From Visual Words to Vision Transformers: Dual Approaches to Water Stress Classification in Maize

Early water stress in maize can quietly reduce growth and yield, making early detection from simple RGB images valuable for timely irrigation. This study compares two fundamentally different representation strategies for maize water stress classification: deep features learned from localized image regions using a custom Swin transformer and handcrafted visual features classified using a backpropagation neural network (BPNN). The Swin transformer learns stress-related representations from localized cropped image regions, with patch-level predictions aggregated into image-level classifications through majority voting, whereas the BPNN relies on handcrafted color- and texture-based descriptors extracted from segmented full-plant images. Both methods are evaluated on a maize water stress image dataset, using an independent test set that is not used during training or cross-validation for the final performance comparison. While the Swin transformer achieved 98% image-level accuracy after aggregation, the BPNN achieved 97% accuracy. Despite relying on fundamentally different representation strategies, both methods achieved high classification performance on the independently held-out test set. The results show that high-accuracy maize water stress classification can be achieved using low-cost RGB imagery and provide insight into the tradeoffs between learned and handcrafted representations in terms of classification performance and computational characteristics. These findings demonstrate the potential of computer-vision approaches for RGB-based agricultural water stress monitoring.