Approaches based on seed and seedling images for classification of the physiological potential of cotton seedsABSTRACT:
Rapid and accurate methods for assessing seed quality have been increasingly demanded by the seed industry. In this context, the X-ray test and computerized seedling image analysis are promising alternatives for use in quality control programs. The objective of this study was to evaluate the efficiency of radiographic images and seedling images combined with interactive machine learning for classifying the physiological potential of cotton seeds. Seeds from eight lots were initially evaluated using germination and vigor tests. Radiographic images were acquired, and the following variables were obtained: area, perimeter, filling, relative density, integrated density, and grayscale. Seedling images captured at two, three, and four days of germination were processed using the GroundEye system, yielding the following variables: hypocotyl length, root length, total seedling length, and growth, uniformity, and vigor indices. Using the Ilastik software, data were generated regarding the percentage of strong normal seedlings, weak normal seedlings, abnormal seedlings, and dead seeds. The X-ray technique showed a relationship with the physiological potential of cotton seeds, in addition to providing information on their internal structures. Shoot length, primary root length, and total seedling length, as well as uniformity, growth, and vigor indices determined by GroundEye, together with data obtained via supervised machine learning using Ilastik, enabled the classification of cotton seed lots (cv. 21065TLP) based on their physiological potential at three days of germination; however, external validation using other cultivars, crop seasons, and regions of the country is required.
Index terms:
Gossypium hirsutum L.; GroundEye; Ilastik; machine learning; physiological quality
RESUMO:
Métodos rápidos e precisos para a avaliação da qualidade das sementes têm sido cada vez mais demandados pela indústria de sementes. Neste contexto, o teste de raios X e a análise computadorizada de plântulas são alternativas promissoras para utilização nos programas de controle de qualidade. O objetivo do trabalho foi avaliar a eficiência de imagens radiográficas e de imagens de plântulas aliadas ao aprendizado interativo de máquina para a classificação do potencial fisiológico de sementes de algodão. Sementes de oito lotes foram inicialmente avaliadas pelos testes de germinação e vigor. Imagens radiográficas foram coletadas, obtendo-se as variáveis: área, perímetro, preenchimento, densidade relativa, densidade integrada e escala de cinza. Imagens de plântulas obtidas aos dois, três e quatro dias. de germinação foram capturadas e processadas em equipamento GroundEye, obtendo-se as variáveis: comprimento do hipocótilo, da raiz e total das plântulas, índices de crescimento, de uniformidade e de vigor. Pelo software Ilastik, foram gerados dados relativos à porcentagem de plântulas normais fortes, normais fracas, anormais e sementes mortas. A técnica de raios X apresenta relação com o potencial fisiológico das sementes de algodão, além de fornecer informações sobre suas estruturas internas. O comprimento da parte aérea, o comprimento da raiz primária e o comprimento total da plântula, bem como os índices de uniformidade, crescimento e vigor determinados pelo GroundEye, juntamente com os dados obtidos por meio de aprendizado de máquina supervisionado utilizando o Ilastik, possibilitaram a classificação dos lotes de sementes de algodão (cv. 21065TLP) quanto ao seu potencial fisiológico aos três dias de germinação. Entretanto, é necessária a validação externa utilizando outras cultivares, safras e regiões do país.
Termos para indexação:
Gossypium hirsutum L.; GroundEye; Ilastik; aprendizado de máquina; qualidade fisiológica
INTRODUCTION
Cotton (Gossypium hirsutum L.) is the main source of commercial fiber for the global textile industry, and Brazil is currently among the five largest producers and the leading exporter worldwide in the 2024/2025 season, surpassing the United States (ABRAPA, 2025). Considering that cotton production is highly technified, the use of high-quality seeds is essential for crop success, ensuring adequate plant stands and uniformity, with positive impacts on yield (Rehman et al., 2020).
Physiological seed potential is routinely evaluated using traditional germination and vigor tests; however, these methods are destructive, time-consuming, and involve human interpretation, which may be subjective, thus limiting more efficient quality control (Elmasry et al., 2019; Xia et al., 2019). In this context, the application of modern technologies based on image analysis and interactive machine learning has emerged as a promising alternative for seed quality assessment, overcoming limitations associated with conventional tests (Xia et al., 2019; Limão et al., 2023). Among these techniques, X-ray analysis and computerized seedling image analysis stand out, offering advantages such as rapid assessment of internal seed structures and seedling performance, reliable results, and the generation of image databases that can be stored and referenced when needed (Gomes-Junior, 2020).
The X-ray test is based on the interaction of electromagnetic waves with the seed, with wavelengths ranging from 0.01 to 10 nanometers and energy levels between 0.1 and 100 keV (Kotwaliwale et al., 2014). X-ray penetration is influenced by the structure, composition, and density of seed tissues; denser tissues absorb more radiation and appear as lighter areas, whereas less dense tissues absorb less radiation and appear darker in radiographic images (Simak and Gustafsson, 1953; ISTA, 2020). This technique enables the non-destructive detection of changes in sample density that may be associated with seed integrity, as demonstrated in several studies involving soybean (Wendt et al., 2014; França-Silva et al., 2023), papaya (Nogueira et al., 2024), crambe (Ribeiro et al., 2021), leucaena (Medeiros et al., 2018), and broccoli (Abud et al., 2018).
In recent years, computerized image analysis has been widely applied in seed technology to assess physiological quality, primarily based on seedling morphological traits. Image acquisition is rapid and can be performed using digital cameras, scanners, or even mobile devices, while interpretation is carried out using specific algorithms capable of identifying relevant attributes and comparing seed lots (Gomes-Junior, 2020). The main advantages of this approach include higher accuracy and speed of analysis, with results obtained within seconds; standardization across laboratories and analysts; reduced human interference; improved reliability; and the generation of image databases for storage and future reference (Gomes-Junior, 2020).
