Genetic dissection of southern corn leaf blight resistance in sweet corn through genome‐wide association studies and genomic selection Authors: Darlon V. Lantican, Juan M. Gonzalez, Marco Antonio Peixoto, Kristen A. Leach, Larissa Carvalho Ferreira, Vitor A. Silva de Moura, Katia V. Xavier, Peter Balint‐Kurti, William Tracy, Marcio F. R. Resende Published: 2026 ABSTRACT AbstractSouthern corn leaf blight (SCLB) is caused by the fungal pathogen Bipolaris maydis (syn. Cochliobolus heterostrophus Drechsler) and is a common disease of fall crops of sweet corn. Phenotyping for SCLB resistance is performed through visual scoring, which is subjective and may limit genetic gain for this quantitative trait. As an alternative, we integrated computer vision (CV)‐based phenotyping, genome‐wide association studies (GWASs), and predictive breeding approaches to dissect the genetic basis of SCLB resistance. We utilized a sweet corn diversity panel with 693 genotypes, for which whole‐genome resequencing produced a high‐density single‐nucleotide polymorphism (SNP) dataset. Broad‐sense heritability for visual scoring ranged from 0.44 to 0.73, while CV‐based phenotyping produced estimates ranging from 0.56 to 0.73 in multi‐environment resistance trials conducted across 5 years and three locations. We performed GWAS using 16,755,210 SNPs and identified 41 associated SNPs. Genomic selection (GS) models on visual scoring phenotypes achieved moderate prediction accuracies under cross‐validation of untested genotypes across characterized environments (0.22–0.47) and high prediction accuracies when predicting tested genotypes in uncharacterized environments (0.49–0.68). Using CV‐based phenotypes for GS, we observed prediction accuracies of 0.45–0.47 under the untested genotypes in the characterized environments cross‐validation scheme and 0.59–0.62 under the tested genotypes in the uncharacterized environments scheme. GS demonstrated reliability for ranking the individuals across a gradient of environments. These findings identify candidate loci and predictive breeding strategies to accelerate the development of resistant sweet corn cultivars. Abbreviations Belle Glade, FL 2023 Belle Glade, FL 2024 best linear unbiased estimate Citra, FL 2019 Citra, FL 2020 computer vision diseased leaf area log‐transformed diseased leaf area Efficient Mixed‐Model Association eXpedited Fixed and Random Model Circulating Probability Unification genotype × environment genomic best linear unbiased prediction genomic selection genome‐wide association study mixed linear model marker‐trait association Clayton, NC 2021 potato dextrose agar root mean square error southern corn leaf blight single‐nucleotide polymorphism INTRODUCTION Sweet corn (Zea mays L.) represents a unique market class of maize harvested at the milk stage, in which recessive mutations in the starch biosynthesis pathway lead to the accumulation of sugars rather than starch in the kernels (Creech, 1968; Finegan et al., 2022; Hu et al., 2021). Sweet corn has become a widely consumed vegetable crop with distinct production and market requirements since its rise in popularity in the early 20th century (Tracy, 2000). Unlike field corn, sweet corn is bred for traits such as ear appearance, tip fill, and shelf life, and is typically grown at lower plant densities to optimize ear quality (Dhaliwal et al., 2021; Kumari & McAvoy, 2025). These considerations make sweet corn breeding more comparable to vegetable breeding than to row‐crop improvement. Southern corn leaf blight (SCLB), caused by the fungal foliar pathogen Bipolaris maydis, is an important disease of maize in hot, humid tropical and subtropical regions (Renfro & Ullstrup, 1976). Race T of B. maydis was responsible for destroying more than 15% of the field corn crop in 1970, one of the most devastating crop epidemics in US history (Balint‐Kurti & Pataky, 2024; Bruns, 2017; Tatum, 1971). The pathogen produces small, diamond‐shaped, yellow‐gray lesions that coalesce under heavy infection, reducing photosynthetic area and limiting ear development (Kumar et al., 2022; Singh et al., 2012). Resistance to SCLB in maize is heritable and quantitative (Chen et al., 2023). Historically, breeders have relied on semiquantitative disease scores based on visual observation to select resistant genotypes, but this approach may limit phenotypic accuracy (Chiang & Bock, 2022). Computer vision (CV) enables machines to simulate human vision through object detection, segmentation, and understanding from digital images (Anagnostopoulos & Krinidis, 2024), offering a standardized, quantitative alternative to semiquantitative visual scoring (Bierman et al., 2019; Sykes et al., 2024). For example, the CV phenotyping package EarCV produced quantitative measurements of ear length and width that were highly concordant with manual measurements while enabling the extraction of traits such as tip fill, taper, curvature, and color that are difficult to quantify manually (Gonzalez et al., 2022). Furthermore, CV‐based approaches can increase phenotyping throughput by reducing per‐plot scoring time or by enabling distributed, parallel data collection by multiple operators, with centralized algorithmic evaluation that reduces interobserver variability. Applying such approaches to SCLB screening could improve both accuracy and efficiency in resistance breeding. Molecular markers are promising tools for increasing genetic gain in traits and have demonstrated positive impacts in sweet corn breeding through genomic selection (GS) (Graciano et al., 2025; Peixoto et al., 2025). However, the effectiveness of GS depends, among other factors, on the quality of the phenotypic data, which remains a bottleneck for achieving optimal genetic gain (Araus et al., 2018). GS leverages genome‐wide marker information to predict resistance without requiring explicit identification of causal loci (Budhlakoti et al., 2022). On the other hand, genome‐wide association studies (GWASs) can identify single‐nucleotide polymorphisms (SNPs) linked to resistance and facilitate the development of DNA markers for marker‐assisted breeding (Liu et al., 2025). Here, we incorporated CV‐based quantitative phenotyping with genomics to facilitate SCLB resistance breeding in sweet corn, motivated by the potential combination of molecular tools with reliable phenotyping. Using disease resistance phenotypes derived from ordinal visual severity ratings and image‐based quantitative estimates of diseased leaf area (DLA), we conducted GWAS with SNPs derived from whole‐genome resequencing. We further evaluated genomic prediction models to accelerate the identification of resistant genotypes for the next breeding cycle. Together, the approaches we implemented in this research provide a framework on how to complement visual scoring with quantitative disease phenotyping and genomic tools to identify disease resistance loci, improve selection efficiency, and fast‐track the development of sweet corn varieties with durable resistance to SCLB. MATERIALS AND METHODS Plant materials experimental design A panel of 693 sweet corn lines encompassing a large portion of the known diversity of sweet corn was screened for resistance to SCLB. This diversity panel includes open‐pollinated varieties and modern inbreds from the 1970s, popular sugary1 (su1) inbreds from the 1960s, shrunken2 (sh2) inbreds historically used in commercial breeding, and temperate and tropical materials. This collection is a compilation of the University of Wisconsin‐Madison Sweet Corn Breeding Program, the University of Florida Sweet Corn Breeding Program, and the US Department of Agriculture–Germplasm Resources Information Network system. For more details, see Colantonio (2023). The diversity panel was established in three locations in the United States over a 5‐year period. Trials were conducted twice in Citra, Florida, during the Spring season (April–July 2019) and Fall season (October 2020–January 2021) (hereafter referred to as CT19 and CT20, respectively; −29.40° N, −82.17°W; planting was delayed to October due to pandemic restrictions); once in Clayton, North Carolina, during the Winter season (February–May 2021) (NC21; −35.67° N, −78.50° W); and twice in Belle Glade, FL, during the Winter season (February–May) in 2023 and 2024 (BG23 and BG24, respectively; −26.66° N, −80.63° W). Plants were grown in single‐row plots with 12 plants per row, spaced 7 in. apart within rows. Row spacing was 36 in. in the Florida trials, whereas in the Clayton, NC trial, plants