Cotton seedling monitoring and growth stage classification integrating deep learning and feature engineering

Early monitoring of crop development is crucial for precision agriculture, particularly for detecting acute stress and ensuring maximum yield potential. This study introduces a robust, high-throughput phenotyping framework that combines deep learning with biologically grounded feature engineering to classify cotton seedling growth stages and detect early stress signals in real-world field conditions. At its core is SeedlingNet, a lightweight residual CNN trained on a rigorously curated dataset of high-resolution UAV and ground-level images. We extract handcrafted phenotypic features like area, greenness, solidity, and texture and merge them into a Composite Stress Index (CSI) that integrates multivariate stress signals in order to improve biological interpretability. By using clustering to divide the field into intervention tiers, this CSI not only makes it possible to identify at-risk seedlings early on, but also facilitates data-driven zone management. According to experimental data, the CSI obtains considerable stage-wise separation and reliable ROC-AUC performance, while the proposed model achieves a classification accuracy of 97.32% with strong F1 and precision-recall balance. The suggested pipeline is ideal for high-throughput phenotyping and adaptive crop management since it can perform supervised classification and unstructured growth pattern identification. By fusing model performance with biological understanding and useful deployability, this work advances the area of image-based plant monitoring. • Developed SeedlingNet, a CNN model that achieved 97.32% accuracy in growth stage classification. • Introduced CSI to detect early stress using phenotypic features and clustering. • Created a diverse cotton seedling dataset via UAV and ground-based imaging. • Integrated deep learning with interpretable traits for precision agriculture use.