser uma região mais sensível ao teor de umidade da planta, permitindo uma melhor investigação da doença devido aos efeitos causados pelo parasita. Destaca-se também a contribuição de uma das bandas do vermelho limítrofe, a Red Edge 3 do sistema Sentinel-2, no aumento do poder dessa detecção, em conjunto com a região espectral do SWIR. O estudo também mostrou que o mapeamento pode trazer indícios de localizações com a ocorrência de nematoides no cafeeiro, se tornando uma ferramenta de ajuda para o fornecimento informações fundamentais no aumento da eficiência do manejo e para alcançar as expectativas de produção agrícola.
Brazil is responsible for much of the world's coffee production. As with other crops, parasite infection is common in coffee crops. Nematodes are soil and root system parasites that affect plant water and nutrient absorption, which can cause water stress and compromising the aerial part development. Plants infected by nematodes have their spectral behavior altered in relation to healthy plants. Specifically regarding to the spectral region of the short-wave infrared (SWIR), a region sensitive to variations in the water content of the plant, it is considered that the occurrence of nematodes in the coffee crop can be detected with greater efficiency. In this sense, the objective of this research was to study and evaluate the potential of multispectral images of the MSI/Sentinel-2 orbital system in the detection of nematode-infected coffee plantation, as well as to develop a method for mapping areas infected by the parasite, based on a machine learning approach. The study was carried out in a coffee area of Cerrado Mineiro where it was found and recorded the occurrence of nematode-infected fields. In order to evaluate the potential of original spectral bands and different types of vegetation indices for the parasite detection, healthy and nematode infected areas were identified and used as references for the realization of two experiments. These experiments helped in the selection process of the most significant features for a comparison of the performance of the classifiers Random Forests (RF) and Artificial Neural Networks (ANNs). The results indicated that the RF showed better indicators of accuracy and shorter processing time. As for the maps generated after the classification, the two classifiers presented similar products, however, with regard to the definition of areas where the assignment of a pixel to a class was more uncertain, the ANN performed better than the RF. Thus, the study developed showed that the use of the SWIR spectral region of the Sentinel-2 system, offered efficiency in the detection of nematodes in coffee plantations, as it is a region more sensitive to the moisture content of the plant, allowing a better investigation of the disease due to the effects caused by the parasite. Also noteworthy is the contribution of one of the Red Edge bands, the Red Edge 3 of the Sentinel-2 system, in increasing the power of this detection, together with the spectral region of the SWIR. The study also showed that the mapping can bring indications of locations with the occurrence of nematodes in the coffee tree, becoming a tool to help provide fundamental information to increase the efficiency of management and to reach the expectations of agricultural production.
Brazil is responsible for much of the world's coffee production. As with other crops, parasite infection is common in coffee crops. Nematodes are soil and root system parasites that affect plant water and nutrient absorption, which can cause water stress and compromising the aerial part development. Plants infected by nematodes have their spectral behavior altered in relation to healthy plants. Specifically regarding to the spectral region of the short-wave infrared (SWIR), a region sensitive to variations in the water content of the plant, it is considered that the occurrence of nematodes in the coffee crop can be detected with greater efficiency. In this sense, the objective of this research was to study and evaluate the potential of multispectral images of the MSI/Sentinel-2 orbital system in the detection of nematode-infected coffee plantation, as well as to develop a method for mapping areas infected by the parasite, based on a machine learning approach. The study was carried out in a coffee area of Cerrado Mineiro where it was found and recorded the occurrence of nematode-infected fields. In order to evaluate the potential of original spectral bands and different types of vegetation indices for the parasite detection, healthy and nematode infected areas were identified and used as references for the realization of two experiments. These experiments helped in the selection process of the most significant features for a comparison of the performance of the classifiers Random Forests (RF) and Artificial Neural Networks (ANNs). The results indicated that the RF showed better indicators of accuracy and shorter processing time. As for the maps generated after the classification, the two classifiers presented similar products, however, with regard to the definition of areas where the assignment of a pixel to a class was more uncertain, the ANN performed better than the RF. Thus, the study developed showed that the use of the SWIR spectral region of the Sentinel-2 system, offered efficiency in the detection of nematodes in coffee plantations, as it is a region more sensitive to the moisture content of the plant, allowing a better investigation of the disease due to the effects caused by the parasite. Also noteworthy is the contribution of one of the Red Edge bands, the Red Edge 3 of the Sentinel-2 system, in increasing the power of this detection, together with the spectral region of the SWIR. The study also showed that the mapping can bring indications of locations with the occurrence of nematodes in the coffee tree, becoming a tool to help provide fundamental information to increase the efficiency of management and to reach the expectations of agricultural production.
Descrição
Palavras-chave
Detecção de doenças, Cultura de café, Sentinel-2, Nematoides, Classificação baseada em aprendizado de máquina, Mapeamento, Disease detection, Coffee plantation, Nematodes, Machine learning classification, Mapping