Edge-Enabled Mobile App for Smart Agriculture Using Multi-Sensor Inputs and a Hybrid CNN–Vision Transformer Model
Résumé
Crop diseases significantly threaten food security, particularly in resource-limited regions. Existing mobile diagnostic tools often lack robustness and fail to integrate environmental context. To address these limitations, this study developed a hybrid CNN-Vision transformer model combining EfficientNetV2, MobileNetV2, and Vision Transformers. The model was trained on Kaggle’s PlantVillage dataset and locally collected field images covering 76 disease and condition classes. Multi-sensor inputs, including soil moisture, temperature, and humidity, were integrated to enhance prediction accuracy. The model achieved 99.2% accuracy, an AUC of 0.999998, and a 69% lower prediction variance than baseline models. Bayesian testing confirmed its superiority over DenseNet50 and other models. TensorFlow Lite was used to optimize the model for deployment on resource-constrained devices and was integrated into AgriScan, a 30.4 MB edge-enabled Android app. AgriScan supports offline inference, delivers real-time predictions with 0.094s latency, and reduces misdiagnoses by 92%. Evaluation on 249 unseen local images confirmed 97.97% accuracy, validating its field applicability. The model achieved a prediction variance of 0.000010, indicating strong confidence consistency across predictions. Benchmarking revealed 4.8× faster inference and 83% lower energy usage than cloud-based alternatives. Cross-device testing confirmed 100% diagnostic consistency, supporting the model’s generalizability. This framework combines edge AI and sensor fusion into a scalable, cloud-independent diagnostic solution for advancing smart agriculture in low-connectivity environments.
Citer ce document
Accès au document
Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter
Voir l'article sur le site de la revueStatistiques
Consultations : 1
Téléchargements : 0