Since the invention of whole genome prediction (WGP) more than two decades ago, breeding programs have established extensive reference populations that are cultivated under diverse environmental conditions. The introduction of the CGM-WGP model, which integrates crop growth models (CGM) with WGP, has expanded the applications of WGP to the prediction of unphenotyped traits in untested environments, including future climates. However, CGMs require multiple seasonal environmental records, unlike WGP, which makes CGM-WGP less accurate when applied to historical reference populations that lack crucial environmental inputs. Here, we investigated the ability of CGM-WGP to approximate missing environmental variables to improve prediction accuracy. Two environmental variables in a wheat CGM, initial soil water content (InitlSoilWCont) and initial nitrate profile, were sampled from different normal distributions separately or jointly in each iteration within the CGM-WGP algorithm. Our results showed that sampling InitlSoilWCont alone gave the best results and improved the prediction accuracy of grain number by 0.07, yield by 0.06 and protein content by 0.03. When using the sampled InitlSoilWCont values as an input for the traditional CGM, the average narrow-sense heritability of the genotype-specific parameters (GSPs) improved by 0.05, with GNSlope, PreAnthRes and VernSen showing the greatest improvements. Moreover, the root mean square of errors for grain number and yield was reduced by about 7% for CGM and 31% for CGM-WGP when using the sampled InitlSoilWCont values. Our results demonstrate the advantage of sampling missing environmental variables in CGM-WGP to improve prediction accuracy and increase the size of the reference population by enabling the utilisation of historical data that is missing environmental records.