Classifying cold‐stress responses of inbred maize seedlings using RGB imaging

Abstract Increasing the tolerance of maize seedlings to low‐temperature episodes could mitigate the effects of increasing climate variability on yield. To aid progress toward this goal, we established a growth chamber‐based system for subjecting seedlings of 40 maize inbred genotypes to a defined, temporary cold stress while collecting digital profile images over a 9‐daytime course. Image analysis performed with PlantCV software quantified shoot height, shoot area, 14 other morphological traits, and necrosis identified by color analysis. Hierarchical clustering of changes in growth rates of morphological traits and quantification of leaf necrosis over two time intervals resulted in three clusters of genotypes, which are characterized by unique responses to cold stress. For any given genotype, the set of traits with similar growth rates is unique. However, the patterns among traits are different between genotypes. Cold sensitivity was not correlated with the latitude where the inbred varieties were released suggesting potential further improvement for this trait. This work will serve as the basis for future experiments investigating the genetic basis of recovery to cold stress in maize seedlings.

Increasing the tolerance of maize seedlings to low temperature episodes could mitigate the 17 effects of increasing climate variability on yield. To aid progress toward this goal, we established 18 a growth chamber-based system for subjecting seedlings of 40 maize inbred genotypes to a 19 defined, temporary cold stress while collecting digital profile images over a 9-day time course. 20 Image analysis performed with PlantCV software quantified shoot height, shoot area, 14 other 21 morphological traits, and necrosis identified by color analysis. Hierarchical clustering of changes 22 in growth rates of morphological traits and quantification of leaf necrosis over two time intervals 23 resulted in three clusters of genotypes, which are characterized by unique responses to cold 24 stress. For any given genotype, the set of traits with similar growth rates is unique. However, the 25 patterns among traits are different between genotypes. Cold sensitivity was not correlated with 26 the latitude where the inbred varieties were released suggesting potential further improvement 27 for this trait. This work will serve as the basis for future experiments investigating the genetic 28 basis of recovery to cold stress in maize seedlings.

Introduction 30
Climate change threatens to negatively impact performance of many important crops, including 31 maize. Extreme heat and drought in maize can cause decreases in yield, especially during later 32 stages of development (Sánchez et al., 2014). One method of avoiding yield losses due to 33 extreme heat and drought late in the season is to plant crops earlier in the season (Kucharik,34 2008); however, earlier planting increases the risk of exposing maize seedlings to low 35 temperature stress conditions. 36 Cold stress is often described as a freezing stress (≤ 0°C) or a chilling stress (generally above 37 0°C and below 15°C) across plant species (Lyons, 1973;Greaves, 1996). Suboptimal 38 temperatures can have multiple impacts on plant growth depending on the severity and 39 developmental time point at which the stress occurs. Effects can range from slight delays in 40 development from growth inhibition to plant death. Other commonly observed stress responses 41 include leaf chlorosis and necrotic lesions (Yadav, 2010). 42 As a species, maize is considered cold-sensitive (Sellschop and Salmon, 1928); however, 43 genetic variation in cold sensitivity exists among inbreds (Greaves, 1996). Several studies have 44 considered maize genotypes that display mild cold sensitive phenotypes to be cold tolerant, 45 despite the lines still being affected by cold stress (Janowiak and Dörffling, 1996;Fracheboud et 46 al., 1999;Sowiński et al., 2005;Wijewardana et al., 2015). However, it is difficult to try to 47 compare levels of sensitivity across studies done under different growth conditions, different 48 temperatures, and at different developmental stages. Also, previous studies have rarely 49 analyzed more than two maize genotypes at a time. Greaves (1996) stated that to improve plant 50 performance under low temperature conditions, genetic variation needed to be characterized for 51 multiple traits, such as levels of tissue injury and growth rates. To identify optimal genetic 52 material for breeding programs interested in maximizing cold tolerance in maize, it is essential 53 to thoroughly characterize the range of cold sensitivity. 54 Many physiological processes in plants are impeded by low temperatures, such as 55 photosynthetic capacity, membrane rigidity, transpiration, and enzyme activity (Marocco et al., 56 2005). Together, these physiological effects of cold stress can result in poor agronomic 57 performance, such as slower emergence, decreased biomass accumulation, reduced growth 58 rates, and leaf chlorosis and necrosis (Miedema, 1982). Relative growth rates (Hetherington 59 and Oquist, 1988;Verheul et al., 1996), electrolyte leakage assays (Capell and Dörffling, 1993), 60 and photosynthesis related measurements (Hetherington and Oquist, 1988;Aguilera et al., 61 1999) have been used in several studies to classify maize seedling responses to cold stress. 