A Unified MPEG–Transformer Framework for Error-Bounded Compression and High-Fidelity Reconstruction of Scientific Spatiotemporal Data

The rapid growth of scientific spatiotemporal data poses increasing challenges for efficient storage and transmission while preserving sufficient reconstruction fidelity for downstream analysis. Existing compression methods remain limited in this setting: lossless approaches often yield low compression efficiency, conventional lossy methods lack flexible local fidelity control, and learning-based schemes may introduce oversmoothing in reconstructed results. To address these issues, we propose a hybrid machine-learning framework for error-bounded compression of gridded scientific data. The proposed framework integrates MPEG-based temporal coding and transformer-based super-resolution reconstruction to exploit temporal correlation and spatial redundancy, and it introduces an error-bounded correction module to explicitly control local reconstruction errors. In addition, a lightweight Dense Residual Swin Transformer (DRCT)-based reconstruction model is employed to enhance long-range dependency modeling and multi-scale feature recovery. Experimental results on Copernicus Marine Service (CMEMS) gridded sea-level data demonstrate that the proposed framework achieves a favorable balance between compression efficiency and reconstruction quality. For 192×192 ADT data, the method reaches a peak PSNR of 38.14 dB with a compression ratio of 594×. With the error-bounded correction module enabled, the reconstructed values are first de-normalized using the recorded normalization parameters, and the local reconstruction error in the original floating-point physical-value domain can then be explicitly controlled by the prescribed absolute error threshold while maintaining a compression ratio of 295×. Additional experiments on SLA further indicate that the proposed framework is not restricted to a single sea-level variable. These results indicate that the proposed framework is a practical and effective solution for compressing large-scale scientific spatiotemporal data with controllable error and high-fidelity reconstruction.