ASHS 2025 Conference

Modern sweetpotato breeding programs evaluate hundreds of genotypes across successive generations to identify lines with superior storage root quality traits. However, traditional phenotyping methods rely on manual storage root evaluation, limiting both the scale and speed of selection. Small Unmanned Aircraft System (sUAS)-based high-throughput phenotyping offers scalable, image-based alternatives that enable breeders to collect highly detailed data with reduced bias, facilitating genomic selection. By linking image-derived phenotypes to genotypic data, these approaches could shorten the breeding cycle by supporting earlier or more optimal selection decisions. In this study, we developed an image-based yield estimation pipeline for early generation and advanced sweetpotato breeding lines using sUAS-based RGB (0.17 cm pixel⁻¹) and multispectral imagery. The pipeline leveraged a previously developed Mask R-CNN segmentation model for sweetpotato storage root detection that was pre-trained using mobile RGB images and fine-tuned using annotated aerial images to optimize performance for sUAS applications. Imagery was acquired in 2024 from two research fields immediately after harvest. Ground truth plot-level root yield was collected using mechanical singulation in an optical sorter (Exeter Engineering). The Mask R-CNN model generated instance masks of individual storage roots directly from plot-level RGB imagery, with root metrics such as length, diameter, and volume estimated using multiple geometrical methods. The model demonstrated strong predictive performance across both locations. Combined-location analysis yielded a correlation coefficient of 0.94 for storage root weight estimation (0.88 and 0.97 for individual locations) with a root mean squared error (RMSE) of 1.24 kg plot⁻¹. Root count estimation achieved a correlation coefficient of 0.78 (0.73 and 0.92 independently) with an RMSE of 11 roots plot⁻¹. These results indicate robust yield estimation across diverse genotypes and field conditions. Furthermore, these findings highlight the potential of combining aerial imagery and deep learning to streamline yield assessment in sweetpotato breeding programs. Future work will focus on enhancing model accuracy by incorporating root feature analysis, quality classifications, and expanded datasets to further support breeding decisions and accelerate selection pipelines.