Development of an algorithm based on meta-learning to classify families of weeds toxic to livestock

Abstract

This work proposes a semantic segmentation algorithm based on Convolutional Neural Networks (CNN) for identifying weeds from the Amaranthaceae, Boraginaceae, and Plantaginaceae families, which can be toxic to livestock. Accurate identification of these plants is crucial for effective control. A dataset of weed images in various contexts, such as pastures, sugarcane fields, and other crops, was collected for the algorithm's development. The images were normalized to a standard size of 704x1056 pixels. The algorithm employs meta-learning techniques, such as MAML, and EfficientNet-B0, a pre-trained feature extractor, within a Feature Pyramid Network (FPN) architecture for semantic segmentation. The Intersection over Union (IoU) metric, also known as mIoU, was used to evaluate the model's performance. During training, the model achieved a loss of 0.007, an mIoU of 0.887, and an accuracy of 0.998. In validation, the results showed a loss of 0.029, an mIoU of 0.851, and an accuracy of 0.996, demonstrating the proposed algorithm's efficiency and high performance.