Segmentation and Classification Networks for Corn/Weed Detection Under Excessive Field Variabilities.

The control of weeds at earlier stages is one of the most relevant tasks in agriculture. However, the detection of plants in environments with uncontrolled conditions is still a challenge. Hence, a deep learning-based approach to address this problem has been proposed in this work. On the one hand, a CNN model based on the UNet architecture has been used to segment plants. On the other hand, a MobileNetV2-based architecture has been implemented to classify different types of plants, in this case, corn, monocotyledon weed species, and dicotyledon weed species. For training the models, a large image dataset was first created and manually annotated. The performance of the segmentation network achieves a Dice Similarity Coefficient (DSC) of 84.27% and a mean Intersection over Union (mIoU) of 74.21%. The performance of the classification model obtained an accuracy, precision, recall, and F 1 -score of 96.36%, 96.68%, 96.34%, and 96.34%, respectively. Then, as the results indicated, the proposed approach is an advantageous alternative for farmers since it provides a way for crop/weed detection in natural field conditions.