62 These approaches require destructive measurements of plants, and therefore necessitate more 63 individuals and space to collect time course data. However, some measurements have been 64 conducted in a non-destructive manner. In field conditions, necrotic injury was visually assessed 65 on a relative scale at a single time point on six genotypes, where lines with the least amount of 66 leaf necrosis were classified as cold tolerant (Janowiak et al., 2003). An alternative approach to 67 destructive and manual assessment is to move to image-based plant phenotyping methods to 68 allow for robust measures of changes in color and growth without destructive, subjective, or 69 labor-intensive techniques. 70 There are a growing number of commercial and custom-built systems for integrated controlled 71 plant growth and imaging (Tisné et al., 2013). Currently available commercial systems can 72 provide valuable insight into variation in plant growth and development, but they tend to have 73 higher costs and infrastructure requirements that limit access to a small number of researchers. 74 Additionally, these systems usually restrict researchers to conducting a single experiment at a 75 time. We sought to develop an image-based approach that could be implemented easily to 76 document morphological traits of maize seedlings at a low cost. This system is not fully 77 automated because it requires manual plant staging. As manual staging of plants is required, 78 we have implemented tools to ensure high-quality standardized images are captured. 79 Many researchers have, or are currently developing, custom low-cost phenotyping platforms to 80 fit their needs. Our system is not necessarily unique in this pursuit and is similar to other 81 recently developed low-cost imaging systems (Knecht et al., 2016;Armoniené et al., 2018;82 Czedik-Eysenberg et al., 2018). Currently, our system is limited to acquiring images of seedlings 83 from one side view image but has the advantage of being scalable. Additionally, This study sought to compare growth rates of 40 diverse maize inbreds in a mild cold stress 99 treatment under controlled growth conditions using image-based phenotyping methods. We 100 established an image acquisition platform, including a system for embedding metadata and 101 sample tracking, to collect high-quality, standardized RGB images of maize seedlings over time. 102 Trait extraction from images was accomplished using PlantCV. This work describes a robust 103 method for analyzing recovery rates across multiple morphological and color-based traits for a 104 large number of maize inbreds that will serve as a foundation for future work uncovering the 105 genetic basis for cold stress recovery in maize. 106 format) and printed on a piece of paper that was mounted in the scene above the plants ( Figure  137 1A). This embedded experimental details and plant identity into the corresponding image data in 138 a machine-readable format. The sample tracking page also contained 24 blue boxes. The user 139 marked the number of blue boxes corresponding to the age of the plants each day they were 140

Methods
photographed. This date/age score was automatically added to the sample tracking data at the 141 time of image acquisition. 142 We refer to the three seedlings in each image as a plot. The staging area consisted of a desk, a 143 4'x6' blue drywall background, three plastic D20T racks (Stuewe and Sons, Inc), nails to hold 144 the QR code and day tracking sheets, and the space to the right of the QR code was used to 145 add color standards but could be used for other information as well. The use of DeePots 146 provided several advantages in the system as plants could be grown at high densities and 147 easily moved from growth chambers to the imaging system. Additionally, DeePots enabled 148 plants to be quickly moved in and out of racks, rotated as necessary to adapt to rotations in 149 growth, and placed in consistent locations each day. An example image acquired using this 150 system is depicted in Figure 1B. This system can be used for imaging a wide variety of plant 151 species to obtain side views of plants over time. The combination of imaging platform size and 152 growth conditions allowed for growth and for data to be collected from 8 to 16 days after sowing 153 (DAS). After this time, the growth of the seedlings began to plateau in some genotypes and the 154 leaves of neighboring plants began to overlap, which hindered proper plant segmentation. 155

Generation of Sample Metadata Tracking Sheets 156
Custom Perl and R scripts were written to create QR code sheets containing metadata and to 157 allow sample tracking over time. Briefly, a Perl script is run that takes in a tab delimited text file 158 containing the desired sample tracking metadata (plot, genotype, treatment, etc). This Perl 159 script outputs an R script that can be run to produce the formatted QR code sheet with 160 embedded metadata and day tracking boxes. These scripts are available at 161 https://github.com/maizeumn/cold-phenotyping. 162

Trait extraction from images using PlantCV 163
