Combination of UAV and deep learning to estimate wheat yield at ripening stage: The potential of phenotypic features

Abstract

A non-destructive, convenient, and low-cost yield estimation at the field scale is vital for precision farming. Significant progress has been made in using UAV-based canopy features to predict crop yield during the mid-growth stages. However, there has been limited effort to explore yield estimation specifically after crop maturity. Researching the effectiveness of artificial intelligence for estimating wheat yield utilizing phenotypic features extracted from UAV images, this study applied a deep learning algorithm (Mask R-CNN) to extract three wheat ear phenotypic features at ripening stage, including ear count, ear size, and ear anomaly index. Subsequently, machine learning algorithms (i.e., multiple linear regression, support vector regression, and random forest regression) driven by ear features were intercompared to obtain the optimal grain yield estimation. Based on the findings, (1) field observed ear count which was linearly associated with grain yield (R2 = 0.93), can be largely detected by UAV images (81 %); (2) Mask R-CNN demonstrated satisfactory performance in ear segmentation, achieving an F1 score of 0.87; (3) random forest regression resulted in the most accurate yield estimation (R2 = 0.86 and rRMSE = 17.53 %), when all three ear phenotypic features were combined. Overall, this study demonstrates that utilizing ear phenotypic features is an alternative approach for estimating wheat grain yield at ripening stage, showing potential as a viable substitute to tedious field sampling methods.

Keywords

• Yield
• Phenotypic features
• Remote sensing
• Deep/machine learning
• Unmanned Aerial Vehicle (UAV)