Prediction of seed germination of Pityrocarpa moniliformis (Benth.) Luckow & RW Jobson using Deep Learning
Forest species, artificial intelligence, convolutional networks, Deep Learning.
The possibility of predicting the germination of a seed lot through the use of Deep Learning has shown potential as a complementary method to the analysis of seed quality. This research evaluated the efficiency of using Deep Learning, with convolutional neural networks, to predict the germination of Pityrocarpa moniliformis (Benth.) Luckow & R. W. Jobson seeds. 1000 seeds were randomly selected from 4 different lots, 250 seeds for each, which were used in germination tests and computational analysis. A scanner was used to capture the images of the seeds, from which the images of each seed were obtained individually. The seed images were used to implement, train and test the convolutional neural networks in the computational algorithm created in this research, aiming at comparing the results obtained from the computational analysis with those of the individual germination of each seed. Therefore, after acquiring the images, the seeds were placed to germinate, identifying each one of the seeds. After the germination period, the seeds were divided into two classes, germinated (0) and non-germinated (1). From the images of the seeds before germination, and with the individual result of each seed, the computational analysis was carried out. The pre-trained networks after 5 epochs of execution indicated a tendency to improve accuracy, however, there were also signs of overfitting, since the performance in the training data was better than the validation data (test). The recall (sensitivity) was greater than 90% in all models for the class of germinated seeds. The recall value was much lower for the non-germinated class, both below 20%. For the customized model, 85% of recall was obtained for the class of non-germinated seeds and 18% for the germinated ones, which may have occurred due to an overlap or inversion in the recognition of the classes. The results of the pre-trained networks and the customized model proved to be promising, both for the training set and the test set, since it is possible to verify the efficiency of the use of Deep Learning in predicting the germination of Pityrocarpa moniliformis seeds, however, the analyzes indicate the need to improve and adjust the pre-processing of the images and require more investigation time and more tests to configure the models.