Raw .nef format RGB files of maize seedlings were converted to .tiff format files using dcraw 164 (https://www.cybercom.net/~dcoffin/dcraw/). Trait measurements were extracted from each .tiff 165 file using PlantCV v3.0.dev2 (Fahlgren et al., 2015;Gehan et al., 2017;166 doi:10.5281/zenodo.1408271). Pixel classification within each image was achieved through the 167 use of the Naive Bayes multiclass training module within PlantCV. This approach allowed for 168 color-based classification of plant tissue into two categories: healthy and necrotic, and therefore 169 quantification of the percent area (number of pixels) corresponding to each of these categories 170 ( Figure 2A). Plant masks were dilated and filled to reduce noise in the segmentation. Initial 171 efforts to classify plant pixels in this manner resulted in soil pixels being included in the plant 172 necrotic category. Ranges of RGB values for necrotic tissue, stem tissue, and soil overlapped 173 and could not be separated using our training set and the Naive Bayes approach. To remove 174 soil pixels from the plant mask, we identified the rack that held each pot using edge detection 175 methods and excluded any pixels below a boundary line to isolate only plant pixels for later trait 176 extraction. The final plant mask and original RGB image were used to measure attributes of the 177 plant object. 178 Our pipeline used PlantCV to measure 16 morphological traits associated with defined objects, 179 such as height, width, area, convex-hull properties, and various measurements of an object-180 bounding ellipse ( Figure 2B) Cold temperatures often result in slower growth and induce leaf necrosis in maize seedlings, but 226 the severity of these effects varies among genetic backgrounds (Greaves, 1996). The goal of 227 this study was to survey the range of cold stress effects on the growth, morphology, and leaf 228 necrosis across various maize genotypes. The first step towards accomplishing this goal was 229 the design of a cold-stress assay that resulted in phenotypic changes across genetic 230 backgrounds, which also allowed for analysis of recovery within the constraints of our image 231 acquisition system. 232 Our system provided an opportunity to measure plant growth and morphology for maize 233 seedlings from 8 to 16 DAS. We conducted several experiments to identify a set of conditions 234 that allowed analysis of variability for responses to cold stress. A variety of temperatures and 235 stress lengths were assessed to determine appropriate cold stress conditions. While 236 temperatures near or below 0°C provided strong stress responses, we found greater 237 experiment-to-experiment variation and some genotype lethality at these temperatures. 238 Therefore, we elected to use a more moderate low temperature condition (6°C day / 2°C night) 239 that was more phenotypically consistent but resulted in more subtle phenotypes than freezing 240 temperatures. We tested the effects of different durations of this cold stress to select a 241 treatment regime that resulted in observable effects on measured traits but also allowed for 242 quantification of stress recovery in the B73 and Mo17 inbreds ( Figure 3). All cold-treated plants 243 were placed in the stress condition two hours after dawn at 9 DAS. Every 24 hours a subset of 244 plants were removed from the cold stress and returned to control conditions in growth chambers 245 for a total of four separate durations of cold treatment ranging from 1 to 4 days. Plants grown 246 under a single 24-hour period of cold treatment had the least amount of growth inhibition in 247 height, area, and width compared to plants grown under control conditions for both genotypes 248 ( Figure 3). Four days of cold stress resulted in the most extreme differences in growth 249 compared to control plants for area, height, and width measurements but only allowed for 3 time 250 points during the recovery period. A 2-or 3-day duration of cold stress had intermediate effects 251 to these two extremes. To collect a maximum amount of time points during recovery, a 2-day 252 cold treatment was chosen for further experiments. Additionally, the selection of a 2-day long 253 cold stress over a 1-day cold stress allowed for the analysis of whether any of our selected 254 genotypes were able to grow during the cold period or if the cold stress conditions resulted in 255 growth arrest for all tested genotypes and the collection of five time points to characterize 256 recovery from stress conditions. 257 Surveying diversity of cold responses using image-based methods 258 To survey the variation in responses to cold stress among diverse genotypes of maize, we 259 implemented a robust cold stress assay that subjected plants to two days of cold stress in 260 growth chamber conditions to analyze cold stress responses in a panel of 40 maize genotypes. 