The field trial was conducted in July, 2015 at the CIAT headquarters in Palmira, Colombia (lat. 3° 29′ N; long. 76° 21′ W; altitude 965 m). Plants grown in a Mollisol (Fluventic Haplustept), clay-loam, pH 7.8, organic matter 18.4 g/kg and bulk density 1.6 g/cm³. The genotypes used in this study were 19 hybrids generated from interspecific crosses (B. brizantha × B. decumbens × B. ruziziensis) in the CIAT Brachiaria breeding program as well as 3 commercial cultivars (B. decumbens cv. Basilisk [CIAT 606]; B. brizantha cv. Marandu [CIAT 6294]; and the interspecific Brachiaria hybrid cv. Caymán [BR02/1752]) as checks. The 19 test hybrids were selected from a larger group of advanced hybrid breeding materials based on their high biomass production after 2 weeks of continuous waterlogging in a previous pot trial (data not shown). The trial was performed as a split plot experiment with control and waterlogged treatments assigned to main plots (8 m width × 36 m long) organized in a randomized complete block design with two replicates and the 22 genotypes assigned to subplots in a randomized complete block design with six replicates per main plot.

Vegetative propagules of each genotype were planted in Jiffy mix (Carefree^®, Jiffy International, Norway) for an establishment period of 2 weeks before being transplanted into the field. Plants were then allowed to grow for an additional period of 2 months under field conditions spaced by 2 m before the experimental treatments were imposed. After the establishment period, waterlogged and control treatments were imposed for 2 weeks. Main plots were leveled to 0° and surrounded by a compacted berms (50 cm height and 60 cm length). The waterlogged treatment was imposed by applying water to a lamina of three centimeters above soil surface by flood irrigation. Waterlogged plots were irrigated daily in order to maintain the water level during the entire 2-week treatment period. Control plots were supplied with adequate water at the beginning of the treatment and the water table was monitored using homemade piezometers. Average temperature was 31/20°C (day/night) and total precipitation was 1.5 mm during the 2 week treatment period.

Initial vigor (IV) before waterlogging treatment was assessed using a five-point, visual scale: level ‘5’ indicated high shoot growth with many leaves and tillers while level ‘1’ represented low shoot growth with fewer tillers and less leaf growth. A second visual evaluation (VE) at harvest was performed only in waterlogged plants. This VE consisted of a waterlogging injury score that was used to characterize the performance of the plants after waterlogging conditions: level ‘5’ represented plants without any visible damage and ‘1’ denoted dead plants. Plant height (PH) was measured manually using a ruler before starting waterlogging treatment and at harvest.

After 2 weeks of treatment, aboveground biomass was harvested at 1 cm above soil surface and oven dried at 60°C for 3 days. Finally, aboveground tissue including leaves and stems were weighed and expressed as shoot biomass (SHB). Then SHB components were separated into dead leaf biomass (DLB) and green leaf biomass (GLB). The waterlogging tolerance coefficient (WTC) was determined according to Liu et al. (2010) and expressed in percentage:

Digital color images of each plant were taken individually before imposing treatments and at harvest from 1.5 m above the soil surface using a Coolpix P6000 camera (Nikon, Japan) with a resolution of 13 megapixels. Digital images were saved in 4224 × 3168 pixel JPG format and analyzed by a set of instructions written in Java and run in ImageJ software (National Institutes of Health, USA). A workflow of the image analysis program is summarized in Figure 1. The main steps of image analysis were performed to classify pixels of the image into its components, namely soil, green leaves, and dead leaves as follows. (1) Image background information was eliminated by using a previous color threshold learning adjusted to remove soil pixels from the image. This process outputs plant silhouettes without background noise. (2) A stretching process was applied to increase color contrast and to reduce correlation among similar pixels. This process enhances subtle differences in hue by creating a false image which will be easily segmented afterward. (3) Images were segmented by a pattern recognition approach using a modified K-means classifier. This non-supervised classifier used Euclidian distance to classify pixels clusters according to classes of interest, e.g., dead leaves vs. green leaves. (4) Segmented pictures were analyzed by counting pixels according to different regions of interest. (5) Data were recorded in.csv format and expressed in percentage of pixels according to different areas of interest in the original picture. For instance, green pixels refers to green areas in the picture and yellow pixels refers to both chlorotic and senescent tissues in the picture. Finally, the number of pixels in each region of interest was converted to mm². Immediately before the conventional destructive phenotyping, we measured the NDVI from 1.5 m above the soil surface using a hand-held device (GreenSeeker^®, Trimble, USA). Projected WTC was also determined following the methodology previously used in the section before and comparing projected green areas under waterlogging and control conditions.

Statistical analyses were conducted using SAS 9.2 (SAS Institute, Inc., USA). Analyses of variance (ANOVA) were performed to test for the effect of treatments and traits measured. Least-square means were calculated for all traits using PROC MIXED, genotypes and treatments were considered as fixed effects and replications as random effects. Least-square means were then used to calculate Pearson’s correlations. Means of measured traits, variation coefficient and Fisher’s protected LSD were also calculated to identify differences among treatments and genotypes.

