Classificação de plantas daninhas em banco de imagens utilizando redes neurais convolucionais
Carregando...
Data
2019-05-28
Autores
Marques Junior, Luiz Carlos
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
As espécies exóticas invasoras, também conhecidas como plantas daninhas, competem por recursos, como sol, água e nutrientes paralelamente a cultura plantada, impondo prejuízos econômicos ao agricultor. Para minimizar este problema, atualmente os agricultores fazem uso de herbicidas para a eliminação e/ou controle das plantas daninhas. O uso de herbicidas depara-se com problemas: i) algumas plantas daninhas são resistentes a aplicação de herbicidas e, ii) quando aplicados em demasia pode-se ter a contaminação da cultura plantada, do lençol freático e dos mananciais como rios e lagos. Nesse contexto, visando o desenvolvimento de ferramentas que permitam a minimização do emprego de herbicidas, novas abordagens que fazem uso de visão computacional e inteligência artificial aparecem como soluções promissoras, agregando novas ferramentas a agricultura de precisão. Dentre essas soluções destaca-se o aprendizado profundo (do inglês Deep Learning), que utiliza as redes neurais convolucionais para extrair características relevantes, principalmente em imagens, dessa maneira, permite por exemplo a identificação e a classificação de plantas daninhas, o que possibilita ao agricultor optar tanto pela eliminação mecânica da planta daninha quanto a aplicação localizada de herbicidas e em quantidades adequadas. A partir deste desafio que é a correta classificação de diferentes espécies de plantas daninhas, especialmente plantas resistentes aos herbicidas comerciais, o objetivo deste trabalho foi aplicar e comparar a performance de quatro arquiteturas de redes neurais convolucionais para a classificação de plantas daninhas de cinco espécies contidas em um banco de imagens desenvolvido para este trabalho. Para isso foi realizado o treinamento e a classificação das espécies nas seguintes arquiteturas de redes neurais convolucionais: VGG16, ResNet50, InceptionV3 e InceptionResNetV2 com 20 épocas de treinamento. Os resultados indicam que a arquitetura InceptionV3 apresenta o melhor desempenho, com 84,73% de exatidão na classificação nas cinco espécies, seguida pela arquitetura InceptionResNetV2 com 82,87%, VGG16 com 80,60%. A arquitetura ResNet50 obteve o pior resultado com 20,00% de exatidão, a rede InceptionV3 foi treinada novamente com 40 épocas, obtendo 88,50%de exatidão.
Exotic invasive species, also known as weeds, compete for resources such as sun, water and nutrients in parallel with the planted crop, imposing economic losses to the farmer. To minimize this problem, farmers are currently using herbicides for the elimination and / or control of weeds.The use of herbicides has problems: i) some weeds are resistant to the application of herbicides and ii) when applied too much can contaminate the planted crop, groundwater and springs such as rivers and lakes. In this context, aiming at developing tools to minimize the use of herbicides, new approaches that make use of computer vision and artificial intelligence appear as promising solutions, adding new tools to precision agriculture. Among these solutions are the Deep Learning, which uses the convolutional neural networks to extract relevant features, mainly in images, thus, allows for example the identification and classification of weeds, which enables the farmer to opt for the mechanical elimination of the weeds as well as the localized application of herbicides and in adequate quantities. From this challenge, which is the correct classification of different weed species, especially plants resistant to commercial herbicides, the objective of this study was to apply and compare the performance of four architectures of convolutional neural networks for classification of weed five species contained in an image bank developed for this work. The training and classification of the species were carried out in the following convolutional neural network architectures: VGG16, ResNet50, InceptionV3 and InceptionResNetV2 with 20 training epochs. The results indicated that the InceptionV3 architecture presented the best performance, with 84.73% accuracy in the classification of the five species, followed by the InceptionResNetV2 architecture with 82.87%, VGG16 with 80.60%. The ResNet50 architecture obtained the worst result with 20.00% accuracy, the InceptionV3 network was trained again with 40 epochs, obtaining 88.50% accuracy.
Exotic invasive species, also known as weeds, compete for resources such as sun, water and nutrients in parallel with the planted crop, imposing economic losses to the farmer. To minimize this problem, farmers are currently using herbicides for the elimination and / or control of weeds.The use of herbicides has problems: i) some weeds are resistant to the application of herbicides and ii) when applied too much can contaminate the planted crop, groundwater and springs such as rivers and lakes. In this context, aiming at developing tools to minimize the use of herbicides, new approaches that make use of computer vision and artificial intelligence appear as promising solutions, adding new tools to precision agriculture. Among these solutions are the Deep Learning, which uses the convolutional neural networks to extract relevant features, mainly in images, thus, allows for example the identification and classification of weeds, which enables the farmer to opt for the mechanical elimination of the weeds as well as the localized application of herbicides and in adequate quantities. From this challenge, which is the correct classification of different weed species, especially plants resistant to commercial herbicides, the objective of this study was to apply and compare the performance of four architectures of convolutional neural networks for classification of weed five species contained in an image bank developed for this work. The training and classification of the species were carried out in the following convolutional neural network architectures: VGG16, ResNet50, InceptionV3 and InceptionResNetV2 with 20 training epochs. The results indicated that the InceptionV3 architecture presented the best performance, with 84.73% accuracy in the classification of the five species, followed by the InceptionResNetV2 architecture with 82.87%, VGG16 with 80.60%. The ResNet50 architecture obtained the worst result with 20.00% accuracy, the InceptionV3 network was trained again with 40 epochs, obtaining 88.50% accuracy.
Descrição
Palavras-chave
Agricultura de precisão, Redes neurais convolucionais, Aprendizado profundo, Plantas daninhas, Precision agriculture, Convolutional neural networks, Deep learning, Weeds