Abstract: Wheat (triticum aestivum l.) breeding technology can face a great challenge on the long cycle, low efficiency, and narrow genetic background. An important breakthrough can be combining the high-throughput phenotyping of in-field wheat and genome-wide association, thereby revealing the genetic variation in dynamic response to environmental stress. Fortunately, the unmanned aerial vehicle (UAV) remote sensing and machine learning can be expected to bridge the genotype–phenotype gap of the wheat in the breeding process. Among them, frost tolerance is an important phenotype target, particularly with the winter survival of wheat in various environments. It is a high demand for the rapid and cost-effective assessment of frost tolerance from the UAV multi-spectral imagery using machine learning. In this study, a genome-wide association study (GWAS) was assessed for the quantitative genomic analysis of wheat frost tolerance. A bi-parental wheat population consisting of 491 doubled haploid lines was also used in four study sites. 491 wheat core materials with a relatively consistent growth stage were selected to obtain their high-density genotype data with the 660 K single nucleotide polymorphism (SNP). The UAV-based multi-spectral imagery of the wheat canopy was collected at the overwintering stage at four experimental sites. At the same time, the wheat in-field phenotypes of frost tolerance were investigated by the wheat breeding experts at the same time. The image pre-processing was performed on the features generation of 16 spectral vegetation indices, including image mosaic, geometric correction, radiometric correction and index calculation. Image segmentation was utilized to obtain the features of the wheat canopy using unsupervised clustering. The features correlation analysis and importance analysis were implemented to compare with the in-field investigation, in order to identify quantitative trait loci (QTL) underlying frost tolerance. A comparison was then made on the evaluation models of wheat freezing injury established by random forests (RF), extreme gradient boosting (XGBoost), gradient boosting decision tree (GBDT), and support vector machine (SVM). The results showed that significantly high accuracy was achieved up to 67.94% of the classifier in the XGBoost, compared with the in-field investigation. The correlation and importance of features were also analyzed during this time. The importance of 22 spectral features to the prediction performance of the classifier was evaluated using the information gain brought by the feature, when the sub node of the classifier split. The results showed that there was the most important for the prediction performance of the classifier in the simplified Canopy Chlorophyll content index (SCCCI) among the 16 spectral features of the wheat canopy. Three QTLs were also closely related to the frost resistance detected by the genome-wide association analysis. The three loci of 2B, 3A, and 5A on chromosome 21 of wheat presented a significant SNP, even exceeding the threshold (-lgP=4). The SNPs were continuously distributed. Therefore, the spectral features using UAV remote sensing can be expected to serve as the wheat frost resistance QTL. The UAV-enabled phenotyping can be an effective, high-throughput, and cost-effective approach to understanding the genetic basis of wheat frost tolerance in genetic studies and practical breeding. This finding can also provide a fast way for the high-throughput phenotyping of wheat frost tolerance for wheat winter survival in the field.