Answers are grounded in this paper and cite their sources.
Accurate and efficient rice plant counting is crucial for precision agriculture, enabling precise assessment of seedling emergence. However, existing methods often struggle with complex field environments, such as varying illumination, plant density, and growth stages, which degrade their performance in plant counting. To improve the accuracy of rice plant counting, we present a novel method GPDS-Net for rice plant counting in paddy fields. Our approach leverages a Global Feature Perception Module to capture multi-scale global plant information. Together with a Feature Aggregation Block, it collectively achieves the global perception of rice plant features. In addition, a Background Suppression Block is designed to enhance the network's attention on rice plant regions while mitigating the effects of background. The integration of background suppression supervision and Dot map supervision forms a dual supervision mechanism, ensuring the effectiveness of the network training. We demonstrated the performance of GPDS-Net on the Unmanned Aerial Vehicle (UAV)-based rice counting (URC) dataset, achieving a Mean Absolute Error (MAE) of 7.0 and a Root Mean Square Error (RMSE) of 9.5. These values represent reductions of 15.7% and 15.2% respectively, compared with the previous state-of-the-art method. Furthermore, we extended the application of GPDS-Net to count other crop objects, such as Maize seedling detection dataset. The results show that GPDS-Net performs well in this application, indicating its potential for application in counting other crop objects.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: