Modeling　inundation　patterns　resulting　from　compound　flooding　induced　by　tropical　cyclones　presents　significant　challenges　due　to　the　complex　interplay　of　drivers　and　features　affecting　inundation　mechanisms.　This　study　introduces　a　machine　learning　framework　designed　to　optimize　the　prediction　of　inundation　depth　by　balancing　model　performance,　computational　costs　and　efforts　for　input　data　retrieval.　Starting　from　a　comprehensive,　physics-informed　identification　of　the　potential　explanatory　variables,　including　features　that　capture　local　flood　dynamics,　as　well　as　topological　and　geographical　characteristics,　the　proposed　methodology　leverages　a　feature　selection　process　based　on　permutation　importance,　which　emphasizes　the　reduction　in　the　number　of　inputs　to　streamline　the　modeling　process　without　compromising　accuracy.　The　framework　has　been　tested　using　Hurricane　Harvey　as　a　case　study.　The　analysis　revealed　performance　in　inundation　depth　prediction　comparable　to　that　of　traditional　hydrodynamic　models　available　in　the　literature.　Results　demonstrated　that　focusing　on　the　most　informative　features　improves　both　model　performance　and　efficiency,　thus　highlighting　the　need　for　careful　feature　selection　for　region-specific　implementation　of　data-driven　approaches　for　inundation　depth　prediction.