UAV-based citrus tree segmentation, counting and yield estimation using lightweight deep learning approaches

In recent years, arid and semi-arid regions have faced severe and persistent drought, with rainfall at most 200 mm per year. In addition, intensive irrigation practices in agribusiness areas aimed at boosting production further increased irrigation water consumption. These practices call for innovative applications that enable optimized yield estimation and tree health monitoring. This paper aims to predict crop yield in a citrus orchard farm using UAV imagery and Deep Learning approaches. It emphasizes the use of a lightweight Tiny U-Net model for tree detection and a CNN-based architecture for crop yield estimation based on vegetation indices and in-situ measurement data. The study was designed to provide a cost-effective solution for precision orchard management and monitoring under climate stress. The CNN model outperformed other machine learning models in yield prediction, achieving the highest coefficient of determination (R2) of 88%. The Tiny U-Net architecture, developed for semantic segmentation and counting, effectively distinguished individual citrus trees and rows. The model reached high accuracy, with overall precision and recall reaching 94.74% and 94.88%, respectively, and maintained a low inference time of 12.55 ms, making it suitable for real-time and on-boarding processing. The segmentation output enabled an accurate counting of both trees and rows, with a R2 exceeding 99%, confirming the reliability of the model for structural orchard analysis. The pipeline supports precision agriculture through reliable and high-resolution yield monitoring, enabling informed decision-making for citrus orchard management and resource optimization.