Reliable seasonal weather forecasts are essential for irrigation management and crop yield prediction, particularly in regions with limited water resources. This study aimed to improve the usability of the North American Multi-Model Ensemble (NMME), an experimental real-time seasonal weather forecast system, for regional crop modelling and irrigation decision-making. Coarse resolution of NMME may introduce bias and uncertainty at regional/local scales. To address this, a statistical downscaling method with bias correction for both mean and variability was used to produce 1-km gridded daily weather projections for temperature and precipitation across the contiguous United States from a representative NMME model, the Canadian Coupled Climate Model version 4 (CanCM4). The daily surface weather and climatological summaries (DAYMET) data were used to calibrate the downscaled hindcast projections of CanCM4. The reliability of downscaled CanCM4 forecasts for local crop modelling was evaluated at lead times of up to six months using a calibrated DSSAT model at a research station in the semi-arid Texas Rolling Plains region. Cross-validation during the hindcast period demonstrated strong forecast skill, with R2 values of 0.72 and 0.71 for maximum and minimum temperatures, respectively. The precipitation forecast remained sensitive to extreme events, with seasonal and annual relative errors of 31% and 1%, respectively. Crop yield predictions had a relative error of 9%, and irrigation water requirements closely matched field observations, outperforming both raw CanCM4 and multi-model mean methods. The downscaling method used in this study significantly improved NMME data reliability, although the degree of improvement may vary with time and location.