Evaluation of drone imagery as a method for selection criteria in soybean breeding

Abstract

Plant breeding is a process of manipulating plants, making them generally more useful. To incorporate beneficial attributes in cultivars, large segregating populations may be necessary. Evaluating large populations may create a bottleneck on plant selection. Sensor technologies have the potential to complement existing phenotyping criteria to improve the rate of genetic gain. This study compared the selection differential in seed yield, maturity, plant height, and lodging among F4-derived soybean lines selected in un-replicated progeny rows based on spectral imagery, visual observations, and random control selections. Spectral imagery was used to calculate a normalized difference red edge (NDRE), a red normalized difference vegetation index (NDVI), a thermal rating (TH), and canopy size (CC) indices for 5338 genotypes in 2017 and 6110 genotypes in 2018. The top 8% of the genotypes based on CC, NDRE, NDVI, TH, progeny row yield (PYLD) and visual (VIS) evaluations, along with random (RAND) selections were advanced to early and late maturing field trials (KPE and KPL) the following years. NDRE, NDVI, and TH selections were measured on mean values (X). Progeny row selections were evaluated in 2018 KPE, KPL, and 2019 KPE trials at three locations and in 2019 KPL trials were evaluated at two locations; all locations using a non-replicated, modified augmented design. Seed yield, maturity, lodging, and plant height were measured on all yield trial plots. Entry means were used to calculate the average seed yield, maturity, lodging, and height for each selection category. Selections based on XNDRE, XNDVI, PYLD, CC, and VIS showed significant yield improvement over RAND selections, however, these observations were not consistent across locations or years. XNDRE and XNDVI showed the greatest consistency across years. Height had shown to be significantly shorter for both XNDRE and XNDVI and lodging had shown to be significantly less severe amongst the XNDRE KPE selections when compared to the random control selections (RAND). This association was supported by a significantly negative correlation between measurements of XNDRE (-.15) and XNDVI (-.07) in 2018 in addition to (-.31), and (-.21), respectively in 2019 to height means. XNDRE had shown a significant negative correlation (-.06) to lodging in 2018 and in 2019, both XNDRE (-.34) and XNDVI (-.21) were significantly correlated to lodging. These patterns were similar in the KPL trials. In 2018, XNDRE accounted for 44% more entries than RAND of the top 30% highest yielding lines, and XNDVI accounting for 47% more entries than RAND. In 2019, XNDVI accounted for 42% more of the top 30% highest yielding lines than RAND in the final population, and XNDRE accounted for 88% more entries than RAND. XNDRE and XNDVI both showed promising results as a selection method. In a program where visual selection is limited by trial size, spectral selection might prove beneficial; however, further research is needed to develop the selection criteria that will produce a consistent positive selection differential. The second experiment consisted of four different locations and 5 different trials (two trials at the same location). Each trial ranged from 10 to 52 entries, set up in a randomized complete block design, planted in 4-row plots 3.7m long, spaced .76m apart. Seed yield and spectral measurements were measured from the center two rows of each plot. MicaSense, Sony, and FLIR cameras were used to make spectral measurements. MicaSense measurements evaluated were blue, green, red, red-edge, near-infrared (NIR), blue normalize difference vegetation index (BNDVI), green normalize difference vegetation index (GNDVI), red normalize difference vegetation index (NDVI), normalized difference red-edge (NDRE), and pigment index (PI). Sony measurements evaluated were blue, green, NIR, BNDVI, GNDVI, and PI. FLIR camera measurement analyzed was thermal (TH). MicaSense BNDVI, GNDVI, BNDVI, and NDVI showed a significant relationship to yield across multiple trials, however, these results showed to be variable, only showing a consistent measurement at two location. Sony cameras BNDVI, and GNDVI measurement had shown a significant relationship to yield across multiple sights as well but altered between positive and negative correlations. No physical plant characteristics were consistently associated with any significant yielding spectral measurements across all trials.