Deep learning techniques for enhancement of weeds growth classification

Automatic weed control is becoming a popular tool for agriculture with its potential for boosting crop productivity. Most of the research in this area has focused on developing deep learning systems for weeds recognition, ignoring the possible different stages of their growth. In the meantime, in order to effectively control weeds, it is useful to know their exact species and growth stage. This study presents another effective method for solving this problem. We use the transfer learning approach applied for Dense Convolutional Neural Network to recognize weed and identify their growth stages. In our work, transfer learning helps to pre-train the deep neural network model for learning useful features such as leaf characteristics directly from the input data representation. We explore various convolutional neural networks (ResNet, MobileNet, Wide ResNet, DenseNet) architectures and deep learning approaches for addressing this problem. We used the publicly available dataset with labeled data for 18 different weed species. The results show that the proposed method has good performance and better results in solving the classification task (71.81% top1 accuracy and 93.45% top3 accuracy) than previous methods in terms of memory and precision. We provide access to our code [1] that includes the whole pipeline and can help with tasks of classification, using the transfer learning datasets.