Machine learning-assisted non-destructive prediction of root architecture in early-stage tomato rootstocks Authors: Hatice Şeyma Yücel Published: 2026 ABSTRACT The architecture of the root system is a primary factor in determining rootstock performance, affecting water and nutrient uptake, biomass accumulation, and overall vigor. However, direct root phenotyping is destructive, labor-intensive, and difficult to do routinely in breeding programs. The present study investigated early root morphological variation among developed interspecific tomato rootstock candidates (Solanum lycopersicum x S. habrochaites). The ability of linear regression and machine learning models to predict root traits from easily measured plant growth parameters was assessed. Nineteen interspecific hybrid rootstock candidates, two commercial rootstocks, and one scion were grown under optimal greenhouse conditions and evaluated at 0, 10, 20, and 30 days after planting. Root length, root surface area, root diameter, and root volume were determined by digital image analysis. In contrast, genotype, plant length, and stem diameter were used as input variables. Significant genotype x sampling date effects were observed for most morphological and biomass traits, indicating dynamic changes in root and shoot development during the first 30 days of growth. The rootstock candidates RSH-17 and RSH-6 generally showed relatively higher root length, surface area, root volume, and biomass accumulation than the commercial rootstocks and scion. XGBoost and OLR were the best predictive models, with R2 values as high as 0.95 for root length, surface area, and volume. Root diameter was predicted less accurately than root length, surface area, and volume, suggesting that it might be a more independent or less variable root trait during early development. Overall, results suggest that vigor-related traits can serve as useful proxies for estimating major root architectural traits in early-stage tomato rootstock selection. Both XGBoost and OLR performed well, suggesting that root and shoot development were highly coordinated under optimal (non-stress) conditions. Hence, predictive modeling may help prioritize promising rootstock candidates before destructive root analysis. However, more validation under stress conditions and for longer periods of development is needed to determine the greater applicability of these models.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12870-026-09332-w. Introduction Tomatoes are among the most widely cultivated vegetables worldwide, including Türkiye [1]. While global tomato production stands at 192,317,973 tons, Türkiye ranks third among tomato producers, behind China and India [1]. Tomatoes can be grown in a wide range of climates and soil types because of their high adaptability. However, biotic and abiotic stress factors limit tomato cultivation [2]. These stress factors negatively affect the physiological and morphological development of tomato plants. They have adverse effects on key production stages, including flowering, fruit setting, and ripening. Under these stress conditions, grafted seedlings are an effective method for increasing resistance to both biotic and abiotic stresses [3]. The first scientific studies on grafting vegetables were conducted in the 1920s to increase resistance to soil-borne diseases [4, 5]. Through grafting, it is possible to leverage the rootstock’s resistance characteristics while preserving the genetic potential of the cultivar (scion) [6]. Therefore, the development of new stress-tolerant rootstocks is critical for sustainable agriculture [7]. The use of grafted seedlings in vegetables is a widely adopted technique, especially to enhance plant growth, development, and productivity under biotic and abiotic stress conditions. Grafting in the Solanaceae family first began with eggplants, followed by tomatoes in the 1960s and later with peppers. Tomatoes are the most grafted species in this family because of their economic importance and area covered [6]. Over the last 20 years, grafting in tomatoes has spread worldwide more rapidly than in other species. This success primarily depends on the performance of the rootstocks used. In recent years, there has been a demand for the use of wild tomato species as rootstocks. Wild tomato species can exhibit different adaptation strategies to biotic and abiotic stressors, reflecting their growing conditions in their habitats. One wild tomato species, S. habrochaites, is considered an important genetic resource because of its strong root system. According to Schwarz et al. [8], S. habrochaites improves root development and water and nutrient uptake in tomato seedlings, thereby promoting growth. However, its direct use is not recommended due to differences in root morphology and vascular tissues. It is usually hybridized with S. lycopersicum to produce commercial rootstocks [9]. Interspecific hybrid rootstocks hybridized from S. habrochaites offer several advantages, including drought tolerance due to a deep and extensive root system, increased biomass and improved fruit quality in grafted plants, resistance to multiple soil-borne pathogens (Fusarium spp., R. solanacearum, nematodes), and osmotic tolerance to salinity stress [10, 11]. These characteristics make S. lycopersicum x S. habrochaites interspecific hybrids among the most widely used combinations in tomato rootstock breeding [10, 12]. Interspecific hybrid rootstocks of S. lycopersicum x S. habrochaites are increasingly used in tomato production in Türkiye. Indeed, 47 interspecific hybrid tomato rootstock varieties are commercially available, including 39 registered and eight production-approved varieties [13]. Using these rootstocks, approximately 100 million grafted seedlings are produced annually in Türkiye. Türkiye has made significant progress in tomato breeding in recent years, and efforts have been made to increase domestic production by breeding several local hybrid varieties [5]. This information shows that the root characteristics of rootstocks are important for better yield under stress conditions. Root structures are controlled by polygenic variation [14] and should be studied in rootstock breeding programs [15]. However, examining the root structure is challenging. Therefore, there are very few studies on rootstock selection based on the phenotypic characteristics of the roots [8, 16, 17]. In recent years, detailed studies on root structures have been conducted using advanced digital imaging systems [18, 19]. In recent years, advances in digital phenotyping and machine learning have enabled the evaluation of root system characteristics using indirect and highly efficient methods [20]. Machine learning can improve prediction accuracy by revealing patterns that are difficult to detect using traditional methods, owing to its ability to capture complex biological relationships. Recent studies have successfully used machine learning models integrated with imaging and remote sensing data to predict the root distribution of field-grown barley [20]. Similarly, in a study conducted on soybean plants under greenhouse conditions, it was demonstrated that complex root architecture variables could be predicted with high accuracy using algorithms such as artificial neural networks, Random Forest, and XGBoost, based on a few easily measurable plant traits [21]. These developments indicate that machine learning-based models can serve as reliable tools for predicting root morphologies. It is imperative that root predictions can be made using machine learning, as root examination can damage plants. In this context, this study aimed to accurately predict root morphology parameters in tomato plants using machine learning models, offering a non-destructive alternative to traditional root phenotyping. Materials and methods The study was carried out in 2020 in a greenhouse located at the Research Site of the Faculty of Agriculture, Ondokuz Mayıs University (41.3640° N, 36.1885° E). In this study, 19 interspecific hybrid tomato rootstocks (S. lycopersicum x S. habrochaites) were used for the experiments. These rootstocks are characterized by their superior traits, including high grafting success rates and resistance to F. oxysporium f.sp. lycopersici (FOL), F. oxysporium f.sp. radicis-lycopersici (FORL), and root-knot nematodes. Two commercial rootstocks and a scion were used as controls for comparison. Seeds from the rootstocks were sown in trays containing a peat: perlite (2:1, v/v) mixture. Vermiculite was used as the covering soil. The macro- and microelements required by the seedlings were regularly supplied during the cultivation period. Seedlings with four true leaves, obtained from the company, were planted in 7-liter plastic pots containing a sterile peat (Klasmann TS1) : perlite (Agroera) (2:1, v/v) mixture. The experiment was conducted in a controlled greenhouse where the temperature ranged between 24 ± 2 °C during the day and 18 ± 2 °C at night and the relative humidity was between 60% and 75%. Irrigation was applied whenever the moisture