From phenome to genome: A cloud-based AI platform for integrative rice grain analysis and genetic mapping to empower grain-focused crop improvement

Rice ( Oryza sativa L. ) is a staple crop of global importance. Accurate assessment of yield-related traits is essential for improving productivity and ensuring global food security. To address the limitations of labour-intensive and often subjective manual grain-based phenotyping, we developed the Rice Grain Phenotyping Analysis System (RGPAS), an open and cloud-based platform that integrates multiple machine learning (ML) and deep learning (DL) models to automate grain-level phenotypic analysis using smartphone-acquired images. The RGPAS platform comprises a digital rice germplasm database of 864 landraces and 176 recombinant inbred lines (RILs) cultivated worldwide, as well as a newly established World Rice Grain Dataset (WRGD) incorporating grain-level morphological and colour characteristics to enable artificial intelligence (AI)-powered analysis. The cloud-based analysis is powered by an optimised RGD-YOLO learning model for accurate grain detection under both occluded and unoccluded conditions together with an EfficientNet-RG classifier for whole-grain classification, resulting in the measurement of 18 grain-level traits such as grain length, grain width, grain number per panicle, and seed-coat colour. Using grain-level phenotypic data quantified in the 2020 and 2022 rice growing seasons, we demonstrated the accuracy, scalability, and robustness of RGPAS-derived traits, enabling reliable assessment of grain diversity and phenotypic variation suitable for genetic mapping in rice. Accordingly, we performed both genome-wide association studies (GWAS; with 864 landraces in total) and quantitative trait locus (QTL) mapping (with 176 RILs) to link the RGPAS-derived traits to specific genomic regions, successfully identifying known regulators (e.g. D61 , GE , GS3 , GW2 , GW5 , GW7 , OsSOC1 , RFT1 , and TGW6 ) and several novel loci. These results validate RGPAS as a reliable and biologically relevant tool for integrative phenotypic and genetic analysis, providing a standardised and accessible solution for breeders and researchers to characterise grains before sowing and after harvest for crop improvement.