GroundEye is a computerized analysis system that provides information on several variables, including those related to seedling performance, such as shoot length, root system length, total seedling length, growth, uniformity, and vigor indices (GroundEye® Manual, 2016; Brandani et al., 2021). The main advantages of this system include rapid analysis of large numbers of seed lots, fast data output, and image storage for reporting and future evaluations (GroundEye® Manual, 2016). Therefore, evaluating its efficiency for cotton seeds is relevant, given that studies have already demonstrated its applicability for soybean (Acha and Vieira, 2020; Brandani et al., 2021), maize (Catão et al., 2020), coffee (Abreu et al., 2016), and vegetable seeds such as cabbage and parsley (Vasconcelos et al., 2018).
Another promising computerized image analysis system is Ilastik, an open-access software based on interactive machine learning that enables users to build models for image analysis and classification (Berg et al., 2019). This software can be used in quality control programs to assist in ranking seed lots according to different levels of physiological potential through the classification of seed and seedling images (Medeiros et al., 2020a. Limão et al., 2023). Limão et al. (2023) reported that Ilastik-derived data enabled the classification of lentil seed vigor at four days of germination. Similar studies have demonstrated the efficiency of this software for soybean (Medeiros et al., 2020a), rice (Yasam et al., 2022), Urochloa decumbens (Souza et al., 2023), and physic nut (Medeiros et al., 2020b).
In this context, the objective of this study was to evaluate the efficiency of radiographic imaging, computerized seedling analysis, and interactive machine learning based on seedling performance for classifying the physiological potential of cotton seeds.
MATERIAL AND METHODS
The study was conducted at the Seed Research Laboratory of the Department of Agronomy at the Universidade Federal de Viçosa, in Viçosa, Minas Gerais, Brazil. Eight cotton seed lots of the cultivar 21065TLP were used.
Initially, seeds from each lot were subjected to the following tests to characterize their initial quality.
Physiological quality characterization
Moisture content: Determined in a forced-air oven at 105 °C for 24 h, according to the Rules for Seed Testing (Brasil, 2025). Two replications of 25 seeds per lot were used, and the results were expressed as percentage on a wet basis.
Germination: Four replications of 50 seeds were placed on paper towel rolls moistened with water in an amount equivalent to 2.5 times the dry paper weight and maintained in a germinator at 25 °C. Evaluations were performed on the fourth and twelfth days after sowing, and the results were expressed as the percentage of normal seedlings on the twelfth day (Brasil, 2025).
First germination count: Conducted together with the germination test, consisting of the percentage of normal seedlings obtained on the fourth day after sowing (Krzyzanowski et al., 2020).
Low-temperature germination: The methodology was similar to that described for the germination test; however, the rolls were placed in plastic bags and maintained in an incubator (BOD type) at 18 °C. A single evaluation was performed on the seventh day after sowing (Dias and Marcos-Filho, 2020). Results were expressed as the percentage of normal seedlings with a length greater than or equal to 4.0 cm.
Accelerated aging: A total of 250 seeds from each lot were distributed on a metal screen placed inside a plastic gerbox containing 40 mL of water at the bottom. The boxes were sealed to achieve approximately 100% relative humidity and maintained in a BOD incubator at 41 °C for 72 h. After this period, the germination test was carried out as previously described, and the percentage of normal seedlings was evaluated on the fourth day after sowing (Marcos-Filho, 2020b).
Seedling emergence and emergence speed index (ESI): This test was conducted in a growth room using plastic trays filled with a soil and sand mixture (1:1). Four replications of 50 seeds were sown at a depth of 1.0 cm. Daily counts were performed, and emergence percentage was calculated based on the total number of emerged seedlings, considering those with plumules above the substrate surface, until stand stabilization (Krzyzanowski et al., 2020). The emergence speed index (ESI) was calculated according to Maguire (1962).
Radiographic analyses
X-ray imaging: Eight replications of 25 seeds per lot were used. Seeds were positioned with the embryonic axis facing downward and fixed in an orderly arrangement on adhesive paper to enable subsequent individual identification in further analyses. The seeds were then placed inside a digital X-ray system, model MX-20 (Faxitron X-ray Corp., Wheeling, IL, USA).
To obtain the radiographic images, the equipment was set to an exposure time of 15 seconds, voltage of 29 kV, focal distance of 34.6 cm, and image contrast calibrated at 6884 (width) × 3847 (center). The radiographic images were processed and analyzed using ImageJ software through the IJCropSeed plugin (Medeiros et al., 2020c).
After processing, the following variables were extracted: Area - Selection obtained in square pixels and subsequently converted to square millimeters (mm²). Perimeter - Length of the external boundary of the selection, expressed in millimeters (mm). Filling - Percentage of the seed completely filled with high-density tissue (%). Relative density - Mean grayscale value of all pixels within the selected area (gray·mm⁻¹). Integrated density - Product of area and mean grayscale value (gray·mm²·pixel⁻¹). Grayscale - Median grayscale value of all pixels within the selected area (gray·pixel⁻¹).
Seedling performance-based image analysis using GroundEye
Initially, a seedling growth test was established using eight replications of 20 seeds per lot. Seeds were distributed along a line drawn in the upper third of the germination paper, in the longitudinal direction. The seeds were positioned with the hilum facing downward, and the rolls were placed in plastic bags and maintained vertically in a germinator at 25 °C for periods of 2, 3, and 4 days. At the end of each period, the seedlings were transferred from the germination paper to a transparent acrylic tray coupled to the GroundEye system, ensuring that they did not overlap. Image acquisition was performed using a high-resolution professional camera integrated into the GroundEye equipment.
The image processing was performed using GroundEye software (version S120-SN 606). Image analysis was configured and calibrated using the HSV model, with saturation ranging from 0.611 to 1.000, brightness from 0.292 to 0.892, hue from 168.1 to 238.1, a distance of 0.030, a minimum size of 0.059 cm², and a maximum size of 2.997 cm². The following variables were obtained: hypocotyl length, root length, and total seedling length, as well as growth, uniformity, and vigor indices, calculated according to the following equations:Growth Index = min (wₕ · lₕ + wᵣ · lᵣ, 1000). Uniformity Index = max [1000 − (wₛₕ · sₕ + wₛᵣ · sᵣ + s_total + wₛᵣ/ₕ · sᵣ/ₕ) − wₘ · number of dead seeds, 0]. Vigor Index = pC · growth + pU · uniformity. Where: lₕ and lᵣ: sample means of hypocotyl length and root length, respectively; sₕ, sᵣ, s_total, and sᵣ/ₕ: standard deviations of hypocotyl length, root length, total seedling length, and the root-to-hypocotyl ratio, respectively; pC and pU: weights assigned to growth and uniformity indices; w: weighting factors for the parameters.
