ARTFEED — Contemporary Art Intelligence

Q-SRDRN: Multi-Quantile Network for Extreme Precipitation Downscaling

other · 2026-05-14

A new deep learning model, Q-SRDRN, addresses the systematic underprediction of heavy-tail precipitation events by using multi-quantile regression with pinball loss at tau values 0.50, 0.95, 0.99, and 0.999. The key innovation is that the primary obstacle is the loss function, not data augmentation. Two CNN-specific design choices—IncrementBound for monotonicity and per-quantile output heads—enable independent processing of bulk and tail distributions. This allows data augmentation via cVAE to complement the median head without contaminating extreme quantiles.

Key facts

  • Deep super-resolution networks under-predict heavy-tail precipitation events
  • Primary obstacle is the loss function, not data augmentation
  • Q-SRDRN uses multi-quantile regression with pinball loss
  • Tau values: 0.50, 0.95, 0.99, 0.999
  • IncrementBound enforces monotonicity while preserving gradient identity
  • Separate per-quantile output heads for bulk and tail detection
  • Data augmentation via cVAE complements the median head
  • Model improves prediction of extreme precipitation events

Entities

Sources