Eggplant Leaf Semantic Segmentation in Aerial Imagery for Precision Agriculture

Eggplant is one of the prominent vegetable crops in the Philippines. Traditional farming practices, like agrochemical spraying, often lack real-time, high-resolution data crucial for informed decisions. Monitoring vegetation indices is vital for crop growth assessment, health monitoring, and precision agriculture. However, optimizing agrochemical spraying for specific crop needs remains a challenge. Unmanned aerial vehicles (UAV s) offer a transformative solution. RGB cameras, designed for visual imagery, may not provide accurate vegetation index measurements. Computer vision, through leaf detection, plays a crucial role in assessing crop health. This study aims to distinguish eggplant leaves from their surroundings. Subsequently, the authors compared two open-source segmentation models, namely U-Net and FPN. Two pre-trained encoders, VGG16 and VGG19, were used in both models to further evaluate their performance. Using a drone-mounted RGB camera, the authors gathered images of mature eggplant leaves. The experimental findings demonstrate the effectiveness of the VGGl6-based FPN architecture for real-time leaf segmentation tasks as it acquires 97.24 % and 96.73 % accuracy with a latency time of 31.5 ms and 32.6 ms, on 0.5 and 0.75 threshold values. The study recommends FPN with VGG16 for its accurate and low-latency performance in time-sensitive leaf segmentation applications.