Improving Future Climate Projections through Machine Learning-Based Downscaling and Bias Adjustment

Abstract Submission: Climate data downscaling is essential for refining Global Climate Models (GCMs) to produce detailed regional and local projections. GCMs, though valuable for global predictions, lack the precision required for localized forecasting. In this study, we employ a hybrid downscaling approach to enhance future precipitation projections for San Diego County. By selecting the four nearest latitude and longitude points around weather gauges and integrating local neighborhood precipitation records, we aim to improve rainfall forecasts at a finer scale.

We applied machine learning techniques, including Random Forest, for downscaling, alongside bias correction methods such as Linear Regression, Probability Density Function (PDF) adjustment, and Quantile Mapping. These tools helped address systematic errors and enhance the model’s accuracy in predicting future rainfall patterns.

In the projections, San Diego typically experiences minimal precipitation during the dry summer months, but in years with above-normal precipitation, the wet months—particularly winter—become significantly wetter. This study highlights how climate change could amplify these seasonal contrasts, with the potential for drier summers and wetter winters, as well as more intense extreme weather events. The refined projections from this hybrid downscaling approach are critical for better understanding future precipitation trends and informing decision-making on climate resilience and resource management. Ultimately, these insights can help prepare the region for the impacts of climate change by providing more accurate, localized climate forecasts.

Learning Objectives/Expected Outcome (Optional) :