Waterlogging treatment had a negative effect on performance of all the genotypes evaluated (Table 1). Mean shoot biomass under waterlogged conditions decreased by 51% compared with drained conditions. The performance of hybrids under waterlogged conditions varied significantly. Two hybrids (BR12/1399; BR12/4951) showed high shoot biomass production under both waterlogged and control conditions (Table 1). Under waterlogged conditions, hybrids BR12/2316, BR12/1176, and BR12/3018 had significantly lower shoot biomass production in comparison with the cultivar Caymán (Table 1). Mean plant height under waterlogged conditions decreased by 23% compared with control conditions (Table 1). After 2 weeks of waterlogging, a set of hybrids (BR12/3436, BR12/1399, BR12/1535, BR12/3377, BR12/4951) were significantly taller than cultivars Basilisk and Marandu (Table 1). Only hybrid BR12/1316 had a significantly lower VE score than the three cultivars tested (Table 1). Hybrid BR12/4856, had a significantly higher waterlogging tolerance coefficient (WTC) than the commercial cultivars (Figure 2).

Harvested shoot biomass under control conditions was significantly correlated with IV (r = 0.58, P < 0.01), plant height before treatment (r = 0.61, P < 0.01), and plant height after treatment (r = 0.56, P < 0.01) (Table 2). Under waterlogged conditions, harvested shoot biomass was positively and significantly associated with IV (r = 0.58, P < 0.01), plant height before treatment (r = 0.57, P < 0.01), plant height after treatment (r = 0.80, P < 0.001) and visual evaluation of symptoms (r = 0.69, P < 0.01) (Table 3).

We evaluated the feasibility of simple and non-destructive phenotyping by comparing it with conventional, more laborious and expensive phenotyping methods. Non-destructive image analysis of projected green area before starting treatment (PGA1) and projected green area after treatment (PGA2) were positively and linearly correlated with harvested shoot biomass among genotypes evaluated under control conditions (r = 0.94, P < 0.001 and r = 0.84, P < 0.001, respectively) (Table 4). Likewise, no strong correlation was found neither between PGA2 and IV (r = 0.52, P < 0.01) nor between PGA2 and both values of plant height before treatment (r = 0.51, P < 0.05) and plant height after treatment (r = 0.51, P < 0.05) (Table 4). NDVI values were not significantly correlated with either PGA1 (r = 0.40, P > 0.05) or with IV (r = 0.34, P > 0.05) under control conditions (Table 4). Plant height before treatment and plant height after treatment under control conditions had highly significant correlations with NDVI values (r = 0.70, P < 0.001 and r = 0.76, P < 0.001, respectively) (Table 4).

Under waterlogging conditions, a significant positive correlation was found between the dead and green leaf area predicted through image analysis and destructively measured dead or green leaf biomass (r = 0.92, P < 0.001 and r = 0.88, P < 0.001, respectively) (Table 5). Normalized difference vegetation index values were also significantly associated with total shoot biomass (r = 0.84, P < 0.001), green leaf biomass (r = 0.85, P < 0.001), plant height after treatment (r = 0.75, P < 0.001), visual evaluation of symptoms (r = 0.85, P < 0.001) and projected green area after waterlogging using digital photography (r = 0.85, P < 0.001) (Table 5). There was no correlation between plant vigor before treatment and projected dead area (r = 0.38) (Table 5). Plant height before starting waterlogging treatment was significantly positively related to projected green area before waterlogging treatment (r = 0.62, P < 0.01). Similarly, plant height after 2 weeks of waterlogging was significantly correlated with projected green area after waterlogging treatment (r = 0.73, P < 0.001) (Table 5). Visual evaluation of symptoms after waterlogging treatment were associated with projected green area after treatment (r = 0.71, P < 0.001) (Table 5). A significant correlation was found (r = 0.7, P < 0.001) between destructively measured WTC and projected WTC calculated from digital images (Figure 3).

Waterlogged soil has a deleterious effect in Brachiaria growth. Major effects of waterlogging in Brachiaria includes decrease in water and nutrient uptake, stomatal closure, reduction of photosynthetic activity, stunted growth and senescence, often leading to plant death (Dias-Filho, 2006). In the present study, significant differences in harvested shoot biomass after waterlogging treatment were found among the genotypes evaluated (Table 1). Differences in biomass production under waterlogging conditions have previously been reported in Brachiaria genotypes (Dias-Filho, 2002). The advanced hybrid selections evaluated in this study showed comparable or even higher shoot biomass production under waterlogging conditions than the commercial cultivars (Table 1). These outstanding hybrids should be an interesting source of alleles to further exploit through breeding.

