ETD

• edge AI computing
• ESP32-S3
• grape leaf disease detection
• image classification
• LoRaWAN
• object detection model
• precision agriculture
• quantization
• wireless sensor networks

Existing vineyard monitoring systems primarily rely on environmental data collection such as temperature, humidity, and pressure transmitted over LoRaWAN networks and providing limited insights into the actual grapevine’s health status. This thesis work focuses on enhancing the vineyard monitoring intelligence by integrating Edge AI computing at sensor level for grape leaf disease detection. Moving beyond the traditional environmental-only monitoring approaches, this research was conducted on developing and optimizing deep learning models (CNN for classification and YOLO for detection) and deploying them directly on resource-constrained edge sensors. This extension enabled the monitoring system to capture and process leaf images to detect diseases directly at the edge, providing real-time disease detection capabilities while adapting to the inherent bandwidth and power limitations of LoRaWAN-based wireless sensor networks. This disease detection feature is implemented using four-stage pipeline architecture and deployed on
ESP32-S3 microcontroller. Stage 1 captures image of vine-trees using the OV3660 camera module operating at VGA resolution with JPEG compression. Stage 2 employs an ESP-optimized YOLO11n model (ESPDet-Pico: 478KB INT8) to localize individual grape leaves within the captured frame, achieving 56.4% mAP@50 for leaf detection. Then these detected leaf regions are cropped and forwarded to Stage 3, where a MobileNetV2 classifier model (2.73 MB INT8) is utilized to identify disease symptoms across the four classes—Black Rot, Esca, Healthy, and Leaf-Blight with 99.8%±0.05 accuracy on individual leaf patches. Stage 4 aggregates the classification results from multiple detected leaves to produce a single output formatted for LoRaWAN transmission. Since each captured frame typically contains multiple grape leaves, the weighted aggregation algorithm combines detection confidence, classification entropy, and spatial quality metrics from up to 15 detected leaves to produce robust per-frame disease diagnosis results. The aggregated classification result is then packaged into LoRaWAN uplink packet (e.g.({"Leaf Blight", 80% confidence}) and transmitted over the WSNs.

To optimize also the power consumption, the system works on duty-cycle operation to capture and process images and apply inferences at defined intervals and then enters into deep sleep mode. Overall this thesis work demonstrated the feasibility of deploying deep learning models on resource-constrained microcontrollers, specifically for disease detection in vineyard monitoring systems and precision agriculture applications at large.