An Algorithmic Framework for Cocoa Ripeness Classification: A Comparative Analysis of Modern Deep Learning Architectures on Drone Imagery

This study addresses the challenge of automating cocoa pod ripeness classification from drone imagery through a comprehensive and statistically rigorous investigation conducted on data collected from Ghanaian cocoa fields. We perform a direct comparison by subjecting a curated set of seven deep learning models to an identical, advanced algorithmic framework. This pipeline incorporates high-resolution (384×384) imagery, aggressive TrivialAugmentWide data augmentation, a weighted loss function with label smoothing, a unified two-stage fine-tuning strategy, and validation with Test Time Augmentation (TTA). To ensure statistical robustness, all experiments were repeated three times using different random seeds. Under these demanding experimental conditions, modern architectures demonstrated strong and consistent performance on this dataset: the Swin Transformer achieved the highest mean accuracy (79.27%±0.56%), followed closely by ConvNeXt-Base (79.21%±0.13%). In contrast, classic architectures such as ResNet-101 (55.86%±4.01%) and ResNet-50 (64.32%±0.94%) showed substantially reduced performance. A paired t-test confirmed that these differences are statistically significant (p<0.05). These results suggest that, within the evaluated setting, modern CNN- and transformer-based architectures exhibit greater robustness under challenging, statistically validated conditions, indicating their potential suitability for drone-based agricultural monitoring tasks.