Automated Approaches for the Early-Stage Distinguishing of Palmer Amaranth from Waterhemp

Abstract

Palmer amaranth is an invasive pigweed species, possessing rapid growth, posing a threat to the economy of crops including corn. Its early detection and mitigation are of utmost importance; however, it is visually similar to waterhemp in the early growth stages. In this study, automated approaches are employed to distinguish palmer amaranth from waterhemp, within two weeks after emergence, from their RGB images. Morphological characteristics of these weeds are estimated and fed to several Machine Learning (ML) models. To improve classification accuracy, RGB images are augmented, and a Convolutional neural network is trained on 16000 images. Validated on images embedded with gaussian noise, it produced a better accuracy compared to ML approaches. Finally, YOLOv5, an object detection algorithm based on transfer learning, is successfully prepared. Tested on synthetic images consisting of both weeds, YOLOv5 successfully detected a significantly high number of palmer amaranth objects while also distinguishing it from waterhemp.