REDUCTION OF SYSTEMATIC ERROR ON IMERG-E RAINFALL ESTIMATION USING BAYESIAN MODEL AVERAGING

• Main Author: Eggy C. P., Alexander
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/68912
• Institution: Institut Teknologi Bandung

The Integrated Multi-satellite Retrievals for GPM (IMERG) satellite rainfall estimate has two products: IMERG-E and IMERG-F. IMERG-E has the potential to be used for early warning operations because of its fast release time, but its accuracy is less than IMERG-F, which involves gauge calibration. The Bayesian Model Averaging (BMA) method offers a calibration solution that can be used in IMERG-E rainfall estimation. In this study, the Bayesian Model Averaging (BMA) method, generally used for ensemble prediction calibration, calibrates or reduces the IMERG-E error. Because it only uses one type of satellite estimation product, the ensemble used is the rain area around the reference station while accommodating the uncertainty of spatial satellite estimation. The results show that the BMA method can reduce the systematic error of IMERG-E and produce a calibration in the form of probability for its uncertainty. BMA reduces the bias and variance components of the Mean Square Error (MSE) systematic error. The configuration of the BMA method greatly affects the calibration results, especially the configuration of spatial sampling or regional rainfall. The Conditional Training Window (CTW) scheme shows the advantages of the Sequential Training Window (STW) scheme in the BMA method. This can be seen in the deterministic evaluation with MSE and the ensemble evaluation with the Verification Rank Histogram (VRH) and Continuous Rank Probability Score (CRPS). In the probabilistic evaluation, the CTW scheme has an Area Under Curve-Relative Operating Characteristic (AUC-ROC), which is larger than the STW scheme, and the smallest resolution attribute is indicated by the Brier Score (BS).