BreedVision--a multi-sensor platform for non-destructive field-based phenotyping in plant breeding

Abstract

To achieve the food and energy security of an increasing World population likely to exceed nine billion by 2050 represents a major challenge for plant breeding. Our ability to measure traits under field conditions has improved little over the last decades and currently constitutes a major bottleneck in crop improvement. This work describes the development of a tractor-pulled multi-sensor phenotyping platform for small grain cereals with a focus on the technological development of the system. Various optical sensors like light curtain imaging, 3D Time-of-Flight cameras, laser distance sensors, hyperspectral imaging as well as color imaging are integrated into the system to collect spectral and morphological information of the plants. The study specifies: the mechanical design, the system architecture for data collection and data processing, the phenotyping procedure of the integrated system, results from field trials for data quality evaluation, as well as calibration results for plant height determination as a quantified example for a platform application. Repeated measurements were taken at three developmental stages of the plants in the years 2011 and 2012 employing triticale (×Triticosecale Wittmack L.) as a model species. The technical repeatability of measurement results was high for nearly all different types of sensors which confirmed the high suitability of the platform under field conditions. The developed platform constitutes a robust basis for the development and calibration of further sensor and multi-sensor fusion models to measure various agronomic traits like plant moisture content, lodging, tiller density or biomass yield, and thus, represents a major step towards widening the bottleneck of non-destructive phenotyping for crop improvement and plant genetic studies.

Figure 1.

Sensor platform during outdoor measurements in the field.

Figure 2.

(a) Sensor platform from rear view and (b) mounting layout of the used sensors.

Figure 3.

System architecture for data collection and analysis.

Figure 4.

Trait calibration and trait determination procedure, each incorporating the three steps of the developed phenotyping process: Data collection on the field (step 1), data processing on a stationary workstation in the lab (step 2) and trait specific calibration/determination in the lab (step 3A/B).

Figure 5.

Information gathered per plot for (a) light curtains, (b) laser distance sensor measuring from top into the plant cover, (c) 3D Time-of-Flight camera measuring from top, and (d) hyperspectral imaging, (exemplarily shown for 970 nm) measuring from top.

Figure 6.

General design of the field trials and moving direction of the platform, the interspaces between the plots and the dimensions of one single yield trial plot. All dimensions are specified in mm, n denotes the number of rows (2011: n = 48; 2012: n = 24) and m the number of plots per row (2011: m = 14; 2012: m = 25).

Figure 7.

(a) Light curtain data with detected start and end of a single plot and (b) histogram of relative errors of repeated plot detections during measurements of the years 2011 and 2012 based on two repetitions of 1,200 samples. MRE_w denotes the mean relative error between repeated plot length measurements.

Figure 8.

(a) Accuracy and (b) technical repeatability of plant height determination based on light curtain data of the year 2011 and 2012. R_c² denotes the coefficient of correlation of calibration, MRE_c the mean relative error of calibration, R_w² the coefficient of correlation of repetition, MRE_w the mean relative error of repetition and Harvest the three different timepoints of reference data collection.