Imputasi Curah Hujan ERA5 Menggunakan Random Forest dan XGBoost di Maluku Utara

Balya Badar Syah (Universitas Internasional Semen Indonesia)
Brina Miftahurrohmah (Universitas Internasional Semen Indonesia)

Progresif: Jurnal Ilmiah Komputer
Vol. 22 No. 3 (2026): Juli

Climate change increases hydrometeorological disaster risks in North Maluku, yet missing BMKG rainfall data and ERA5's spatial bias hinder precise risk analysis. This study applied point-based Statistical Downscaling using Random Forest (RF) and Extreme Gradient Boosting (XGBoost), with ERA5 precipitation as the single predictor, to impute missing rainfall at four stations. XGBoost achieved lower MSE and RMSE at three stations (Sultan Babullah: MSE 1017.29, RMSE 31.89, MAE 12.31, R² 0.0351; Gamar Malamo: MSE 513.96, RMSE 22.67, MAE 9.91, R² -0.0141; Oesman Sadik: MSE 163.59, RMSE 12.79, MAE 6.27, R² 0.0369), with NRMSE-based accuracy of 83.89 – 87.56%, though RF retained slightly lower MAE. RF was selected at Emalamo (MSE 163.75, RMSE 12.79, MAE 5.62, R² -0.0517, accuracy 85.78%). The models imputed 298 missing days, though the univariate predictor underestimated rainfall above 40 mm. The resulting continuous dataset offers a scientific basis for regional disaster mitigation planning. Key Word: ERA5; Imputation; Statistical Downscaling; Random Forest; XGBoost   Abstrak Perubahan iklim meningkatkan risiko bencana hidrometeorologi di Maluku Utara, sementara kekosongan data curah hujan BMKG dan bias spasial ERA5 menghambat analisis risiko presisi. Penelitian ini menerapkan Statistical Downscaling berbasis titik menggunakan Random Forest (RF) dan Extreme Gradient Boosting (XGBoost), dengan presipitasi ERA5 sebagai prediktor tunggal, untuk mengimputasi kekosongan data curah hujan di empat stasiun. XGBoost menghasilkan MSE dan RMSE lebih rendah di tiga stasiun (Sultan Babullah: MSE 1017,29, RMSE 31,89, MAE 12,31, R² 0,0351; Gamar Malamo: MSE 513,96, RMSE 22,67, MAE 9,91, R² -0,0141; Oesman Sadik: MSE 163,59, RMSE 12,79, MAE 6,27, R² 0,0369), dengan akurasi berbasis NRMSE 83,89 – 87,56%, meski RF tetap mencatat MAE sedikit lebih rendah. RF terpilih di Emalamo (MSE 163,75, RMSE 12,79, MAE 5,62, R² -0,0517, akurasi 85,78%). Model berhasil mengimputasi 298 hari data kosong, meski prediktor univariat underestimate curah hujan di atas 40 mm. Basis data historis kontinu menjadi landasan saintifik bagi mitigasi bencana daerah.