Machine-learning bias correction cuts Ina-Flows SST forecast error by 29.83% at Karimun Jawa and 26.10% at Bira, with the best algorithm differing by site
Synopsis
Using MAWS observations, this study evaluated the bias in BMKG Ina-Flows sea surface temperature forecasts and compared three machine-learning bias-correction algorithms (SVR, LSTM, Bi-LSTM) at Karimun Jawa and Bira, finding a warm bias at both sites, with LSTM best at Karimun Jawa (RMSE 0.207, MAE 0.182, MBE -0.171, a 29.83% RMSE reduction) and Bi-LSTM best at Bira (RMSE 0.218, MAE 0.173, MBE 0.011, a 26.10% RMSE reduction), indicating that correction effectiveness is location-dependent and algorithm choice should rest on local validation.
Interpretation
The study confirms a systematic warm bias in BMKG Ina-Flows sea surface temperature forecasts at both the Karimun Jawa and Bira MAWS sites. Prior evaluation of Ina-Flows output bias had not been carried out at these two specific sites; this work grounds the bias diagnosis in local observations. Based on paired Ina-Flows output and MAWS observations from 1 December 2023 to 7 December 2024, with MBE and related metrics quantifying the direction of bias.
Machine-learning bias correction substantially reduces Ina-Flows SST forecast error: LSTM lowers RMSE to 0.207 at Karimun Jawa (a 29.83% reduction), and Bi-LSTM lowers RMSE to 0.218 at Bira (a 26.10% reduction). The work compares SVR, LSTM, and Bi-LSTM side by side within one data framework and reports the best choice per site, rather than presenting a single model. 1–7 December 2024 serves as an independent test period, with RMSE, MAE, and MBE as evaluation metrics and data up to 30 November 2024 used for training and validation.
The best bias-correction algorithm varies by location: LSTM suits Karimun Jawa while Bi-LSTM suits Bira, showing that correction effectiveness depends on site characteristics. This comparison frames 'algorithm selection requires local validation' as an actionable conclusion rather than assuming one algorithm is universally best. The two sites show different best algorithms and different MBE signs (Karimun Jawa -0.171, Bira 0.011) under the same time window and metric set.
Perspective
The results target operational and research users of Ina-Flows output and MAWS observations, under the data conditions of the Karimun Jawa and Bira sites from December 2023 to December 2024. What it supports is training machine-learning correction models on local observations and selecting algorithms per site, rather than treating any single algorithm as universally best. For teams aiming to extend the correction workflow to other MAWS sites, the work offers a reproducible evaluation framework and metric combination.
The test period covers only 1–7 December 2024, so whether the correction holds over longer periods and across seasons remains to be seen. Whether the two-site conclusions extend to other sea areas or MAWS sites requires further local validation. In addition, the available text is the abstract and references, lacking details from the body such as feature construction, hyperparameter settings, and per-site error decomposition, so judging the reproducibility of the method details remains an open question.