Phenotypic differences in plant height after waterlogging treatment were found among the evaluated plants (Table 1). Taller plants have an advantage over short plants under waterlogging conditions, as more water is required to fully cover the plants. It is widely known that plants respond to flooding or submergence conditions by growing hyponastically or changing growth habit and leaf angle (reviewed by Polko et al., 2011) as mechanisms to escape the anoxic environment. Therefore, plants with an erect growth habit might be more resistant to waterlogging conditions. Despite the mean plant height decreased under waterlogging conditions, we found a significant and strong correlation between plant height and shoot biomass under waterlogging conditions (r = 0.80, P < 0.001) (Table 3). This suggested that plants that were able to grow hyponastically tended to produce more biomass under waterlogging conditions. This finding suggests that plant height could be used as a simple trait in screening schemes aimed at selecting waterlogging tolerant grasses.

The WTC represents a calculation of the biomass produced under waterlogging against that produced under control conditions, and has been employed as a practical screening method to test waterlogging tolerance in maize (Liu et al., 2010). One hybrid (BR12/4856) had significantly superior WTC compared to the three commercial cultivars (Figure 2). However, this hybrid was not among the best performers in terms of overall shoot biomass produced under waterlogging conditions (Table 1), suggesting that its high WTC was influenced by its reduced growth under control conditions. Small plants require less metabolic demands for maintaining growth in comparison with bigger plants, thus it would represent a competitive advantage under waterlogging conditions. Similarly, a significant and strong correlation between shoot biomass and dead leaf biomass (r = 0.86, P > 0.0001) under waterlogging condition was found (Table 3), indicating that plants showing high biomass also had more proportion of dead leaves. Genotypes showing high WTC also had high projected WTC (Figure 3), indicating the feasibility of image analysis as a phenotyping tool. Nonetheless, no substantial differences between evaluated hybrids were found suggesting that calculation of WTC is not efficient for screening purposes in Brachiaria.

In the present field study, a strong correlation between analysis of regular RGB images and harvested aboveground biomass was found (Tables 4 and 5). The correlations between projected green and dead shoot area obtained from digital photographs of plants subjected to 2 weeks of waterlogging treatment and destructively harvested green and dead leaf biomass were highly positive and significant (r = 0.88, P < 0.0001 and r = 0.92, P < 0.0001, respectively) (Table 5), indicating the potential use of this approach as non-destructive phenotyping tool. This approach could also be used to determine the negative effect of waterlogging and could provide an estimate of acclimation ability -or lack of it-, as the stress continues. In a field experiment, Lootens et al. (2016) found a strong correlation (r = 0.82) between leaf growth predicted by digital photography and its respective destructive measurement in a Lolium perenne germplasm panel. Other studies in controlled greenhouse conditions have found similar correlations between areas estimated by image analysis and harvested biomass (Tackenberg, 2007; Rajendran et al., 2009; Fanourakis et al., 2014; Hairmansis et al., 2014; Joalland et al., 2016).

Although plants under control conditions were vigorous with a considerable amount of overlapping leaves (Figure 1), the correlation between projected green area and harvested shoot biomass was also highly positive and significant (r = 0.84, P < 0.001) (Table 4). This indicates that pictures collected with a regular 13 megapixels’ camera located 1.5 m above the plants give adequate resolution for phenotyping purposes. Image analysis of vigorous plants with overlapping leaves can lead to underestimations of projected areas (Fanourakis et al., 2014; Joalland et al., 2016). To address this problem, Tessmer et al. (2013) have proposed a functional analysis involving plant area estimation, growth modeling and analysis which takes overlapping leaves into account in small rosette plants. In the same way, image analysis accuracy should be improved by taking into account plant area and plant age (Golzarian et al., 2011).

Under waterlogged conditions we found significant correlations among projected green area, shoot biomass and green leaf area. NDVI values were also well-correlated with shoot biomass and green leaf biomass. Thus non-destructive phenotyping through analysis of digital photographs is a more accurate method to estimate total shoot biomass and green leaf biomass than visual estimation. As a consequence of waterlogging, all of the evaluated plants showed green and dead leaves, hindering the conventional harvest and visual scoring method. In this sense, automation of phenotyping by using complementary tools such as NDVI and image analysis offers potential benefits in labor savings and accuracy because they are not affected by human fatigue or bias. Non-destructive phenotyping through image analysis could also be used to quantify the proportion of aboveground biomass affected by waterlogging in early stages. This in turn could give information about which species can be targeted to be planted in zones with different periods of waterlogging. In the same way, analysis of growth rate under waterlogging conditions could also be performed using the proposed phenotyping technique. The advantages of image analysis are that the measurements obtained are quantitative, non-destructive, rapid, and can be used as a proxy for forage yield (Hairmansis et al., 2014). Digital photography could be used as a powerful tool for increasing selection accuracy by increasing the number of replicates evaluated and could also be applied to large mapping populations to identify genes underlying variation in senescence under stress conditions (Furbank and Tester, 2011). Thus the use of non-destructive phenotyping approaches should facilitate rapid and reliable field evaluation of a greater number of genotypes without compromising selection accuracy of a traditional visual evaluation and destructive type of analysis. This in turn will support the on-going Brachiaria breeding efforts and accelerate genetic gain in improving forage yield under waterlogging and also other abiotic stress factors.