Seedling classification using Ilastik software
The images obtained, as described for the GroundEye analysis, were imported into the Ilastik software. For segmentation, the pixel classification workflow was used, defining two classes: “seed or seedling” (region of interest) and “background” (region to be excluded). The probability of each pixel belonging to the semantic segmentation classes was estimated and illustrated in Figure 2. Subsequently, a model was developed to classify four seedling categories: (a) strong normal seedlings (well-formed, without defects, showing vigorous shoot and root growth); (b) weak normal seedlings (morphologically well-formed but presenting minor defects that still classify them as normal, or with smaller size compared to strong normal seedlings); (c) abnormal seedlings (with severe defects, such as absence or deformation of essential structures, according to Brasil (2025)); and (d) dead seeds, according to the standard illustrated in Figure 1 (Limão et al., 2023; Souza et al., 2023). The trained classifier was applied to all images, generating data on the number of strong normal seedlings, weak normal seedlings, abnormal seedlings, and dead seeds. A schematic overview of the main steps of the interactive method for classifying cotton seeds and seedlings is presented in Figure 2.
Standard classification of strong normal seedlings, weak normal seedlings, abnormal seedlings, and dead cotton seeds using the Ilastik-based classification model.
Schematic representation of the training and classification steps of interactive machine learning for assessing the physiological potential of different cotton seed lots at 2 (a), 3 (b), and 4 (c) days after sowing. Strong normal seedlings (dark green), weak normal seedlings (light green), abnormal seedlings (neon blue), and dead seeds (purple). Original images of seedlings on a blue background (I). Subsequently, rendering was applied for segmentation and identification of the region of interest (ROI) to generate the probability map (II). After ROI identification, the software detects individual seedlings and seeds using the probability map (III), and finally, classification prediction for each seedling or seed is performed based on the predefined training color classes (IV).
Experimental design and statistical analysis
The moisture content data was not subjected to formal statistical analysis; only the mean values were calculated. The experiment was conducted using a completely randomized design (CRD) with four replicates. Data were tested for error normality using the Shapiro-Wilk test and for homogeneity of variances using Bartlett’s test. Subsequently, the data were subjected to analysis of variance (ANOVA), and the mean values obtained for each batch were compared using Tukey’s test (p < 0.05). Principal component analysis (PCA) was also performed using data from all tests related to the initial physiological potential characterization, as well as variables derived from radiographic images and seedling growth and classification analyses. All statistical analyses were performed using R software version 4.3.3 (R Core Team, 2025).
RESULTS AND DISCUSSION
Seed moisture content among the lots ranged from 9.4 to 10.3% (Table 1). This is a fundamental characteristic to be considered during test procedures, as it ensures standardization of evaluations and the reliability of results (Marcos-Filho, 2015).
The evaluation of the physiological potential of the seed lots showed that lots 1 and 5 exhibited the best performance in germination, first germination count, low-temperature germination, accelerated aging, emergence, and emergence speed index tests (Table 1). In contrast, lots 4 and 8 showed the poorest performance. In the low-temperature germination (LTG) test, lots 1 and 5 maintained superior performance, not differing significantly from lots 2, 6, and 7, but presenting higher values than those observed for lots 4 and 8, with lot 4 standing out as the lowest-performing.
As observed in the present study, the seed lots differed in terms of physiological potential. The evaluation of seed physiological potential is essential to estimate the performance of seed lots under both field and storage conditions; therefore, the combined use of multiple tests allows for the identification of differences in seed vigor (Marcos-Filho, 2020a), enabling the classification of lots into performance-based groups (Table 1), which is fundamental for quality control in seed companies.
Based on the results of variables obtained from radiographic images, such as area (Figure 3a), perimeter (Figure 3b), and filling (Figure 3c), which are directly related to seed size and the integrity of the seed coat and embryo, the absence of significant differences among the lots indicates a high level of physical uniformity. This result suggests that all lots have high physical quality, with well-formed seeds and no apparent structural damage, making it difficult to associate these variables with possible differences in seed physiological potential (Table 1). In this context, the physical characteristics observed through X-ray imaging, particularly those related to seed coat integrity and the organization of internal tissues, may be associated with seed physiological potential, as structural alterations such as malformations, mechanical damage, or tissue irregularities may reflect conditions that affect seed physiological performance. However, in the present study, the seed lots exhibited similar physical patterns and high structural homogeneity, which may have contributed to the absence of marked differences among the materials regarding the characteristics obtained from X-ray image analysis and the physiological attributes evaluated. Furthermore, it is important to emphasize that physical characteristics, especially seed size, are not always directly associated with physiological performance, such as germination and vigor, as previously demonstrated by Marcos-Filho et al. (2000). Thus, although radiographic analysis is important for assessing seed structural integrity, its integration with physiological test results allows for a more comprehensive evaluation of seed quality.
Values of area (a), perimeter (b), filling (c), relative density (d), integrated density (e), and mean grayscale (f) for eight cotton seed lots. Means followed by the same letter do not differ according to Tukey’s test (p < 0.05). Error bars represent the standard deviation.
For variables related to internal seed structures, such as relative density (Figure 3d), integrated density (Figure 3e), and mean grayscale (Figure 3f), it was possible to distinguish lots with lower physiological potential from those with medium and high potential. Studies with crambe (Ribeiro et al., 2021), broccoli (Abud et al., 2018), and leucaena (Medeiros et al., 2018) have shown that the use of tissue density parameters and grayscale values obtained from radiographic images is promising and provides relevant information regarding seed physiological quality, as seeds with lower density tend to exhibit reduced vigor, resulting in abnormal seedlings, as observed in Figure 4.