content of the growing medium fell below about 50%. Each pot was irrigated on a routine basis and 100 ml of Hoagland nutrient solution was provided at the beginning of each week to ensure uniform and optimal nutrient supply to all genotypes. Plants were grown under ambient light conditions with no supplemental lighting. The pots were positioned according to the Completely Randomized Design (CRD) with three replicates and twelve plants per replicate. To reveal the relationship between plant growth and root architecture, stem diameter, plant length, root and plant dry weight, and quantitative root analyses were performed on days 0, 10, 20, and 30 after planting. For each sampling day, three plants from each replicate were analyzed. The day on which the seedlings were planted was considered day 0. On day 0, plants were didn’t planted and analyzed upon arrival. Plant length was measured using a ruler from the root collar to the growing tip, and stem diameter was measured using calipers 5 cm above the soil surface. For root and plant dry weight values, the roots, stems, and leaves of the plants were dried in an oven at 80 °C for 72 h after analysis [19]. Subsequently, the root dry weights (g) were determined by weighing them on a precision balance (0.001 g). The WinRhizo root analysis program (ver. 2013, Regent Instruments, QC, Canada) was used to examine the root architecture [16, 17, 22]. For this purpose, the roots of tomato plants were carefully washed and dried with paper towels to avoid damaging them. The roots were then placed on A3-sized acetate and carefully separated for detailed detection by the device. The roots prepared for scanning were placed in a scanner (Epson Expression 10 000XL, Epson America Inc., Long Beach, CA, USA) and transferred to a computer environment at 400 dpi. The tomato varieties were examined for the following phenotypic root traits [23–25].Total root length (cm): The total lengths of roots in all diameter classes were determined for all tomato rootstocks in the study.Root surface area (cm²): The surface area of the outer perimeter of all roots was calculated as the root surface area (projected areas).Average root diameter (mm): The average root diameter was calculated by individually examining all root extensions in tomato rootstocks using the WinRhizo program.Total root volume (cm³): It was computed by WinRhizo as a function of root length and root diameter in the roots of the quantitatively analyzed tomato plants [26]. Total root length (cm): The total lengths of roots in all diameter classes were determined for all tomato rootstocks in the study. Total root length (cm): The total lengths of roots in all diameter classes were determined for all tomato rootstocks in the study. Root surface area (cm²): The surface area of the outer perimeter of all roots was calculated as the root surface area (projected areas). Root surface area (cm²): The surface area of the outer perimeter of all roots was calculated as the root surface area (projected areas). Average root diameter (mm): The average root diameter was calculated by individually examining all root extensions in tomato rootstocks using the WinRhizo program. Average root diameter (mm): The average root diameter was calculated by individually examining all root extensions in tomato rootstocks using the WinRhizo program. Total root volume (cm³): It was computed by WinRhizo as a function of root length and root diameter in the roots of the quantitatively analyzed tomato plants [26]. Total root volume (cm³): It was computed by WinRhizo as a function of root length and root diameter in the roots of the quantitatively analyzed tomato plants [26]. Correlation analysis and principal component analysis To examine the relationships among root and shoot morphological traits, correlation and multivariate analyses were performed prior to machine learning modeling. For this purpose, Pearson’s r was chosen because it can accurately measure the interaction between variables when the data exhibit a normal distribution. Pearson’s correlation coefficients were calculated using numerical variables to quantify the strength and direction of the pairwise associations between root length, root surface area, root diameter, root volume, plant stem diameter, plant length, plant dry weight, and root dry weight. Correlation matrices were visualized as heatmaps to facilitate the identification of highly correlated traits and potential multicollinearity among the variables. Principal component analysis (PCA) was conducted to reduce data dimensionality and identify the main sources of variation within the dataset. Prior to the PCA, all variables were mean-centered and scaled to unit variance to eliminate scale effects [27]. The analysis was performed using a covariance matrix, and eigenvalues, explained variance ratios, and variable loadings were extracted. Principal components with the highest explained variance were retained and loading plots were generated to assess the contribution and grouping of individual traits along the principal axes. PCA was used as an exploratory tool to summarize complex trait interrelationships and support the interpretation of correlation patterns and subsequent machine learning analyses. Modeling procedure Four machine learning algorithms were used to evaluate the ability to predict root traits based on vigor related traits. For this purpose, Extreme Gradient Boosting (XGBoost), Random Forest (RF), Gaussian Process Regression (GPR), and an artificial neural network-based Multilayer Perceptron (MLP) were used [28–31]. To objectively assess the model performance, the dataset was randomly split into training and testing sets. 70% of the observations were used for model training, and the remaining 30% were reserved for testing the model. Data partitioning was performed using the “createDataPartition” function of the caret package in R [32, 33]. To minimize overfitting and improve model generalization, a repeated k-fold cross-validation strategy (five folds with three repeats) was applied during model training. All preprocessing and model tuning steps were performed exclusively within the training subset to avoid data leakage. The final model evaluation was performed using the independent 30% test set, which was not seen by the models during the training or tuning process. As the MLP and GPR are scale-sensitive algorithms, continuous predictors were mean-centered and scaled. Model development and tuning were performed using the caret framework in R, which provided a consistent interface for model fitting, hyperparameter optimization, and performance evaluation. Genotype, plant length and stem diameter, were used as input variables in the models, and root morphological traits, including root length, surface area, diameter, and volume were predicted. Gaussian Process Regression models were implemented using a radial basis function kernel, allowing flexible non-linear mapping between predictors and response variables. The input features were standardized prior to model fitting. Random Forest models were constructed as ensemble learners composed of 500 regression trees. At each split, a subset of the predictor variables was randomly sampled to reduce the correlation among trees and enhance predictive stability. Extreme Gradient Boosting models were implemented using a regularized boosting framework. To achieve robust predictive performance, a grid search was used to optimize the hyperparameters. The search space included tree depth values of 3, 6, and 9; learning rates (η) of 0.01, 0.05, and 0.1; subsampling ratios of 0.6, 0.7, and 0.8; L1 regularization coefficients (α) of 0 and 0.1; and L2 regularization coefficients (λ) of 1 and 1.5. The optimal model configuration was determined as a tree depth of 3, a learning rate of 0.05, and a subsampling ratio of 0.8. Multi-Layer Perceptron models were developed as feedforward neural networks with a single hidden layer. To determine the optimal model architecture, the number of hidden neurons and the weight decay parameter were optimized using a grid search strategy. The number of hidden units was tested at 3, 5, 7, and 9, and the weight decay parameter was evaluated at 0, 0.001, 0.01, and 0.1. An early-stopping criterion based on validation error was used to prevent overfitting and improve generalization to unseen data. The predictor variables were standardized prior to training, and a linear output activation function was employed to accommodate continuous response variables. Ordinary Linear Regression (OLR), based on the ordinary least squares method, was used as a baseline linear model to compare the performance of advanced machine learning algorithms. To objectively evaluate the model’s generalizability and predictive power, the dataset was split into a 70% training set and a 30% independent test set using the caret package [33]. The performance of the OLR model was evaluated using comprehensive metrics, including the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), and the results were compared with those of other non-linear models. The model performance was evaluated using multiple complementary metrics, including the mean absolute error (MAE) (Eq. (1)), root mean square error (RMSE) (Eq. (2)), and coefficient of determination (R²) (Eq. (3)).