Tile-Based Recursive Forecasting of Satellite-Derived Sea Surface Temperature Using ConvLSTM

With advances in satellite remote sensing and deep learning, high-resolution Sea Surface Temperature (SST) forecasting has become increasingly feasible over large spatial domains.However, such applications require substantial computational resources, while achieving reliable forecast performance remains a significant challenge.To address these limitations, this study proposes a tile-based framework to apply a Convolutional Long Short-Term Memory (ConvLSTM) model for high-resolution SST forecasting over large spatial domains.Tile-based training and forecasting with overlap-based stitching were used to reduce computational memory requirements and reconstruct continuous SST fields.Then, spatial bias correction was applied to reduce systematic forecast errors.Model performance was evaluated through recursive forecasting for lead times up to 30 days, with statistical variability assessed using bootstrap-based 95% confidence intervals.After spatial bias correction, the overall root mean squared error (RMSE) decreased by approximately 29%, from 1.5974°C to 1.1342°C, while the coefficient of determination (R²) increased from 0.8606 to 0.9204.The D+01 forecast achieved an RMSE of 1.13°C and an R² of 0.9163.Performance remained relatively stable through D+07, with RMSE below 2.15°C and R² above 0.65, but degraded substantially after D+14.These results demonstrate the tile-based framework's computational scalability for high-resolution SST forecasting over a large spatial domain while reducing the Graphics Processing Unit (GPU) memory requirements associated with full-domain processing.This framework is expected to support near-real-time ocean monitoring, fisheries disaster response, and marine and fisheries policy applications.