Abstract
Traditional statistical downscaling methods often fail to accurately predict extreme climate events because they focus on conditional means. This paper presents a Smoothed Quantile Regression (LSQR) framework and its semi-supervised extension (LSSQR) to address this gap. By incorporating a smoothness constraint and leveraging unlabeled test data, the proposed models provide more accurate predictions of the magnitude and frequency of local temperature extremes.
Introduction
Global Climate Models (GCMs) are the primary tools for climate projection, yet their coarse spatial resolution limits their utility for local-scale impact assessment. Statistical downscaling is used to bridge this gap by relating large-scale atmospheric predictors to local-scale surface variables. However, standard techniques like Multiple Linear Regression (MLR) are optimized for average conditions and frequently underestimate the intensity of extreme events, such as heatwaves and cold snaps. While Quantile Regression (QR) can target specific distribution tails, it often suffers from high variance and overestimation. This research introduces a regularized approach to quantile regression designed specifically for the robust downscaling of climate extremes.
Methodology
The core of the proposed approach is the integration of spatial and physical consistency into the quantile regression framework.
Smoothed Quantile Regression (LSQR)
Standard QR minimizes an asymmetrically weighted sum of absolute errors. LSQR enhances this by adding a smoothness penalty term. This penalty is based on the principle that similar atmospheric states (predictors) should result in similar local climate responses. The similarity is quantified using a Radial Basis Function (RBF) kernel, which acts as a regularizer to prevent the model from producing erratic or physically implausible extreme predictions.
Semi-Supervised Extension (LSSQR)
The LSSQR model further improves performance by utilizing semi-supervised learning. In many climate modeling scenarios, predictor data for future periods (the test set) is available even when the corresponding local observations are not. LSSQR "harvests" the distribution of these unlabeled predictors to better calibrate the model's response. By understanding the manifold of the predictor space, the model can more accurately distinguish between normal variability and true extreme events.
Experimental Evaluation
The models were tested using daily temperature data from 37 weather stations across Canada, spanning the years 1961 to 2001. The predictor set included 26 coarse-scale variables from the NCEP re-analysis dataset, such as sea-level pressure, humidity, and airflow strength. To account for seasonal variations in climate dynamics, separate models were trained for Winter, Spring, Summer, and Fall.
Key Findings
The performance of LSSQR was compared against MLR, Ridge Regression, and standard QR across several metrics:
- Magnitude Accuracy: LSSQR demonstrated superior accuracy in predicting the magnitude of extremes, outperforming baseline methods in 85% of the weather stations based on Root Mean Square Error (RMSE) for tail data.
- Frequency and Timing: Standard QR frequently over-predicted the occurrence of extremes, leading to high false-alarm rates. In contrast, LSSQR achieved a better balance of precision and recall (F-measure), with predicted frequencies closely aligning with historical observations.
- Predictive Stability: While MLR consistently "smoothed over" the peaks of extreme events and standard QR often overshot them, LSSQR provided a stable and accurate tracking of actual temperature extremes.
Conclusion and Future Directions
The Smoothed Quantile Regression framework offers a robust solution for downscaling extreme climate events. By combining the tail-sensitivity of quantile regression with smoothness regularization and semi-supervised learning, the model provides a more realistic representation of climate risks. Future work will explore the application of non-linear kernels to capture more complex atmospheric interactions and adapt the framework for zero-inflated variables such as precipitation.
Summary updated