What type of study is this?

September 10, 2025

Machine Learning Models for Daily Rainfall Forecasting in Northern Tropical Africa Using Tropical Wave Predictors

Key Points

Machine learning models significantly outperform traditional rainfall forecasting methods in northern tropical Africa.
The study identified tropical depression and near-wave predictors as critical for enhancing meteorological model accuracy.
Gradient-boosting regression selected important predictors from seven wave types to enhance daily rainfall forecasts.
Probabilistic forecasts created via machine learning demonstrate notable skill over various benchmarks in tropical Africa.

Abstract

Abstract Numerical weather prediction models often struggle to outperform simple climatology-based precipitation forecasts in northern tropical Africa, even after statistical postprocessing. Artificial intelligence (AI)-based methods show promise, but accurately incorporating precipitation remains challenging. This study explores, for the first time, the predictive value of synoptic–planetary-scale forcing associated with African easterly waves (AEWs) and other tropical waves using machine learning (ML). We apply a gamma regression and convolutional neural network (CNN), trained on wave predictors derived from satellite-based rainfall estimates (GPM IMERG), to predict daily rainfall during the July–September monsoon season. Predictor variables capture local amplitude and phase information from seven wave types at the target 1° × 1° degree grid box and adjacent upstream and downstream neighbors, with important predictors selected via gradient-boosting regression. We find downstream predictors contributing most significantly to model skill, highlighting robust predictability of rainfall suppression during dry wave phases. Tropical depression (AEWs) emerges as the most influential predictor, while mixed Rossby–gravity, Kelvin, and inertio-gravity waves show importance in specific subregions. We transform deterministic ML predictions into probabilistic forecasts using the novel Easy Uncertainty Quantification (EasyUQ) and compare them against three benchmarks: extended probabilistic climatology with a 15-day window (EPC15) derived from GPM IMERG, European Centre for Medium-Range Weather Forecasts (ECMWF) operational ensemble (ENS) forecasts, and probabilistic forecasts based on EasyUQ applied to the ENS control member (CTRL EasyUQ). ENS forecasts exhibit substantial miscalibration compared to EPC15, while CTRL EasyUQ provides minor improvements over land. In contrast, the simpler, cost-effective ML-based forecasts significantly outperform all benchmarks across tropical Africa, highlighting their operational potential for African weather services.

AIに質問

Bookmark