LSTM-Based Regional Sea Surface Temperature Prediction Model for Climate Projection in North Maluku

Abstract

North Maluku Province, as an archipelagic region, is highly vulnerable to the impacts of climate change, particularly the increasing frequency of extreme precipitation events, while facing limitations in observational data. This study aims to develop a regional sea surface temperature (SST) prediction model using a Long Short-Term Memory (LSTM) neural network to generate more accurate SST projections by applying bias correction to outputs from the BNU-ESM global climate model. The model was trained using daily SST data from 1980 to 2005, with ERA5 reanalysis data as the reference and BNU-ESM outputs as predictors. Evaluation on the test data showed strong performance, with the best model (1-layer LSTM, 80 epochs, batch size 1024, learning rate 0.003) achieving a Mean Absolute Error (MAE) of 0.56°C, Mean Squared Error (MSE) of 0.57, and Mean Absolute Percentage Error (MAPE) of 2.09%. These results demonstrate the model's reliability in capturing SST patterns and its potential use for precipitation modeling, early warning systems, and climate adaptation strategies in data-scarce regions.