Microsoft has revealed advancements in its weather prediction capabilities, significantly improving the accuracy of forecasts for cloud cover and precipitation. These updates are integrated into the Weather from Microsoft Start platform, utilizing artificial intelligence to merge data from radar and satellite sources.
Users can access this enhanced weather information through its integration into Windows 10, Windows 11, Microsoft Edge, Bing, and the Bing and Microsoft Start mobile apps.
According to an independent study commissioned by Microsoft, Weather from Microsoft Start has been recognized for its leading forecast accuracy.
AI-Driven Precipitation Nowcasting
Since 2021, Weather from Microsoft Start has operated a short-term precipitation nowcasting model powered by generative AI. This model, updated every two minutes, provides hyper-local forecasts at a 1-kilometer resolution for up to four hours ahead. The integration of radar and satellite data addresses the issue of limited weather radar hardware in various regions, enhancing the overall accuracy of predictions.
The updated model is four times larger than its predecessor and predicts both simulated radar and satellite reflectivity. This dual approach fills data gaps and improves forecast reliability. The radar channel model was given six times more weight during AI training compared to the satellite model, reflecting the higher importance of radar-derived data. Microsoft employed an adversarial learning approach, using a generative adversarial model (GAN) to enhance the realism of predictions. The spatial and temporal discriminators improve visual fidelity and temporal consistency, respectively.
The new model has unlocked the ability for users to experience continuous cloud and precipitation forecasts and maps. Simulated radar reflectivity is evaluated by checking precision and recall for different reflectivity thresholds indicative of varying rainfall. Satellite image predictions are compared against persistence using metrics such as MSE, MAE, PSNR, MS-SSIM for similarity, and FID scores for sharpness. This comprehensive approach ensures that Weather from Microsoft Start provides more precise and reliable weather information globally.
Enhanced Forecast Accuracy
Internal testing on benchmarks such as the SEVIR dataset shows that Microsoft Start’s model ranks near the top, providing forecasts up to twice as far out as other generative AI models like DGMR (2021) and PreDiff (2023). The model’s training loss function includes pixel-wise regression loss and adversarial loss, with the α parameter tuned to balance missed rain instances and rain bias. The use of L1 loss instead of L2 prevents the model from being overly penalized for missing extreme precipitation conditions.
Productionizing a global forecast model with up-to-the-minute data involves challenges such as managing high latency and segmentation effects. The generator architecture meets conditions of translation equivariance, spatially unconstrained operations, and low memory footprint, allowing flexibility in window sizing during training and inference. This has enabled Microsoft to provide accurate forecasts even during satellite data outages.
