Empirical Reconstruction and a Pre-Specified GeoAI Evaluation Framework for Episodic Karst-Lake Drawdowns: Lake Jackson, Florida

Episodic karst-lake drawdowns can outpace field observations. This Technical Note combines a satellite reconstruction of Lake Jackson, Florida, with a prespecified framework for evaluating future drawdown-anticipation models. We processed 333 Landsat 5 TM and Landsat 7 ETM+ Collection 2 Level-2 candidate scenes from 1984–2003; 167 observations passed coverage and quality screening. During the 1999–2001 drawdown, open-water area fell from a pre-event mean of 1061 ha to 64.9 ha on 3 July 2000, a loss of 93.9%. Comparison with in situ stage across 56 paired observations yielded an empirical stage–area relationship (R2 = 0.67; RMSE = 136 ha). Coverage-threshold sensitivity and precipitation-lag screening informed the subsequent evaluation design. That design specifies four lake-surface classes, objective event extraction, statistical and machine-learning baselines, spatially structured challengers, leave-one-drawdown-event-out validation, and a prespecified model-promotion rule. Predictive model comparisons and the planned groundwater, graph, precipitation, soil-moisture, and morphometric sensitivity tests have not been executed. The demonstrated contribution is a satellite reconstruction with an independent stage comparison; predictive skill remains to be evaluated.