Remote-sensing-assisted and in situ-constrained reconstruction of 4-hourly chlorophyll-a dynamics in inland lakes using vision Mamba and temporal attention

Chlorophyll a (Chl-a) is widely used to indicate eutrophication and bloom risk in inland waters, but daily Sentinel-3 observations alone cannot resolve transient sub-daily variability represented in partially observed 4-hourly in situ Chl-a records. This study therefore develops VGTA (Vision Mamba–GKOA–TCN-Attention), a staged remote-sensing-assisted and in situ-constrained framework that integrates daily Sentinel-3 OLCI-derived priors with partially observed 4-hourly in situ Chl-a sequences and binary observation masks. A 20-dimensional OLCI-derived feature space provides spectral–spatial priors for Chl-a estimation and bloom-risk state identification. GKOA optimizes TCN-Attention to reconstruct masked values within partially observed 4-hourly in situ Chl-a sequences. Vision Mamba outperformed CNN– and Transformer-based models in Chl-a estimation, achieving 94.8% accuracy, a Macro-F1 score of 0.909, and a Bloom-F1 score of 0.870. Ablation experiments confirmed the contributions of bidirectional contextual modeling, positional embedding, selective state-space modeling, and gating. Under Missing Completely at Random, block-missing, and real missing scenarios, GTA improved reconstruction performance by 12.71–55.54% relative to the best baseline models. Uncertainty increased once continuous gaps reached seven days, defining a practical boundary of data-driven reconstruction. Chl-a fluctuations and bloom-risk transitions can unfold within a day, and daily satellite snapshots may blur their timing or miss short events. Reconstruction of masked time steps within partially observed 4-hourly in situ Chl-a sequences can support earlier identification of bloom-risk transitions, more timely field sampling and management responses, and temporally finer monitoring of inland waters.