Automated crop plant counting from very high-resolution aerial imagery

Abstract

Knowing before harvesting how many plants have emerged and how they are growing is key in optimizing labour and efficient use of resources. Unmanned aerial vehicles (UAV) are a useful tool for fast and cost efficient data acquisition. However, imagery need to be converted into operational spatial products that can be further used by crop producers to have insight in the spatial distribution of the number of plants in the field. In this research, an automated method for counting plants from very high-resolution UAV imagery is addressed. The proposed method uses machine vision—Excess Green Index and Otsu’s method—and transfer learning using convolutional neural networks to identify and count plants. The integrated methods have been implemented to count 10 weeks old spinach plants in an experimental field with a surface area of 3.2 ha. Validation data of plant counts were available for 1/8 of the surface area. The results showed that the proposed methodology can count plants with an accuracy of 95% for a spatial resolution of 8 mm/pixel in an area up to 172 m². Moreover, when the spatial resolution decreases with 50%, the maximum additional counting error achieved is 0.7%. Finally, a total amount of 170 000 plants in an area of 3.5 ha with an error of 42.5% was computed. The study shows that it is feasible to count individual plants using UAV-based off-the-shelf products and that via machine vision/learning algorithms it is possible to translate image data in non-expert practical information.

Plant phenotyping relevance

UAV画像から個体数という植物形態・個体状態を自動抽出する手法を開発し、精度検証まで行っており、フェノタイピング手法が中心である。

Code and data availability

The article describes UAV spinach imagery, manual annotations, and an AlexNet transfer-learning counting algorithm, but contains no data or code availability statement, no public repository, and no author-provided URL for imagery, annotations, models, or scripts. Field location is even withheld for confidentiality. All

No evidence-backed public reproduction asset is currently recorded.