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

Abstract

Effective weed control in crop fieldsat an early stage is a crucial aspect of modernagriculture. Nonetheless, detecting and identifying theseplants in environments with unpredictable conditionsremain a challenging task for the agricultural industry.Thus, a two-stage deep learning-based methodologyto effectively address the issue is proposed in thiswork. In the first stage, multi-plant image segmentationis performed, whereas regions of interest (ROIs) areclassified in the second stage. In the segmentationstage, a Deep learning model, specifically a UNet-likearchitecture, has been used to segment the plants withinan image following two approaches: resizing the imageor dividing the image into patches. In the classificationstage, four architectures, including ResNet101, VGG16,Xception, and MobileNetV2, have been implemented toclassify different types of plants, including corn and weedplants. A large image dataset was used for training themodels. After resizing the images, the segmentationnetwork achieved a Dice Similarity Coefficient (DSC)of around 84% and a mean Intersection over Union(mIoU) of around 74%. On the other hand, when theimages were divided into patches, the segmentationnetwork achieved a mean DSC of 87.48% and amIoU of 78.17%. Regarding the classification, the bestperformance was achieved by the Xception network witha 97.43% Accuracy. Then, According to the results, theproposed approach is a beneficial alternative for farmersas it offers a method for detecting crops and weedsunder natural field conditions.