were spaced 2 ft apart within rows and rows were spaced 38 in. apart. The experiments were laid out in an α‐lattice incomplete‐block design, with one hybrid and two additional inbreds used as experimental checks within each incomplete block. The experiments in Citra had two replicates, while those in Clayton, NC, and Belle Glade had three replicates. For statistical modeling, each unique combination of year and location was treated as a distinct environment, resulting in a total of five environments. SCLB inoculation Protocols for disease inoculation differed among trial locations and years. In CT19, plants were exposed to natural B. maydis disease pressure. In the same location the following year (CT20), plants received both natural and artificial epiphytotics of B. maydis. Artificial inoculation was performed by growing a single‐spore isolate of B. maydis on potato dextrose agar (PDA) and inducing sporulation by maintaining high humidity at 25°C. A spore suspension (∼1000 spores/mL) was used as inoculum, applied twice, 1 week apart, using a backpack sprayer at a constant spray rate of 20 gallons per acre when the checks had four to six fully developed leaves (V4–V6 stage). Meanwhile, the NC21 trial was conducted using a local single‐spore population of B. maydis grown on PDA and used to inoculate sterile sorghum grains. A total of 20–30 inoculated sorghum grains were dropped into the corn whorl when the checks were at the V4–V6 stage (Sermons & Balint‐Kurti, 2018). In BG23 and BG24 trials, plants were exposed to natural SCLB disease pressure. Disease resistance data collection We employed a semiquantitative, visual rating score‐based pipeline and a quantitative, CV‐based pipeline to screen the tested genotypes for SCLB resistance. Disease scoring occurred when at least half of the checks were shedding pollen. Here, we define anthesis as the stage at which approximately 50% of the tassel has shed pollen. SCLB resistance in the CT19 and CT20 environments was assessed by visually scoring three plants per row on a scale from 1 to 5, where 1 indicated complete resistance, 2 indicated high resistance, 3 indicated moderate resistance, 4 indicated high susceptibility, and 5 indicated complete susceptibility. In the NC21 environment, the phenotype was assessed using a 1–9 scale, where 1 indicated complete susceptibility, and 9 indicated complete resistance (Sermons & Balint‐Kurti, 2018). For BG23 and BG24 trials, plots were rated on a scale of 0–5, with 0 indicating complete resistance and 5 complete susceptibility. The NC21 environment was scored twice, 2 weeks apart: once when Yellow‐Stone (a commercial hybrid used as a check) started shedding, and again when at least half of the Yellow‐Stone checks were shedding pollen. We also used inbred checks to monitor disease pressure. Days to anthesis for each row were recorded in CT19. The CV‐based phenotyping was conducted for CT19 and NC21 trials. Two ear leaves and two V max leaves were harvested from each row and laid flat against a uniform‐colored background. V max refers to the youngest leaf showing symptoms of SCLB. A Nikon model D750 was used with a Nikon DX lens, 0.98 ft aperture, 52 mm diameter, and ISO 1200, mounted to the ceiling with a double‐jointed ball‐head clamp. A color checker was included in all images for color normalization, along with a QR code to identify the experimental plot (Figure 1). The images taken for each plot were saved in JPG format and analyzed using a CV method. Here, we refer to the CV method as a custom Python script written using the OpenCV library version 4.5.0, a set of tools for CV (Bradski, 2000). The Python script used for leaf CV image analysis is available in our public GitHub repository (https://github.com/Resende‐Lab/SCLB‐Disease). Leaf imaging set up with QR‐coded plot IDs (bottom right) and color checker for computer vision phenotyping of southern corn leaf blight disease severity in sweet corn. In the CT19 environment, a black cloth attached to a wooden board was used as the background. A wooden frame was used to clamp the leaves down to ensure they lay flat during imaging. The wooden board was placed inside a custom‐built light box, which diffused ambient light to prevent shadows in the images. In the NC21 environment season, a white trifold was used as the background. The leaves were placed on the trifold, and the camera was mounted on the ceiling. Due to consistent lighting in the room and the relatively flat position of the leaves, the use of a light box or wooden frame was no longer required. A CV‐based method was used to segment leaf tissue and to identify disease lesions within the segmented leaf area (Figure 2). Color channel conversions and segmentation algorithms were applied to isolate regions of interest in the image. Segmentation was performed using either Otsu's method or K‐means clustering. Otsu's method (Otsu, 1979) calculates an intensity threshold that minimizes the weighted within‐class variance when separating grayscale pixel values into two classes. First, a QR scan was performed, and then the QR code tag was removed by using the QR scan coordinates to find the largest connected component in that vicinity. The tag was then segmented out using K‐means clustering. A K‐means approach was chosen because the tags were orange in CT19 and NC21. The hue channel was used to segment and remove the orange square, which was made from construction paper (color reference not used in this analysis). The wooden frame in the CT19 analysis was removed by segmenting with Otsu on the blue channel. The background was separated from the leaves by performing Otsu segmentation on the green channel. Finally, lesions were separated from the leaves by performing Otsu segmentation on the hue channel. A connected component analysis on the lesion image matrix was used to extract descriptive features for each segmented lesion. The area was calculated using Green's theorem (Lang, 1987). Dominant colors in the RGB, HSV, and LAB color spaces were extracted using K‐means clustering and by averaging all the foreground pixel intensities. The DLA trait was calculated by dividing the total number of lesion pixels by the total number of pixels in the leaf. All DLA calculations were log‐transformed (referred to as DLA_Log). A computer vision pipeline used to calculate diseased leaf area (DLA) from image capture of sweet corn leaves. Phenotypic data analysis Phenotypic data within each plot were averaged prior to statistical analysis, and a two‐stage analysis was performed (Holland & Piepho, 2024). We used the restricted maximum likelihood and best linear unbiased prediction methodology (Henderson, 1974; Patterson & Thompson, 1971) for evaluating the traits in both single‐environment and multi‐environment analyses. The following linear mixed model was used to analyze each environment separately: (1) y=Xr+Zg+Wb+e where y is a vector of the phenotypic performance; r is the vector of fixed effect of replications, checks within incomplete blocks and covariates for developmental group based on anthesis added to the overall mean; g is the vector of fixed effects of the genotypes; b is the vector of random effects of the incomplete blocks, where b ∼ N(0, σb2 I), given σb2 is the incomplete block variance and I is an identity matrix; and e is the vector of the residuals, where e ∼ N(0, σe2 I), given σe2 is the residual variance. X, Z, and W are the incidence matrices for r, g, and b, respectively. Variance components were calculated based on Equation (1) except that the genotype term was set as random effect, where g ∼ N(0, σg2 I), given σg2 is the genotypic variance. Broad‐sense heritability (H 2) for each environment was calculated as the proportion of genotypic variance over the total phenotypic variance (Piepho & Möhring, 2007), as per the equation below: (2) H2=σg2σg2+σb2+σe2 A multi‐environment model was also used to estimate best linear unbiased estimates (BLUEs) based on the following linear mixed model: (3) y=Xr+Zg+Wb+Ti+e where y is a vector of the phenotypic performance; r is the vector of fixed effects of replications within environment, checks within incomplete blocks within environment, and covariates for developmental group based on anthesis added to the overall mean; g is the vector of the fixed effects of the genotypes; b is the vector of random effects of the incomplete blocks within season, where b ∼ N(0, σb2 I), given σb2 is the incomplete block variance and I is an identity matrix; i is the vector genotype‐by‐environment (G × E) interaction effects, assumed as random, where i∼ N(0, σg×e2 I), where σg×e2 is the genotype‐by‐environment variance; and e is the vector of the residuals, where e ∼ N(0, σe2 I). X, Z, W, and T are the incidence matrices for r, g, b, and i, respectively. For visual phenotyping, multi‐environment BLUEs were calculated from CT19, CT20, NC21, BG23, and BG24 (hereafter referred to as 5‐year). In contrast, multi‐environment BLUEs from CV‐based phenotyping included only two environments (CT19 and NC21), hereafter referred to as 2‐year. BLUEs were calculated via the ASREML package (Gilmour et al., 2015) in R and used for second‐stage analyses (GWAS and GS). For GWAS, we used single‐environment and multi‐environment BLUEs derived from both visual and CV‐based phenotyping. The multi‐environment BLUEs accounts for systematic differences among environments, including differences in disease pressure, environmental conditions, and scoring scales used during phenotyping for GWAS. For GS, single‐environment BLUEs for both visual scoring and CV‐based phenotyping were used, but the environment main effect was included in the prediction model to account for scale and environmental differences across trials. The data were organized into a genotype‐by‐environment matrix and transformed into long format for modeling, ensuring that only non‐missing records were retained for GS. Genome‐wide variant calling and filtering We performed whole‐genome re‐sequencing of 693 genotypes from the SweetCAP diversity panel at ∼8.45× sequencing depth on the Illumina NovaSeq platform (PE 2 × 150 bp). Sequencing reads were aligned to the sweet corn reference genome Ia453‐sh2 (Hu et al., 2021) using BWA‐MEM (Li, in press) with default parameters. Genome‐wide variants were detected using two variant callers: FreeBayes (Garrison & Marth, in press) and GATK (McKenna et al., 2010; Van der Auwera & O'Connor, 2020). Intersecting variants from both callers were identified, and filters were applied according to GATK Best Practices (DePristo et al., 2011) using BCFtools (Danecek et al., 2021). Further variant filtering was based on a minor allele frequency threshold of 0.01, missingness of 30% across variants, and linkage disequilibrium pruning as previously described (Colantonio, 2023). A final set of 16,755,210 variants was used for GWAS. From this set, 200,000 SNPs were randomly sampled and subjected to additional filtering (see Section 2.7.2 for details), resulting in a dataset of 128,202 SNPs for GS. Genome‐wide association studies We employed two GWAS approaches to control false positives or spurious associations: (i) Fixed and Random Model Circulating Probability Unification (FarmCPU) and (ii) Efficient Mixed‐Model Association eXpedited (EMMAX). These methods used both multi‐season and single‐environment BLUEs from the first‐stage analysis, along with the 16‐million‐SNP dataset. We implemented FarmCPU using the GAPIT version 3 package (Wang & Zhang, 2021) and mixed linear model (MLM) using the EMMAX software (Kang et al., 2010). FarmCPU was implemented in GAPIT after modifying the R code to subset 1 million of the 16 million SNPs for principal component analysis, while retaining the full 16‐million‐SNP dataset for association testing. Three principal components (PC1–PC3) were calculated and included as fixed covariates in downstream analyses to account for population structure (complete details can be found in Supporting Information S1). The customized GAPIT script is available in our GitHub repository (https://github.com/Resende‐Lab/SCLB‐Disease). Population structure and unequal relatedness were accounted (Colantonio, 2023), and a genomic relationship matrix (kinship) was estimated via the VanRaden method (VanRaden, 2008). The MLM/EMMAX and FarmCPU also included starch biosynthetic mutations (su1, sh2‐R, and sh2‐i) as covariates, to avoid spurious associations with such mutations as previously demonstrated for tocotrienols and certain carotenoids (Baseggio et al., 2019, 2020). Significance thresholds were established using a Bonferroni correction based on the SNP markers (−log10[0.05/16,755,210]). Additionally, we computed the significance threshold adjusted for the effective number of independent SNPs calculated using the Genetic type 1 error calculator (M. X. Li et al., 2012), corresponding to −log10(0.05/4,719,846.15) = 7.97. Finally, genes located within ±500 kb of a significant SNP were designated as part of the associated locus. The B73 RefGen v5 in JBrowse at MaizeGDB was used to identify candidate genes in the region. Candidate genes were identified based on the close proximity to the associated SNP locus and corresponding literatures supporting roles in plant resistance. Genomic prediction models Environmental data acquisition and processing Planting dates were April 4, 2019 (Citra; CT19); October 6, 2020 (Citra; CT20); February 17, 2021 (North Carolina; NC21); May 24, 2023 (Belle Glade, BG23); and May 16, 2024 (Belle Glade, BG24), with a 72‐day interval to harvest—an interval appropriate for sweet corn production. Site‐specific elevation data (21, 21, 109, 3, and 3 m, respectively) were recorded. Daily weather data (e.g., temperature, radiation, dew point) were obtained using the get_weather function from the envirotypeR package (Costa‐Neto et al., 2021) over the period from planting to harvest. Simultaneously, soil properties were acquired via the get_soil function. Missing temperature values were imputed with the corresponding means, and the daily observations were aggregated into interval metrics (over windows of 0, 7, 14, 28, 42, 56, and 70 days) using the EnvRtype package (Costa‐Neto et al., 2021). Temperature parameters were calculated using the suggested cardinal limit (T_base1 = 8°C, T_base2 = 45°C, T_opt1 = 30°C, T_opt2 = 37°C) according to the established threshold for maize (Costa‐Neto et al., 2021). Radiation and atmospheric variables (adjusted for elevation) were similarly computed. Finally, date components (year, month, day) were extracted to support subsequent analyses. An environmental kernel was then computed from the processed data to generate an environmental matrix (WMat) that captures the dynamic aspects of the target environments. Genomic data preparation A set of 200,000 SNPs was randomly sampled from the whole‐genome resequencing data. Quality control was applied by allowing up to 30% missing data and setting a minor allele frequency threshold (≥0.01), totaling 128,202 SNPs used in the subsequent analyses. A genomic relationship matrix (G_matrix) was computed using VanRaden (2008) method in the AGHmatrix package (Amadeu et al., 2016). In addition, missing SNP values were imputed with the column means to derive the G_matrix and the marker‐based design matrix (M_matrix). These matrices were aligned with the phenotypic data by ordering individuals consistently across datasets. Models for genomic prediction To evaluate the application of GS for SCLB resistance prediction, we implemented a reaction norm model (Jarquin et al., 2014), a standard genomic framework for multi‐environment trials. To this end, we implemented two distinct statistical models. First, a genomic best linear unbiased prediction (GBLUP) model was used in the BGLR package (Pérez & de los Campos, 2014), reported as follows: (4) y=u+ZEβ+ZGa+gw+e where y is the vector of BLUEs estimated in the first stage of the analysis; β is the vector of environment effects, where β ∼ N(0, Iσβ2), being σβ2 the environmental variance; a is the vector of genetic effects, being a ∼ N(0, G⊗σa2), being σa2 the genetic variance; gw is the interaction between the genotype and the environment, with gw=N(0,[(ZGGZ′G)∘((ZEWZ′E)],σgw2, where represents the Hadamard product between genomic and environment matrices with variance σgw2 due to G × E, and W is the environmental similarity matrix coming from the covariates in each environment; and e is the residual effect (random), where e ∼ N(0, Iσres2), with σres2 is the residual variance. ZE and ZG are incidence matrices connecting the information of the factors with the vector of BLUEs. For comparison, a second set of reaction norms models was implemented using BayesB (Meuwissen et al., 2001), where marker effects were estimated rather than individual predictions. The complete model was implemented as follows: (5) y=u+ZEβ+Xm+gw+e where X is the matrix for the predictors (i.e., the marker matrix for each individual in y, M_matrix), m is a vector with the effects of each SNP (markers), m ∼ N(0, Iσm2), being σm2 the variance of nonzero markers (π), which follow an inverse χ2 distribution. Derivations of the Models (4) and (5) explained above were also implemented for comparison and are described in Supporting Information S2. Specifically, these reduced models exclude either the environmental main effects or the environmental covariance term to evaluate the relative contribution of genomic main effects, environmental effects, and genotype‐by‐environment (G × E) interactions to predictive performance. In each model, genotypes were randomly partitioned into fivefolds, and the untested genotypes in characterized environments cross‐validation (Jarquin et al., 2017) process was repeated over 10 replicates. For each fold, BLUEs values of test genotypes were masked and predicted using MCMC (10,000 iterations, burn‐in of 1000, thinning of 10). Prediction accuracy was assessed by Pearson correlation, and root mean square error (RMSE) was calculated for each environment individually. In addition to the untested genotype in a characterized environment cross‐validation scheme, a tested genotype in uncharacterized environments cross‐validation scheme (Jarquin et al., 2017) was implemented to evaluate the GS models. Under this true validation scheme, each environment was sequentially treated as the target (i.e., all phenotypes from that environment were masked) while models were trained using data from the remaining environments. Five suites of reaction norms similar to those above were implemented for each of the two independent sets of models (GBLUP, BayesB). Prediction accuracy and RMSE were obtained with the same settings in untested genotype in characterized environment cross‐validation, and computed for each target environment. The datasets and code used in this paper can be found in the GitHub repo: https://github.com/Resende‐Lab/SCLB‐Disease. RESULTS Quantitative genetics parameters A total of 536, 461, 464, 101, and 86 genotypes from the SweetCAP diversity panel lines were screened for SCLB resistance in CT19, CT20, NC21, BG23, and BG24, respectively. All rows were evaluated for SCLB disease severity using an ordinal visual rating scale. Additionally, the log‐transformed diseased leaf area (DLA_Log) was calculated for the CT19 and NC21 environments using a quantitative CV‐based approach (Figure 3). A summary of the SCLB resistance data is presented for each environment (Figure S1). In general, a wide range of phenotypic variability was observed. Single‐environment analyses revealed moderate to high broad‐sense heritability (H 2) estimates for sweet corn SCLB resistance. Heritability values ranged from 0.32 to 0.73 (Table 1). CV‐based methods for assessing SCLB resistance resulted in comparable or slightly higher broad‐sense heritability estimates relative to visual scoring within environments where both phenotyping methods were applied (CT19 and NC21). In CT19, CV‐based phenotyping produced a slightly higher heritability estimate than visual scoring (0.56 vs. 0.52), whereas in NC21, both methods yielded similar heritability estimates (0.73). Representative maize leaves from the diversity panel sorted by diseased leaf area (DLA) quantified using computer vision. Leaf pairs display the original image (left) and the computer vision (CV)‐derived DLA segmentation (right), ordered from lowest to highest disease severity. TABLE 1Mean, broad‐sense heritability (H 2) estimates, and standard errors (in parentheses) for single‐environment and multi‐environment analyses of South corn leaf blight resistance.EnvironmentPhenotyping methodMean a Intercept b H 2 5‐yearVisual scoring3.212.180.32 (0.03)2‐yearComputer vision1.832.380.50 (0.04)CT19Visual scoring2.932.040.52 (0.04)CT19Computer vision1.210.350.56 (0.04)CT20Visual scoring3.452.710.44 (0.03)NC21Visual scoring3.663.050.73 (0.02)NC21Computer vision2.552.460.73 (0.02)BG23Visual scoring2.281.110.54 (0.06)BG24Visual scoring2.633.330.47 (0.08) Note: Visual scoring refers to disease ratings obtained through subjective visual assessment using a standard scale; computer vision (CV) refers to phenotypes derived from the log‐transformed diseased leaf area (DLA_Log) extracted via CV‐based phenotyping.Abbreviations: BG23, Belle Glade, FL 2023; BG24, Belle Glade, FL 2024; CT19, Citra, FL 2019; CT20, Citra, FL 2020; NC21, Clayton, NC 2021. a Arithmetic mean of the response used in the model (visual score or DLA_Log). b Mixed model intercept for each analysis.John Wiley & Sons, Ltd. Mean, broad‐sense heritability (H 2) estimates, and standard errors (in parentheses) for single‐environment and multi‐environment analyses of South corn leaf blight resistance. Note: Visual scoring refers to disease ratings obtained through subjective visual assessment using a standard scale; computer vision (CV) refers to phenotypes derived from the log‐transformed diseased leaf area (DLA_Log) extracted via CV‐based phenotyping. Abbreviations: BG23, Belle Glade, FL 2023; BG24, Belle Glade, FL 2024; CT19, Citra, FL 2019; CT20, Citra, FL 2020; NC21, Clayton, NC 2021. Arithmetic mean of the response used in the model (visual score or DLA_Log). Mixed model intercept for each analysis. We explored phenotypic correlation between environments and phenotyping methods (Figure S2). Correlations between SCLB visual scores BLUEs across the five single environments ranged from 0.45 to 0.68. Phenotypic correlation between the CT19 and NC21 environments for the CV‐based method was 0.58. The correlation between the SCLB visual score and the CV‐based method CT19 was 0.57. Notably, we observed a high correlation (0.87) between the visual scoring and CV‐based BLUEs in NC21. Nonetheless, differences in the distributions of the two BLUE sets from visual scoring and CV suggest that, while the phenotypes are strongly linearly correlated, the two phenotyping methods capture disease severity on different scales and with different distributions (Figure S3). Partitioning of phenotypic variance based on a multi‐environment model highlighted the genetic control of SCLB resistance and differences in the distribution of variance components associated with each phenotyping method (Figure 4). Under visual scoring, a large proportion of the observed phenotypic variance was largely contributed to the environment component, followed by the genotypic component. CV‐based phenotyping exhibited a variance structure characterized by a relatively larger contribution of genotypic variance compared with environmental variance within the environments evaluated. The contribution of genotype × environment interaction variance remained low (approximately 10%) under both phenotyping methods. Estimating variance components for visual scoring using only the environments in which CV‐based phenotyping was conducted (CT19 and NC21) revealed reduced genotypic and environmental variance, but increased genotype × environment interaction variance, relative to the CV‐based method (Figure S4). Proportion of phenotypic variance in southern corn leaf blight resistance explained by genotype (σg2​), environment (σenv2​), genotype × environment (σg×e2​), block (σb2​), and residual error (σe2​) in the multi‐environment model using visual scoring phenotyping (A; five environments: Citra, FL 2019 [CT19]; Citra, FL 2020 [CT20], Clayton NC [NC21], Belle Glade, FL 2023 [BG23]; Belle Glade, FL 2024 [BG24]) and computer vision phenotyping (B; two environments: CT19; NC21). GWAS reveals SNPs associated with SCLB We tested the association of 16,755,210 SNPs derived from whole‐genome re‐sequencing of the SweetCAP diversity panel to BLUEs obtained from both visual scoring and CV‐based phenotyping. Analyses were conducted separately for single‐season BLUEs and multi‐season BLUEs to capture both season‐specific and across‐season associations. The genetic diversity of the panel was assessed using principal component analysis (Figure S5), which indicated the absence of strong population structure. We applied the FarmCPU model and identified a total of 21 SNPs significantly associated with resistance to SCLB (Table 2). TABLE 2Significant single‐nucleotide polymorphisms (SNPs) associated with resistance to southern corn leaf blight detected by Fixed and Random Model Circulating Probability Unification (FarmCPU).SNPPhenotyping methodEnvironmentEffect p‐value−log10(P)Linked candidate genes or proteins a Chr1_199765263VisualCT19−0.26729.54E‐1413.02030672Ethylene‐responsive transcription factor ERF112 (Ma et al., 2024)Chr1_232188031CVNC210.29906.94E‐098.1587Glutaredoxin23 (Bleau & Spoel, 2021)Chr2_6906594VisualCT190.14493.01E‐098.5217Precursor