Transfer learning-enhanced deep learning for tree crown geometric analysis and crop yield estimation using UAV imagery

Abstract

Estimating tree yields is essential for optimizing productivity in precision agriculture, where informed decisions rely on accurate measurements of tree characteristics. Crown parameters such as diameter, radius, depth, base height, ratio, predicted area, shape, and volume play a key role in assessing canopy biovolume, which refers to the above-ground space occupied by the tree canopy and serves as a proxy for biomass. This study proposes an innovative methodology for estimating the biovolume and predicting the productivity of trees from multi-view images captured by drones. Trees are classified into two main geometric categories (oval and round) before being analyzed using advanced segmentation and geometric analysis techniques. A pre-trained segmentation model, FastSAM, was refined through YOLOv11-specific fine-tuning to optimize the detection and segmentation of olive tree crowns, allowing precise crown isolation and extraction of essential geometric parameters. The geometric analysis of aerial view contours relies on the Convex Hull method to derive shape parameters, while tree heights are determined from side views using a 2D triangle projection technique. These characteristics, combined with photogrammetry principles, are then used to estimate volume and predict the weight of each individual tree. The extracted data, along with field measurements, were used in regression models to establish correlations between crown size, height, and volume. The segmentation results showed an accuracy of 97.90 % for top-view images and 97.60 % for front views. Canopy volume estimation achieved 96.45 % accuracy for oval crowns and 95.36 % for round crowns. Productivity prediction using regression models yielded an R² of 95.84 % without interaction terms and 97.78 % with interactions. Additionally, the relative error in crop yield estimation was 6.05 %. By integrating multi-view drone imagery, fine-tuning of pre-trained models, advanced geometric analysis, and 2D projection techniques, this methodology provides a robust, accurate, and scalable solution for enhancing precision agriculture practices, enabling efficient prediction of tree yields and biomass.