A Multiscale Precipitation Forecasting Framework: Linking Teleconnections and Climate Dipoles to Seasonal and 24‐hr Extreme Rainfall Prediction

Abstract

We develop a hybrid statistical forecasting model for the simultaneous season‐ahead forecasting of both seasonal rainfall and the 24‐hr maximum rainfall for the upcoming season, using predictors identified through the Shared Reciprocal Nearest Neighbor approach. The model uses a generalized linear regression and a four‐parameter Beta distribution model for downscaling extremes using the predictors that were identified. A cross‐validation experiment for the last four decades in both Han‐River and Geum‐River watersheds, South Korea was performed to test the efficacy of the model. The leave‐one‐out cross‐validated seasonal precipitation forecast demonstrates a correlation that ranges from 0.69–0.78 to 0.68–0.76 for the Han‐River and Geum‐River watershed, respectively. Similarly, for the 24‐hr maximum rainfalls in the upcoming season, the cross‐validated correlations between the predicted and the observed values range from 0.67–0.73 to 0.50–0.63, for the two river basins. A discussion of the potential causes of the skill is offered.

A comparative study of extensive machine learning models for predicting long‐term monthly rainfall with an ensemble of climatic and meteorological predictors

“…The first one is physical numerical models based on the physical principles requiring a thorough description of the physical and dynamical processes of interactions between oceans, land, and atmosphere (Busuioc et al, 2001; Haupt et al, 2017). These models are driven by a huge amount of current and historical meteorological data and are built with systems of dynamic weather equations (Kim et al, 2020; Yu et al, 2015; Yu et al, 2016). However, most of these models are limited to computational capacity, efficiency and resolution (Abbot & Marohasy, 2012; Haidar & Verma, 2018).…”

Reconstruction of Continuous High-Resolution Sea Surface Temperature Data Using Time-Aware Implicit Neural Representation

“…We use Charbonnier Loss as the loss function for the optimization [31,32]. The batch size used in our experiment is 16.…”

A Novel Spatial Downscaling Approach for Climate Change Assessment in Regions With Sparse Ground Data Networks

“…The proposed modeling framework is demonstrated through a Leave‐One‐Out CV (LOOCV) evaluation in South Korea and climate change scenarios simulated by three different RCMs informed by the HadGEM2‐AO GCM (Y.‐T. Kim, So, et al., 2020; Magnusson et al., 2020; Park et al., 2016; Sivula et al., 2020).…”

“…Our approach assumes that persistence or low-frequency variability attributes in the precipitation simulations are unbiased, and the main bias resides in the probability distribution of the rainfall amounts, largely a result of the use of point observed data instead of compatible gridded fields. The proposed modeling framework is demonstrated through a Leave-One-Out CV (LOOCV) evaluation in South Korea and climate -T. Kim, So, et al, 2020;Magnusson et al, 2020;Park et al, 2016;Sivula et al, 2020).…”