Open Access

ARTICLE

Rice Spike Identification and Number Prediction in Different Periods Based on UAV Imagery and Improved YOLOv8

Fuheng Qu¹, Hailong Li^1,*, Ping Wang², Sike Guo², Lu Wang², Xiaofeng Li^3,*

1 College of Computer Science and Technology, Changchun University of Science and Technology, Changchun, 130022, China
2 Jalaid Banner National Modern Agricultural Industrial Park Management Center, Hinggan League, 137600, China
3 State Key Laboratory of Black Soils Conservation and Utilization, Northeast Institute of Geography and Agroecology, Chinese Academy of Sciences, Changchun, 130102, China

* Corresponding Authors: Hailong Li. Email: ; Xiaofeng Li. Email:

Computers, Materials & Continua 2025, 84(2), 3911-3925. https://doi.org/10.32604/cmc.2025.063820

Received 24 January 2025; Accepted 28 May 2025; Issue published 03 July 2025

Abstract

Rice spike detection and counting play a crucial role in rice yield research. Automatic detection technology based on Unmanned Aerial Vehicle (UAV) imagery has the advantages of flexibility, efficiency, low cost, safety, and reliability. However, due to the complex field environment and the small target morphology of some rice spikes, the accuracy of detection and counting is relatively low, and the differences in phenotypic characteristics of rice spikes at different growth stages have a significant impact on detection results. To solve the above problems, this paper improves the You Only Look Once v8 (YOLOv8) model, proposes a new method for detecting and counting rice spikes, and designs a comparison experiment using rice spike detection in different periods. The method improves the model’s ability to detect rice ears with special morphologies by introducing a Dynamic Snake Convolution (DSConv) module into the Bottleneck of the C2f structure of YOLOv8, which enhances the module’s ability to extract elongated structural features; In addition, the Weighted Interpolation of Sequential Evidence for Intersection over Union (Wise-IoU) loss function is improved to reduce the harmful gradient of low quality target frames and enhance the model’s ability to locate small spikelet targets, thus improving the overall detection performance of the model. The experimental results show that the enhanced rice spike detection model has an average accuracy of 91.4% and a precision of 93.3%, respectively, which are 2.3 percentage points and 2.5 percentage points higher than those of the baseline model. Furthermore, it effectively reduces the occurrence of missed and false detections of rice spikes. In addition, six rice spike detection models were developed by training the proposed models with images of rice spikes at the milk and wax maturity stages. The experimental findings demonstrated that the models trained on milk maturity data attained the highest detection accuracy for the same data, with an average accuracy of 96.2%, an R squared (R²) value of 0.71, and a Root mean squared error (RMSE) of 20.980. This study provides technical support for early and non-destructive yield estimation in rice in the future.

---

Keywords

---