Radiographic image of the seed, heatmap based on seed relative density, and seedling image obtained from each cotton seed (a - no damage and with a well-defined and fully developed embryo; b - seed coat cracks and embryo damage).
Thus, it can be emphasized that the variables relative density, integrated density, and mean grayscale may be associated with the physiological potential of cotton seeds, as observed in Figure 5. This technique has been used to evaluate morphometric attributes and embryo integrity associated with the physiological quality of seeds from several species, such as soybean (Wendt et al., 2014), crambe (Ribeiro et al., 2021), okra (Santos et al., 2020), broccoli (Abud et al., 2018), papaya (Nogueira et al., 2024), and leucaena (Medeiros et al., 2018). The use of the X-ray test in cotton seed analysis made it possible to identify intact seeds without any damage to the seed coat or embryo (Figure 4a) and to distinguish them from those presenting some type of injury or damage to the seed coat, which ultimately results in embryo damage (Figure 4b).
In technical terms, X-rays pass through the seed and undergo attenuation proportional to tissue density: denser regions (such as well-formed embryos and intact storage tissues) absorb more radiation and appear lighter in the image, whereas less dense regions (such as empty spaces, mechanical damage, or deteriorated tissues) appear darker (Figures 4a and 4b). Thus, radiographic density acts as an indirect indicator of the internal organization of the seed. Seeds with higher density exhibit better embryo formation and a greater amount of reserves, which favor early germination metabolism by ensuring the availability of energy and substrates for seedling growth. In contrast, seeds with lower density often show deterioration, cellular disorganization, embryo damage, or the presence of void spaces, factors that impair essential physiological processes (Campos et al., 2022).
Moreover, seed deterioration is associated with biochemical changes such as membrane degradation, loss of cellular compartmentalization, and depletion or degradation of reserves. These processes reduce internal density and, consequently, the capacity for X-ray absorption. As a result, less dense seeds tend to exhibit lower germination and vigor, establishing the relationship between radiographic density and physiological quality (Campos et al., 2022).
As observed in the principal component analysis (Figure 5), there is a clear association between the vectors related to physiological variables, relative density, integrated density, and mean grayscale, and lots 1 and 5, which showed the highest physiological potential, as indicated in Table 1 and Figures 3d, 3e, and 3f. In contrast, lots 4 and 8 are positioned in the negative scores and are not associated with these physiological and physical variables.
Principal component analysis of physiological variables: germination (G), first germination count (FGC), low-temperature germination (LTG), accelerated aging (AA), emergence (E), and emergence speed index (ESI), and variables obtained from the X-ray test: area (AREA), perimeter (PER), filling (FILL), relative density (RD), integrated density (ID), and mean grayscale (GRAY) for eight cotton seed lots.
Overall, the combination of these morphometric variables enables the identification of well-formed seeds and the integrity of their tissues. The use of X-ray analysis serves as a complementary test to traditional methods for evaluating physiological seed quality, providing information on internal morphology and seed integrity (Medeiros et al., 2020a and Medeiros et al., 2020b). This type of assessment is highly advantageous for seed companies in quality control programs, as such damage cannot be detected through visual inspection alone, in addition to reducing analysis time and improving the efficiency of information acquisition.
The results of computerized seedling analysis using the GroundEye software revealed significant differences in shoot length (SL) among seed lots as early as the second day of germination. Lots 1 and 5 showed superior performance, while lots 2, 3, 4, and 8 exhibited the lowest shoot growth, and lots 6 and 7 presented intermediate values (Table 2). However, on the third day of germination, it was possible to classify the lots into three distinct levels based on SL, with lots 1, 5, and 6 showing the highest values, and lots 4 and 8 the lowest, consistent with the results observed on the second day (Table 2). By the fourth day, the lots could only be grouped into two vigor levels, indicating that the longer growth period reduced differentiation among them. Lots 1 and 5 remained with the highest SL. Interestingly, lots 2, 3, 6, and 7 did not differ statistically from lots 1 and 5, and these, in turn, also did not differ from lots 4 and 8 (Table 2).
For root length (RL), lots 1, 5, and 6 were superior to the others, while lots 4 and 8 showed the lowest values, and lots 2, 3, and 7 were intermediate at two days after germination (Table 2). On the third day, lots 1, 5, and 6 maintained the highest RL, while lot 8 presented the lowest values, with the remaining lots showing intermediate performance. By the fourth day, only lots 1 and 5 remained superior, whereas lots 4 and 8 had the lowest RL, similar to the pattern observed on the second day, with the other lots remaining intermediate (Table 2).
Regarding total seedling length (TSL), at two days it was possible to classify the lots into three levels of physiological potential, with lots 1 and 5 showing the highest performance and lots 4 and 8 the lowest, while lots 2, 3, 6, and 7 were intermediate (Table 2). At three days, a clearer separation among vigor levels was observed, although the ranking of the lots remained the same. By the fourth day, lots 1 and 5 maintained the highest performance, while lot 8 showed the lowest TSL, followed by lots 4 and 7 (Table 2). Considering that rapid results from vigor tests are essential for decision-making regarding seed lot management and allocation, it is recommended that SL, RL, and TSL evaluations be performed at three days after sowing, as this time point allowed effective classification of the lots into different levels of physiological potential.
More vigorous seeds typically exhibit greater reserve accumulation, higher metabolic activity, and better structural integrity, enabling rapid germination and early seedling growth. Therefore, traits associated with seedling growth are often used as indicators of physiological potential, reflecting the actual capacity for rapid and uniform seedling establishment in the field (Krzyzanowski et al., 2020). GroundEye has proven effective in determining seed vigor in maize (Andriazzi et al., 2020). In soybean seeds, data obtained using the GroundEye system were effective for vigor assessment as early as two days, reducing by six days the time required to obtain results compared to the standard germination test (Brandani et al., 2021). Studies evaluating image analysis techniques for physiological potential assessment have also been conducted in lentil (Limão et al., 2023) and chickpea (Araújo et al., 2021) seeds using SAPL® software, demonstrating high accuracy of the technique.
