Loading…
A rotated rice spike detection model and a crop yield estimation application based on UAV images
•Precise detection and counting with a rotated rice spike detection model.•Detection of rice spikes by using UAV images.•Angle classification for improved accuracy of rice spike detection.•Three mechanisms for enhanced rice spike feature extraction.•The use of rotated rice spike detection for rice y...
Saved in:
Published in: | Computers and electronics in agriculture 2024-09, Vol.224, p.109188, Article 109188 |
---|---|
Main Authors: | , , , , , , , , , , , |
Format: | Article |
Language: | English |
Subjects: | |
Citations: | Items that this one cites |
Online Access: | Get full text |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | •Precise detection and counting with a rotated rice spike detection model.•Detection of rice spikes by using UAV images.•Angle classification for improved accuracy of rice spike detection.•Three mechanisms for enhanced rice spike feature extraction.•The use of rotated rice spike detection for rice yield estimation.
Accurately detecting and counting rice spikes via unmanned aerial vehicles (UAVs) in field environments is an important aspect of rice research. Due to the flexible and elongated features of rice spikes and the dense and overlapping arrangement of spikes in fields, detecting spikes in UAV images comes with significant difficulties and challenges. In this study, a rotated rice spike detection model was proposed to achieve precise detection and counting of spikes, and the model was validated for field estimation of rice yields. The circular smooth label (CSL) method was designed to introduce spike orientation into the you only look once version 5 (YOLOv5) model; efficient attention mechanisms, shuffle attention (SA) and gather excite attention (GEA) were fused, and the GSConv convolution replacement strategy was used. With these methods, spike orientation was classified, resulting in directional detection boxes that more closely adhered to spike contours, reduced detection of overlapping spikes in the field, reduced redundant information in detection boxes, and improved robustness of spike detection in complex field environments, thereby increasing spike detection accuracy. The experimental results showed that the rotated rice spike detection model outperformed traditional detection algorithms, with an average precision of 95.6%. In the field estimation validation experiment of rice yields, the estimation error obtained by using the proposed rotated rice spike detection algorithm was as low as 1.4%, and the overall estimation error did not exceed 11.7%. These experimental results demonstrate the accuracy and feasibility of the proposed model, which could have a positive impact on the use of artificial intelligence in rice research. |
---|---|
ISSN: | 0168-1699 1872-7107 |
DOI: | 10.1016/j.compag.2024.109188 |