261 Within our selected genotypes, there were representatives from multiple heterotic groups and 262 genotypes that resulted from breeding programs at very distinct latitudes that likely faced 263 variable levels of early season cold stress ( Figure 4A; Supplemental Table 1 methods with a 2-day cold stress as the treatment group. This 2-day cold treatment was able to 269 recapitulate phenotypes observed in field grown plants that experience cold stress, such as leaf 270 necrosis, chlorosis, and growth inhibition. Additionally, our image-based data collection method 271 enabled the capture of nine time points during the early developmental stages of maize plants 272 ( Figure 4C, D). Images of Mo17 seedlings grown under control conditions ( Figure 4C) and cold-273 treatment conditions ( Figure 4D) provide examples of the developmental time points captured 274 for each seedling and the degree of growth inhibition achieved in our assay. We collected data 275 on three biological replicates that represented different grow-outs. For each biological replicate, 276 we measured traits for six individuals exposed to control conditions and six individuals exposed 277 to a cold stress. In total, this dataset contained images for nine consecutive days of growth for 278 ~18 plants per genotype per treatment resulting in ~12,000 images of ~1,400 plants. 279 Impact of cold stress on leaf necrosis across genotypes 280 One of the more noticeable effects of cold stress in maize seedlings is the appearance of leaf 281 necrosis ( Figure 5). Cold temperatures can cause leaf tissue to wilt, and over the course of 282 several days this wilted tissue can die, resulting in changes in leaf color from green to brown 283 and texture from healthy leaves with high turgor to dehydrated, dead leaf tissue (Guye et al., 284 1987). The Naive Bayes color-based classification module that was trained and implemented 285 within PlantCV classified each plant pixel into a healthy or necrotic category that enabled the 286 quantification of necrosis as a percentage of area belonging to each category for every plant at 287 each time point. The pipeline output included images of each seedling indicating the category 288 each pixel was classified into by color (Figure 2A). 289 Among the genotypes surveyed, we observed substantial variation for this response ( Figure 5A, 290 Supplemental Figure 2). Some genotypes, such as Oh43 and NC350, did not exhibit any 291 changes in the proportion of necrotic tissue following a cold stress ( Figure 5A). For other 292 genotypes, such as MoG or Ki11, a significant portion of the plant exhibited necrosis. A 293 comparison among these genotypes also revealed variability in the level of necrotic tissue 294 present within the control plants. This resulted from healthy portions of the lower stem having 295 color values with a high probability of being classified as necrosis. This occurred at different 296 frequencies among genotypes. To control for this, we focused on comparing the amount of 297 necrotic tissue in control plants compared to cold stressed plants within each genotype. We also 298 noted variability in the total percent of necrotic tissue during our time series. The percent 299 necrotic value often peaked at 13 DAS followed by a gradual decline. This was a result of new 300 growth of healthy/green tissue following the cold stress, resulting in the overall percent necrotic 301 area decreasing. The percentage of plant area that exhibited necrosis was finite and underwent 302 a predictable sequence of color changes. Therefore, a time-course analysis is unnecessary for 303 this phenotype within our experiments. 304 Accordingly, the percent necrotic tissue was assessed for all 40 genotypes at 13 DAS ( Figure  305 5B). Seven genotypes had a significant increase in tissue classified as necrosis relative to the 306 controls. For many of the other genotypes, the cold treatment group never accumulated an 307 amount of necrotic tissue greater than the amount of tissue misclassified as necrotic within the 308 control group. It is worth noting that all genotypes survived the cold stress and continued 309 growth. Even the most severe necrosis responses only resulted in the loss of healthy tissue for 310 two or three leaves. 311 Impact of cold stress on plant morphology across genotypes Genotypes were compared on a single day for quantification of necrotic tissue, however the 313 morphological traits changed at different rates among genotypes, and therefore the entire time 314 course of data collected was utilized (Supplemental Figures 3-7). For example, plant area did 315 not increase during the cold treatment for any genotype (Supplemental Figure 3). Yet, the rate 316 at which area increased following the cold stress period was faster for some genotypes 317 compared to others ( Figure 6A, Supplemental Figure 3). Growth rates across traits are more 318 similar within genotypes than among genotypes. The set of traits that are most affected by 319 stress vary across genotypes. Values for all analyzed morphological traits did not appear to 320 increase during the cold treatment for any genotype but recovery rates following the stress 321 varied. Because genotypes are not growing during the cold stress period, they are delayed in 322 development compared to plants grown under control conditions. Because of the cold-induced 323 delay in growth, a more equivalent comparison than comparing control and cold-stressed data 324 at the same time point was to compare cold-stressed data to control data at a time point two 325 days earlier. Therefore, we chose to compare measurements for morphological traits with a 326 time-shifted approach, comparing the cold-stress plant measurements to the control plant 327 measurements collected 2 days prior for each genotype ( Figure 6B). This allowed comparisons 328 of control and cold stressed plants at more similar developmental stages. Additionally, to 329 minimize the effects of individual plant size on comparisons, we calculated growth rates during 330 two intervals among the time-shifted data ( Figure 6C). Within each interval, we calculated the 331 log 2 fold change in growth rates between treatments for each genotype for each trait. 