elicitor peptide1 (Zelman & Berkowitz, 2023)Chr2_163317806VisualCT190.29182.01E‐109.695860707NAC transcription factor 5 (Dong et al., 2024; Nuruzzaman et al., 2013)Chr2_186708747VisualCT19−0.15406.57E‐1110.18231778Myb‐related protein 306 (Biswas et al., 2023)Chr2_207842723VisualCT19−0.38641.10E‐1110.95938305MYB‐transcription factor 108 (Biswas et al., 2023)Chr3_197281858CVNC21−0.19368.09E‐109.092148568Cytokinin oxidase 4 (Hönig et al., 2026; Gupta et al., 2020)Chr3_36210548VisualCT19−0.22136.83E‐1211.16533662Pectin methylesterase inhibitor 2 (Molina et al., 2024)Chr3_51138707CVNC21−0.57432.15E‐098.668442685Calcineurin B‐like‐interacting protein kinase1 (Vishwakarma et al., 2026)Chr3_52630814VisualCT19−0.16895.35E‐109.272046511Polygalacturonase 12 (D'Ovidio et al., 2004)Chr4_71671611VisualCT190.20964.65E‐098.3322Chr4_122511958CVNC210.47683.59E‐1211.44456825Cytokinin response regulator 7 (Gupta et al., 2020)Chr4_45406703CVNC21−0.23497.90E‐1110.10216085Chr5_180910047CVNC21−0.43001.75E‐1514.7579769Prohibitin 3 (Nadimpalli et al., 2000)Chr5_206599174CVNC210.41361.37E‐109.863080807Basic helix‐loop‐helix transcription factor 53 (Ding et al., 2023)Chr6_10182302CVNC21−0.19725.56E‐1211.2549003Chr6_6091658VisualCT190.19204.07E‐109.390080018NAC transcription factor 62 (Dong et al., 2024)Chr7_81785418VisualCT19−0.30412.26E‐1211.64631102Chr8_163565779CVNC21−0.20571.11E‐098.956618482Basic helix‐loop‐helix transcription factor 54 (Ding et al., 2023)Chr8_77002294VisualCT190.25632.14E‐1110.67032933Calcium‐dependent protein kinase 23 (Deng et al., 2025)Chr8_79315562CVNC21−0.34361.00E‐1615.99877044Protein phosphatase homolog 15 (Barman & Ray, 2020)Chr9_14620318VisualCT190.16124.68E‐109.329530747ARR‐B‐transcription factor 2 (Falconieri et al., 2022)Chr10_2537545CVNC21−0.41116.01E‐1110.22127345Rust resistance‐like protein RP1‐4 (Pataky & Campaña, 2007)Chr10_48218503CVNC210.37046.20E‐109.207930032Abbreviations: CT19, Citra, FL 2019; CV, computer vision; NC21, Clayton, NC 2021. a Genes that are within ± 500 kbp of the SNP region.John Wiley & Sons, Ltd. Significant single‐nucleotide polymorphisms (SNPs) associated with resistance to southern corn leaf blight detected by Fixed and Random Model Circulating Probability Unification (FarmCPU). Abbreviations: CT19, Citra, FL 2019; CV, computer vision; NC21, Clayton, NC 2021. Genes that are within ± 500 kbp of the SNP region. The FarmCPU model identified a total of 24 significant SNPs across phenotyping methods and seasons (Figure 5; Table 2). Significant associations were detected only in the CT19 and NC21 environments, which also exhibited the highest heritability estimates among the environments evaluated. In the CT19 environment, 12 SNPs were detected using visual scoring phenotypes, while no marker‐trait associations (MTAs) were found using CV‐based phenotyping. On the other hand, in NC21, we also identified 12 SNPs using CV‐based phenotypes, whereas no associations were detected in the visual scoring phenotypes. We also did not detect any significant MTAs using single‐environment BLUEs from CT20, BG23, and BG24, or using multi‐environment BLUEs from the five visual scoring phenotyping environments (CT19, CT20, NC21, BG23, and BG24) and the two CV‐based phenotyping environments (CT19 and NC21). Hence, identification of significant MTAs appears to depend on the specific phenotype–environment combination, which suggests that differences in disease pressure, environmental conditions, trait expression, measurement resolution, and genotype‐by‐environment interactions may impact results. Manhattan plot of a genome‐wide association study of southern corn leaf blight resistance using Fixed and Random Model Circulating Probability Unification (FarmCPU) across phenotyping methods and seasons. Significant single‐nucleotide polymorphisms (SNPs) exceeding the −log10(P) threshold are shown for single‐environment best linear unbiased estimates (BLUEs) from (A) CT19 visual scoring phenotyping, and (B) NC21 computer vision (CV)‐based phenotyping. Peaks above the threshold highlight associations. Dashed line indicates Bonferroni‐corrected significance threshold (−log10[0.05/16,755,210] = 8.53), while the solid line represents significance threshold adjusted for the effective number of independent SNPs calculated using the Genetic type 1 error calculator (GEC; M. X. Li et al., 2012), corresponding to −log10(0.05/4,719,846.15) = 7.97. bHLH24, basic helix‐loop‐helix transcription factor 54 (Ding et al., 2023); CPK23, calcium‐ dependent protein kinase 23 (Deng et al., 2025); ERF112, ethylene‐responsive transcription factor 112 (Ma et al., 2024); PHB3, prohibitin 3 (Nadimpalli et al., 2000); PMEI12, pectin methylesterase inhibitor 2 (Molina et al., 2024). Despite this environment‐specific pattern, comparison across phenotyping methods and environments revealed consistent signals in specific chromosomal regions. On chromosome 3, SNP Chr3_52630814 was identified using visual scoring phenotyping in CT19, and SNP Chr3_51138707 was identified using CV‐based phenotyping in NC21, and they were mapped within approximately 1.5 Mb of each other. Similarly, on chromosome 8, signals detected by visual scoring in CT19 (Chr8_77002294) and by CV‐based phenotyping in NC21 (Chr8_79315562) were approximately 2.3 Mb apart. One significant MTA detected by FarmCPU (Chr10_2537545 in NC21 CV‐based phenotyping) was linked to a resistance (R)‐gene rust resistance‐like protein RP1‐4 (Pataky & Campaña, 2007). No CT19 visual‐scoring MTAs were directly physically located to genes with classical R‐gene annotation. However, when considering the gene's defense‐related functions, we observed that 11 of 12 CT19 visual scoring SNPs and 9 of 12 NC21 CV‐based SNPs were located near genes with potential roles in disease resistance, including transcriptional regulation, hormone signaling, redox regulation, cell wall modification, and pathogen‐response signaling (Table 2). The FarmCPU MTA at ∼207 kb of chromosome 5 was also previously detected as a region associated with SCLB resistance in a corn NAM population (Kump et al., 2011). Similarly, Kump et al. (2011) also identified a QTL region for SCLB resistance at position ∼181 Mb, which was also detected in this current study. The MLM/EMMAX, on the other hand, did not detect significant SNPs that passed the Bonferroni‐corrected significance threshold. However, visual inspection of the Manhattan plots revealed regions with clusters of SNPs forming tower‐like peaks just below the threshold (Figure S6). Although these SNPs did not pass the significance threshold, their importance warrants further investigation since Bonferroni correction is a conservative threshold, as previously observed (Gulisano et al., 2023). Among these suggestive associations, one (Chr1_30951358) was identified using multi‐season BLUEs from visual scoring phenotyping, while four (Chr2_106950334, Chr5_220087088, Chr8_179436624, and Chr10_79763266) were detected from the CT19 CV‐based BLUEs. We then compared the significant SNPs identified by FarmCPU with the suggestive SNPs from MLM/EMMAX. While most signals appeared model‐specific, consistent associations emerged on chromosomes 5 and 8. On the distal end of chromosome 5, FarmCPU identified SNPs between positions ∼181 and 207 Mb, whereas MLM/EMMAX detected a putative association at position ∼220 Mb. Similarly, on chromosome 8, FarmCPU detected a signal at 163.6 Mb, while MLM/EMMAX identified a SNP peak at position ∼179.4 Mb. Genomic selection across cross‐validation schemes based on visual scores We evaluated the performance of GS for predicting SCLB resistance using visual scoring phenotyping under different cross‐validation schemes. Under the untested genotype in a characterized environment cross‐validation scheme, prediction accuracies were moderate, ranging from 0.22 to 0.47 (Figure 6). Prediction accuracy varied across environments, with higher values in CT19, CT20, and NC21 compared to those in BG23 and BG24. This observation can be attributed to the size of the training population: the CT19, CT20, and NC21 environments have 536, 461, and 464 genotypes, respectively, whereas only 101 and 86 genotypes were tested in BG23 and BG24. Genotype overlaps across all environments show that 46 genotypes were phenotyped across all five environments, while BG23 and BG24 had only five and one unique genotypes, respectively, with most of their entries overlapping with the larger CT19, CT20, and NC21 (Figure S7). Across all environments tested, both GBLUP and BayesB resulted