A hierarchical framework for cotton boll opening rate monitoring using UAV RGB imagery

The cotton boll opening rate (CBOR) is a critical agronomic indicator determining yield and fiber quality. This study developed a hierarchical framework integrating unmanned aerial vehicle (UAV) RGB imagery to automate CBOR classification. To address blurred boll edges and complex backgrounds, an improved TransUNet augmented with the Convolutional Block Attention Module (CBAM) and Dynamic Snake Convolution (DySnakeConv) was proposed to extract the cotton pixel ratio (CPR). CPR, combined with vegetation indices and texture features, served as inputs for CBOR classification. Two classifiers, Random Forest (RF) and Support Vector Machine (SVM), were evaluated. Results indicated that the enhanced TransUNet significantly outperformed baseline models (e.g., U-Net, DeepLabV3+), achieving state-of-the-art segmentation (Precision 92.15%, Recall 91.93%, F1-score 92.04%, IoU 85.27%). Correlation analysis revealed that CPR served as a robust quantitative proxy for CBOR across multiple sampling dates. In classification tasks, RF consistently outperformed SVM in overall F1-scores across three dates (80.78% vs. 72.74%; 84.32% vs. 80.54%; 83.09% vs. 78.42%), likely due to its robustness in handling high-dimensional feature interactions. These findings validate a practical UAV-based framework for accurate CBOR monitoring, providing actionable support for precision harvest scheduling and variable-rate defoliant application.