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MAE=\frac{1}{n}{\sum\:}_{i=1}^{n}\left|{Y}_{i}-\widehat{{Y}_{i}}\right|$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:RMSE=\sqrt{\frac{1}{n}{\sum\:}_{i=1}^{n}{\left({Y}_{i}-\widehat{{Y}_{i}}\right)}^{2}}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{2}=1-\frac{{\sum\:}_{i=1}^{n}{\left({Y}_{i}-\widehat{{Y}_{i}}\right)}^{2}}{{\sum\:}_{i=1}^{n}{\left({Y}_{i}-\stackrel{-}{Y}\right)}^{2}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}$$\end{document}: Observed values; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{Y}_{i}}$$\end{document}: Predicted values; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{-}{Y}$$\end{document}: Mean of the observed values; n: Total number of observations. Experimental design and statistical analysis Statistical analyses were conducted using R software (version 4.5.1 for macOS) [32]. Prior to parametric analysis, the normality of the data was assessed using the Shapiro–Wilk test, and the homogeneity of variances was verified via Levene’s test. While most morphological parameters exhibited a normal distribution (P > 0.05), the root diameter and root dry weight data underwent a log10 transformation to meet the normality assumption (Table S1). Following verification of normality, a two-way ANOVA was used to determine the main and interaction effects of genotype and sampling date on the measured morphological parameters. Differences among the genotype and sampling date interaction means were further evaluated using Duncan’s multiple range test at P < 0.05. Hyphenated labels such as “e–g” denote overlapping significance groups from e to g and were used to avoid writing all letters separately. Results and discussion In this study, statistically significant differences were found in the phenotypic root characteristics (root length, average root diameter, root surface area, and root volume) that depended on the root system architecture of 21 rootstocks (19 developed rootstock and two commercial variety) and one scion. Quantitative root analysis was conducted at 10-day intervals. These analyses revealed that depending on the genotype, some periods showed faster growth in the root characteristics examined, whereas other periods showed slower growth. In this study, root length was measured for rootstock candidates and commercial rootstocks on days 0, 10, 20, and 30, and statistically significant differences were observed between genotypes at all measurement times (P < 0.001) (Fig. 1). Fig. 1Root length (cm) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) Root length (cm) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) All genotypes exhibited increased root length over the trial period. However, significant variations were noted among genotypes in terms of the rate of increase and total root length. At the beginning of the experiment (day 0), the root length values ranged from 423.47 to 781.62 cm. The highest initial root length was observed in the RSH-17 genotype (781.62 cm), which was significantly higher than that of all other genotypes. The RSH-6 (680.02 cm) and RSH-2 (668.02 cm) genotypes followed, whereas the scion exhibited the shortest root length (423.47 cm). It was determined that root length values increased to the range of 1039.36–1794.42 cm on the 10th day of treatment. During this period, the RSH-17 genotype reached the highest root length of 1794.42 cm. After the RSH-17 genotype, the RSH-6 (1722.89 cm), RSH-2 (1716.62 cm), and RSH-9 (1709.29 cm) genotypes showed the highest root length values. In contrast, the RSH-16 genotype (1043.90 cm) and scion (1039.36 cm) exhibited the shortest root lengths. When the measurements on day 20 were examined, it was observed that the increase in root length continued, and the differences between genotypes became even more noticeable. During this period, the RSH-17 genotype reached the highest root length value (3191.54 cm). The RSH-6 genotype also showed a similarly high performance (3024.57 cm). The RSH-2 (2896.22 cm), RSH-9 (2863.81 cm), and RSH-1 (2832.61 cm) genotypes followed this trend, whereas the scion showed the lowest value (1899.96 cm). The highest root length on day 30 was observed in the RSH-6 genotype (3996.80 cm), which was statistically equivalent to that of the RSH-17 genotype (3976.66 cm). The RSH-2 (3655.51 cm), RSH-9 (3533.90 cm), RSH-1 (3502.20 cm), and RSH-10 (3444.81 cm) genotypes followed this group. Commercial rootstocks remained in the range of 2802.00–3037.72 cm during this period, whereas the scion showed the lowest root length value (2298.41 cm). Under abiotic stress conditions, such as low phosphorus levels or water restriction, the development of plant root systems into deeper soil layers is crucial for the effective uptake of water and plant nutrients. In this context, a high total root length is considered a fundamental characteristic that increases the plant’s capacity to adapt to limited resource conditions [22, 34–37]. Furthermore, long root systems confer a significant advantage in terms of nitrate uptake [15]. The findings of this study are consistent with those of previous studies and indicate that the RSH-17 and RSH-6 genotypes exhibited higher total root lengths at all measurement times. On the 30th day, the RSH-6 (3996.80 cm) and RSH-17 (3976.66 cm) genotypes reached the highest root lengths, which were statistically different from those of the commercial rootstocks and rootstock candidates. Studies have indicated that among plants with the same root biomass, those with longer root systems can access deeper soil layers more easily and thus obtain water and nutrients more efficiently [22, 23]. In this context, our study suggests that the RSH-17 and RSH-6 genotypes, which exhibit high root length, have the potential to utilize deeper soil layers under low-nutrient and water stress conditions and may have potential advantages for stress tolerance. Root surface area was measured for all genotypes and commercial rootstocks in the trial on days 0, 10, 20, and 30 of the trial. Statistically significant differences among genotypes were observed at all measurement times (P < 0.001). In general, a significant increase in the root surface area was observed over time across all genotypes. However, the rate and magnitude of this increase differed, particularly among genotypes (Fig. 2). Fig. 2Root surface area (cm2) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) Root surface area (cm2) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) At the onset of the experiment (day 0), the root surface area values ranged from 88.27 to 122.69 cm². The genotypes RSH-17 (122.69 cm²), RSH-6 (121.74 cm²), and RSH-2 (116.45 cm²) exhibited the highest initial values, suggesting more vigorous root system establishment in the early stages. In contrast, the scion was statistically categorized in the lowest group, showing the minimum root surface area (88.27 cm²). Measurements taken on day 10 showed a significant increase in root surface area values and more apparent differences between the genotypes. During this period, the RSH-17 genotype reached the highest value (502.95 cm²) and was statistically different from all other genotypes. The RSH-6 (444.48 cm²), RSH-2 (441.49 cm²), and RSH-9 (430.75 cm²) genotypes were statistically placed in the upper group, with the highest root surface area values, whereas the scion had the lowest value during this period (179.73 cm²). On the 20th day, measurements showed that the increase in root surface area continued, and the RSH-17 genotype maintained its superiority (741.99 cm²). The RSH-6 (730.73 cm²) and RSH-2 (697.76 cm²) genotypes maintained their high performance during this period, followed by RSH-9 (678.47 cm²) and RSH-1 (675.22 cm²) genotypes. However, the scion had the lowest root surface area (389.30 cm²). When the data obtained on the 30th day were examined, it was found that the RSH-17 genotype had the highest value (932.92 cm²). The RSH-6 (882.15 cm²) and RSH-2 (878.07 cm²) genotypes were also in the top statistical group for high root surface area during this period, followed by RSH-9 (848.27 cm²), RSH-1 (842.83 cm²), and RSH-10 (834.76 cm²) genotypes. Commercial-1 (708.95 cm²) and commercial-2 (654.59 cm²) had lower values, whereas the scion had the lowest root surface area value during this period, as it did during all measurement periods (497.60 cm²). One of the most important root architectural characteristics is the effect of water and nutrient uptake in plants, which is the total root surface area. A high root surface area positively affects water and nutrient uptake by increasing the contact area between plants and soil solution, especially under abiotic stress conditions, such as drought and high temperatures [19, 38]. The results of this study are consistent with the relevant literature and indicate that