Deep learning-based spectral classification of spatial twins in three cotton breeding lines

Authors: RAM, BILLY - North Dakota State University; ZHANG, YU - North Dakota State University; VILLAMIL-MAHECHA, MARIA - North Dakota State University; McCarty Jr, Jack; Jenkins, Johnie; SUN, XIN - North Dakota State University

Submitted to: Journal of Agriculture and Food Research
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/13/2025
Publication Date: 10/16/2025
Citation: Ram, B.G., Zhang, Y., Villamil-Mahecha, M., Mccarty Jr, J.C., Jenkins, J.N., Sun, X. 2025. Deep learning-based spectral classification of spatial twins in three cotton breeding lines. Journal of Agriculture and Food Research. 24(2025):102454. https://doi.org/10.1016/j.jafr.2025.102454.
DOI: https://doi.org/10.1016/j.jafr.2025.102454

Interpretive Summary: The rising popularity of cotton as a personal clothing material has made fire resistance testing crucial. Present methods of fire-resistant testing in cotton involve destructive open flame and shouldering tests of cotton fabric textiles. For a successful breeding program geared to development of cotton varieties that can be made into fire retardant cloth textiles a method of selecting plants based on lint properties of individual plants is highly desirable. Currently none exists. However, current advancements in computer vision and sensor technologies may offer solutions. This study explores the efficiency of hyperspectral sensors for spectral classification of three breeding lines of cotton fibers, a cultivar, a fire resistant line and a fire susceptible line. Spectral data was collected in two ranges of 300–1000 nm and 900–1700 nm. Neural architectural search and hyperparameter tuning are two emerging areas of research in deep learning that facilitate the optimum deep learning model architecture and learning parameters by streamlining trial and error. A one-dimensional convolutional neural network served as the foundational model in our study. Utilizing this, various architectural designs, and parameters to determine the optimal model for cotton classification were explored. This systematic approach allowed fine-tuning of model to achieve superior performance in cotton classification tasks. Base model and hyper-tuned model were trained using both wavelength ranges and the performance of hyper-tuned model in the range of 900–1700 nm was found to be considerably better than the base model achieving a test accuracy of 96% compared to 83% accuracy of base model for identifying the three lines of cotton in this study. This study offers an innovative approach that should be explored with a larger set of plants available in a breeding population that is segregating for the fire retardant trait.

Technical Abstract: The increasing popularity of cotton as a personal clothing material has made fire resistance testing of cotton fibers crucial. Current methods for testing fire resistance in cotton textiles involve destructive open flame and smoldering tests. Due to advancements in spectroscopic sensor technologies and deep learning (DL) frameworks, there is a need for non-destructive flame resistance testing methods. This study investigates the effectiveness of hyperspectral sensors for spectral classification of lint from three breeding lines of cotton fibers: SG 747 (a commercial cultivar), RIL 385 (a fire-resistant breeding line), and RIL 225 (a fire-susceptible breeding line). Spectral data was collected in the visible near-infrared range (300–1000 nm) and the shortwave infrared region (900–1700 nm). A one-dimensional convolutional neural network (1D-CNN) was the foundational model in this study. Machine learning frameworks, including neural architectural search, hyperparameter tuning, and distributed learning, were applied to select an optimized DL model architecture and parameters. This systematic approach enabled fine-tuning of the model to achieve superior performance in cotton classification tasks. The performance of the hyper-tuned model in the 900–1700 nm range was found to be higher than that of the base model, achieving a test accuracy of 86% compared to the base model's 84% accuracy. This study demonstrates the potential for non-destructive fire resistance testing in cotton fibers with the successful implementation of deep learning modeling for spectral classification of spatial twins.