Potato (Solanum tuberosum) breeding remains unusually time- and resource-intensive. New cultivars can take 10 to 15 years to develop, and selection must account for complex polyploid inheritance, low propagation rates, and strong environmental effects on yield. Genomic selection can accelerate breeding by predicting performance from DNA markers, while high-throughput phenotyping can rapidly measure crop responses in the field. Yet genomic prediction for highly quantitative potato traits often reaches only modest accuracy, and aboveground images do not always translate reliably into belowground tuber performance. Previous work has also relied heavily on multispectral imaging. Given these challenges, a deeper investigation is needed into whether inexpensive, repeated field imaging can complement genomic information for more accurate potato selection.
Published (DOI: 10.1093/hr/uhag175) online on April 30, 2026, in Horticulture Research, the study was conducted by researchers from the Department of Plant Breeding at the Swedish University of Agricultural Sciences (SLU) in Alnarp, Sweden. The team evaluated whether genomic single nucleotide polymorphism (SNP) data and phenomic traits derived from unmanned aerial vehicle (UAV) red–green–blue (RGB) imagery could be combined to predict agronomic performance. Their analysis focused on tuber yield by size class and starch content, while also testing whether longitudinal image information could capture environmental variation that genomic data alone may miss.
The researchers evaluated 256 potato breeding clones and cultivars across three location-year environments in southern Sweden: Helgegården in 2020 and 2021, and Mosslunda in 2021. Conventional measurements included total tuber weight, tuber weight in four size classes, and fresh starch content. Repeated UAV flights collected RGB images, from which canopy coverage, plant height, chlorophyll-related measurements, and vegetation indices were extracted. To represent crop development across the season rather than relying on a single flight, the team summarized image-derived traits using the area under the curve (AUC). They then compared genomic, phenomic, and integrated prediction strategies using regression and machine-learning methods. For kernel-based models, genomic best linear unbiased prediction (G-BLUP), phenomic best linear unbiased prediction (P-BLUP), and combined genomic-plus-phenomic best linear unbiased prediction (G+P-BLUP) were evaluated with leave-one-out and repeated five-fold cross-validation. The strongest advantage of integration appeared for total tuber weight and larger tubers. In five-fold testing, G+P-BLUP significantly outperformed single-source models for most of these traits, with prediction accuracy for total weight and tubers larger than 60 millimeters exceeding a squared Pearson correlation coefficient (R²) of 0.3—roughly twice the performance of G-BLUP alone. By contrast, starch content and smaller-tuber traits were predicted as well as or better by genomic information alone.
The authors said the results show why field images should be viewed as more than a faster substitute for manual measurements. Because repeated imagery captures how plants develop under real field conditions, it can add information about environmental responses that DNA markers do not directly encode. They said this complementary signal was most useful for yield-related traits, where growing conditions strongly shape performance. At the same time, the trait-specific results matter: adding more data is not automatically better. For starch content and small-tuber weight, genomic information already captured much of the useful predictive signal, so phenomic data added little.
The findings point to a staged strategy for potato improvement. Early in a breeding program, inexpensive high-throughput phenotyping (HTP) could help screen large numbers of clones before breeders invest in more costly genotyping or intensive field measurements. At intermediate stages, genomic and phenomic information could be combined when both are available, improving selection for yield while making better use of data already collected. Standard RGB cameras also lower the technological barrier compared with multispectral systems. As field sensors, robotics, and artificial intelligence become more common, integrated prediction frameworks could help breeders allocate resources more efficiently, select promising clones sooner, and accelerate genetic gain without assuming that one prediction model or data source works equally well for every trait.
###
References
DOI
10.1093/hr/uhag175
Original Source URL
https://doi.org/10.1093/hr/uhag175
Funding information
Nordic PPP Sustain Potato project; Formas – The Swedish Research Council for Environment, Agricultural Sciences, and Spatial Planning.
About Horticulture Research
Horticulture Research is an open access journal of Nanjing Agricultural University and ranked number one in the Horticulture category of the Journal Citation Reports ™ from Clarivate, 2023. The journal is committed to publishing original research articles, reviews, perspectives, comments, correspondence articles and letters to the editor related to all major horticultural plants and disciplines, including biotechnology, breeding, cellular and molecular biology, evolution, genetics, inter-species interactions, physiology, and the origination and domestication of crops.