The results for the uniformity index (UI), growth index (GI), and vigor index (VI) allowed classification of the lots into vigor levels, highlighting superior performance for lots 1 and 5 at two, three, and four days after germination (Table 2). It is noteworthy that lot 6 showed similar performance to lots 1 and 5 at two and three days for UI, GI, and VI (Table 2). Lots 4 and 8 consistently showed the lowest performance for UI, GI, and VI at all evaluation times, with a classification pattern similar to that observed for shoot growth, particularly at two and three days (Table 2).
The UI, GI, and VI proved effective in distinguishing levels of physiological potential in soybean (Brandani et al., 2021) and maize seeds (Andriazzi et al., 2020). In lentil (Limão et al., 2023) and chickpea (Araújo et al., 2021), data obtained using SAPL software were also efficient in separating seed lots according to vigor at three and four days, respectively.
Uniformity, growth, and vigor indices are fundamental for estimating the physiological quality of seeds. These indices, derived from seedling growth tests, constitute an important tool for application in seed quality control programs (Silva et al., 2017; Pereira et al., 2020).
Thus, computerized seedling image analysis can provide quantitative measurements of seedling growth characteristics, such as seedling length, which are commonly associated with seed physiological potential. The use of image-based measurements may reduce the influence of evaluator subjectivity and improve the standardization of measurements; however, the ability of these parameters to discriminate seed lots depends on the magnitude of variation in seedling growth and physiological quality among the materials evaluated (Gomes-Junior, 2020).
Principal component analysis (Figure 6) further reinforces the results presented in Table 2, highlighting lots 1 and 5 at the extremes of the negative scores of PC1, associated with variables related to initial physiological potential characterization and those obtained from the seedling growth test, as well as lot 6, which also showed good performance in this test. Conversely, lots 4 and 8 were positioned in the positive scores of PC1, showing no association with physiological potential and growth variables, while the remaining lots were located in intermediate positions (Figure 6).
Principal component analysis of physiological variables: germination (G), first germination count (FGC), low-temperature germination (LTG), accelerated aging (AA), emergence (E), and emergence speed index (ESI), and variables obtained from the seedling growth test using the GroundEye software: shoot length (SL), root length (RL), total seedling length (TSL), uniformity index (UI), growth index (GI), and vigor index (VI) at 2, 3, and 4 days after germination for eight cotton seed lots.
The percentages of strong normal seedlings, weak normal seedlings, abnormal seedlings, and dead seeds generated by the lots at 2, 3, and 4 days after sowing are presented in Figure 7. The most vigorous lots showed a higher percentage of strong normal seedlings, with well-developed shoots and primary roots. In contrast, less vigorous lots exhibited higher percentages of weak normal seedlings, abnormal seedlings, and dead seeds.
Percentage of strong normal seedlings, weak normal seedlings, abnormal seedlings, and dead seeds at two (a), three (b), and four (c) days after sowing, according to data obtained using the Ilastik software. Means followed by the same letter do not differ according to Tukey’s test (p < 0.05). Error bars represent the 95% confidence interval.
However, greater reliability and sensitivity in distinguishing among classes were observed from the third day of germination onward, enabling the ranking of lots according to their levels of physiological potential. Lots 1 and 5 stood out with the highest percentage of strong normal seedlings, followed by lots 2, 6, and 7 at intermediate levels, while lots 3, 4, and 8 showed the poorest performance in this class and higher frequencies of weak normal and abnormal seedlings (Figure 7b). For the dead seed class, only lot 4 differed from the others by not presenting this category (Figure 7b).
At two days, lots 1 and 5 had a higher percentage of strong normal seedlings compared to lots 4 and 8, not differing from the remaining lots. At three and four days, lots 1 and 5 were superior to the others, while lots 4 and 8 were inferior at three days, and only lot 4 remained inferior at four days (Figure 7).
Ilastik is an image processing software based on interactive machine learning that allows users to dynamically classify and analyze images according to their objectives (Berg et al., 2019). Its application opens valuable opportunities across various fields of scientific research involving image analysis, offering flexibility and reproducibility of results (Berg et al., 2019; Dietz et al., 2020). The use of image analysis has been steadily increasing, driving scientific and technological advances in seed quality evaluation. Limão et al. (2023) evaluated the classification of lentil seed vigor based on seedling image analysis and interactive machine learning using Ilastik at 3, 4, 5, and 10 days after germination, concluding that the data obtained were sensitive for developing highly accurate models for vigor assessment at 4 days. In Urochloa decumbens, Ilastik was also effective in classifying seeds according to physiological potential with only 7 days of germination (Souza et al., 2023).
The application of machine learning techniques using data generated by the Ilastik software to classify the physiological quality of soybean seeds and seedlings was evaluated by Medeiros et al. (2020a), who concluded that this tool was accurate in identifying seed damage and classifying seedlings according to vigor.
The developed models can be refined and further trained within the Ilastik software, with the possibility of expanding the dataset (images) and treatments by using all available and necessary tools within the platform. These tools include features such as batch processing, probabilistic models, and user-friendly graphical interfaces, enabling fast and interactive training as well as the extraction of accurate information from images (Berg et al., 2019).
The efficiency of seedling classification using Ilastik can also be confirmed through principal component analysis (Figure 8), in which lots 1 and 5 are directly associated with vectors related to physiological potential at the extremes of the negative scores, followed by lots 6, 7, 2, and 3, which showed intermediate performance. In contrast, lots 4 and 8 are positioned in the opposite direction, within the positive scores, directly associated with vectors of abnormal and weak seedlings at 2, 3, and 4 days (Figure 8).
Principal component analysis of physiological variables: germination (G), first germination count (FGC), low-temperature germination (LTG), accelerated aging (AA), emergence (E), and emergence speed index (ESI), and seedling classification obtained using the Ilastik software: strong normal seedlings (STR), weak normal seedlings (WEAK), abnormal seedlings (ABN), and dead seeds (DEAD) at 2, 3, and 4 days after germination for eight cotton seed lots.
Thus, it can be stated that the parameters obtained through the classification of images generated by X-ray imaging, computerized seedling analysis using GroundEye, and Ilastik are promising tools for obtaining rapid and accurate results, with potential practical applications in seed quality control programs.
