Comprehensive Image Analysis of Seed and Plant for Classification of Cotton Genotypes Using Deep Learning Methodologies

This study presents a novel deep learning based framework for the classification of cotton genotypes using three uniquely curated image datasets that differ in complexity and biological relevance: Dataset I contains high-resolution non-overlapping seed images; Dataset II comprises overlapping seed images that simulate realistic seed handling conditions; and Dataset III includes whole-plant images captured under field conditions across multiple growth stages. The integration of both seed- and plant-level datasets enables a comprehensive evaluation of genotype recognition at different phenological stages, an approach not reported in prior cotton studies. Two convolutional neural network architectures, AlexNet and ResNet, were systematically evaluated following standardized preprocessing protocols, targeted data augmentation, and class balancing to improve model generalization. ResNet consistently demonstrated superior performance over AlexNet, achieving high classification accuracy across all datasets, with F1 scores of 0.98 for both Dataset I and Dataset III, and 0.97 for Dataset II, even under the challenging conditions posed by overlapping seed images. The results demonstrate that residual learning enhances feature extraction of subtle genotype-specific traits such as seed coat texture, hilum orientation, and leaf morphology. This work provides the first benchmark for cotton genotype classification across both seed and plant imagery in the Indian context and establishes a reproducible pipeline suitable for deployment in breeding programs, germplasm management, and intellectual property protection.