Perbandingan Metode Koreksi Bias Deep Learning dan Empirical Quantile Mapping pada Prediksi Sub-seasonal to Seasonal di Pulau Jawa
Date
2023-05-29Author
Budi Raharja, Adyaksa
Faqih, Akhmad
Mudzakir Setiawan, Amsari
Metadata
Show full item recordAbstract
Climate information and predictions on the sub-seasonal to seasonal (S2S) scale are important and useful for planning and decision-making in various fields such as disaster management, health, energy, hydrology and agriculture. However, until now, the prediction accuracy on the S2S scale is still relatively low compared to weather and seasonal predictions. The existence of bias and uncertainty from the global climate model (GCM) output is a factor causing the low accuracy of S2S predictions. Problems related to the bias of the GCM output can be overcome through bias correction. This study compares the performance of conventional bias correction methods, empirical quantile mapping (EQM) and DL based on convolutional neural networks (CNN) on S2S rainfall prediction.
Daily rainfall observation data from 477 stations and rain posts spread across the island of Java were selected based on the level of data availability of more than or equal to 80% for a period of 20 years (1999-2019). The rain post data is then used to build grid-based daily rainfall data with a resolution of 0.05x0.05° through multiple linear regression methods and residual kriging with CHIRPS satellite data predictors, latitude, longitude and elevation. The resulting grid data is then used to perform bias corrections on the output of the Sub-seasonal Experiments (SubX) model. The output of the SubX model consists of 7 (seven) individual models, including CCSM4, CFSv2, FIMr1p1, GEFS, GEPS6, GEOS_v2p1, and NESM. Each model has a member number that varies from 1 to 21.
The empirical quantile mapping (EQM), convolutional autoencoder (CAE) and UNET method is then applied to bias correct the SubX model output. The evaluation was carried out on aggregate weekly rainfall during the training period, namely 1999 – 2016, through kfold cross-validation with k=3, and during the testing or unseen data period, namely 2018 – 2019. The performance indicators used to evaluate the bias correction method are the correlation coefficient, root mean square error (RMSE) and continuous ranked probability skill scores (CRPSS). The DL-based bias correction method for the individual SubX model outputs generally performs better than the EQM method. It is indicated by the average value of the correlation coefficient and CRPSS in the output of the individual SubX- corrected DL model, which is generally higher than the results of the EQM correction. The performance advantage of this DL-based bias correction method consistently occurs, both in the training period and the testing period. In the individual SubX models, the DL-based bias correction method increased the correlation coefficient by up to 16%, decreased the average RMSE by 26.5%, and increased the average CRPSS by up to 95.1% in the training period. However, the EQM-based bias correction decreases the average correlation coefficient value in most models, reduces the average RMSE value to 18.5% and increases the CRPSS value to 68.3%. During the testing period, the DL-based bias correction method succeeded in increasing the average correlation coefficient value by up to 15%, reducing the average RMSE value by up to 6.9% and increasing the CRPSS value by up to 59%. However, in the EQM bias correction, the average correlation coefficient only increases to 3.5%, the average RMSE decreases to 8.7%, and the average CRPSS increases to 17%.
The lowland areas of the island of Java generally have better S2S prediction performance than the highlands. Relatively the same thing is shown by the results of corrected S2S prediction performance using the DL and EQM approaches. It is caused by the output characteristics of global climate models, which generally have a low spatial resolution, so they cannot capture orographic rainfall variability, which has not been properly handled by the bias correction method in this study.