Multi-Horizon Forecasting of Surface Water Across Australian Waterbodies

Australia's surface water swings between wet and dry extremes, yet existing satellite products map past behaviour rather than forecast it, leaving water managers to react rather than plan. We frame surface water dynamics as a multi-horizon forecasting problem, training one global LightGBM model per horizon to predict percentage wet surface area for all 313,389 Digital Earth Australia waterbodies at nine horizons from 1 to 36 months, and ablating four optional feature groups across all 16 combinations against a climatology baseline. Skill decays from 0.405 at one month to 0.119 at three years but stays positive throughout, showing that long-range forecasting is achievable. Per-waterbody seasonal normals contribute most but are not irreplaceable: models omitting them recover much of the same information from lags, rolling statistics and static attributes, and a three-group model comes within 0.008 skill of the full model. Feature importance shifts systematically from recent observations to seasonal normals as horizon lengthens.