Classification of Weeds in Agricultural Crops with Ensembles of Convolutional Neural Networks


The present work aimed at proposing and evaluation ensembles of convolutional neural networks to address the classification of agricultural crops versus weeds from images of seedlings in their early stages. To do so, there was a training data preparation phase, adoption of five convolutional neural network architectures and the proposition of three ensemble with different voting strategies. Upon considering the performance of individual networks, an accuracy of 95.77% was reported for MobileNet, meanwhile the ensemble with Support Vector Machine smart voting had accuracy of 97.08%. The results obtained consider a Computational Vision task that enhances the development of digital agriculture, favoring a better yield of agricultural production.

In: XVI Encontro Nacional de Inteligência Artificial e Computacional