Supervised and Self-Supervised Computer Vision Approaches for Weed Detection and Yield Prediction in Soybean

Abstract

Effective weed management relies on efficient monitoring of dominant weed species and their spatial distribution within a field. Traditional manual monitoring methods are often inefficient, particularly in large farm areas and unfavorable weather conditions. These practices can also be inaccurate due to limited spatial coverage and human subjectivity. Recent advancements in Unmanned Aerial Systems (UAS) and machine learning technologies can provide timely data on weed distribution and density, aiding growers in developing short-, and long-term strategies for weed management. This research study compares the performance of state-of-the-art classification and object detection deep learning models using UAS-acquired RGB (red, green, blue) images from various growth stages of soybean and common ragweed. The Vision Transformer achieved 97.95% test accuracy, while the MLP-Mixer reached 96.92% accuracy, with the former being 2.8 times faster. YOLOv6 outperformed YOLOv5 in detecting common ragweed in soybean, achieving a mean average precision (mAP) of 81.5% with an average inference speed of 7.05 milliseconds at an IoU of 0.5. Furthermore, the study aimed to utilize the digital library representations of different conditions to predict soybean yield. A self-supervised learning-based approach was proposed for early yield prediction and prediction models achieved a coefficient of determination up to 0.80 and a correlation coefficient of 0.88 between predicted and actual yield, with a root mean square error below 0.45 tons per hectare. This research on weed identification and yield estimation has the potential to enable more effective and cost-efficient site-specific weed management practices.

Keywords: common ragweed, deep learning, precision weed management, self-supervised, soybean