Development of a low-cost automated greenhouse imaging system with machine learning-based processing for evaluating genetic performance of drought tolerance in a bentgrass hybrid population

Extreme weather events due to climate change threaten sustainable crop production and quality, and developing crop varieties with resilient yield and quality traits through breeding is needed. One of the major issues for turfgrass management is excessive water use for irrigation. In the face of climate change, water availability is becoming increasingly limited and more costly, and water conservation in turfgrass culture has become extremely important; development of drought tolerant turfgrass germplasm is therefore critical. Precise phenotypic assessment of large mapping populations, comprising a few thousand plants, is time-consuming and labor-intensive. Depending on the trait being phenotyped, results may be subjectively assessed and variable due to environmental effects. This is particularly the case when evaluating genetic responses to stresses which may involve multiple measurements over time as well as a range of induced stress levels. A machine learning (ML)based imaging analysis system offers an efficient and precise means to capture temporal progression of stress symptoms within a large genetic population that can be interpreted through computer vision algorithms. In this study, we developed an automatic image analysis system through ML methods that can automatically (1) capture images, (2) segment an image containing numerous plants into sub-images with individual turfgrass plants, (3) label each image with the corresponding sample information, (4) quantify stress symptoms (i.e. percent green cover), and (5) classify breeding lines based on the progression pattern of drought stress symptoms. This system enabled effective and precise evaluation of genetic performance of drought tolerance based on temporal progression of drought symptoms in a large turfgrass hybrid population (about 1,000 lines including three replications), with a processing time to quantify drought stress symptoms from 345,600 images was accomplished within 30 min. Such a machine learning (ML)-based imaging processing/analysis platform along with a low-cost automated greenhouse-based RGB imaging system can significantly boost the effectiveness of germplasm evaluation, quantitative trait locus (QTL) mapping and candidate gene identification to develop potential molecular markers that will aid in faster development of improved germplasm.