In cereals, the number of new leaves each plant produces is used to study the periodic events that make up the biological life cycle of crops. Traditional methods of determining leaf number involve manual counting, which is time consuming, labor intensive, and usually associated with large uncertainties due to small sample sizes. Therefore, it is difficult to obtain accurate estimates of some traits by manually counting leaves.
However, traditional methods have been improved by technology. Deep learning has made it possible to estimate the number of plants (and leaves of these plants) in an area using object detection and segmentation algorithms. However, there are obstacles to using these algorithms. They count the tips of the leaves, which appear small in the image, proving difficult to detect. As a result, deep learning techniques often perform poorly in real field conditions. Aiming to solve this problem, a multinational research team developed a self-supervised leaf tip counting method based on deep learning technology. This resulted in highly accurate wheat leaf counts. The research was led by his Professor Shouyang Liu from Nanjing Agricultural University and was published online in his Plant Phenomics on March 20, 2023.
Professor Liu said of their research: Using the Digital Plant Phenotyping Platform (D3P), we simulated a large and diverse dataset of RGB images and the corresponding leaf tip labels of wheat seedlings. Over 150,000 images were generated with over 2 million labels. ”
The researchers used domain adaptation, in which a neural network trained on a ‘source’ dataset is applied to a ‘test’ dataset, also called a ‘target’ dataset. This was achieved through deep learning techniques that mimic the neural processes used by the human brain and use algorithms inspired by their structure and function.
The researchers then collected 2,763 RGB images of young wheat fields from 11 locations across 5 countries. Various means were used to create a robust and reliable source dataset. Images of different types of cameras, different imaging angles and different soil background/light conditions were used. In addition to capturing images of the fields, the team also generated simulated wheat images and automatically annotated them using D3P. Domain adaptation was used to improve the realism of these images, which were then used to train deep learning models.
Six combinations of deep learning models and domain adaptation techniques were used in this study. His Faster-RCNN model using CycleGAN adaptation technique showed the best performance. This is due to its high coefficient of determination (R2 = 0.94) — a measure that determines the model’s goodness of fit — and the optimal root-mean-square error (RMSE = 8.7) — a standard way of measuring a model’s error in predicting quantitative data.
Furthermore, we found that of the three factors in evaluating the performance of the leaf counting model, light conditions were the most important. On the other hand, we found that leaf texture and soil brightness were less important to performance, but we found that combining all three elements significantly improved the realism of the image. The results also revealed that a spatial resolution greater than 0.6 mm per pixel was required to accurately identify leaf tips.
Explaining the implications of their research, Professor Liu said: We also used domain adaptation techniques to make the images more realistic. ”
The research team has made the trained network available at https://github.com/YinglunLi/Wheat-leaf-tip-detection to facilitate further research in this area. Kudos to the team for their notable contributions!