CONCLUSIONS
The X-ray technique is associated with the physiological potential of cotton seeds and provides information on their internal structures. At three days of germination, seedling length, shoot length, and primary root length, as well as the uniformity, growth, and vigor indices determined using the GroundEye system, are efficient for classifying the physiological potential of seed lots. This classification was also achieved by evaluating strong normal, weak normal, and abnormal seedlings through supervised machine learning using the Ilastik software.
ACKNOWLEDGMENTS
The authors thank the Universidade Federal de Viçosa (UFV), the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq), the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES; Finance Code 001), and the Fundação de Amparo à Pesquisa do Estado de Minas Gerais (FAPEMIG).
REFERENCES
-
ABRAPA. Associação Brasileira dos Produtores de Algodão. Relatório de Safra Brasil 2025. https://app.powerbi.com/view?r=eyJrIjoiNDFhMDlhZjYtZTljNi00NjNmLWIwMDEtOGY4OGE1OTE5ZmE1IiwidCI6IjRhMDk1OGIzLTg4MWQtNDBmYS05NTU1LTIwODQ1MzdhYWZkMyJ9&pageName=ReportSection10c4aa5d90c47641852d
» https://app.powerbi.com/view?r=eyJrIjoiNDFhMDlhZjYtZTljNi00NjNmLWIwMDEtOGY4OGE1OTE5ZmE1IiwidCI6IjRhMDk1OGIzLTg4MWQtNDBmYS05NTU1LTIwODQ1MzdhYWZkMyJ9&pageName=ReportSection10c4aa5d90c47641852d -
ABREU, L.A.S.; ANDRADE, D.B.; MARQUES, E.R.; ASSIS, J.G.R.; LOPES, C.A.; CARVALHO, M.L.M. Computerized analysis in the physiological quality of coffee seeds. International Journal of Current Research, v.8, n.11, p.40820-40823, 2016. https://www.journalcra.com/sites/default/files/issue-pdf/17672.pdf
» https://www.journalcra.com/sites/default/files/issue-pdf/17672.pdf - ABUD, H.F.; CICERO, S.M.; GOMES-JUNIOR, F.G. Radiographic images and relationship of the internal morphology and physiological of broccoli seed. Acta Scientiarum. Agronomy, v.40, n.1, p.1-9, 2018. https://doi.org/10.4025/actasciagron.v40i1.34950
-
ACHA, A.J.; VIEIRA, H.D. Digital image processing of coated perennial-soybean seeds and correlation with physiological attributes. Journal of Seed Science, v.42, e202042004, 2020. https://doi.org/10.1590/2317-1545v42227516
» https://doi.org/10.1590/2317-1545v42227516 -
ANDRIAZZI, C.V.G.; ROCHA, D.K.; SOUZA, I.C.P.; BICHUETTE, S.A. Avaliação da qualidade fisiológica de sementes de milho por meio da análise de imagens. Revista Agro, v.14, p.1-13, 2020. https://doi.org/10.18227/1982-8470ragro.v14i0.6750
» https://doi.org/10.18227/1982-8470ragro.v14i0.6750 -
ARAÚJO, J.O.; DIAS, D.C.F.S.; MEDEIROS, A.D.; SILVA, L.J.; NASCIMENTO, W.M. Chickpea seed vigor evaluated by computerized seedling analysis. Semina: Ciências Agrárias, v.42, n.1, p.71-80, 2021. https://doi.org/10.5433/1679-0359.2021v42n1p71
» https://doi.org/10.5433/1679-0359.2021v42n1p71 -
BERG, S.; KUTRA, D.; KROEGER, T.; STRAEHLE, C.N.; KAUSSLER, B.X.; HAUBOLD, C.; SCHIEGG, M.; ALES, J.; BEIER, T.; RUDY, M.; EREN, K.; CERVANTES, J.I.; XU, B.; BEUTTENMUELLER, F.; WOLNY, A.; ZHANG, C.; KOETHE, U.; HAMPRECHT, F.A.; KRESHUK, A. ilastik: interactive machine learning for (bio)image analysis. Nature Methods, v.16, n.12, p.1226-1232, 2019. https://doi.org/10.1038/s41592-019-0582-9
» https://doi.org/https://doi.org/10.1038/s41592-019-0582-9 -
BRANDANI, E.B.; FERREIRA, W.F.S.; JESUS, F.F.J.S.; MATTIONI, N.M.; MARQUES, É.A.; SILVA, N.O.S.; VILELA, M.S. Image analysis for the evaluation of soybean seeds vigor. Acta Agronómica, v.70, n.3, p.311-316, 2021. https://doi.org/10.15446/acag.v70n3.92093
» https://doi.org/10.15446/acag.v70n3.92093 - BRASIL. Ministério da Agricultura, Pecuária e Abastecimento. Regras para Análise de Sementes Brasília: MAPA, 2025.
-
CAMPOS, L.V.; RODRIGUES, A.A.; SALES, J.F.; RODRIGUES, D.A.; VASCONCELOS FILHO, S.C.; RODRIGUES, C.L.; VIEIRA, D.A.; CASTRO, S.T.; RUBIO-NETO, A. Radiographic imaging as a quality index proxy for Brachiaria brizantha seeds. Plants, v.11, n.8, p.1014, 2022. https://doi.org/10.3390/plants11081014
» https://doi.org/10.3390/plants11081014 -
CATÃO, H.C.R.M.; CAIXETA, F.; LOPES, A.M.; NERY SILVA, F.A.; SÁ JÚNIOR, A. de. Antioxidant activity and physiological performance of popcorn seed after saline stress and analysis of seedling images. Ciência e Agrotecnologia, v.44, e005020, 2020. https://doi.org/10.1590/1413-7054202044005020
» https://doi.org/10.1590/1413-7054202044005020 - DIAS, C. F. S.; MARCOS-FILHO, J. Teste de germinação a baixa temperatura In: KRZYZANOWSKI, F. C.; VIEIRA, R. D.; MARCOS-FILHO, J.; FRANÇA NETO, J. B. (Eds.). Vigor de sementes: conceitos e testes Londrina: ABRATES, 2020. p. 317-329.
