Background: Metabolomics is one of the most widely used omics tools for deciphering the functional networks of the metabolites for crop improvement. However, it is technically demanding and costly.Results: We propose a relatively inexpensive approach for metabolomics analysis in which metabolomics is combined with hyperspectral imaging via machine learning. This approach can be used to target important steps in flavonoid and lipid biosynthesis in rice. We extract 1848 hyperspectral indices and 887 metabolites from polished grains of 533 Oryza sativa accessions. Hyperspectral indices are then linked to metabolites through correlation analysis and modelling. Based on this, a total of 554 metabolites and 1313 hyperspectral indices are identified for further genome-wide association study (GWAS). By GWAS, we detect 17,509 significant locus-trait associations with 2882 single nucleotide polymorphism (SNPs). Colocalization analysis links these SNPs to the corresponding metabolites and hyperspectral indices. We detect 6415 pairs of metabolites and hyperspectral indices within a linkage disequilibrium of 300 kb in the Oryza sativa genome. We then characterize 1761 candidate genes colocalized to these loci by transcriptomic analysis. We further verify novel candidate genes encoding a novel flavonoid (LOC_Os09g18450) and a flavonoid/lipid (LOC_Os07g11020) respectively by gene editing and overexpression in rice.Conclusion: Our findings indicate that hyperspectral imaging combined with machine learning methods could serve as a powerful tool for quickly and inexpensively assessing crop metabolites.