Dynamic Bias Adjustment of Integrated Multi-Satellite Rainfall Estimates using Convolutional Long Short-Term Memory Networks

Abstract

Satellite precipitation products (SPPs) are increasingly utilized for their higher spatio-temporal resolution; however, they often exhibit biases compared to ground observations, which hinders accurate local impact assessments. To address this, various statistical bias adjustment methods have been developed, ranging from simplistic linear scaling to more sophisticated quantile and Cumulative Distribution Function transferring (CDFt) [1] mapping techniques. However, many of these methods assume stationarity in relationships, leading to inaccuracies in regions with significant climate variability. Some alternative approaches, instead, like Equidistant CDF Matching (ECDFM) [2] and Quantile Delta Mapping (QDM) [3], aim to capture non-stationarity in precipitation data. Additionally, to deal with zero-inflated datasets, methods range from simple methods, such as positive correction and threshold adaptation [4], to more complex techniques including left tail censoring, singularity stochastic removal [5], and zero-truncated hurdle modeling [6]. In the field of atmospheric sciences, phenomena such as climate variables and precipitation often exhibit teleconnections, where changes in one region can influence conditions in distant areas. Additionally, climate variability often demonstrates scale-invariance and long-term memory, indicating that past states of the system influence future states. Recent advancements in deep learning, particularly Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, offer promising avenues for capturing spatio-temporal spatial dependencies [7, 8]. In this study, we propose a two-stage bias adjustment framework for refining and assessing distribution-based bias adjustment techniques applied to the Global Precipitation Measurement (GPM) Integrated Multi-satellitE Retrievals for GPM (IMERG) satellite rainfall estimates. By integrating multiple bias adjustment methodologies and leveraging advanced deep learning techniques, we aim to reduce systematic bias within satellite precipitation estimates.