Agricultural scientists evaluating colored cotton cultivars have demonstrated that advanced machine learning architectures can predict leaf area nondestructively with greater precision than conventional allometric equations. Conducting field trials at the didactic garden of the Federal Rural University of the Semi-Arid in Mossoró, Rio Grande do Norte, Brazil, the research team analyzed how simple, non-invasive linear dimensions can replace destructive harvesting techniques. Leaf surface area constitutes an essential physiological index for monitoring vegetative vigor, photosynthetic interception, transpiration dynamics, and overall crop productivity throughout developmental phases. In traditional agronomic assessments, calculating leaf area requires detaching leaves for laboratory scanning or planimeter measurements, a laborious practice that inherently destroys plant tissue and prevents repetitive evaluations of individual specimens across their full growth cycle. To overcome these historical limitations, the researchers formulated and rigorously evaluated mathematical modeling strategies capable of estimating the canopy surface of five distinct cultivars of naturally colored cotton using only maximum leaf length and maximum leaf width.
Naturally colored cotton, classified botanically as Gossypium hirsutum, represents an environmentally strategic crop that provides naturally pigmented fibers in shades such as green and brown, thereby bypassing chemical dyeing processes and mitigating effluent pollution in industrial textile processing. Despite growing commercial interest and environmental value, particularly across semiarid cultivation zones, accurate phenotyping protocols have lacked calibrated machine learning frameworks specifically tuned for colored cultivars. The experimental investigation took place between November 2024 and January 2025 within an area characterized by a BSh-type semiarid climate under the Köppen-Geiger classification and containing Eutrophic Red-Yellow Argisol soil. Meteorological monitoring during this cultivation window documented a mean ambient air temperature of 26.8 degrees Celsius, total accumulated rainfall of 147.8 millimeters, and an average atmospheric relative humidity of 67.0 percent. The study evaluated five registered Brazilian colored cotton cultivars obtained from the university active germplasm bank: BRS Safira, BRS Rubi, BRS Topázio, BRS Jade, and BRS Verde, establishing three independent fifteen-meter experimental rows per cultivar with drip irrigation supplying 9.12 millimeters of water per plant daily.
Eighty-five days after initial planting, the investigators conducted an extensive sampling procedure, collecting 1,334 healthy and fully expanded leaves evenly distributed across the lower, middle, and upper thirds of the plant canopy to capture morphological diversity across maturity levels. The collected cohort comprised 231 leaves from BRS Safira, 340 from BRS Rubi, 266 from BRS Topázio, 288 from BRS Jade, and 209 from BRS Verde. Transported in isothermal containers to prevent desiccation, all sampled leaves were digitized alongside a calibrated one-square-centimeter reference scale using a flatbed optical scanner. Subsequent image processing via the open-source software ImageJ generated empirical benchmark values for maximum leaf length, maximum leaf width, and observed total leaf area. Morphometric analysis demonstrated pronounced structural heterogeneity across the sampled foliage, with leaf lengths spanning from 1.429 centimeters in BRS Jade to 8.835 centimeters in BRS Topázio, and widths extending from 1.194 centimeters to 7.606 centimeters. Cultivar mean lengths ranged between 4.438 centimeters for BRS Safira and 4.973 centimeters for BRS Verde, while mean widths ranged from 3.767 centimeters to 4.180 centimeters across the same respective cultivars.
The distribution of individual leaf surface areas demonstrated substantial dispersion, ranging from a minimum of 1.010 square centimeters in BRS Safira to an observed maximum of 34.427 square centimeters in BRS Rubi, reflecting cultivar-specific canopy architecture. Average leaf areas per cultivar extended from 10.409 square centimeters in BRS Safira to 12.483 square centimeters in BRS Topázio, accompanied by high coefficients of variation reaching between 45.7 percent and 56.8 percent across the cultivars. Frequency classification revealed that over eighty percent of all sampled leaves across the five cultivars occupied small to intermediate surface area classes between 1.01 and 20.00 square centimeters. Specifically, small leaves measuring between 1.01 and 10.00 square centimeters comprised 50.29 percent of BRS Rubi, 49.35 percent of BRS Safira, 45.83 percent of BRS Jade, and 41.62 percent of BRS Verde foliage. Conversely, intermediate leaves measuring between 10.01 and 20.00 square centimeters were dominated by BRS Topázio at 51.87 percent, while large foliage exceeding 20.01 square centimeters remained comparatively rare, with BRS Verde exhibiting the highest proportion at 16.26 percent.
