Corn/Weed Plants Detection Under Authentic Fields based on Patching Segmentation and Classification Networks

Effective weed control in crop fields at an early stage is a crucial aspect of modern agriculture. Nonetheless, detecting and identifying these plants in environments with unpredictable conditions remain a challenging task for the agricultural industry. Thus, a two -stage deep learning -based methodology to effectively address the issue is proposed in this work. In the first stage, multi -plant image segmentation is performed, whereas regions of interest (ROIs) are classified in the second stage. In the segmentation stage, a Deep learning model, specifically a UNet-like architecture, has been used to segment the plants within an image following two approaches: resizing the image or dividing the image into patches. In the classification stage, four architectures, including ResNet101, VGG16, Xception, and MobileNetV2, have been implemented to classify different types of plants, including corn and weed plants. A large image dataset was used for training the models. After resizing the images, the segmentation network achieved a Dice Similarity Coefficient (DSC) of around 84% and a mean Intersection over Union (mIoU) of around 74%. On the other hand, when the images were divided into patches, the segmentation network achieved a mean DSC of 87.48% and a mIoU of 78.17%. Regarding the classification, the best performance was achieved by the Xception network with a 97.43% Accuracy. Then, According to the results, the proposed approach is a beneficial alternative for farmers as it offers a method for detecting crops and weeds under natural field conditions.