Yang et al. (2026) Efficient spectral-temporal reconstruction of long-term satellite time series via temporal segments and mask-informed embedding

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Date: 2026-05-09
• Authors: Jie Yang, Xin Huang
• DOI: 10.1016/j.rse.2026.115470

Research Groups

• School of Remote Sensing and Information Engineering, Wuhan University, China.

Short Summary

The paper proposes the Mask-informed Spectral-temporal Transformer (MISTR), a framework designed to reconstruct missing values in long-term satellite time series (STS) by utilizing mask-informed embeddings and adaptive spectral-temporal blocks.

Objective

• To develop a robust and computationally efficient method for the continuous reconstruction of long-term satellite time series that maintains spectral consistency and handles complex missingness caused by atmospheric interference and sensor limitations.

Study Configuration

• Spatial Scale: 30 Landsat ARD tiles distributed over China.
• Temporal Scale: 1997 to 2023.

Methodology and Data

• Models used: Mask-informed spectral-temporal Transformer (MISTR), featuring:
  • Mask-informed embedding (MIE) module: Generates embeddings robust to missing data.
  • Temporal-spectral adaptive (TSA) blocks: Models temporal dynamics and spectral dependencies using a gate mechanism for information flow control.
  • Temporal segments: Partitioning multi-spectral sequences into parallel, single-band segments.
• Data sources: Landsat ARD (Analysis Ready Data), with additional validation/adaptation using Sentinel-2 and MODIS data.

Main Results

• Reconstruction Performance: MISTR achieves competitive accuracy and shows consistent improvements over the strongest baseline models, particularly for multi-year time series.
• Efficiency: The framework demonstrates superior computational efficiency compared to existing Transformer-based architectures.
• Downstream Utility: The learned spectral-temporal representations are stable and transfer effectively to land cover classification, change detection, and vegetation phenology tasks.
• Versatility: The model exhibits cross-sensor flexibility (Sentinel-2 and MODIS) and supports joint-sensor collaborative reconstruction under sparse data conditions.

Contributions

• Introduces a novel partitioning strategy and a mask-informed embedding module to explicitly handle the high proportion of missing values in STS.
• Enhances the balance between spectral consistency and temporal dynamics through the TSA block's adaptive gate mechanism.
• Provides a scalable solution for generating seamless, analysis-ready long-term satellite records across different sensor platforms.

Funding

• Not specified in the provided text.

Citation

@article{Yang2026Efficient,
  author = {Yang, Jie and Huang, Xin},
  title = {Efficient spectral-temporal reconstruction of long-term satellite time series via temporal segments and mask-informed embedding},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115470},
  url = {https://doi.org/10.1016/j.rse.2026.115470}
}