To evaluate predictive capabilities, the researchers partitioned the aggregate dataset of 1,334 leaf observations, assigning seventy percent of the records to model calibration and reserving thirty percent exclusively for independent validation testing. The modeling architecture juxtaposed classical allometric linear regression against three sophisticated machine learning paradigms: support vector regression, the adaptive neuro-fuzzy inference system, and a deep multilayer perceptron neural network. Model accuracy was systematically quantified through statistical criteria comprising the coefficient of determination, root mean square error, mean absolute error, mean absolute percentage error, and Willmott index of agreement. In linear regression modeling, allometric formulations evaluated ten mathematical iterations incorporating linear dimensions, additive sums, exponentiations, and dimensional products while testing multicollinearity through variance inflation factors and tolerance values. Although collinearity diagnostics confirmed that length and width could be integrated without confounding variance inflation, and residual autocorrelation was ruled out via Durbin-Watson statistics ranging between 1.8967 and 2.0506, Kolmogorov-Smirnov testing showed that linear model residuals deviated from normal distributions.
Among the empirical linear regressions, the allometric equation formulated from the direct product of leaf length and leaf width, designated as model LRM-4, achieved the strongest performance within the linear class during validation. This bivariate formulation attained an independent test coefficient of determination of 0.9783, a Willmott index of agreement of 0.9944, a root mean square error of 0.9213, a mean absolute error of 0.6558, and a mean absolute percentage error of 6.9263 percent. Conversely, linear iterations relying upon quadratic transformations or square root extractions demonstrated inferior accuracy, with the model based on squared products yielding a test determination coefficient of only 0.8800 alongside elevated residual dispersion. Furthermore, three-dimensional response surface plots highlighted irregular and segmented prediction surfaces generated by the linear regression equations, indicating that rigid parametric structures struggled to accommodate the biological nuances and subtle contour irregularities characterizing leaf margins across diverse colored cotton genotypes.
In contrast, supervised machine learning algorithms exhibited superior flexibility when processing nonlinear interactions between foliar dimensions and observed surface areas. The adaptive neuro-fuzzy inference system, implemented within MATLAB utilizing a first-order Takagi-Sugeno configuration and hybrid gradient descent with least squares optimization, evaluated four distinct membership function geometries. The configuration utilizing triangular membership functions, designated ANFIS-1, outperformed Gaussian, trapezoidal, and spline alternatives, recording a validation coefficient of determination of 0.9793 and training error metrics including a mean absolute error of 0.5579 and mean absolute percentage error of 5.2327 percent. Similarly, the deep multilayer perceptron architecture, implemented through TensorFlow using rectified linear units across two hidden layers, was evaluated across five structural configurations. The optimized network architecture possessing sixteen artificial neurons in each hidden layer, configured as a 2-16-16-1 system, achieved an identical validation determination coefficient of 0.9793, an agreement index of 0.9947, a reduced test root mean square error of 0.8990, and a test mean absolute percentage error of 5.6627 percent.
The highest overall predictive precision emerged from support vector regression utilizing a nonlinear radial basis function kernel, implemented in Python via the Scikit-learn computational library. While an alternative linear kernel support vector model produced satisfactory metrics, the radial basis function formulation achieved superior generalization, securing an initial training determination coefficient of 0.9816 and maintaining a test coefficient of determination of 0.9790 alongside minimal residual spread. Standardized paired Student t-tests conducted between observed leaf areas and model estimations confirmed absence of statistically significant differences across all evaluated methods, yielding probability values consistently greater than or equal to 0.09. Visual inspection of residual scatter plots demonstrated homoscedasticity without systematic error drift across the range of leaf dimensions. Furthermore, contour maps and three-dimensional response surfaces confirmed that the radial basis function support vector algorithm produced exceptionally smooth, continuous mathematical topologies that closely mirrored observed empirical morphology across all five colored cotton varieties.
