Optimizing high-resolution multi-view drone imaging for detecting foreign grains in gluten-free oat production fields

Abstract

To reduce the high cost of manually detecting and removing gluten-containing grains from oat crops, drone imaging and deep learning can be used to automate the detection process. In a previous work, a multi-image object detection approach was proposed utilizing high-resolution RGB images captured by a drone using multi-view technology, including nadir and four oblique angles. The images were georeferenced using bundle block adjustment (BBA), and a semi-supervised object detection model (Faster R-CNN) was trained to identify foreign grains. The detector outputs were projected into ground coordinates using a photogrammetric technique. These coordinates were then analyzed using a clustering approach to generate a detection map of barley plant locations. In this study focused on three main objectives. First, it aimed to optimize parameters related to the clustering phase. Second, it evaluated drone data capture settings by assessing whether fewer images could maintain acceptable detection accuracy to reduce flight time. Third, it tested whether direct georeferencing could produce results comparable to those obtained using BBA-based georeferencing. The study showed that using fewer images—for example, two view angles and a side overlap of 80%—could maintain good detection accuracy (omission error of 0.14 and commission error of 0.27). This setup would reduce data collection time from 100 min/ha to 40 min/ha—a substantial improvement for practical field operations. Direct georeferencing showed promising practical results, even though error statistics increased slightly compared to BBA-based georeferencing. These improvements could significantly reduce data capture and processing time, representing a meaningful step toward a practical, cost-effective solution for end-users aiming to detect weedy foreign barley in gluten-free oat production fields.