Deep learning and georeferenced RGB-D imaging for hydroponic strawberry yield mapping

Abstract

Yield mapping in agricultural crops remains a significant challenge, particularly in uncontrolled environments. This study evaluates four instance segmentation algorithms: YOLOv8n, YOLOv8s, YOLOv8m, and YOLOv8l, along with a low-cost GNSS RTK system to detect and count strawberries in a hydroponic environment. A depth camera is used to remove background information from nonrelevant furrows, improving fruit detection accuracy. The low-cost RTK receiver, configured in Base Rover mode, provides centimeter-level precision and enables the generation of detailed yield maps that can be seamlessly integrated into commercial systems to increase growers' yield and profit. Data were collected in the municipality of Arcabuco, Boyacá (Colombia), resulting in 8848 images processed after augmentation. Among the models evaluated, YOLOv8l achieved the highest performance with a maximum F1-Score of 0.9295 and a mAP50 of 0.9689 during validation. Furthermore, in the fruit counting process - evaluated against manual counts - the same model achieved a R2 of 0.9997 and a mean relative error (MRE) of 1.5511%. In general, this work presents a systematic methodology for the extraction and visualization of information in fruit crops using computer vision and deep learning, showcasing a robust yield mapping system. The approach integrates pre-processing and post-processing steps, as well as 2D–3D image acquisition, georeferencing, and processing technologies, offering thus a novel solution for accurate and efficient hydroponic strawberry yield mapping.