To establish an objective hierarchical classification among competing algorithms, the authors computed the standardized ranking performance index developed by Aschonitis, which integrates bounded agreement metrics and unbounded residual error magnitudes into a normalized scale ranging from zero to one. Under this rigorous multi-criteria ranking, the support vector regression model featuring the radial basis function kernel achieved an undisputed score of 1.0000, ranking first among all examined computational approaches. Deep multilayer perceptron architectures secured the subsequent positions, with the 2-16-16-1 configuration attaining a ranking index of 0.9973 to place second, followed by a sixteen-and-eight neuron configuration ranking third with an index of 0.9924. While the premier linear regression model achieved respectable statistical agreement, its overall standardized ranking placed seventh, trailing behind the nonlinear artificial intelligence frameworks due to elevated percentage errors and mathematical inflexibility when confronted with extreme foliar geometries.
To translate these experimental findings into practical agronomic management, the researchers deployed the trained radial basis function support vector regression model within an interactive, publicly accessible web application hosted on cloud infrastructure. Built with a Python Flask framework and responsive frontend styling, the software interface allows farmers, agronomists, and plant breeders to input field-measured leaf length and width obtained with a simple ruler without detaching foliage from living plants. The platform incorporates input validation thresholds aligned with experimental boundaries, dynamically warning users when submitted dimensions exceed calibrated ranges between 1.429 and 8.835 centimeters for length or 1.134 and 7.606 centimeters for width. Immediate real-time execution outputs estimated single-leaf area, tracks measurement logs, and aggregates cumulative canopy values. While noting that future validation under differing environmental conditions and additional cultivars will expand operational scope, the authors emphasize that deploying machine learning models into accessible digital interfaces delivers an efficient, cost-effective tool for nondestructive crop phenotyping and precision irrigation management.
Subject of Research: Nondestructive leaf area prediction in colored cotton cultivars using machine learning algorithms
Article Title: Nondestructive prediction of leaf area in colored cotton cultivars: a comparative approach using machine learning models with an interactive web interface
Article References: da SILVA RIBEIRO, J. E., da SILVA, A. G. C., de ALMEIDA OLIVEIRA, P. H., dos SANTOS COÊLHO, E., FAGUNDES, R. I., MELO, R. S., de SOUZA MORAIS, C. D., SANTOS, D. M., da SILVEIRA, L. M., & BARROS JÚNIOR, A. P. (2026). Nondestructive prediction of leaf area in colored cotton cultivars: a comparative approach using machine learning models with an interactive web interface. Journal of Cotton Research, 9(1), Article 11. https://doi.org/10.1186/s42397-025-00256-8
Image Credits: AI Generated
DOI: 10.1186/s42397-025-00256-8
Keywords: colored cotton, Gossypium hirsutum, leaf area prediction, support vector regression, adaptive neuro-fuzzy inference system, deep multilayer perceptron, linear regression, nondestructive phenotyping, radial basis function, plant morphometry, precision agriculture, crop canopy analysis, digital image processing, ImageJ, Flask web application, semiarid agronomy, standardized ranking performance index
Cite Scienmag News
Teresa Odom. (November 21, 2026). Machine Learning Models Accurately Predict Colored Cotton Leaf Area Without Plant Damage. Scienmag. https://scienmag.com/machine-learning-models-accurately-predict-colored-cotton-leaf-area-without-plant-damage/
Teresa Odom. "Machine Learning Models Accurately Predict Colored Cotton Leaf Area Without Plant Damage." Scienmag, 21 November 2026, https://scienmag.com/machine-learning-models-accurately-predict-colored-cotton-leaf-area-without-plant-damage/. Accessed 21 November 2026.
Teresa Odom. "Machine Learning Models Accurately Predict Colored Cotton Leaf Area Without Plant Damage." Scienmag. November 21, 2026. https://scienmag.com/machine-learning-models-accurately-predict-colored-cotton-leaf-area-without-plant-damage/








