An Efficient Multimodal Framework for Barley Drought Stress Detection on Resource-Constrained Devices

Drought stress significantly impacts barley (Hordeum vulgare L.) production, necessitating early and accurate detection systems for precision agriculture. Traditional monitoring approaches rely on manual inspection or single-modality sensing, which often fail to capture the complex physiological responses to water deficit. This study presents a novel multimodal deep learning framework that integrates RGB imaging with environmental sensor data (temperature and humidity) for real-time drought stress classification in barley plants. The proposed architecture employs EfficientNetV2-S for visual feature extraction, coupled with a dedicated sensor encoding branch, unified through a cross-modal attention mechanism and gated multimodal fusion strategy. To address the computational constraints of agricultural IoT systems, we implemented comprehensive CPU optimization techniques and model compression via TensorFlow Lite INT8 quantization, achieving a 68.5% reduction in training time and 90% reduction in model size. The system was validated on a custom greenhouse dataset (379 samples, 80/20 split) and the PlantVillage dataset (26,000 images, binary reformulation). A 10-seed evaluation protocol demonstrated that the full multimodal model achieves 98.3 ± 1.5% accuracy, outperforming both an image-only baseline (97.4 ± 1.8%) and a sensor-only MLP (73.8 ± 3.5%). Across seeds, the model also achieved an F1-score of 98.34 ± 1.48% and ROC-AUC of 99.93 ± 0.13%. Ablation analysis with ANOVA (F(4,36) = 4.44, p = 0.005) confirmed that multimodal fusion improves accuracy by 0.92% over image-only models, with the full gated cross-modal attention mechanism outperforming all simplified baselines, including AgriFusionNet (75.22%), Shallow CNN (92.54%), Logistic Regression multimodal (92.11%), and Random Forest multimodal (89.91%). These results further show that relying on environmental data alone is insufficient, reinforcing the benefit of multimodal fusion. External validation on PlantVillage achieved 99.97% accuracy, demonstrating strong generalization capabilities. The optimized model operates efficiently on CPU-only hardware (training time: 9.1 min/epoch), making it suitable for edge deployment in resource-constrained agricultural environments. This work demonstrates that a low-cost, CPU-compatible multimodal deep learning system can reliably detect drought stress in barley under real greenhouse conditions and provides a practical and scalable solution for early stress monitoring in precision agriculture.