Author(s): Raden Surya Pratama

Abstract: Early detection of abiotic and biotic stress is critical for protecting yield and quality in high-intensity aeroponic production systems, where plants are highly sensitive to short-term disruptions in misting, nutrient supply and microclimate. This study proposes and evaluates a deep learning-based leaf image analysis pipeline for early stress detection in an IoT-enabled aeroponic greenhouse cultivating high-value fruit vegetables. A total of 18,720 RGB leaf images were acquired in situ from aeroponic towers under controlled non-stress conditions and four induced stress types (nutrient deficiency, water/misting interruption, heat stress and biotic stress), each annotated at pre-stress, early-stress and overt-stress stages by expert agronomists and plant pathologists. After leaf segmentation and standardised pre-processing, several convolutional and transformer architectures were fine-tuned and compared, with EfficientNet-B3 emerging as the best-performing model. On a held-out test set, EfficientNet-B3 achieved 94.5% overall accuracy, macro-F1 of 0.93, macro-averaged AUC of 0.98 and Cohen’s kappa of 0.92 for multi-class stress classification. Compared with an IoT-only threshold-based monitoring scheme and a classical random forest baseline using handcrafted image features, the deep learning model showed significantly higher sensitivity to early-stress states and reduced misclassification, particularly for water/misting and nutrient-related stress. Time-to-detection analysis indicated that the proposed pipeline detected stress on average 19 hours earlier than IoT thresholds and approximately 27 hours earlier than expert visual inspection, with even larger gains for water/misting stress episodes. Class activation map visualisation confirmed that the network focused on physiologically meaningful leaf regions, enhancing interpretability and supporting agronomic trust. When integrated into the aeroponic IoT platform, model-driven alerts enabled timely corrective actions that reduced progression to overt stress without compromising yield, demonstrating the practical value of deep learning-based leaf image analysis as a core component of smart, resilient aeroponic crop management.

DOI: 10.33545/27076571.2025.v6.i2c.206

Pages: 254-261 | Views: 75 | Downloads: 44

Download Full Article: Click Here