the RSH-17, RSH-6, and RSH-2 genotypes exhibited higher root surface areas at all measurement times. At 30 days, the RSH-17 genotype reached the highest root surface area, whereas the RSH-6 and RSH-2 genotypes also showed statistically superior performance compared to commercial rootstocks. As stated by Johnson et al. [23], a greater distribution of roots in the soil provides an important advantage for the effective use of applied fertilizers. In this context, the RSH-17 and RSH-6 genotypes, which showed a higher root surface area in our study, may be advantageous for more efficient resource use and greater stress tolerance, especially under conditions of limited water and nutrients. Root diameter was measured for all genotypes and commercial rootstocks on days 0, 10, 20, and 30 after planting. Statistically significant differences between genotypes were observed at all measurement times (day 0: P < 0.001; day 10: P < 0.01; days 20 and 30: P < 0.001). In general, a gradual increase in root diameter was observed throughout the trial period for all genotypes. However, the level and continuity of this increase varied significantly between genotypes (Fig. 3). Fig. 3Root diameter (mm) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) Root diameter (mm) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) The root diameter values at the start of the experiment (day 0) ranged from 0.46 to 0.58 mm. The highest initial root diameters were observed in the scion (0.58 mm), RSH-5 (0.57 mm), RSH-16 (0.57 mm), and RSH-18 (0.57 mm), whereas the RSH-17 genotype had the lowest root diameter (0.46 mm). On day 10, the root diameter increased to 0.55–0.65 mm. During this period, the scion reached the highest root diameter (0.65 mm), which was statistically the highest. The RSH-5 (0.63 mm), RSH-16 (0.63 mm), and RSH-18 (0.63 mm) genotypes also showed high root diameter values. In contrast, the RSH-17 genotype exhibited the smallest root diameter during this period (0.55 mm). On the 20th day, the measurements showed that the increase in root diameter continued, and the differences between the genotypes became more prominent. The highest root diameter was observed in the scion (0.77 mm). The RSH-16 (0.73 mm), RSH-5 (0.70 mm), and RSH-18 (0.70 mm) genotypes followed the scion, whereas the RSH-17 genotype showed the lowest value (0.57 mm). On the 30th day, the root diameter values ranged from 0.62 to 1.06 mm. The scion exhibited the largest root diameter (1.06 mm). This was followed by RSH-5 (1.03 mm), RSH-16 (1.03 mm), and RSH-18 (1.00 mm). In contrast, the lowest root diameter was measured in the RSH-17 genotype (0.62 mm). Eissenstat [39] reported that the average root diameter is an important indicator of lateral rooting tendency and that root diameter directly affects the root’s ability to absorb water and nutrients. Similarly, Pereira-Dias et al. [40] found that roots with diameters of less than 1 mm in grafted pepper plants showed approximately 4–5 times greater nutrient uptake efficiency than roots with diameters greater than 1 mm. The findings of this study are consistent with the literature and show significant differences in root diameter among genotypes. Genotypes with average root diameters below 1 mm on days 0 and 10 also had higher root surface areas and total root lengths. The development of a broad, extensive fibrous root system with balanced root diameters in the later stages by the RSH-17 and RSH-6 genotypes suggests that these genotypes have a functional root architecture for nutrient and water uptake. In contrast, the fact that higher root diameters were observed in some genotypes at later stages did not parallel increases in root surface area or root length, indicating that root diameter alone is not sufficient to explain the effectiveness of the root system. In this study, root volume was measured for all genotypes and commercial rootstocks on days 0, 10, 20, and 30, and differences between genotypes were determined at each measurement time. In terms of root volume, differences between genotypes were statistically significant on day 0 (P < 0.05), but not on day 10 (P > 0.05), and were significant on days 20 and 30 (P < 0.001) (Fig. 4). Fig. 4Root volume (cm3) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) Root volume (cm3) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) At the start of the experiment (day 0), the root volume values ranged from 1.18 to 1.64 cm³. The highest initial root volume was detected in the RSH-17 genotype (1.64 cm³), followed by RSH-2 (1.50 cm³), RSH-6 (1.49 cm³), and RSH-1 (1.48 cm³). In contrast, the scion (1.18 cm³) and RSH-16 (1.18 cm³) genotypes had the lowest root volumes among the genotypes. These results indicate that there are limited but significant differences in root volume among genotypes during the early period. On day 10, root volume values ranged from 4.42 to 5.92 cm³, and the differences between genotypes were not statistically significant. The highest root volume during this period was observed in the RSH-17 genotype (5.92 cm³), followed by RSH-2 (5.87 cm³), and RSH-6 (5.77 cm³). The lowest values were detected in the scion (4.42 cm³) and RSH-16 (4.72 cm³). When the measurements on the twentieth day were examined, it became evident that differences between genotypes in root volume emerged. During this period, the RSH-17 genotype reached the highest root volume (16.33 cm³) and was statistically in the upper group. The RSH-2 genotype also showed a high root volume (14.69 cm³). The scion (9.45 cm³) and RSH-16 (9.67 cm³) genotypes had the lowest root volumes among the genotypes. The data obtained on the thirtieth day revealed the final root volume potential of the genotypes. The highest root volume values were observed in the RSH-17 (24.14 cm³) and RSH-2 (23.08 cm³) genotypes, which were statistically distinct from the others. These were followed by the genotypes RSH-6 (19.89 cm³), RSH-9 (19.50 cm³), RSH-1 (19.25 cm³), and RSH-10 (19.25 cm³). The lowest root volume values were detected in the scion (16.16 cm³), RSH-16 (16.34 cm³), and RSH-5 (16.76 cm³) genotypes. The root volume values of the commercial rootstocks remained in the range of 17.47–17.86 cm³ during this period. Root volume is an important parameter to consider in production [38, 41]. Root volume is a fundamental indicator of the area of the root system in the soil and, consequently, its capacity to access water and nutrients. In this context, Karaağaç et al. [42] reported that plants develop distinct root adaptation strategies, as evidenced by changes in total root length and root volume in response to different stress conditions. Our study revealed significant differences in root volume between genotypes, which increased over time. These results paralleled the genotypic differences in root surface area. Genotypes with a high root surface area and volume are considered to have a more effective root architecture, as they increase the area of contact with the soil. Li et al. [25] reported that genotype was a determinant of root volume in lettuce plants, with 45-day-old plants ranging from 1.64 to 7.24 cm³. Sarıbaş et al. [16] measured root volume values in eggplant rootstock candidates of 40.5–96.8 cm³ at the end of the 50th day. Similarly, Karaağaç et al. [42] reported that root volume values in pepper genotypes of C. annuum, C. baccatum, and C. chinense ranged from 33.15 to 40.55, 39.03 to 73.91, and 39.05 to 71.38 cm³, respectively. The quantitative differences between the root volume values obtained in our study and those reported in the literature are due to species and genotype differences, as well as differences in the developmental stage and the number of days over which measurements were made. However, when root volume and root surface area are evaluated together, genotypes with strong root systems have the potential to utilize soil resources more efficiently. In this study, the stem diameters of the rootstocks were measured on days 0, 10, 20, and 30, and statistically significant differences were observed between the genotypes at all measurement times (day 0: P < 0.001; days 10 and 20: P < 0.01; day 30: P < 0.001). A steady increase in stem diameter was observed across all genotypes throughout the trial period according to the genotypes (Fig. 5). Fig. 5Plant stem diameter (mm) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) Plant stem diameter (mm) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) At the initial stage (day 0), the stem diameter values ranged from 2.12 to 3.72 mm. The highest stem diameter was observed in the RSH-17 genotype (3.72 mm), which was significantly greater than that of all other genotypes. The RSH-6 (3.22 mm), RSH-2 (3.05 mm), and RSH-9 (3.05 mm) genotypes also exhibited higher values. In contrast, the scion (2.12 mm), RSH-16 (2.16 mm), and RSH-5 (2.18 mm) genotypes had the lowest stem diameters. These results revealed significant differences among genotypes in the early stages. It was determined that the stem diameter values increased to 3.98–4.76 mm on the 10th day. During this period, the RSH-17 