Leaf counting made efficient with a high-throughput AI approach

In cereal crops, the number of new leaves produced by each plant is used to study the cyclical events that make up the biological life cycle of a crop. The traditional method for determining leaf count involves manual counting, which is slow, labor-intensive, and usually associated with significant uncertainties due to the small sample sizes involved. It is therefore difficult to obtain accurate estimates of some traits by manually counting leaves.

However, the traditional methods have been improved with technology. Deep learning has made it possible to use object detection and segmentation algorithms to estimate the number of plants (and the leaves on those plants) in an area. However, there is an obstacle to using these algorithms. They count the tips of the leaves, which look small in photos, and are hard to spot. Thus, deep learning methods often fail to perform in actual field conditions.

With the aim of solving this problem, a multinational research team developed a self-supervised leaf tip counting method based on deep learning techniques, which yielded wheat leaf count with high accuracy. The study was led by Professor Shuyang Liu of Nanjing Agricultural University and is published online in plant phenomics On March 20, 2023.

Speaking about their work, Prof. Liu says: “We have developed a high-throughput method for counting the number of leaves on wheat plants by detecting leaf tips in RGB (red, green, and blue) images.” The Digital Plant Virtual Modeling Platform (D3P) was used to simulate a large dataset. and a variety of RGB images and corresponding leaf-tip labels of wheat seedlings. Over 150,000 images created, with over 2 million labels.”

The researchers used domain conditioning – where a neural network trained on the “source” data set is applied to a “test” data set, also referred to as the “target” data set. This was achieved through deep learning techniques that simulate the neural processes used by the human brain and use algorithms inspired by their structure and function.

Next, the researchers collected 2,763 RGB images of young wheat fields from 11 sites spread across five countries. A variety of metrics were used to create a robust and reliable source dataset – different types of cameras, different shooting angles, and images with varied soil backgrounds/lighting conditions were used. Besides taking field images, the team also created simulated wheat images, which were automatically annotated using D3P. Domain conditioning was used to improve the realism of these images, which were then used to train deep learning models.

Six sets of deep learning models and domain conditioning techniques were used in this study; The Faster-RCNN model with CycleGAN adaptive technology showed the best performance. This was evident from the high coefficient of determination (R² = 0.94) – a measure of model fit – and Root Optimum Square Error (RMSE = 8.7) – a standard way to measure a model’s error in predicting quantitative data.

Moreover, among the three factors evaluated for the performance of the leaf-counting models, light condition was found to be of paramount importance. On the other hand, leaf texture and soil brightness were found to be less important for performance, but the combination of all three factors was found to significantly improve the realism of the images. The results also revealed that a spatial resolution higher than 0.6 mm per pixel was required to ensure accurate identification of leaf tips.

Explaining the implications of their study, Professor Liu explains, “The resulting proposed deep learning method looks very attractive because it eliminates the tedious, costly, and sometimes imprecise manual labeling task by simulating images for which labels are generated automatically. The images were also made more realistic using techniques Field conditioning.

The research team made the trained networks Available here To facilitate further research in this field.