-
DIETZ, C.; RUEDEN, C.T.; HELFRICH, S.; DOBSON, E.T.A.; HORN, M.; EGLINGER, J.; EVANS, I.E.L.; MCLEAN, D.T.; NOVITSKAYA, T.; RICKE, W.A.; SHERER, N.M.; ZIJLSTRA, A.; BERTHOLD, M.R.; ELICEIRI, K.W. Integration of the ImageJ ecosystem in the KNIME analytics platform. Frontiers of Computer Science, v.2, p.8, 2020. https://doi.org/10.3389/fcomp.2020.00008
» https://doi.org/10.3389/fcomp.2020.00008 -
ELMASRY, G.; MANDOUR, N.; WAGNER, M.H.; DEMILLY, D.; VERDIER, J.; BELIN, E.; ROUSSEAU, D. Utilization of computer vision and multispectral imaging techniques for classification of cowpea seeds. Plant Methods, v.15, p.24, 2019. https://doi.org/10.1186/s13007-019-0411-2
» https://doi.org/https://doi.org/10.1186/s13007-019-0411-2 -
FRANÇA-SILVA, F.; GOMES-JUNIOR, F.G.; REGO, C.H.Q.; MARASSI, A.G.; TANNÚS, A. Advances in imaging technologies for soybean seed analysis. Journal of Seed Science , v.45, e202345022, 2023. https://doi.org/10.1590/2317-1545v45274098
» https://doi.org/https://doi.org/10.1590/2317-1545v45274098 - GOMES-JUNIOR, F. G. Análise computadorizada de imagens de plântulas In: KRZYZANOWSKI, F. C.; VIEIRA, R. D.; MARCOS-FILHO, J.; FRANÇA NETO, J. B. (Eds.). Vigor de sementes: conceitos e testes . Londrina: ABRATES , 2020. p. 141-177.
- ISTA. INTERNATIONAL SEED TESTING ASSOCIATION. lnternational Rules for Seed Testing Bassersdorf: ISTA, 2020. 320p.
- KRZYZANOWSKI, F.C.; FRANÇA-NETO, J.B.; GOMES-JUNIOR, F.G.; NAKAGAWA, J. Testes de vigor baseados em desempenho de plântulas In: KRZYZANOWSKI, F. C.; VIEIRA, R. D.; MARCOS-FILHO, J.; FRANÇA NETO, J.B. (Eds.). Vigor de sementes: conceitos e testes . Londrina: ABRATES , 2020. p.79-127.
-
KOTWALIWALE, N.; SINGH, K.; KALNE, A.; JHA, S.N.; SETH, N.; KAR, A. X-ray imaging methods for internal quality evaluation of agricultural produce. Journal of Food Science and Technology, v.51, n.1, p.1-15, 2014. https://doi.org/10.1007/s13197-011-0485-y
» https://doi.org/https://doi.org/10.1007/s13197-011-0485-y -
LIMÃO, M.A.R.; DIAS, D.C.F.S.; ARAUJO, J.O.; SOARES, J.M.; NASCIMENTO, W.M.; SILVA, L.J. Classification of lentil seed vigor based on seedling image analysis techniques and interactive machine learning. Journal of Seed Science , v.45, n.1, e20238712, 2023. https://doi.org/10.1590/2317-1545v45277692
» https://doi.org/https://doi.org/10.1590/2317-1545v45277692 -
MAGUIRE, J. D. Speed of germination - Aid in selection and evaluation for seedling emergence and vigor. Crop Science, v.2, n.2, p.176, 1962. https://acsess.onlinelibrary.wiley.com/doi/abs/10.2135/cropsci1962.0011183X000200020033x
» https://acsess.onlinelibrary.wiley.com/doi/abs/10.2135/cropsci1962.0011183X000200020033x - MANUAL GROUNDEYE® . Tbit Tecnologia e Sistemas, Lavras. 2016.88p.
- MARCOS-FILHO, J. Fisiologia de sementes de plantas cultivadas Piracicaba: FEALQ, 2015. 660p.
- MARCOS-FILHO, J. Testes de vigor: importância e utilização In: KRZYZANOWSKI, F. C.; VIEIRA, R. D.; MARCOS-FILHO, J.; FRANÇA NETO, J.B. (Eds.). Vigor de sementes: conceitos e testes . Londrina: ABRATES , 2020a. p. 17-72.
- MARCOS-FILHO, J. Teste de envelhecimento acelerado In: KRZYZANOWSKI, F. C.; VIEIRA, R. D.; MARCOS-FILHO, J.; FRANÇA NETO, J.B. (Eds.). Vigor de sementes: conceitos e testes . Londrina: ABRATES , 2020b. p. 182-244.
