Peer review

Open Access

The rapid development of unmanned aerial vehicles (UAVs) and imaging technologies has opened new research avenues for precision agriculture, particularly in the context of plant phenotyping where their utilization has been intensive over the last decade. This review focuses on the interplay of machine learning, UAV-based multispectral imagery and plant phenotyping. We systematically reviewed the current literature to catalog and assess the variety of machine learning methodologies applied to multispectral UAV data for the prediction of key phenotypic traits such as biomass, yield and nitrogen. In this study, we conducted a comprehensive meta-analysis to analyze the relationship between the machine learning model performance and variables such crop type, the type of aerial phenotyping platform, the phenological stage, etc A trait-based comparison of the efficiency and popularity of machine learning algorithms was conducted. Our findings showed that the multiple linear regression is the most effective model in predicting biomass while artificial neural networks showed up as the top performing algorithm in determining nitrogen content. Random forest was identified as the most popular algorithm in estimating those key phenotypic traits. The best combinations of UAV and sensors that significantly enhance model performance for predicting critical agronomic traits were thoroughly examined. Results highlighted, for instance, that pairing the DJI 2 UAV with Micasense sensor led to better machine learning performance in predicting biomass while Parrot Sequoia was identified as the most efficient multispectral sensor to phenotype leaf nitrogen content. Ultimately, the challenges and future research prospects of UAV-based predictions related to the phenotype data variability, the choice of UAV platform, the model complexity and interpretability are discussed. Since previous studies described the broad applications of UAVs and sensors in agriculture, this review aimed to provide a targeted, systematic and quantitative analysis of optimal use of machine learning algorithms and UAV-based multispectral imagery for plant phenotyping.

Text in English

Excellence in Agronomy Bill & Melinda Gates Foundation (BMGF) CGIAR Trust Fund

https://hdl.handle.net/10568/175812