Weed Localization: Comparison of Different Transfer Learning Models with U-Net

Unwanted plants called weeds challenges the crops for nutrients, water, sunshine, and space in agricultural areas, which lowers crop quality and yield. Farmers have used a variety of techniques and tools to get rid of weeds for millennia. Herbicides are currently being used by farmers to manage weeds, but they are negatively affecting agricultural productivity. Farmers seek to use fewer herbicides to boost crop productivity. To overcome this, precision agriculture is used. To effectively identify weeds in rice fields, unmanned aerial vehicle (UAV) system-based imaging has been used. The usage of UAV and deep learning models has provided effective results in weed detection in the fields. Therefore, in this work, the performance of the UNet model and modified UNet using transfer learning models are analysed to detect weeds using UAV images collected from a rice crop field. U-Net, UNet-VGG16, UNet-VGG19, and UNet-ResNet50 have been compared with the different batch sizes i.e. 16, 32, and 64. Among this UNet-VGG19 95.51% with a batch size of 16 outperformed. Precision, Recall, and F1-score were also considered to analyze the results.