genotype reached its highest value (4.76 mm). The RSH-6 (4.73 mm) and RSH-2 (4.72 mm) genotypes exhibited similar stem diameter values. The scion had the lowest stem diameter during this period (3.98 mm). The stem diameter values of the commercial rootstocks were determined as follows: Commercial-1 (4.30 mm) and Commercial-2 (4.18 mm). On day 20, the differences between the genotypes persisted, with the RSH-17 genotype maintaining its superiority (8.51 mm). The RSH-6 genotype also showed a high value (7.37 mm). Commercial rootstocks showed stem diameter values in the range of 6.46–6.52 mm. The highest stem diameter on day 30 was observed in the RSH-17 genotype (9.87 mm), which was statistically different from all other genotypes. The RSH-6 (9.53 mm) and RSH-2 (9.20 mm) genotypes were in the upper statistical group, followed by the RSH-9 (8.97 mm) and RSH-1 (8.87 mm) ones. Commercial rootstocks remained in the range of 8.07–8.37 mm during this period and lagged behind high-performing genotypes. The lowest stem diameters were observed in the scion (7.56 mm) and RSH-16 (7.67 mm) genotypes. Stem diameter, plant length, and dry matter content are important factors in determining seedling quality [43]. Stem diameter is a growth criterion directly related to basic physiological processes, such as plant mechanical strength and capacity to transport water and nutrients [44]. In a study examining the effects of grafted and ungrafted seedlings on plant development, yield, and quality, measurements taken on days 25 and 50 showed that grafted plants had higher values than the control group in terms of stem diameter, plant length, and number of leaves [45]. Similarly, Yarşi and Rad [46] found that the rootstock diameters of grafted plants were thicker than those of the control group, while Kurum [47] reported that the rootstock diameter in cucumbers varied depending on the rootstock genotype used. The stem diameter results of this study are consistent with those in the literature and show statistically significant differences between genotypes at all measurement times. Specifically, on days 20 and 30, the RSH-17 genotype had the highest stem diameter and was statistically distinct from the other genotypes, whereas the RSH-6 and RSH-2 genotypes were also ranked in the upper statistical group. These findings reveal that stem development varies significantly across rootstock genotypes and that rootstock candidates with strong stem structures can positively support plant vegetative development. Plant length was measured on days 0, 10, 20, and 30 for genotypes and commercial rootstocks, and statistically significant differences were observed between genotypes at all measurement times (P < 0.001). A steady increase in plant length was observed for all the genotypes (Fig. 6). Fig. 6Plant length (cm) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) Plant length (cm) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) At the start of the trial (Day 0), the plant length ranged from 10.40 to 16.75 cm. The highest initial plant length was observed in the RSH-17 genotype (16.75 cm), which was statistically significant. The RSH-6 (16.18 cm) and RSH-2 (14.94 cm) genotypes also showed high initial values, similar to those of RSH-17. In contrast, the scion had the lowest plant length (10.40 cm), followed by RSH-16 (11.54 cm) and RSH-5 (11.68 cm). On day 10, the plant length increased to 15.30–19.73 cm. During this period, the RSH-17 (19.73 cm) and RSH-6 (19.63 cm) genotypes attained the highest values. The RSH-2 genotype followed them at 19.30 cm. In contrast, the scion showed the lowest plant length during this period (15.30 cm). The plant length values of the commercial rootstocks were determined as follows: Commercial-2 (15.90 cm) and Commercial-1 (16.57 cm). On the 20th day, the measurements showed that the plant length continued to increase, and the differences between the genotypes became more noticeable. During this period, the RSH-17 genotype reached the highest plant length (32.93 cm). The RSH-6 genotype also showed a high value (31.27 cm). The RSH-2 (29.90 cm), RSH-9 (29.77 cm), and RSH-1 (29.57 cm) genotypes followed these. In contrast, the scion (23.33 cm) and RSH-16 (23.47 cm) genotypes had the shortest lengths. The plant length values for the commercial rootstocks remained in the range of 25.13–27.30 cm during this period. The highest plant length on the 30th day was observed in the RSH-17 genotype (51.67 cm), which was significantly different from all other genotypes. This was followed by RSH-6 (49.67 cm), RSH-2 (49.33 cm), RSH-1 (49.0 cm), and RSH-9 (49.0 cm). The plant length values of the commercial rootstocks ranged from 44.67 to 45.67 cm and were lower than those of the high-performing candidates. The lowest plant lengths were observed in the scion (40.33 cm) and RSH-16 (43.0 cm) genotypes. The effect of rootstock use on plant length has been demonstrated in several previous studies. Khah [48] reported that grafted plants had statistically higher plant lengths than control plants. den Nijs [49] noted that the positive effect on plant length and leaf formation in cucumber lines grafted onto low-temperature-tolerant rootstocks (Cucurbita ficifolia and Sicyos angulatus) varied according to the rootstock genotype. Similarly, Bletsos et al. [50] stated that the longer plant length and thicker stem structure observed in eggplant plants grafted onto Verticillium dahliae Kleb.-resistant rootstocks (S. torvum and S. sisymbriifolium) resulted in longer plant length and thicker stems in grafted eggplant plants, which was attributed to the stronger root systems of the rootstocks used. Oztekin and Tuzel [51] investigated the effects of different rootstocks on plant length under salt stress conditions in greenhouse tomato cultivation. They found that rootstock use generally increased plant length; however, the effect varied according to the rootstock genotype. The plant length results obtained in this study are consistent with those in the literature and show statistically significant differences between genotypes at all measurement times. In particular, the RSH-17 genotype exhibited the highest plant length values from the early stages and was statistically distinct from the other genotypes at 51.67 cm on day 30. The RSH-6 (49.67 cm) and RSH-2 (49.33 cm) genotypes also ranked in the upper group with high values for plant length. In contrast, the commercial rootstocks showed shorter plant lengths. These findings reveal that plant length development is largely dependent on the rootstock genotype and that rootstock candidates with strong root systems can better support shoot development [52, 53]. The dry weights of the genotypes and commercial rootstocks were determined on days 0, 10, 20, and 30. Statistically significant differences were found between the genotypes at all measurement times (day 0: P < 0.05; day 10: P < 0.01; days 20 and 30: P < 0.001). A steady increase in plant dry weight was observed across all genotypes throughout the trial period; however, the importance of this increase varied significantly among the genotypes (Fig. 7). Fig. 7Plant dry weight (g) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) Plant dry weight (g) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) At the beginning of the study (day 0), the plant dry weight ranged from 0.37 to 0.54 g. The highest initial dry weight was observed in the RSH-17 genotype (0.54 g), which was statistically the highest. The RSH-2 (0.48 g), RSH-6 (0.48 g), and RSH-1 (0.46 g) genotypes also showed high initial values, similar to RSH-17. In contrast, the scion had the lowest initial dry weight (0.37 g). On the 10th day, the plant dry weight values increased to 0.46–0.80 g. During this period, RSH-17 attained the highest dry weight (0.80 g), which was statistically different from that of the other genotypes. The RSH-6 genotype also showed a higher value (0.73 g) for this trait. The RSH-1, RSH-2, and RSH-9 genotypes were in the same statistical group, with similar dry weights (0.69 g each). The lowest values were observed in the scion (0.46 g) and RSH-16 (0.53 g). On the 20th day, the differences between the genotypes increased significantly. During this period, the RSH-17 genotype achieved the highest plant dry weight (6.01 g) and was statistically distinct from all other genotypes. The RSH-6 (4.19 g) and RSH-2 (4.07 g) genotypes followed this group, whereas the RSH-1 (3.97 g), RSH-9 (4.02 g), and RSH-10 (3.94 g) genotypes followed the RSH-6 and RSH-2 genotypes. In contrast, the scion (2.55 g) and RSH-16 (2.62 g) genotypes had the lowest dry weights. The dry weight values of the commercial rootstocks remained in the range of 2.98–3.47 g during this period. The highest plant dry weight on the 30th day was observed in the RSH-17 genotype (15.56 g). The RSH-6 genotype also showed a high value (13.63 g) for this trait. These were followed by RSH-2 (12.71 g), RSH-9 (11.73 g), and RSH-1 (11.72 g). In contrast, the scion (8.39 g) and RSH-16 (8.99 g) genotypes had the lowest dry weights. Commercial rootstocks