Automatic rape flower cluster counting method based on low-cost labelling and UAV-RGB images

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Flowering period is critical for rape plant growth and yield prediction.
• Manual counting of rape flower clusters is labor-intensive and time-consuming.
• Unmanned aircraft vehicles (UAVs) offer a potential solution for automated field counting.

Purpose of the Study:

• To develop a deep learning-based method for counting rape flower clusters using UAV imagery.
• To address the limitations of manual counting and object detection methods.
• To establish an efficient and accurate system for crop yield estimation.

Main Methods:

• Developed a deep learning counting method framed as a density estimation problem.
• Trained deep neural networks (RapeNet and RapeNet+) using annotated datasets (RFRB and RFCP).
• Evaluated performance using metrics like accuracy (Acc) and relative root mean square error (rrMSE).

Main Results:

• The RapeNet series achieved high accuracy (up to 0.9538) and low error (rrMSE as low as 5.61).
• Performance was robust across different resolutions and datasets.
• RapeNet series demonstrated superior performance compared to state-of-the-art counting approaches.

Conclusions:

• The proposed deep learning method provides accurate and efficient counting of rape flower clusters.
• RapeNet series offers significant technical support for agricultural crop statistics.
• This approach enables better yield prediction and farm management.