Predição da germinação de sementes de soja usando aprendizado de máquina profundo

Abstract

The evaluation of the quality of the soybean seed is based on the visual inspection of the seedlings after the germination period, being a subjective and time-consuming process, in addition to being destructive. The present work proposes the use of deep machine learning techniques to predict the germination of soybean seeds from seed images. With this, greater agility is sought in the process of evaluating the quality of soybean seeds, in addition to characterizing a non-destructive seed evaluation process. To achieve this aim, a database was built containing approximately ten thousand images of seeds with the information whether they germinated or not germinated. After generating the image database, pre-trained convolutional neural network models were selected (VGG19, ResNet50 and InceptionV3) were selected, as well as the creation of an architecture for the seed classification task. After training and adjusting the architectures, a comparative analysis based on performance was performed using training graphs, confusion matrices and the following metrics: the final accuracy obtained during training, the accuracy of the test data, specificity and sensibility. The results obtained demonstrated the feasibility of applying convolutional neural networks for the task of predicting the germination of soybean seeds, indicating that the pre-trained architectures obtained superior results when compared with the model created, according to the following test accuracy averages : VGG19: 95.42%; Resnet50: 90.80%; Inception: 94.78 and model created: 87.44%.