Abstract:
Cotton species determination has major implications for agriculture, textile manufacturing, and
luxury product design. A brief yet deployable deep learning pipeline is proposed for accurately classifying
four cotton species of commercial importance: G. arboreum, G. barbadense, G. herbaceum, and G. hirsu
tum. A curated dataset of 20,850 high-resolution images collected from public repositories and controlled
photography was standardized and further augmented to promote generalization. For fiber texture, boll
morphology, and vein patterns, the model utilizes ResNet50 in conjunction with a Convolutional Block At
tention Module (CBAM) to emphasize spatial and channel-wise features. Transfer learning with ImageNet
weights, coupled with suitable regularization, established robust performance with a Top-1 test accuracy
of 94.2% and a macro F1-score of 0.938. Grad-CAM visualizations confirm that the model attended to
discriminative fiber regions, enhancing interpretability. A mobile-friendly Streamlit web application delivers
real-time species predictions with an average latency of 0.84 seconds and maintains 93.4% accuracy under
f
ield trials. Predicted species map to luxury product recommendations via a transparent rule-based mapper
that synthesizes species identity and an image-derived quality score to suggest product categories (e.g., pre
mium apparel, high-end home textiles, artisanal linens) with brief reasoning behind each recommendation.
This integrated pipeline connects farmers, agronomists, and textile professionals to translate species-level
identification into informed product decisions, quality assurance, and value-added marketing. The study
demonstrates that attention-augmented residual networks constitute an effective solution for fine-grained
agricultural image classification and its real-world industry implementation.
