Distinguishing seedling volunteer corn from soybean through greenhouse color, color-infrared, and fused images using machine and deep learning

Abstract

Volunteer corn (VC; Zea mays L.), as a weed in corn-soybean (Glycine max (L.) Merr.) rotation, has negatively impacted soybean production by reducing yield, lowering grain quality, and increasing the production costs. To assess the level of VC infestation to guide chemical applications, farmers usually rely on field visual observation. Since the field visual inspection is labor-intensive, time-consuming, and subjective, an automatic solution to differentiate VC from soybeans is required. As a first step toward that goal, this study aimed to develop a solution to differentiate those crops at seedling stage under greenhouse conditions. Color RGB and color-infrared (CIR) images of VC and soybean seedlings were collected in a greenhouse. A fused image dataset was created by blending the RGB and CIR image datasets. The top 20 extracted relevant features and the image datasets were fed into four different machine learning (ML) classifiers and four deep learning (DL) algorithms, respectively. All ML classifiers resulted in the highest accuracies on the fused images, with support vector machine (SVM) outperforming the other classifiers. Similarly, all the DL algorithms had a superior performance on the fused images, with GoogLeNet as the best algorithm. Overall, GoogLeNet was selected for its high accuracy 99.9%, reasonable computation time (0.02 s per plant), and simple and direct application. The application of fused images along with GoogLeNet can be used as a novel tool to automatically distinguish VC from soybean. Future research should focus on field testing this methodology in a real-time mode.