An Evolutionary Process-Embedded Spatiotemporal Interpolation Method for Marine Environmental Fields

In the geographic world, phenomena such as mesoscale ocean eddies exhibit continuous and gradual changes. Due to limitations in remote sensing observation technology, a contradiction exists between discrete observational data and these evolving phenomena. While spatiotemporal interpolation is crucial for bridging this gap, existing single-model methods fail to account for continuous process characteristics, making it difficult to obtain consistent datasets. To address this, this paper proposes an evolutionary process-embedded marine spatiotemporal interpolation model (EPMSIM) by integrating deep learning and geostatistics. EPMSIM first decomposes marine time-series fields into trend, seasonal, and evolutionary components using seasonal and trend decomposition using loess (STL). A convolutional bidirectional long short-term memory (ConvBiLSTM) model is designed to reconstruct the trend and seasonal components, while a process-based spatiotemporal dynamic tracking interpolation method (PSDTIM) reconstructs the evolutionary component. Finally, these components are additively coupled for interpolation. A case study on sea surface temperature (SST)-based mesoscale eddies shows that EPMSIM outperforms traditional geostatistical and deep learning-based baseline models in terms of root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and structural similarity index measure (SSIM). These results confirm the model’s effectiveness and feasibility in capturing the continuous evolution of marine phenomena and generating high-quality spatiotemporal datasets.