Pre-harvest loss quantification in grain crops

Abstract

Traditional methods for measuring pre-harvest loss, such as using quadrats, are labor-intensive and provide sparse data coverage. This study proposes an automated approach that leverages computer vision to replace and enhance the current method, using advanced imaging technologies and deep learning methodologies to detect and quantify pre-harvest losses in grain crops. Specifically, the methodology employs a camera mounted on the front snout of a ground vehicle, allowing continuous image capture along the crop rows. By automating image collection and analysis, this approach provides denser spatial coverage across the field, reduces errors associated with manual sampling and human judgment, and significantly accelerates the process compared to traditional quadrat sampling. In addition, this approach offers the potential to segregate different types of pre-harvest loss, such as natural shattering versus losses caused by mechanical disturbance, providing a level of granularity not achievable with conventional quadrat methods. By leveraging state-of-the-art object detection architectures the system is designed to handle the complex visual environment of the field floor, where grains may be obscured by crop residue, shadows, and similarly-colored objects such as stones. This capability represents a significant advancement over traditional methods, which cannot distinguish between these different loss sources. The images were annotated using the Segment Anything Model (SAM) to ensure consistency and accuracy across the dataset. Several state-of-the-art models were trained and evaluated on the collected data, including Mask RCNN, YOLOX, DETR, and a modified YOLOv8-p2. The modified YOLOv8-p2 model, which incorporated a p2 head to improve the detection of smaller objects, outperformed the others, yielding the highest Precision, Recall, and F1 scores on both the soybean (Precision = 0.727, Recall = 0.694, F1 = 0.710) and wheat (Precision = 0.709, Recall = 0.688, F1 = 0.698) datasets. Integrating the 850 nm NIR image channel did not produce a meaningful boost in performance, as evidenced by the soybean (Precision = 0.741, Recall = 0.689, F1 = 0.715) and wheat (Precision = 0.729, Recall = 0.690, F1 = 0.709) results. This research demonstrates that it is possible to integrate a vision system on the header of a combine and identify the initial shedding loss in the field. Future work will focus on refining the models further, exploring their applicability to other crop types, and integrating real-time processing and automation in data collection and analysis.

Plant phenotyping relevance

穀粒の収穫前損失という植物・作物状態を、車載カメラとコンピュータビジョンで自動検出・定量する手法を開発し、複数モデルで性能評価しており、表現型取得が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.