Ocean salinity regulates seawater density, water-mass transformation, and large-scale circulation, yet its spatio-temporal variability remains difficult to project because in situ observations are sparse and often collected along drifting platforms. In coastal and offshore waters, strong gradients driven by river discharge, tidal mixing, and mesoscale transport further complicate projection from irregular observations. We develop a geometric deep learning framework for projecting salinity variability from sparse drifter observations across regional and global scales. By linking observations through spatial proximity and along-trajectory continuity within a grid-based representation, our solution captures both local interactions and broader spatial organization. Applied to multi-year global drifter archives and high-resolution regional datasets, it reduces mean absolute error by up to 70% relative to classical interpolation and deep learning baselines, with improvements exceeding 50% in dynamically complex nearshore regions. The projected salinity fields also retain variability relevant to chlorophyll dynamics, highlighting the value of our approach for marine monitoring.