Maize emergence rate and leaf emergence speed estimation via image detection under field rail-based phenotyping platform

Abstract

Accurate and efficient acquisition of maize emergence rate and leaf emergence speed in the field is essential for detecting seed quality, evaluating crop field management plans, and yield assessment. This study constructs a system solution to obtain the maize seedling emergence rate and leaf emergence speed based on the field rail-based phenotyping platform and convolutional neural network. Firstly, we use the field rail-based phenotyping platform to collect a high-temporal sequence visible light images of maize plant during the seedling stage. In the first stage, an improved Faster R-CNN is used to detect maize seedlings in the plot images, and the plant ROI area is cropped as the input for the second stage network. In the second stage, the best performing ResNeSt network out of four backbone networks is chosen, using the Mask R-CNN model to segment the leaves of the input plant image, which is then used to calculate the number of leaves. We propose a quantification index for leaf emergence speed based on a weighted average combination of leaf numbers. Using the method described in this paper, we analyzed the plant images from 52 inbred lines plots of over seven consecutive days. The experimental results show that when the Intersection Over Union (IOU) is 0.50, the bbox_mAP of the maize seedling detection model is 0.969, with an accuracy rate of 99.53%. Compared with manual counting, the calculated R2 is 0.997 and RMSE is 43.382. The segm_mAP of the plant leaf segmentation model is 0.942. The differences in emergence rate and leaf emergence speed across 52 inbred lines were compared, providing new phenotyping reference indices for further exploring the genotypic differences affecting seed emergence and leafing.