How Much Did it Rain? Predicting Real Rainfall Totals Based on Radar Data

TL;DR

This paper explores machine learning models, including a k-nearest-neighbor predictor, to accurately forecast rainfall totals from radar data, demonstrating competitive performance in a global challenge.

Contribution

It introduces a comprehensive evaluation of parametric and non-parametric models for rainfall prediction using radar data, highlighting the effectiveness of a k-nearest-neighbor approach.

Findings

The k-nearest-neighbor model achieved top performance.

Model training took approximately six days.

The approach was competitive in a worldwide competition.

Abstract

We applied a variety of parametric and non-parametric machine learning models to predict the probability distribution of rainfall based on 1M training examples over a single year across several U.S. states. Our top performing model based on a squared loss objective was a cross-validated parametric k-nearest-neighbor predictor that took about six days to compute, and was competitive in a world-wide competition.