Seagrass beds underpin coastal biodiversity and deliver critical ecosystem services, including shoreline stabilization, habitat provision and blue carbon storage. Despite their growing importance for coastal conservation and climate mitigation, seagrasses remain difficult to map consistently in optically shallow waters, where satellite reflectance is strongly modulated by tidal stage, water-column optical properties, water depth and benthic background. Here we develop a tide-conditioned temporal deep-learning framework for high-resolution mapping of coastal seagrass beds from multi-temporal 3-m PlanetScope imagery. The framework combines multispectral features and spectral indices with tidal-state and local water-condition variables, enabling the model to account for observation-dependent variability rather than treating multi-date spectral differences as seagrass change. A convolutional recurrent architecture with temporal attention integrates neighbourhood-scale spatial–spectral information and observations acquired under different tidal conditions to estimate pixel-level seagrass occurrence probability. The model is further refined through weighted fine-tuning using existing reference data, additional manually interpreted seagrass and non-seagrass labels, and object-derived pseudo-positive samples from segmentation objects intersecting manual seagrass observations. Across validation and case-study analyses, the framework generated spatially coherent seagrass probability maps and extracted high-resolution seagrass distributions under variable tidal and shallow-water conditions. These results demonstrate the potential of combining high-resolution satellite observations, tide-condition information and temporal deep learning to produce scalable spatial data for coastal seagrass mapping, monitoring applications and subsequent blue carbon assessment.