remained in the range of 9.71–10.29 g during this period and were ranked behind all rootstock candidates. It has been reported that dry matter accumulates primarily in basic organs, such as roots and stems, during the early stages of predominantly vegetative growth [44]. The root dry weight results obtained in this study showed significant differences between genotypes as the development period progressed. Although no statistically significant difference was observed between genotypes in terms of root dry weight on day 0, these differences became significant on days 10 and 20 and were highly significant on day 30. In particular, the RSH-17 genotype exhibited the highest root dry weight at all measurement times. The RSH-6 genotype followed RSH-17 with a high root dry weight value (4.06 g), whereas the commercial rootstocks showed lower root dry weight values. Sarıbaş et al. [16] reported that rooting levels in eggplant rootstock candidates varied depending on the genotype and that the root dry weight of rootstocks was higher than that of the control plants. However, these researchers emphasized that the root dry weight parameter alone is insufficient for evaluation in rootstock breeding programs and must be considered in conjunction with other root architecture characteristics, such as root surface area, length, and volume. The findings of this study also support the need to interpret root dry weight in conjunction with other root parameters when evaluating a strong, functional root system, despite its importance as an indicator of root development. Root dry weight values for the genotypes and commercial rootstocks were determined at 0, 10, 20, and 30 days, and differences among genotypes varied with the measurement time. No statistically significant differences were detected among genotypes in root dry weight at day 0 (P > 0.05); however, differences became significant at days 10 and 20 (P < 0.05). By day 30, the differences were highly significant (P < 0.001). Throughout the experimental period, a consistent increase in root dry weight was observed in all genotypes; however, the magnitude of this increase varied substantially among the genotypes (Fig. 8). Fig. 8Root dry weight (g) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) Root dry weight (g) of interspecific tomato rootstock candidates and commercial rootstocks measured at 0, 10, 20, and 30 days after planting (Error bars represent ± SE. Means followed by different letters differ significantly according to Duncan’s multiple range test at p ≤ 0.05) At day 0, the root dry weight values ranged from 0.08 to 0.15 g, and no statistically significant differences were observed between genotypes. However, the RSH-17 genotype showed the highest initial value (0.15 g), followed by RSH-6 (0.14 g) and RSH-2 (0.14 g). The lowest root dry weight was observed in the scion (0.08 g). On day 10, the root dry weight values ranged from 0.16 to 0.31 g, and the differences between genotypes were significant. During this period, the RSH-17 genotype achieved the highest root dry weight (0.31 g). The RSH-6 genotype also showed a similar value (0.29 g) for the same parameter. In contrast, the scion had the lowest root dry weight (0.16 g). Commercial rootstocks were generally ranked in the middle and lower groups during this period (0.21–0.23 g). On day 20, the measurements showed that the increase in root dry weight continued, and the differences between the genotypes became more pronounced. The highest root dry weight values were observed in RSH-17 (1.53 g), RSH-6 (1.42 g), and RSH-2 (1.42 g). In contrast, the scion (0.83 g) had the lowest weight. The root dry weights of the commercial rootstocks remained in the range of 1.07–1.16 g during this period. Data obtained on day 30 showed that the RSH-17 (4.45 g) genotype achieved the highest root dry weight and was statistically different from all other genotypes. The RSH-6 genotype also showed a high root dry weight (4.06 g). This was followed by RSH-2 (3.07 g), RSH-1 (2.93 g), RSH-9 (2.93 g), and RSH-10 (2.92 g). The lowest root dry weights were observed in the scion (2.24 g) and RSH-16 (2.37 g). The root dry weights of the commercial rootstocks remained in the range of 2.52–2.70 g during this period. Performance of the MLP and machine learning models When the models’ overall performance was evaluated collectively, it was determined that all models achieved varying degrees of success in predicting root morphological characteristics (Table 1). Among the non-linear models, XGBoost demonstrated the highest overall prediction accuracy and showed low error values across many features. MLP and GPR models also showed strong predictive capacity, with high R² values; however, their performance varied across certain root features. The Random Forest model tended to show lower predictive performance than the other non-linear models. The linear model, OLR, generally produced results quite close to those of the non-linear models and even showed performance similar to or better than the most successful models for some parameters. This indicates that a linear structure can strongly explain certain relationships in the dataset and that OLR can be considered not merely a simple reference model but also a competitive predictive model. However, the general dominance of XGBoost may be explained by its capacity to provide a more flexible modeling approach in capturing complex, possibly non-linear relationships between root morphological traits and descriptive variables. Overall, the results indicate that non-linear models could offer more predictive power. Nevertheless, also in the context of this study, the OLR model provides a robust benchmark. Table 1Prediction performance of the models for root length, root surface area, root diameter and root volumeParametersPredictive ModelsR2RMSEMAETrainTestTrainTestTrainTestRoot Length (cm)XGBoost0.990.9495.65255.6671.34189.74GPR0.980.91149.73313.14117.24261.48Random Forest0.890.82319.75476.33259.84398.48MLP0.980.93112.56287.5087.81211.75OLR0.960.92183.27265.93149.09209.14Root Surface Area (cm2)XGBoost0.990.9523.8461.3117.1748.44GPR0.930.9063.2880.7652.1566.83Random Forest0.900.8280.28114.7265.8599.07MLP0.980.9227.6377.7420.1757.78OLR0.960.9050.0476.7139.5956.44Root Diameter (mm)XGBoost0.960.680.020.070.020.04GPR0.930.610.040.090.040.06Random Forest0.850.490.050.090.040.06MLP0.960.580.020.080.020.05OLR0.910.700.040.060.030.04Root Volume (cm3)XGBoost0.990.950.551.590.420.98GPR0.950.921.511.871.241.53Random Forest0.860.822.133.081.922.24MLP0.980.780.733.430.542.77OLR0.980.950.891.380.671.09 Table 1Prediction performance of the models for root length, root surface area, root diameter and root volumeParametersPredictive ModelsR2RMSEMAETrainTestTrainTestTrainTestRoot Length (cm)XGBoost0.990.9495.65255.6671.34189.74GPR0.980.91149.73313.14117.24261.48Random Forest0.890.82319.75476.33259.84398.48MLP0.980.93112.56287.5087.81211.75OLR0.960.92183.27265.93149.09209.14Root Surface Area (cm2)XGBoost0.990.9523.8461.3117.1748.44GPR0.930.9063.2880.7652.1566.83Random Forest0.900.8280.28114.7265.8599.07MLP0.980.9227.6377.7420.1757.78OLR0.960.9050.0476.7139.5956.44Root Diameter (mm)XGBoost0.960.680.020.070.020.04GPR0.930.610.040.090.040.06Random Forest0.850.490.050.090.040.06MLP0.960.580.020.080.020.05OLR0.910.700.040.060.030.04Root Volume (cm3)XGBoost0.990.950.551.590.420.98GPR0.950.921.511.871.241.53Random Forest0.860.822.133.081.922.24MLP0.980.780.733.430.542.77OLR0.980.950.891.380.671.09 Prediction performance of the models for root length, root surface area, root diameter and root volume Root length prediction performance The XGBoost model achieved the highest test performance for root length, with R² = 0.94, RMSE = 255.66, and MAE = 189.74 (Fig. 9). The MLP model also demonstrated a similarly high prediction success rate, achieving an R² of 0.93 on the test set. The OLR model, with an R² of 0.92, demonstrated that the linear relationship is also quite strong. In contrast, the Random Forest model performed worse than the other models, with an R² of 0.82 on the test set. In their study, Akyüz and Cemek [54] found that XGBoost outperformed the MLP and GPR models, achieving the highest R2 value of 0.90 in predicting the root characteristics of aronia plants subjected to salt stress. In addition, Şimşek et al. [55] reported that XGBoost also performed better than other models for predicting root length (R² = 0.87) in a lavender micropropagation study. MLP model by Tütüncü et al. [56] with R² value of 0.99 was found as the most reliable model for predicting root length in myrtle plants under cadmium stress. In another aronia study, the RF model (R2 = 0.87) had better performance than the MLP and SVM models in predicting root length [57]. Fig. 9Comparison of observed and predicted root length values using OLR and various machine learning models (XGBoost, GPR, Random Forest, and MLP) (The solid line represents the 1:1 identity line (y = x), serving as a reference for ideal model performance. All data points shown in the plot are derived exclusively from the independent test