StalkNet: A Deep Learning Pipeline for High-Throughput Measurement of Plant Stalk Count and Stalk Width

Abstract

Recently, a body of computer vision research has studied the task of high-throughput plant phenotyping (measurement of plant attributes). The goal is to more rapidly and more accurately estimate plant properties as compared to conventional manual methods. In this work, we develop a method to measure two primary yield attributes of interest; stalk count and stalk width that are important for many broad-acre annual crops (sorghum, sugarcane, corn, maize for example). Prior work of using convolutional deep neural networks for plant analysis has either focused on object detection or dense image segmentation. In our work, we develop a novel pipeline that accurately extracts both detected object regions and dense semantic segmentation for extracting both stalk counts and stalk width. A ground-robot called the Robotanist is used to deploy a high-resolution stereo imager to capture dense image data of experimental plots of Sorghum plants. We ground-truth validate data extracted using two humans who assess the traits independently and we compare both accuracy and efficiency of human versus robotic measurements. Our method yields R-squared correlation of 0.88 for stalk count and a mean absolute error of 2.77 mm where average stalk width is 14.354 mm. Our approach is 30 times faster for stalk count and 270 times faster for stalk width measurement.