Benchmarking of deep learning methods enables more

In crop breeding, plant phenotyping is the detailed study of the characteristic “visible” or phenotypic traits of a plant. This includes counting the number of plants produced by the crossbreeding experiment and grading the features displayed by the progeny or progeny. Offspring with the desired traits are then crossed to produce the next generation of crops, and the process is repeated to increase crop varieties. Traditional methods of plant phenotyping typically lack scalability, precision and are very labor intensive. This imposes certain bottlenecks on crop breeding programs.

But with technological advances and the need for adequate global food security to sustain a growing population, new methods are slowly taking center stage. There are image-based techniques that capture, extract features using machine learning tools, and compare the results to available databases to perform phenotyping tasks in much less time and with higher accuracy. Currently, most machine learning approaches adhere to supervised learning frameworks using labeled datasets, which can be costly and time consuming. Self-supervised learning (SSL) is a machine learning method that reduces the need for labeled data. Despite the proliferation of SSL research, his application of SSL to image-based plant phenotyping tasks is lacking.

In a new study, a research team led by Associate Professor Ian Stavness of the University of Saskatchewan, Canada, evaluated the performance of two SSL methods to improve plant phenotypes. Studies arguing that self-supervised methods may be more sensitive to pre-training dataset redundancy than supervised methods include: Published in Volume 5 of plant phenomics April 3, 2023Associate Professor Stavness explains:These results highlight the importance of paying attention to dataset redundancy when training models for plant phenotypic tasks, especially when using the SSL method.”

Using wheat as a model crop, this study compared a conventional supervised (pre-training) method with two SSL methods: Momentum Contrast (MoCo) v2 and High Density Contrast Learning (DenseCL). All learning methods were subjected to his four phenotyping tasks: wheat head detection, plant entity detection, wheat spikelet counting, and leaf counting. The team found that supervised pre-training produced the best performing models on all tasks except leaf counting.

Contrasting SSL methods, unlike supervised methods, relied heavily on large labeled and annotated databases to perform phenotypic tasks. The algorithm is trained to attract positive samples and move away negative samples, thus increasing the strength of positive samples and training the algorithm to identify more of those samples. MoCo v2 works on optimizing sample global image-level features, while DenseCL focuses on local pixel-level features. Both methods showed comparable performance in the context of internal representations of training models for the desired task.

A specialized and diverse dataset is essential for the pre-training algorithm to perform well. However, in the face of redundancy in large datasets, self-supervised methods perform better in terms of accuracy and sensitivity. The conclusions of this study are often based on empirical observations with little theoretical justification, so the authors would like to extend the study to crop breeding trials. large-scale commercialization.

Associate Professor Stavness concludes:SSL could potentially be used to learn a richer representation of plant phenotypes by aligning them with genotypic and environmental data in the joint embedding space. We hope that this benchmarking or evaluation study will guide practitioners in developing better SSL methods for image-based plant phenotyping. ”

These findings highlight that while artificial intelligence can make our lives easier in some ways, it still cannot dominate humans.

###

reference

author

Franklin C. Ozidi, Mark G. Eramian, Ian Stavness

Affiliation

Department of Computer Science, University of Saskatchewan, Saskatchewan, Canada