Uso de redes neurais profundas na avaliação do ataque de mosca minadora ao tomateiro

Abstract

Use of deep neural networks to evaluate leaf-miner flies attack on tomato plants. Adviser: Domingos Sárvio Magalhães Valente.One of the biggest problems of global agriculture is with regard to pests and diseases.Among crops, tomato culture is one of the most susceptible to the attack of pests and diseases. In tomato, one of the main pests are the leaf minadora flies of the genus Liriomyza, especially of the species sativae. The detection and quantification of the severity of the infestation is important to define the moment of control and define the effectiveness of the control systems. Severity quantification is performed by field sampling manually by trained technicians. However, manual evaluation requires trained technicians, and even then, can lead to estimation errors due to the subjectivity of the evaluation process. Thus, the objective of this work was the development of an artificial intelligence model for detection and automatic estimation of the severity of symptoms of the attack of the minadora fly in tomato leaves. The data set for this study gathered 1932 images captured in field conditions containing a leaf with the pest symptom in evidence. The three classes were manually annotated in all images, the background, tomato leaf and leaf symptom of the minadora fly. Three architectures and four different backbones were compared for multiclass semantic segmentation tasks using accuracy, precision, recall and IoU metrics. Sequentially, the severity of the symptom in the mask predicted by the model was estimated using the manually annotated mask for quantification. Comparing all models and dorsal spines used, the U-Net model with Inceptionv3 backbone achieved a better average IoU result reaching 77.71%, followed by the FPN model with Densenet121 as backbone, reaching the mean IoU of 76.62%. Analyzing the IoU separately for the class of the symptom under study, the FPN model with densenet121 backbone achieved a result of 61.02%, followed by the LinkNet model also with Densenet121 as backbone with a result of IoU of 60.99%. To estimate the severity, the FPN model obtained a better result when compared to the others, with the dorsal spines ResNet34 and DenseNet121 being the ones that presented the lowest RSME value, which also confirms the UOU values found for these models. The computational experiments demonstrated in this research were promising, mainly due to the ability of models to automatically segment small objects in images with desharping lighting and complex background conditions, mainly with the use of a database with class unbalance.