Deep learning: as the new frontier in high-throughput plant phenotyping

With climate change and ever-increasing population growth, the pace of varietal development needs to be accelerated in order to feed a population of 10 billion by 2050. Non-invasive high-throughput plant phenotyping (HTP) using advanced imaging technology has capabilities to boost the varietal development process. The tremendous data generated with sensor aided HTP have created the big data and problem in the downstream data analysis pipeline. The higher-level abstraction achieved on high dimensional data by multiple hidden layers for function approximation have made deep learning applications in HTP of significant interest. Application of deep learning models to enhance image-based throughput in phenotyping is an emerging and dynamic area of research in plant phenomics. In this comprehensive review we highlighted the recent developments in the field of deep learning application for HTP. The deep learning principles are described and contextualized relative to machine learning and conventional computer vision algorithms. Novel and emerging deep learning applications are identified. Recommendations are provided with the intent of choosing the most suitable models and training strategy for the capturing and predicting sensor-based phenotyping traits. It also includes steps and suggestions for the development and eventual deployment of such models for multi-task phenotyping. Public datasets have been identified and these datasets are reported which can be used for model training and benchmarking. Overall, this study provided a comprehensive overview of deep learning, it's application in plant phenomics, potential barriers and scope of improvement.

Automated summary

This comprehensive review focuses on the recent developments in deep learning application for non-invasive high-throughput plant phenotyping. It describes the principles of deep learning and contextualizes them in comparison to traditional algorithms. The review identifies novel and emerging deep learning applications, provides recommendations for choosing suitable models and training strategies for capturing and predicting sensor-based phenotyping traits, and includes steps and suggestions for the development and deployment of multi-task phenotyping models. The review also reports public datasets that can be used for model training and benchmarking.