dataset) Comparison of observed and predicted root length values using OLR and various machine learning models (XGBoost, GPR, Random Forest, and MLP) (The solid line represents the 1:1 identity line (y = x), serving as a reference for ideal model performance. All data points shown in the plot are derived exclusively from the independent test dataset) Root surface area prediction performance Among the models to predict root surface area, the XGBoost model again demonstrated the highest test performance (Fig. 10). The XGBoost model was the most successful, achieving R² = 0.95, RMSE = 61.31, and MAE = 48.44 on the test set. The MLP model ranked second with R² = 0.92, while the GPR and OLR models performed similarly on the test set. For both models, the test R² was calculated as 0.90. The Random Forest model, however, demonstrated more limited success in predicting Root Surface Area, achieving an R² value of 0.82 on the test set. Similarly, in two independent studies on aronia, Akyüz and Cemek [54] and Yaman et al. [57] reported that XGBoost outperformed the other predictive models. Also, Okumuş et al. [58] found that XGBoost minimized error margins when predicting the effects of environmental stresses such as salt stress on root morphology. These findings indicate that tree-based models are more reliable for predicting biological variations in root morphological traits, such as surface area. However, the relatively strong performance of the OLR model implies that the root surface area may be easier to learn or more linear than the root length. Fig. 10Comparison of observed and predicted root surface area values using OLR and various machine learning models (XGBoost, GPR, Random Forest, and MLP) (The solid line represents the 1:1 identity line (y = x), serving as a reference for ideal model performance. All data points shown in the plot are derived exclusively from the independent test dataset) Comparison of observed and predicted root surface area values using OLR and various machine learning models (XGBoost, GPR, Random Forest, and MLP) (The solid line represents the 1:1 identity line (y = x), serving as a reference for ideal model performance. All data points shown in the plot are derived exclusively from the independent test dataset) Root diameter prediction performance The models’ performance in predicting root diameter was lower than for other root characteristics (Fig. 11). The highest test R² value for this parameter was obtained with the OLR model, which demonstrated the best overall performance (R² = 0.70, RMSE = 0.06, MAE = 0.04). The XGBoost model demonstrated prediction accuracy close to that of the OLR model, with an R² of 0.68 on the test set. In contrast, the test R² values for the GPR, MLP, and Random Forest models were 0.61, 0.58, and 0.49, respectively. These results indicate that the root diameter variable can be adequately explained using linear models and that complex machine learning models do not provide a significant advantage for this parameter. Duarte et al. [59] reported that none of the models was effective in predicting average root diameter in soybean with R2 of 0.10 for the linear model, RF, and ANN, 0.07 for SVM, and 0.08 for XGBoost. The authors explained this result by the failure of the input variables to explain variation in average root diameter adequately, and by the weak relationships between root diameter and the other variables. Similarly, Akyüz and Cemek [54] found higher prediction errors for traits such as average root diameter and root volume in aronia plants. Furthermore, the relatively low correlation between mean root diameter and plant biomass needs further investigation. The diameter was averaged across the entire root system, and the high number of fine roots may have masked the strong effect of thicker, structural roots on biomass. Fig. 11Comparison of observed and predicted root diameter values using OLR and various machine learning models (XGBoost, GPR, Random Forest, and MLP) (The solid line represents the 1:1 identity line (y = x), serving as a reference for ideal model performance. All data points shown in the plot are derived exclusively from the independent test dataset) Comparison of observed and predicted root diameter values using OLR and various machine learning models (XGBoost, GPR, Random Forest, and MLP) (The solid line represents the 1:1 identity line (y = x), serving as a reference for ideal model performance. All data points shown in the plot are derived exclusively from the independent test dataset) Root volume prediction performance The XGBoost and OLR models achieved the highest test performance in root volume prediction (Fig. 12). Both models achieved an R² value of 0.95 on the test set. However, when error metrics are considered, the OLR model has lower RMSE (1.38) than the XGBoost model, while the XGBoost model has lower mean absolute error (MAE) (0.98). The GPR model also demonstrated high predictive accuracy (R² = 0.92). The test R² values for the Random Forest and MLP models were 0.82 and 0.78, respectively, indicating that these two models performed worse than the others in root volume prediction. Duarte et al. [59] demonstrated that root volume was a significant factor in modeling root architecture in soybeans and emphasized the usefulness of machine learning models for estimating root-related traits. On the other hand, some error rates were higher for root volume prediction than other parameters in the aronia study of Akyüz and Cemek [54]. The good performance of the test for root volume (R2 = 0.95) in the present study indicates that root volume was well associated with the explanatory variables in the dataset and that both XGBoost and OLR were able to capture this relationship. Fig. 12Comparison of observed and predicted root volume values using OLR and various machine learning models (XGBoost, GPR, Random Forest, and MLP) (The solid line represents the 1:1 identity line (y=x), serving as a reference for ideal model performance. All data points shown in the plot are derived exclusively from the independent test dataset) Comparison of observed and predicted root volume values using OLR and various machine learning models (XGBoost, GPR, Random Forest, and MLP) (The solid line represents the 1:1 identity line (y=x), serving as a reference for ideal model performance. All data points shown in the plot are derived exclusively from the independent test dataset) In summary, the current findings indicate that root morphological traits can be well predicted by both machine learning and standard linear regression models. The high R² values for root length, root surface area, and root volume indicate that these parameters are strongly related to other morphological traits in the dataset, and that XGBoost captured these relationships particularly well. On the contrary, the lower prediction accuracy for root diameter may suggest that this trait is less variable, less dependent or more sensitive to measurement precision than other root traits. It is noteworthy that the present study was not conducted under abiotic or biotic stress conditions. Thus, the relationships among root parameters might have followed a more regular, developmental pattern in time, depending on the sampling day. The PCA results supported this interpretation, showing a progressive separation of the samples along PC1 by sampling time, indicating a strong directional growth gradient. This structured developmental pattern might explain why linear regression models successfully predict root traits, especially Root Length, Root Surface Area and Root Volume. Under stress conditions, additional factors such as genotype-specific stress responses, physiological damage, growth inhibition and adaptive mechanisms can lead to more complex, non-linear relationships among root traits. The good performance of linear regression in the present study may thus be related to the controlled, non-stressful growth conditions in which root development proceeded in a rather regular manner over time. Hence, the performance of the model in root architecture studies should be evaluated not only in terms of overall prediction accuracy but also considering the biological and statistical nature of each root trait and the experimental context in which the data were generated. As the XGBoost model showed better prediction results it was analyzed for variable importance for root morphological traits (Table 2). The variable importance analysis indicated that plant length and stem diameter were the main predictors of root morphology traits, whereas genotype contributed little. Plant length was the most important variable for root length, root surface area, and root diameter, with importance values of 0.502, 0.603, and 0.649, respectively. Stem diameter was the most important predictor of root volume, with an importance value of 0.504, closely followed by plant length. These results show that vegetative growth parameters are highly correlated with root system development. The low importance values for genotype suggest that genotypic differences contributed less