332 Clustering genotypes based on leaf necrosis and morphology phenotypes 333 Because of intrinsic morphological differences among genotypes, growth rates manifest in 334 different patterns across traits. Therefore, a phenotype fingerprint, or phingerprint, captures the 335 unique aspects of cold response in each genotype. The time-shifted and normalized data from 336 plant morphology phenotypes and the percent necrosis data from day 13 was used to cluster 337 genotypes using hierarchical clustering (Figure 7). This resulted in three clusters. Overall 338 patterns of growth inhibition were fairly similar across all genotypes. However, subtle differences 339 among degrees of fold change and the pattern of changes across different morphological traits 340 helped define genotypes into different clusters. One cluster was characterized by more subtle 341 changes in growth rates compared to controls of morphological traits over the two defined 342 intervals. A second cluster was dominated by higher degrees of necrosis than controls on day 343 13. A third cluster was characterized by moderate levels of necrosis on day 13 and a smaller 344 fold change in growth rate of height compared to controls. Our initial hypothesis was that 345 genotypes may cluster together based on some attributes, such as population group, market 346 class, kernel type, or latitude. This did not appear to be the case. Latitude did not have a 347 significant correlation with any of the log2 fold changes in growth rates for any traits during either 348 interval. Therefore, cold sensitivity during early development was likely not a target of breeding 349 programs for these genotypes. 350

Discussion 351
Improving cold tolerance in maize remains a challenge, despite nearly 100 years of studies. 352 Phenotyping approaches can enable new insights into cold stress responses in plants. This 353 study successfully used image-based phenotyping methods to characterize how maize 354 seedlings respond and recover from a cold stress event. Our approach allowed comparisons of 355 multiple genotypes over time and for quantitative measurements of cold stress responses, such 356 as area, height, width, and development of leaf necrosis. The quantification of these traits from 357 images facilitated nondestructive measurements to be made over time and enabled our analysis 358 of recovery rates. The use of hierarchical clustering allowed for a methodical approach to 359 compare the ability of genotypes to recover from cold stress. 360 As with most image-based time course phenotyping methods, there were several limitations to 361 our approach. The patterns of clustering were influenced by the choice of time intervals and 362 included traits. Every inbred in our study is sensitive to cold, so differences among that 363 sensitivity can be subtle. The quantification of the leaf necrosis phenotype provided a 364 straightforward definition of the most cold sensitive genotypes within our assay, however the 365 morphological traits captured subtle variations in the diverse patterns of cold response. 366 Additionally, although our assay recapitulates phenotypes observed in field-grown plants that 367 experience cold stress (chlorotic bands on leaves and leaf necrosis), caution must be made 368 when extrapolating controlled-condition studies to field conditions. Finally, the performance of 369 inbreds does not always predict combining ability for hybrids, so future studies could include 370 analysis of heterosis among various hybrids for recovery from cold stress. 371 Many studies suggested flint genotypes are more cold tolerant than dent genotypes across a 372 number of different cold assays and developmental stages (Bhosale et al., 2007;Riva-Roveda 373 et al., 2016). With our approach, flint and dent genotypes did not classify into separate clusters, 374 although dent genotypes did appear to have more severe necrosis than flint genotypes. Ki11, a 375 flint genotype, exhibits the highest percentage of necrosis of all genotypes in our study.
Other 376 examples violating the assumption that dent genotypes are cold sensitive exist, such as a 377 favorable allele for cold tolerance being identified in a European dent line (Strigens et al., 2013). 378 Additionally, in a large study comparing cold tolerance of flint and dent varieties of European 379 origin, genotypes with a high degree of cold tolerance were found among both dent and flint 380 groups (Revilla et al., 2014). The same study found no strong pattern among cold tolerance and 381 latitude of geographical origin, which is also consistent with our results. 382 Assigning a latitude and other attributes, such as kernel texture and population group, to maize 383 inbreds can be a difficult task.  Cold stressed seedlings were transferred to a separate incubator with a 6°C/2°C day/night cycle 535 on day 9 after sowing, and stressed for the indicated amount of time (1, 2, 3, or 4 days), then 536 returned to the control-temperature growth chamber. Error bars represent standard error of the 537 mean of 6 plants. 538