WS-YOLO: An Agronomical and Computer Vision-Based Framework to Detect Drought Stress in Lettuce Seedlings Using IR Imaging and YOLOv8

Lettuce (Lactuca sativa L.) is highly susceptible to drought and water deficits, resulting in lower crop yields, unharvested areas, reduced crop health and quality. To address this, we developed a High-Throughput Phenotyping platform using Deep Learning and infrared images to detect stress stages in lettuce seedlings, which could help to apply real-time agronomical decisions from data using variable rate irrigation systems. Accordingly, a comprehensive database comprising infrared images of lettuce grown under drought-induced stress conditions was built. In order to capture the required data, we deployed a Raspberry Pi robot to autonomously collect infrared images of lettuce seedlings during an 8-day drought stress experiment. This resulted in the generation of a database containing 2119 images through augmentation. Leveraging this data, a YOLOv8 model was trained (WS-YOLO), employing instance segmentation for accurate stress level detection. The results demonstrated the efficacy of our approach, with WS-YOLO achieving a mean Average Precision (mAP) of 93.62% and an F1 score of 89.31%. Particularly, high efficiency in early stress detection was achieved, being a critical factor for improving food security through timely interventions. Therefore, our proposed High-Throughput Phenotyping platform holds the potential for high-yield lettuce breeding, enabling early stress detection and supporting informed decision-making to mitigate losses. This interdisciplinary approach highlights the potential of AI-driven solutions in addressing pressing challenges in food production and sustainability. This work contributes to the field of precision agricultural technology, providing opportunities for further research and implementation of cutting-edge Deep Learning techniques for stress detection in crops.