LKNet: Enhancing rice canopy panicle counting accuracy with an optimized point-based framework.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Location-based methods for rice panicle counting are often underestimated compared to detection-based techniques.
• Existing model architectures limit the full potential of location-based rice panicle counting.
• Accurate panicle counting is crucial for rice breeding programs.

Purpose of the Study:

• To introduce LKNet, an innovative location-based model for enhanced rice panicle counting.
• To improve the performance of panicle counting across diverse rice varieties and growth stages.
• To address the limitations of current model architectures in location-based counting.

Main Methods:

• Developed LKNet based on the P2Pnet location-based framework.
• Reconstructed the localization loss function as a predictive probability distribution to minimize manual labeling influence.
• Implemented dynamic receptive field adaptation using large kernel convolutional blocks for varied panicle types.

Main Results:

• Achieved state-of-the-art performance on the Diverse Rice Panicle Detection dataset.
• Demonstrated effective accommodation of panicle morphology variations on a custom dataset, with R² values from 0.903 to 0.989.
• Validated LKNet's performance on multiple publicly available counting task datasets.

Conclusions:

• LKNet significantly enhances the precision of location-based rice panicle counting.
• The model's adaptability makes it suitable for diverse panicle types and growth stages.
• LKNet shows strong potential for application in precision agriculture and rice breeding programs.