Analyzing the Impact of Training Data Selection in Deep Learning for Sargassum Detection

This study proposes the hypothesis that the selection of images for the training of artificial neural networks used for detecting Sargassum using quality metrics, blur, BRISQUE, and entropy, significantly influences their performance. To test this hypothesis, five neural network architectures specifically tailored for the task were used. The results showed that images selected using the blur metric provided the best performances in most architectures, while the entropy metric tended to produce the worst results. In addition, hypothesis tests such as the Student's t-tests were applied to determine the statistical significance of the observed differences in network performance during experimentation. The results highlight the importance of selecting image quality metrics in the neural network training process and suggest that using appropriate metrics can significantly improve model performance, even with limited data sets. The study provides evidence of the influence of image quality metrics on the performance of neural networks for Sargassum detection. The results obtained validate the initial hypothesis and provide a guide for future research and practical applications in the field of environmental monitoring and other areas related to computer vision.