Training Weed Recognition Models for Integrated Weed Management in Precision Agriculture Through Advanced Remote Sensing Technologies

In precision agriculture, efficient weed management is crucial to contrast the significant competitive impact of weeds on crops. This study explores the integration of advanced remote sensing technologies, specifically UAVs, for SiteSpecific Weed Management (SSWM). Conducted in a soybean field, UAV surveys at varying altitudes (10 m, 12.5 m, 15 m, and 30 m) captured high-resolution images to train a deep-learning model for weed classification. Using ArcGIS Pro (c) software, a UNet based semantic segmentation model was developed and validated with expert-labeled data. The model, trained at 15 m, achieved an overall accuracy of 88.3% and demonstrated varying accuracies when applied to images from other altitudes. Results showed species-specific accuracies ranging from 56.3% to 94.1%, indicating that flight altitude significantly influences classification performance. The findings suggest the need for altitude-specific models for optimal weed identification. Despite current recognition algorithms' limitations, this study marks a pioneering effort to adapt pre-trained models across different imaging conditions. Future research should focus on refining these models to enhance their applicability and accuracy in diverse agricultural settings, ultimately contributing to more sustainable and precise weed management practices.