An unsupervised automatic measurement of wheat spike dimensions in dense 3D point clouds for field application

Abstract

An accurate measurement of field-grown wheat traits, including spike number, dimension and volume are essential for crop phenotyping and yield analysis. A high-throughput method to image field-grown wheat in three dimensions is presented with an accompanying unsupervised measuring method to obtain individual wheat spike data. Images are captured using four structured light scanners on a field mobile platform, creating dimensionally accurate sub-millimetre resolution 3D point clouds for a 4.5 m 3 volume in less than 10 s. The unsupervised method analyses each trial plot's 3D point cloud, containing hundreds of wheat spikes, calculating the average size of the wheat spike and total spike volume per plot. The analysis utilises an adaptive k -means algorithm with dynamic perspectives, to fit each spike's shape and measures the dimensions with a random sample consensus algorithm. The method generates small cuboids to fit all the wheat spikes and estimate the total spikes volume. Experimental results show that the proposed algorithm is a reliable tool for identifying spikes from wheat crops and identifying individual spikes. Compared with the manual measurement, according to the results of five scenes, the average error rate in the number of spikes, spikes' height and spikes' width in tests were 16.27%, 5.24% and 12.38% respectively. • Laser imaging technology for automatic field phenotyping application. • High-throughput field capture platform was constructed. • Unsupervised algorithm allows for wheat dimensions measurement in 3D point clouds. • Automatic measurement can estimate wheat spike size, volume and number of spikes. • Framework was evaluated by comparing the experiment with manual measurement.