Transforming rice phenotyping: Advanced deep learning models enhance panicle analysis and nitrogen impact studies

Rice is crucial for global food security, providing sustenance for half of the world's population. Its production, particularly influenced by environmental factors during the heading-flowering stage, affects crucial growth traits. Traditional phenotyping methods are inefficient for large-scale analysis, necessitating advanced, accurate monitoring solutions. Recent advancements in computer vision and machine learning, especially deep learning, have improved plant phenotyping, with methods like the scale-invariant feature transform (SIFT) algorithm and convolutional neural networks aiding in rice panicle analysis. However, these techniques face limitations in capturing the dynamic growth of rice panicles over time. Addressing this gap requires combining field cameras with deep learning for detailed, real-time monitoring.

In June 2023, Plant Phenomics published a research article entitled by "Analyzing Nitrogen Effects on Rice Panicle Development by Panicle Detection and Time-Series Tracking".

In this study, researchers developed a pipeline utilizing YOLO v5, ResNet50, and DeepSORT models to automatically extract detailed panicle traits from time-series images. This method was tested for its ability to detect subtle differences in panicle development under varying nitrogen treatments. Results showed high accuracy in panicle counting (R² = 0.96, RMSE = 1.73) and precise estimation of the heading date (absolute error of 0.25 days). Moreover, the method facilitated the analysis of flowering duration and individual panicle flowering times. This analysis  revealed that increased nitrogen leads to more panicles, longer flowering durations, and earlier flowering initiation. The panicle detection model, evaluated against different nitrogen treatments, maintained consistent accuracy across years, indicating its universality for different rice varieties. It also effectively detected panicles with varying shapes, colors, and textures. For panicle classification, the ResNet 50-based model distinguished between vigorous and non-vigorous flowering panicles with high accuracy. This study facilitated the analysis of the rice flowering process and heading date identification, aligning closely with manual counts and field observations. Furthermore, the method effectively identified sensitive flowering changes due to environmental factors like temperature and humidity. The tracking of individual panicles revealed that higher nitrogen application led to earlier flowering initiation and longer flowering periods. The model exhibited robust performance in tracking rice panicles,, with approximately 70% of panicles tracked continuously and completely, despite environmental changes. The study also highlighted the effect of nitrogen on rice heading and flowering, indicating potential impacts on grain filling. It was observed that nitrogen application increases the panicle number but also affects grain filling initiation and duration.

In conclusion, the proposed pipeline demonstrates a non-destructive, accurate, and efficient approach to obtaining panicle traits. It opens new avenues for analyzing rice phenotypes under different nitrogen treatments and environmental conditions, thereby aiding in the advancement of agronomic research and cultivation practices.

###

References

Authors

Qinyang  Zhou¹, Wei  Guo², Na  Chen¹, Ze  Wang¹, Ganghua  Li¹,  Yanfeng  Ding¹, Seishi  Ninomiya^1,2*, and Yue  Mu^1*

Affiliations

¹College of Agriculture, Academy for Advanced Interdisciplinary Studies, Collaborative Innovation Center for  Modern  Crop  Production  co-sponsored  by  Province  and  Ministry,  Nanjing  Agricultural  University,  Nanjing  210095,  China.  

²Graduate  School  of  Agricultural  and  Life  Sciences,  The  University  of  Tokyo,  1-1-1 Midori-cho, Nishi-Tokyo, Tokyo 188-0002, Japan.

About Seishi  Ninomiya & Yue  Mu

Seishi  Ninomiya: He is a professor of College of Agriculture at Nanjing  Agricultural  University. He also belongs to Graduate  School  of  Agricultural  and  Life  Sciences,  The  University  of  Tokyo. His research areas: 1) multi-platform image-based crop phenotyping research, including crop growth monitoring, flower, fruit and other organ detection, and crop (and fruit tree) 3D model construction and analysis; 2) Field Server (Field Server) research, including the collection of images, local meteorological data, sampling, and so on; 3) Agricultural Big Data research, including the construction of Agricultural Big Data (G+E+M+P), and the application of machine learning and other technologies for analysis.

Yue  Mu: Her research focuses on plant (including horticulture and crop) phenotyping using machine learning and computer vision techniques.