Artificial Intelligence and Machine Learning in Plant Breeding: Data-driven Frameworks for Genomic Prediction, High-throughput Phenotyping and Accelerated Cultivar Development
Abhishek
*
Genetics and Plant Breeding, University of Horticultural Sciences, Bagalkot, India.
Kiran Mahavir Magdum
ICAR- Indian Institute of Horticultural Research, Bengaluru, India.
Santosh Mavinali
Fruit Science, University of Horticultural Sciences, Bagalkot, India.
Savitha S. Desai
P.G. Department of Studies and Research in Biotechnology, KLES P.C. Jabin Science College, Karnataka, India.
Arvind Patel
Genetics and Plant Breeding, University of Horticultural Sciences, Bagalkot, India.
Sudharshan D. Kove
Department of Biotechnology, Sant Gadge Baba Amravati University, Amravati, India.
Akshay Sawant
Vegetable Science, University of Horticultural Sciences, Bagalkot, India.
*Author to whom correspondence should be addressed.
Abstract
Plant breeding is increasingly a data-intensive science. Modern programmes generate genome-wide marker data, multi-environment trial records, remote-sensing and sensor streams, and high-resolution images at a scale that exceeds the capacity of conventional statistical workflows. Artificial intelligence (AI) and machine learning (ML) offer a set of tools for extracting predictive signal from such data. This review surveys the application of ML and deep learning across the breeding cycle, beginning with the foundations of the main algorithm families and moving to genomic selection, where linear mixed models remain the benchmark against which random forests, support vector machines, neural networks and convolutional architectures are measured. It then considers image-based high-throughput phenotyping and disease detection, modelling of genotype-by-environment interaction and integration with crop growth models, yield prediction, optimisation of training populations and trial design, and the coupling of predictive models with genome editing. Deep learning is most firmly established for image- and sensor-based phenotyping, although generalisation across sites remains a recognised weakness, whereas the advantage of ML over well-specified linear models for genomic prediction of complex traits is modest and context dependent. Key obstacles include small sample sizes relative to the number of predictors, limited model interpretability, data heterogeneity, and unequal access to computing and expertise.
Keywords: Artificial intelligence, machine learning, deep learning, genomic selection, high-throughput phenotyping, plant breeding, genotype × environment interaction