in comparable prediction accuracies. However, GBLUP produced slightly higher and more stable correlations across replicates, whereas BayesB occasionally yielded greater variability across folds. Prediction accuracy under untested genotype in characterized environment cross‐validation scheme for southern corn leaf blight (SCLB) resistance (visual scoring phenotyping) across five environments using whole‐genome regression (genomic best linear unbiased prediction [GBLUP], BayesB). Reaction‐norm variants per approach: Mod1, G; Mod2, G + E; Mod3, G + E + KW; Mod4, G + E + G × E (KGW); Mod5, G + G × E (KGW). Models 4 and 5 explicitly include G × E interaction through the KGW kernel derived from genomic relationships and environmental covariates. E, environmental main effects (BRR on W); G, genomic main effects (KG); KGW, interaction kernel KG × KW; KW, RKHS kernel from W. To further evaluate the robustness of GS across environments, we next examined the tested genotypes in uncharacterized environments scheme, which predicts tested genotypes in new environments, where higher accuracies are generally observed in comparison to untested genotypes in characterized environments cross‐validation scheme. Under the tested genotypes in uncharacterized environments cross‐validation scheme, accuracies ranged from 0.49 to 0.68 (Figure 7). Across environments, no differences were observed between the two models: GBLUP was observed to have consistent prediction accuracies, ranging from 0.49 to 0.68, whereas BayesB results were similarly consistent, spanning from 0.49 to 0.67. The results suggest that Bayesian variable selection did not provide a clear advantage over the simpler GBLUP model in this dataset. Prediction accuracy under tested genotypes in uncharacterized environments cross‐validation scheme for southern corn leaf blight (SCLB) resistance (visual scoring phenotyping) across five environments using whole‐genome regression (genomic best linear unbiased prediction [GBLUP], BayesB). Reaction‐norm variants per approach: Mod1, G; Mod2, G + E; Mod3, G + E + KW; Mod4, G + E + G × E (KGW); Mod5, G + G × E (KGW). Models 4 and 5 explicitly include G × E interaction through the KGW kernel derived from genomic relationships and environmental covariates. E, environmental main effects (BRR on W); G, genomic main effects (KG); KGW, interaction kernel KG × KW; KW, RKHS kernel from W. To test whether incorporating environmental covariates could improve GS performance, we evaluated models that accounted for genotype × environment (G × E) interactions, a W‐matrix constructed from a comprehensive set of soil and weather covariates. Environmental data included soil clay content across multiple depth layers (0–200 cm) and time‐resolved atmospheric variables describing temperature, radiation, humidity, precipitation, vapor pressure deficit, and evapotranspiration, summarized across multiple intervals spanning early to late crop development. Environmental kinship analysis revealed two distinct environmental clusters, with the three Florida environments (CT19, BG23, and BG24) grouping together and showing high similarity, while CT20 and NC21 formed a separate cluster distinct from the other environments (Figure S8). Across both untested genotypes in a characterized environment and the tested genotype in an uncharacterized environment cross‐validation schemes, the inclusion of the W‐matrix did not significantly improve predictive ability relative to the genotypic main effects alone (Figures 6 and 7). In several environments, accuracies were nearly identical between the two models, while in others, the W‐matrix marginally reduced performance. These findings suggest that, at least for the environments and traits evaluated here, modeling G × E through environmental covariates (Models 4 and 5) does not enhance prediction accuracy. Importantly, this result demonstrates that the simplest GS framework, using only the G‐matrix modeled via GBLUP (Mod1), is sufficient for practical implementation in SCLB resistance breeding, without the added computational complexity of modeling G × E. Genomic selection on CV‐based phenotypes We also compared the differences between the prediction accuracies of visual scoring and CV‐based phenotyping to assess how the two methods affected the prediction under two cross‐validation schemes. Here, we only modeled the genomic main effect under GBLUP and BayesB. Using the untested genotype in a characterized environment cross‐validation scheme, we observed mean prediction accuracies of 0.45 and 0.451 in GBLUP and BayesB, respectively (Figure 8). These values were approximately 36.4%–36.7% higher than the corresponding accuracies we obtained using visual scoring phenotypes. In NC21, the prediction accuracy values were approximately 0.47, which was comparable to those obtained from similar visual‐scoring models, with only an approximately 4% advantage. Prediction accuracy under untested genotype in characterized environment cross‐validation scheme (A) and tested genotypes in uncharacterized environments cross‐validation scheme (B) for southern corn leaf blight (SCLB) resistance (computer vision [CV]‐based phenotyping) across two environments using whole‐genome regression (genomic best linear unbiased prediction [GBLUP], BayesB) modeling only the genomic main effect. We evaluated the prediction accuracies of CV‐based phenotyping under the tested genotypes in an uncharacterized environment cross‐validation scheme. Under this scheme, we observed a similar trend to the untested genotypes in the characterized environments scheme, where CV‐based phenotyping showed a clear advantage in CT19 and comparable prediction accuracies in NC21 relative to visual phenotyping. In CT19, prediction accuracies from CV‐based phenotyping were approximately 0.60, which was about 22.45% higher than visual scoring (∼0.49 for both Model 1 of GBLUP and BayesB). In NC21, prediction accuracies from CV‐based phenotyping were approximately 0.62, which was comparable to the values obtained using visual scoring phenotypes, approximately 0.63. DISCUSSION SCLB significantly affects sweet corn production, especially in warm areas. Hence, breeders aim to develop genotypes with resistance to SCLB to provide a sustainable and robust solution against the negative impact of this disease. Although breeding for SCLB resistance has been largely successful, it may suffer from subjective differences between scorers (Poland & Nelson, 2011). Here, we developed a CV‐based phenotyping approach generating broad‐sense heritability (H 2) estimates comparable to visual scoring performed by experienced scorers. We deployed this approach with a marker‐based pipeline, and compared it with semiquantitative visual scoring to identify SNPs associated with resistance through GWASs. We also demonstrated the potential of GS to fast‐track the breeding of SCLB‐resistant cultivars, complementing marker‐assisted breeding efforts. The use of a CV‐based approach as a tool to estimate SCLB scores In our study, heritability estimates for visual disease scoring ranged from 0.44 to 0.73 across environments, while CV‐based phenotyping yielded estimates of 0.57 and 0.73 in the two environments where the CV approach was implemented. These results suggest that CV‐based phenotyping captures a substantial genetic contribution for SCLB resistance that is comparable to visual scoring within the same experimental environment. We also observed a moderate to high correlation between BLUEs derived from CV‐based phenotyping and visual disease scores, supporting the quantitative validity of the CV approach. While visual scoring remains an effective approach, it depends heavily on the experience and expertise of the individual performing the assessment (Poland & Nelson, 2011). While effective, we also acknowledge some limitations of the current CV‐based approach to SCLB phenotyping. Our current CV setup requires additional upfront effort for leaf harvesting, image acquisition, and the processing pipeline, which can be time‐consuming compared to traditional visual scoring. The current work demonstrated the feasibility of the CV phenotyping approach, but future work using in‐field approaches, such as in situ leaf imaging, mobile devices, drones, and robotic systems, can reduce handling time and enable the scalable collection of high‐resolution phenotypes (Araus & Cairns, 2014). Given ongoing advances in plant imaging, automation, and image analysis, the CV‐based phenotyping framework implemented in this study has the