In plant breeding, the selection of the best individuals is mainly based on phenotyping records. Because phenotyping is costly and time consuming, predictive tools such as Genomic selection (GS) have been developed in order to select among unphenotyped candidates. GS allows predicting the target traits for the selection candidates using the phenotypes of a training set and genotypic information collected on the training set and the selection candidates. Despite a good potential of the method to assist breeders in their selection choices, the cost of the genotyping still remains expensive, as GS requires to genotype each year the new selection candidates. In 2018, Rincent et al. developed a new, low cost, and high throughput method to predict the target trait of unobserved selection candidates. This method called phenomic selection (PS) is similar to GS, but genotyping is replaced by near infrared spectroscopy (NIRS). NIRS has the main advantage of being affordable, and already routinely applied on the selection candidates for many species such as wheat. GS has been well studied for twenty years, and many factors influencing its predictive ability are well understood. In PS, little is known about the factors influencing the predictive abilities, and about its performance relative to GS. We conducted the analyses on several datasets, corresponding to breeding lines drawn from the first or second years of trial evaluation from two breeding companies and one research institute in France. We evaluated several factors affecting PS predictive abilities including the possibility of combining spectra collected in different environments or at different steps of the breeding program. Contrary to genotypic data, near infrared spectra are indeed influenced by both the genotype and the environment. Thus, a selection candidate can be characterised by a multitude of spectra measured in different environments. The statistical model used was a simple H-BLUP model, reaching prediction ability from 0.26 to 0.62.Our results showed that the environment in which the NIR spectra was collected had an impor-tant impact on predictive ability and this impact was specific to the trait considered. Among all the models tested, combining NIR spectra from different environments were the best PS models and were at least as accurate as GS in most of the datasets. We finally tested a model which gathered NIRS and molecular marker effects. This model, GH-BLUP, was the best model of all, regardless of the trait or dataset, with prediction abilities reaching 0.31 to 0.73. In this study we showed that PS could be a great support tool for breeders to improve wheat breeding programs and could efficiently replace or complement GS..