Automation of Leaf Counting in Maize and Sorghum Using Deep Learning

Abstract

Leaf number and leaf emergence rate are phenotypes of interest to plant breeders, plant geneticists, and crop modelers. Counting the extant leaves of an individual plant is straightforward even for an untrained individual, but manually tracking changes in leaf numbers for hundreds of individuals across multiple time points is logistically challenging. This study generated a dataset including over 150,000 maize and sorghum images for leaf counting projects. A subset of 17,783 images also includes annotations of the positions of individual leaf tips. With these annotated images, we evaluate two deep learning-based approaches for automated leaf counting: the first based on counting-by-regression from whole image analysis and a second based on counting-by-detection. Both approaches can achieve RMSE (root of mean square error) smaller than one leaf, only moderately inferior to the RMSE between human annotators of between 0.57 and 0.73 leaves. The counting-by-regression approach based on CNNs (convolutional neural networks) exhibited lower accuracy and increased bias for plants with extreme leaf numbers which are underrepresented in this dataset. The counting-by-detection approach based on Faster R-CNN object detection models achieve near human performance for plants where all leaf tips are visible. The annotated image data and model performance metrics generated as part of this study provide large scale resources for the comparison and improvement of algorithms for leaf counting from image data in grain crops.

PhenoTrack3D: an automatic high-throughput phenotyping pipeline to track maize organs over time

“…7B: RMSE = 1.29 vs RMSE = 2.62 for Phenomenal). Our method therefore avoids the bias that may occur in other leaf counting methods [Miao et al, 2021; Zhou et al, 2021; Souza and Yang, 2021] that do not take into account the disappearance of bottom leaves due to senescence. However, this trait was still consistently underestimated (bias = −1.09, Fig.…”

TasselLFANet: a novel lightweight multi-branch feature aggregation neural network for high-throughput image-based maize tassels detection and counting

“…While the above detection methods are designed for specific cultivars in particular environments, they may not accurately distinguish tassels at early tasseling stages, making it difficult to continuously monitor maize breeding requirements. Miao et al (2021) found that the convolutional neural network-based regression counting method had poor accuracy and high bias for plants with extreme leaf counts, while the count-by-detection method based on the Faster R-CNN object detection model achieved near-human performance for plants where all leaf tips are visible. However, the two-stage detection network used in the count-by-detection method ignores the real-time requirements of field applications.…”

The field phenotyping platform's next darling: Dicotyledons

“…This focus on monocotyledons probably occurred because their morphological structure is relatively simple, and the difficulty of image acquisition and data analysis is relatively low. Leaf counting has been realized in maize, sorghum, and other monocotyledons over the entire growth period (Miao et al, 2021). However, studies on the leaves of dicotyledonous species such as soybean and cotton have focused only on comprehensive indicators such as canopy coverage and compactness due to the severe occlusion between leaves and complex plant types (Moreira et al, 2019;Li et al, 2020), which has led to the loss of many details.…”