potential to become more practical and higher‐throughput in future applications. Further limitation concerns disease specificity under complex field conditions. Although we conducted CV‐based phenotyping under conditions where SCLB was the dominant foliar disease, we recognize that CV‐derived phenotypes should be interpreted cautiously in mixed infection environments. Under such conditions, it may be difficult for our CV method to reliably distinguish SCLB‐specific necrosis from damage caused by other diseases (e.g., northern corn leaf blight) or other factors (senescence) (Mutka & Bart, 2015). Although limitations related to throughput and disease specificity remain, the CV‐based phenotyping enabled us to estimate the percent DLA, generating a continuous phenotype rather than discrete classes, which can retain within‐class variation that visual scales may not resolve (Mutka & Bart, 2015). Notably, we observed that genotypic classification based on visual scores exhibited variability in CV‐derived BLUEs, supporting the notion that CV phenotyping captures differences not resolved by an ordinal rating scale (Figure S3). Such quantitative phenotypes are better aligned with the assumptions of GWAS and GS models, which benefit from continuous, high‐resolution trait measurements. Importantly, this gain in phenotypic resolution was achieved while maintaining heritability estimates comparable to those obtained using conventional visual ratings. Furthermore, if the CV method is further developed to enable higher‐throughput implementation, it could be utilized to support breeding decision‐making and may ultimately contribute to genetic gain. The use of GWAS for associating genomic regions with SCLB resistance The second point we explored was integrating the CV‐based and visual scoring phenotyping approaches with GWAS to pinpoint genomic regions associated with the phenotypes. We identified 10 SNPs in visual scoring, while 12 SNPs were detected in CV‐based phenotyping. Notably, significant associations were observed only in CT19 and NC21 environments. In CT19, all significant SNPs were identified exclusively using visual scoring, while in NC21, significant associations were detected only when using CV‐derived phenotypes. Thus, the CV‐derived phenotypes should be interpreted as independent quantitative phenotypes that complement visual scoring. Each phenotyping approach captured specific MTAs that were not detected by the other, and these associations varied across environments. Genes linked to the significant SNPs detected by GWAS include several candidates with previously reported roles in plant defense response (Table 2). These include transcription factors (e.g., NAC, MYB, bHLH, ARR‐B families), kinases and signaling components (e.g., calcium‐dependent protein kinases, calcineurin B‐like interacting kinases), hormone‐related regulators (ethylene and cytokinin pathways), and cell wall‐associated enzymes such as pectin methylesterase inhibitors and polygalacturonases. Collectively, these functional categories are consistent with mechanisms underlying quantitative, polygenic resistance to foliar fungal pathogens, supporting the biological plausibility of the detected associations. While the candidate genes we detected through GWAS represent resistance gene candidates for breeding, it is still necessary to validate their actual involvement in SCLB resistance. In particular, validation through approaches such as overexpression, gene knockout, or reverse genetics, as well as fine‐mapping experiments, would provide evidence before actual implementation in the sweet corn breeding program (Mores et al., 2021). Once validated, gene‐specific markers can be developed and integrated into routine marker‐assisted selection using cost‐effective SNP assays (Dubiela et al., 2019; Fresnedo‐Ramírez et al., 2017; Zhong et al., 2018). Application of molecular breeding via genomic selection in developing SCLB‐resistant genotypes Another important result was that the use of GS based on visual scoring phenotypes can be effective even under a simple model, reaching relatively high predictive accuracies (0.22–0.47 under cross‐validation of untested genotypes in characterized environments; 0.53–0.70 under cross‐validation of tested genotypes in uncharacterized environments) for predicting individuals’ performance for SCLB scores. For cross‐validation of untested genotypes in characterized environments, we observed that the environments CT19, CT20, and NC21 have higher prediction for untested genotypes in the known environments cross‐validation scheme as compared to BG23 and BG24. The reduced sample size in later years limited the ability of the models to capture sufficient genetic variance, which affects predictive performance (Asoro et al., 2011; Wientjes et al., 2016; Norman et al., 2018). Also, our method implemented a cross‐validation scheme, which masks the entire genotypes across environments, so prediction for BG23 and BG24 heavily relied on the limited number of training genotypes available, which can explain the lower accuracies observed. In contrast, high prediction accuracies under cross‐validation of tested genotype in uncharacterized environments demonstrate the potential of GS for predicting SCLB resistance in untested environments, even in BG23 and BG24, which is important for deployment in breeding programs where field resources are often limited. This scheme masks entire environments rather than specific genotypes across environments. As a result, BG23 and BG24 recovered high prediction accuracies by leveraging phenotypic information from the larger CT19, CT20, and NC21 environment, with which they share substantial genotypic overlap (Figure S7). While we tested whether modeling G × E interactions could improve the prediction of SCLB visual scores, there is no clear advantage over the simplest model. We then hypothesize that this is likely due to the low contribution of G × E variance in our dataset, which only accounts for 10% of the total phenotypic variance (Figure 4). However, it can be possible that the contribution of G × E would be more significant in trials conducted across a broader range of environments or locations, as our current study was limited to Florida and North Carolina. A similar observation was reported in a study on Fusarium head blight in wheat, where incorporating G × E in the model only slightly increased the prediction accuracy, with a G × E variance component of 3%–5% (Brault et al., 2025). The authors attributed this result to controlled inoculation, reducing the impact of the environment. In our study, most of the trials were artificially inoculated with the SCLB pathogen, which corroborates their findings. Furthermore, our variance component analysis showed that the genotypic effect (G) accounted for the majority of the variance component relative to the G × E; hence, modeling G only resulted to a high prediction accuracy. Also, the substantial genotypic overlap across environments of genotypes shared may limit the environmental contrast available for G × E to contribute significantly in the prediction model. Nevertheless, we observed comparable prediction accuracies to those reported for quantitative disease resistance in wheat stripe rust, where GS models reached ∼0.7 accuracy and outperformed marker‐assisted schemes in the presence of both major and minor genes (Merrick et al., 2021). Although we detected environment‐specific MTAs in our GWAS study, this does not contradict the results from modeling G × E in GS. GWAS detects MTAs with environment‐specific effects, while our GS implementation aimed to determine whether modeling G × E improves prediction accuracy. We hypothesize that while GS is not affected by the GE, slight changes in the phenotypic scores across environments also affected the GWAS result, primarily since each environment was evaluated by different raters. Hence, environment‐specific MTAs can be detected in GWAS even if G × E models provide only limited improvement in GS, if any at all. Implementing various GS models for the visual scoring revealed that the “tested genotypes in uncharacterized environments” cross‐validation scheme performs better than the “untested genotype in characterized environment” cross‐validation scheme in our SweetCAP diversity panel. This scenario contradicts the previous observation of Roorkiwal et al. (2018), who noted that the former scheme is more difficult to predict than the latter. However, this observation is not universal and can be data‑ and