to predicting root traits than morphological variables related to plant growth in the present dataset. This could also be caused by the fact that they are close relatives from the same hybridization study. However, these findings need to be regarded as model-based associations rather than causal relationships. Table 2Variable importance for root morphological traits predicted by XGBoost (Gain values)Input VariableRoot LengthRoot Surface AreaRoot VolumeRoot DiameterGenotype0.0150.0130.0040.061Plant Length0.5020.6030.4910.649Stem Diameter0.4820.3830.5040.289 Table 2Variable importance for root morphological traits predicted by XGBoost (Gain values)Input VariableRoot LengthRoot Surface AreaRoot VolumeRoot DiameterGenotype0.0150.0130.0040.061Plant Length0.5020.6030.4910.649Stem Diameter0.4820.3830.5040.289 Variable importance for root morphological traits predicted by XGBoost (Gain values) Principal component analysis indicated a clear temporal separation among sampling days for root morphology traits. The first and second principal components explained 91.5 and 5.3% of the total variation, respectively, implying that the first two components explained ~ 96.8% of the total variability (Fig. 13). In similar multivariate analyses of crop traits, two components often capture approximately 70–80% of the variation [60]. Samples collected at 0, 10, 20, and 30 days were progressively distributed along PC1, indicating a strong time-dependent developmental gradient. In general, the 0-day samples were located on the negative side of the PC1. In contrast, the 30 days samples were placed on the positive side, that is, higher values for root and shoot growth traits at later sampling. Generally, the loading vectors of root length, root surface area, root volume, plant length, stem diameter, and dry weight-related traits were positively correlated with PC1, indicating that these variables played a strong role in the temporal differentiation among the samples. This pattern implies a tight coupling between root system development and vegetative growth during the experimental period. Genotype-level variation was observed within each sampling day; however, the overall clustering pattern was largely influenced by sampling time rather than genotype. This result is consistent with the variable importance analysis, which showed that plant length and stem diameter were the main predictors of root morphological traits. In contrast, the contribution of genotype was relatively low. Therefore, the PCA results support the conclusion that root morphological development is primarily related to plant growth progression over time. Fig. 13Principal component analysis (PCA) biplot showing the relationships among root and shoot morphological traits in tomato genotypes. RL: Root Length; RSA: Root Surface Area; RD: Root Diameter; RV: Root Volume; PSD: Plant Stem Diameter; PL: Plant Length; PDW: Plant Dry Weight; RDW: Root Dry Weight, R: RSH, C: Commercial, and S: Scion Principal component analysis (PCA) biplot showing the relationships among root and shoot morphological traits in tomato genotypes. RL: Root Length; RSA: Root Surface Area; RD: Root Diameter; RV: Root Volume; PSD: Plant Stem Diameter; PL: Plant Length; PDW: Plant Dry Weight; RDW: Root Dry Weight, R: RSH, C: Commercial, and S: Scion Correlation analysis revealed strong and predominantly positive relationships among the measured root and shoot traits (Fig. 14). Root length showed very high correlations with root surface area (r = 0.99), root volume (r = 0.97), and plant stem diameter (r = 0.97), indicating that increases in root system extension were closely associated with overall root size and shoot structure development. Similarly, root surface area was strongly correlated with root volume (r = 0.96) and plant stem diameter (r = 0.97), further confirming the integrated nature of the root morphological traits. Plant length exhibited high positive correlations with both root- and shoot-related parameters (r = 0.91–0.97). Biomass-related traits, including plant dry weight and root dry weight, were also highly correlated with root volume (r = 0.93–0.95) and plant length (r = 0.97), reflecting the close linkage between root system development and biomass accumulation. Plants with thicker stems often have thicker and more voluminous roots, which may reflect a common developmental or genetic control over the overall robustness of the plant structure. These positive correlations are consistent with the fundamental principles of plant growth and have been widely reported. For example, Chen et al. [61] found that root dry weight was significantly and positively correlated with plant length and crop dry weight in wheat. In contrast, root diameter showed comparatively lower, although still positive, correlations with other traits (r = 0.68–0.81), indicating that root thickness varied more independently than root length, surface area, or volume. It is important to note that the correlations reported in this study were observed in relatively young plants under controlled (non-stressed) conditions. The root-to-shoot ratio typically varies as plants progress through various developmental stages, and environmental stressors can vary these physiological relationships [62]. Thus, the correlations observed may differ in mature plants or under field conditions, although these data are a useful baseline for early-stage selection. Furthermore, the relatively low correlation between mean root diameter and plant biomass needs further investigation. The diameter was averaged across the entire root system, and the high number of fine roots may have masked the strong effect of thicker, structural roots on biomass. Fig. 14Pearson correlation heatmap showing the relationships among root and shoot morphological traits in tomato plants. RL: Root Length; RSA: Root Surface Area; RD: Root Diameter; RV: Root Volume; PSD: Plant Stem Diameter; PL: Plant Length; PDW: Plant Dry Weight; RDW: Root Dry Weight Pearson correlation heatmap showing the relationships among root and shoot morphological traits in tomato plants. RL: Root Length; RSA: Root Surface Area; RD: Root Diameter; RV: Root Volume; PSD: Plant Stem Diameter; PL: Plant Length; PDW: Plant Dry Weight; RDW: Root Dry Weight Conclusion This study shows that root morphological traits of developed interspecific tomato rootstock candidates can be predicted with high accuracy by both machine learning and linear modelling approaches during the early stage of plant development. Regarding the non-linear models, the XGBoost model demonstrated the best performance in terms of overall predictive performance, thus corroborating its capability to capture complex relationships between vigor-related growth parameters and root architecture. However, the OLR model also showed similar results for some root traits. This similarity can be explained by the strong correlations between plant length, stem diameter, biomass related parameters and root morphological traits under the experimental conditions. The study was performed in optimal greenhouse conditions, without biotic or abiotic stress, which resulted in a relatively coordinated shoot growth and root development, enabling both linear and non-linear models to explain well the observed variation. The study was limited to the first 30 days of plant development corresponding to the early developmental stage of tomato rootstock candidates. During this period genotypic differences in root architecture may be less pronounced than at later growth stages or in stress conditions. However, results indicated that some interspecific rootstock candidates had promising root growth patterns and may have value for rootstock breeding programs. Our results also support the use of easily measurable morphological traits as indirect predictors of root architecture, especially when direct root phenotyping is destructive, labor intensive or unsuitable for large scale selection. Consequently, the implementation of these machine learning models offers a non-destructive pathway for root phenotyping. The present study has a limitation that the experiment was performed under optimum growing conditions and did not involve any abiotic or biotic stress treatments. Hence, the predictive relationships found in this study should be interpreted mainly in the context of early growth conditions without stress. Another limitation is the lack of phenological records and Growing Degree Days (GDD), which could have provided a more accurate developmental basis for genotypic comparison. Future studies should validate these rootstock candidates under different stress conditions like drought, salinity, nutrient limitation and soil-borne disease pressure over longer developmental periods. Moreover, the incorporation of GDD and detailed phenological observations in future modeling frameworks would improve the biological interpretation of genotype performance and increase the reliability of rootstock selection under variable environmental conditions. Supplementary Information Supplementary Material 1. Supplementary Material 1.