-
MARCOS-FILHO, J.; NOVEMBRE, A.D.L.C.; CHAMMA, H.M.C.P. Tamanho da semente e o teste de envelhecimento acelerado para soja. Scientia Agricola, v.57, n.3, p.473-482, 2000. https://doi.org/10.1590/S0103-90162000000300016
» https://doi.org/https://doi.org/10.1590/S0103-90162000000300016 -
MEDEIROS, A.D.; ARAÚJO, J.O.; LEÓN, M.J.Z.; SILVA, L.J.; DIAS, D.C.F.S. Parameters based on X-ray images to assess the physical and physiological quality of Leucaena leucocephala seeds. Ciência e Agrotecnologia , v.42, n.6, p.643-652, 2018. https://doi.org/10.1590/1413-70542018426023318
» https://doi.org/https://doi.org/10.1590/1413-70542018426023318 -
MEDEIROS, A.D.; CAPOBIANGO, N.P.; SILVA, J.M.; SILVA, L.J.; SILVA, C.B.; DIAS, D.C.F.S. Interactive machine learning for soybean seed and seedling quality classification. Scientific Reports, v.10, n.1, p.11267, 2020a. https://doi.org/10.1038/s41598-020-68273-y
» https://doi.org/https://doi.org/10.1038/s41598-020-68273-y -
MEDEIROS, A.D.; PINHEIRO, D.T.; XAVIER, W.A.; SILVA, L.J.; DIAS, D.C.F.S. Classificação de qualidade de sementes de pinhão manso utilizando imagens radiográficas e aprendizado de máquina. Industrial Crops and Products, v.146, e112162, 2020b. https://doi.org/10.1016/j.indcrop.2020.112162
» https://doi.org/https://doi.org/10.1016/j.indcrop.2020.112162 -
MEDEIROS, A.D.; SILVA, L.J.; SILVA, J.M.; DIAS, D.C.F.S.; PEREIRA, M.D. IJCropSeed: An open-access tool for high-throughput analysis of crop seed radiographs. Computers and Electronics in Agriculture, v.175, p.105555, 2020c. https://doi.org/10.1016/j.compag.2020.105555
» https://doi.org/https://doi.org/10.1016/j.compag.2020.105555 -
NOGUEIRA, G.A.; XAVIER, G.F.; SILVA, T.R.B.; BACCHETTA, G.; TORRES, S.B.; PEREIRA, M.D.; DUTRA, A.S.; ALVES, C.Z. X-rays for determination of papaya seeds viability. Revista Ciência Agronômica, v.55, n.1, e20238712, 2024. https://doi.org/10.5935/1806-6690.20240021
» https://doi.org/https://doi.org/10.5935/1806-6690.20240021 -
PEREIRA, M.D.; REIS, J.A.; FERRARI, C.S.; VALE, A.M.P.G. Processamento digital de imagens de plântulas na avaliação do vigor de sementes de Moringa oleifera Lam. Ciência Florestal, v.30, n.2, p.291-306, 2020. https://doi.org/10.5902/1980509825750
» https://doi.org/https://doi.org/10.5902/1980509825750 -
R CORE TEAM. R: A language and environment for statistical computing 2025. https://www.R-project.org/
» https://www.R-project.org/ -
REHMAN, A.; KAMRAN, M.; AFZAL, I. Production and processing of quality cotton seed. In: AHMAD, S.; HASANUZZAMAN, M. (Eds.). Cotton production and uses Singapore: Springer, p.547-570, 2020. https://doi.org/10.1007/978-981-15-1472-2_27
» https://doi.org/https://doi.org/10.1007/978-981-15-1472-2_27 -
RIBEIRO, A.S.; NERI, T.F.S.; MEDEIROS, A.D.; MILAGRES, C.C.; SILVA, L.J. Use of radiographic images for rapid and non-destructive assessment of crambe seed quality. Journal of Seed Science , v.43, e202143001, 2021. https://doi.org/10.1590/2317-1545v43239136
» https://doi.org/https://doi.org/10.1590/2317-1545v43239136 -
SANTOS, R.F.; GOMES-JUNIOR, F.G.; MARCOS-FILHO, J. Morphological and physiological changes during maturation of okra seeds evaluated through image analysis. Scientia Agricola , v.77, n.3, e20180297, 2020. https://doi.org/10.1590/1678-992X-2018-0297
» https://doi.org/https://doi.org/10.1590/1678-992X-2018-0297 -
SILVA, P.P.; BARROS, A.C.S.A.; MARCOS-FILHO, J.; GOMES-JUNIOR, F.G.; NASCIMENTO, W.M. Assessment of squash seed vigor using computerized image analysis. Journal of Seed Science , v.39, n.2, p.159-165, 2017. https://doi.org/10.1590/2317-1545v39n2171177
» https://doi.org/https://doi.org/10.1590/2317-1545v39n2171177 -
SIMAK, M.; GUSTAFSSON, Å. X-ray photography and sensitivity in seed tests. Hereditas, v.39, n.3-4, p.458-468, 1953. https://doi.org/10.1111/j.1601-5223.1953.tb03430.x
» https://doi.org/https://doi.org/10.1111/j.1601-5223.1953.tb03430.x -
SOUZA, L.R.; LIMÃO, M.A.R.; PINHEIRO, D.T.; PERIS, G.C.O.; DIAS, D.C.F.; DIAS, L.A.D. Near infrared spectroscopy and seedling image analysis to evaluate the physiological potential of Urochloa decumbens (Stapf) R.D. Webster seeds. Journal of Seed Science , v.45, e202277021, 2023. https://www.scielo.br/j/jss/a/6f5FyDzQyg4JkgY49Kbnb9L/?format=html⟨=en
» https://www.scielo.br/j/jss/a/6f5FyDzQyg4JkgY49Kbnb9L/?format=html⟨=en -
VASCONCELOS, M.C.; COSTA, J.C.; OLIVEIRA, A.M.S.; GRANJA, J.A.A.; GUIMARÃES, R.M. Radiography and biometric analysis of broadleaf vegetable seeds. Amazonian Journal of Agricultural and Environmental Sciences, v.61, 2018. http://dx.doi.org/10.22491/rca.2018.2885
» https://doi.org/http://dx.doi.org/10.22491/rca.2018.2885 -
WENDT, L.; GOMES-JUNIOR, F.G.; ZORATO, M.F.; MOREIRA, G.C. Avaliação do potencial fisiológico de sementes de soja por meio de imagens. Pesquisa Agropecuária Tropical, v.44, n.3, p.280-286, 2014. https://doi.org/10.1590/S1983-40632014000300011
» https://doi.org/https://doi.org/10.1590/S1983-40632014000300011 -
XIA, Y.; XU, Y.; LI, J.; ZHANG, C.; FAN, S. Recent advances in emerging techniques for non-destructive detection of seed viability: A review. Artificial Intelligence in Agriculture, v.1, p.35-47, 2019. https://doi.org/10.1016/j.aiia.2019.05.001
» https://doi.org/https://doi.org/10.1016/j.aiia.2019.05.001 -
YASAM, S.; NAIR, A.; KUMAR, K.P. Supervised learning-based seed germination ability prediction for precision farming. Soft Computing, v.26, p.13133-13144, 2022. https://doi.org/10.1007/s00500-022-06910-6
» https://doi.org/https://doi.org/10.1007/s00500-022-06910-6
