Title : Beyond the monsoon baseline: Evaluating transparent AI for coastal environmental governance in Bangladesh
Abstract:
Coastal governance in Bangladesh is mostly reactive. One reason is that ocean data come from many sources and are hard to turn into decisions. AI can help, but only if it is honest about how well it works. Many studies report high accuracy without comparing their models with a simple seasonal baseline. We built and tested an open and transparent decision support system (AI-DSS) for the northern Bay of Bengal (88-93°E, 20-23°N). Ocean reanalysis, satellite, weather, cyclone, and sea-depth data for 2003-2024 were placed on one 1/12° grid (1,261 cells, 332,904 cell-months, 32 predictors). Four models (XGBoost, LightGBM, random forest, and logistic regression) forecast two events one month ahead: a high chlorophyll-a bloom, used as a sign of water-quality stress, and a high-wave state, used as a sign of erosion pressure. The models were trained on 2003-2017 and tested on 2018–2024. We compared them with climatology and persistence. We explained them with SHAP, made risk maps, ran a warming scenario, and added a seven-part governance framework. The region warmed by 0.39 °C per decade (p = 0.0002), and sea level rose by 0.045 m per decade (p = 0.003). Extreme events also became more common: bloom months rose from about 10% to 14.1%, and high-wave months rose to 13.3%. On the test years, ROC-AUC was 0.718–0.744 for blooms and 0.635-0.770 for high waves. Climatology alone scored 0.733 and 0.762. So the models matched the seasonal baseline but did not clearly beat it. SHAP showed that sea level pressure and month dominated the high-wave model (66.8% of importance), which means the models mostly learned the monsoon cycle. A common spatial test inside the training years gave scores up to 0.984, against 0.728 for the same model on the time test. This shows how easily skill can be overstated. In the warming scenario (SST +1 °C, salinity −1 psu), the mean risk index rose from 0.093 to 0.106, with the highest bloom risk near the Meghna estuary. At a one-month lead, seasonal climatology is a strong benchmark. A defensible coastal AI-DSS must report skill against it, test over time and not only over space, show its uncertainty, and include a governance layer. In our system, four of seven safeguards are in place, while accountability and stakeholder participation are still proposed. We recommend using such outputs as advisory screening layers, for example, to decide where to focus field monitoring. They are not hazard warnings. The open and low-cost workflow can be repeated for other river deltas facing similar pressures.
Keywords: decision support system; coastal governance; tabular machine learning; Bay of Bengal; sustainability intelligence.
