LKNet Sets New Benchmark for Accurate Rice Panicle Counting Across Growth Stages Using Advanced Deep Learning

A new deep learning model called LKNet has set a new benchmark for accurate rice panicle counting across different growth stages, offering a powerful tool for precision agriculture and crop phenotyping. Developed by Song Chen’s team at the Chinese Academy of Agricultural Sciences, LKNet improves upon existing methods by integrating large-kernel convolutional blocks (LKconv) and a novel localization loss function to overcome persistent challenges in automated panicle detection. Traditional approaches to rice panicle counting—such as detection-based, density-based, and location-based models—each face limitations. Detection models struggle in dense canopies due to overlapping panicles, while density-based methods are vulnerable to background noise. Location-based models like P2PNet provide interpretability and simplicity but are constrained by fixed receptive fields and sensitive to annotation inaccuracies, especially when panicle structures vary across rice varieties and growth stages. To address these issues, LKNet enhances the P2PNet framework with dynamic receptive field adaptation and a more flexible loss function. This allows the model to better capture spatial patterns and adapt to morphological changes in panicles throughout the growing season. The model was tested on UAV imagery and multiple crop datasets, demonstrating superior performance and robustness. In benchmark evaluations, LKNet achieved a mean absolute error (MAE) of 48.6 and root mean square error (RMSE) of 77.9 on the high-density SHTech PartA crowd dataset—outperforming both P2PNet and the detection-based PSDNN_CHat model. On PartB, it matched state-of-the-art results with minimal error. In cross-domain assessments focused on rice panicle counting, LKNet delivered an RMSE of just 1.76 and an R² of 0.965, significantly outperforming models that excel on larger targets like maize tassels. When tested on a rice canopy dataset collected at 7 meters above ground, LKNet maintained consistently high accuracy across three panicle types—compact, intermediate, and open—with R² values exceeding 0.98. However, accuracy slightly decreased at later growth stages due to increased occlusion and structural variability, highlighting a remaining challenge in highly crowded fields. Ablation studies confirmed that the LKconv backbone significantly boosted both accuracy and efficiency. The model reduced RMSE from 2.821 to 0.846 while cutting parameter count by nearly 50%. Among different kernel configurations, a sequential large-kernel module enhanced with an attention mechanism achieved the highest performance, with an R² of 0.993. Visualization of class activation maps revealed that LKNet better localizes panicles in complex scenes, exhibiting broader focus areas and improved suppression of background interference compared to P2PNet. These results underscore LKNet’s ability to adapt to real-world variability in field conditions. By enabling precise, high-throughput panicle counting without the need for labor-intensive manual annotations, LKNet supports critical applications such as yield prediction, breeding selection, and phenotypic analysis. Its advanced architecture and loss function make it a transformative tool for UAV-based crop monitoring, paving the way for more efficient and data-driven agricultural practices.