population‑dependent. For instance, the “tested genotypes in uncharacterized environments” scheme was significantly better than the “untested genotype in characterized environment” scheme, regardless of the models used or the population background in GS studies on stripe‑rust resistance in spring wheat (Semagn et al., 2022). A similar trend was also observed in durum wheat for predicting yield potential, drought stress, and heat stress (Sukumaran et al., 2018). Some features of our dataset may have driven this observation. One reason could be that, because we used a diversity panel, the “untested genotype in characterized environment” scheme cannot borrow information from related individuals due to the inherent diversity of our materials (Lorenz & Smith, 2015; Edwards et al., 2019). Moreover, under the “tested genotypes in uncharacterized environments” scheme, the model can borrow information from the same genotypes evaluated in other environments to predict performance in the masked environment, supported by the substantial genotypic overlap across environments and the range of environmental relationships documented in our panel (Jarquín et al., 2017). To our knowledge, there are still no published reports on the application of GS for SCLB resistance in either sweet corn or field corn. Together with simulation work showing that GS can simultaneously improve quantitative race‐nonspecific resistance while maintaining race‐specific R‐genes in the background (Michel et al., 2023), our results provide the first empirical evidence that GS is feasible for SCLB resistance in sweet corn and can be expected to efficiently capture both polygenic and potential major‐effect components of resistance. GS facilitates genetic gain by addressing all components of the breeder's equation (ΔG =i×r×σAL). By utilizing genomic information, GS increases the reliability of selection (r) relative to traditional phenotypic selection. Simulating recurrent selection in self‐pollinated crops shows that integrating GS in a two‐part breeding strategy can double genetic gain per unit cost by reducing cycle length (L), increasing selection intensity (i), and maintaining additive variance (σA ) (Gaynor et al., 2017). Similarly, the Big BIT maize experiment demonstrates real‐life application of GS for early‐stage testing in commercial pipelines (Jines et al., 2025). Beyond ranking individual candidates, GS can also be used to predict the performance of specific crosses that maximize long‐term genetic gain. For instance, SimpleMating (Peixoto et al., 2025) implements optimum cross‐selection based on predicted cross means, within‐family variance, and explicit constraints on co‐ancestry, enhancing long‐term genetic gains while controlling inbreeding. Applying this approach to sweet corn SCLB resistance would allow us to move from identifying “best genotypes” to explicitly selecting the “best crosses” for resistance while managing inbreeding. In practice, this would enable capturing both polygenic (horizontal) resistance and major‐effect alleles (vertical resistance) into elite sweet corn genotypes with durable, quantitative resistance to SCLB in a predictive‐breeding scheme. CONCLUSION We developed a CV‐based phenotyping tool that provides a standardized and repeatable approach for screening resistance to SCLB in sweet corn. Using this tool alongside visual scoring, GWAS identified candidate genes detected across single‐ and multi‐environment analyses, including transcription factors, kinases, and signaling components, hormone‐related regulators, and cell wall‐associated enzymes such as pectin methylesterase inhibitors and polygalacturonases, which represent strong targets for marker development. Finally, we conclude that GS can be effective, achieving high accuracies, which offers an alternative to traditional methods of selection for disease resistance traits. Hence, this paper demonstrates the potential of integrating quantitative phenotyping and predictive breeding to accelerate the development of SCLB‐resistant sweet corn. AUTHOR CONTRIBUTIONS Darlon V. Lantican: Data curation; formal analysis; investigation; methodology; validation; visualization; writing—original draft. Juan M. Gonzalez: Data curation; formal analysis; investigation; methodology; software; validation; visualization; writing—original draft. Marco Antonio Peixoto: Conceptualization; data curation; formal analysis; investigation; methodology; supervision; validation; writing—review and editing. Kristen A. Leach: Investigation; methodology; supervision; writing—review and editing. Larissa Carvalho Ferreira: Investigation; methodology; writing—review and editing. Vitor A. Silva de Moura: Investigation; methodology; writing—review and editing. Katia V. Xavier: Investigation; methodology; resources; supervision; writing—review and editing. Peter Balint‐Kurti: Data curation; investigation; methodology; writing—review and editing. William Tracy: Conceptualization; funding acquisition; project administration; resources; writing—review and editing. Marcio F. R. Resende Jr.: Conceptualization; funding acquisition; methodology; project administration; resources; supervision; writing—review and editing. CONFLICT OF INTEREST STATEMENT The authors declare no conflicts of interest. Supporting information Supplemental Figure S1. Phenotypic distribution of southern corn leaf blight (SCLB) resistance in the SweetCAP diversity panel. Manual disease severity ratings are shown for Citra, Florida in 2019 (a), Citra, Florida in 2020 (b), Clayton, North Carolina in 2021 (c), Belle Glade, Florida in 2023 (d), and Belle Glade, Florida in 2024 (e). Distributions of the logarithm of diseased leaf area (DLA) obtained from computer vision (CV)‐based phenotyping are presented for Citra, Florida in 2019 (f) and Clayton, North Carolina in 2021 (g). Supplemental Figure S2. Phenotypic correlations (Pearson's r) among visual and computer vision (CV)‐based disease resistance scores across environments and combined BLUEs for southern corn leaf blight (SCLB) in the SweetCAP diversity panel. Supplemental Figure S3. Scatterplot and density comparison of BLUEs from visual scoring and CV‐based phenotyping for southern corn leaf blight resistance in the NC21 environment (Pearson's r = 0.87). (A) Linear relationship between BLUEs; (B) Differences in the distribution of BLUE values between phenotyping methods. Supplemental Figure S4. Proportion of phenotypic variance in southern corn leaf blight resistance explained by genotype (σg2​), environment (σenv2​), genotype × environment (σg×e2​), block (σb2​), and residual error (σe2​) in the multi‐environment model using visual scoring phenotyping for environments where computer vision phenotyping were also collected (CT19; NC21). Supplemental Figure S5. Principal component analysis of the SweetCAP diversity panel. Supplemental Figure S6. Manhattan plot suggestive of marker‐trait associations (MTAs) for southern corn leaf blight (SCLB) resistance using the linear mixed model (MLM/EMMAX) for (A) 5‐year multi‐environment BLUEs from visual scoring phenotyping, and (B) CT19 single‐environment BLUEs from CV‐based phenotyping. Peaks above the threshold highlight associations. Dashed line indicates Bonferroni‐corrected significance threshold (‐log10[0.05/16,755,210] = 8.53) while the solid line represents significance threshold adjusted for the effective number of independent SNPs calculated using the Genetic type 1 error calculator (GEC; Li et al., 2012), corresponding to −log10(0.05/4,719,846.15) = 7.97. The rectangles indicate cluster of SNPs forming tower‐like peaks below the threshold, harboring candidate genes related to disease resistance. PERK9: Protein kinase superfamily 9 (Invernizzi et al., 2022); CCP23/CCP24: cysteine protease 23/24 (Niño et al., 2014); def3: defensin‐like protein 3 (Dowd and Johnson, 2018); NPR1: non‐repressor of pathogenesis‐related gene homolog 1 (Narváez et al., 2020; Zavaliev & Dong, 2023); NAC61: NAC transcription factor 61 (Nuruzzaman et al., 2013; Dong et al., 2024). Supplemental Figure S7. Venn diagram showing the genotype overlaps across the five environments (CT19, CT20, NC21, BG23, and BG24) of the SweetCAP diversity panel. Supplemental Figure S8. Environmental kinship matrix across the five SCLB evaluation environments. Supplemental File S1. Complete description of SNP sub‐setting and selection of principal components for genome‐wide association studies. Supplemental File S2. Complete description of the GBLUP and BayesB model implemented in BGLR.