Banana holes are karst depressions that have primarily been reported from strandplains within the Bahamian archipelago. Banana holes have been hypothesized to form by downward dissolution in the vadose zone and in the phreatic zone by mixing dissolution and/or spatial variability in organic carbon inputs to the water table. While vadose models have been unable to explain overhanging roofs common in banana holes, phreatic models require anomalously high dissolution rates. In this study, we develop a new model for banana hole formation based on field observations, an airborne LiDAR (light detection and ranging) survey, and published geochemical data on San Salvador Island, Bahamas. We detected 3356 depressions in LiDAR data consistent with banana hole morphologies in Marine isotope substage (MIS) 5e and MIS 9/11 strandplains. No banana holes were detected in Holocene strandplains. All banana holes were found in swales between ridges. Of the banana holes found in MIS 5e strandplains, 109 had floor elevations between 6 and 19 m above sea level. These banana holes could not have formed in the phreatic zone because the MIS 5e freshwater lens reached a maximum elevation of 6 m above modern sea level. We also observed banana holes filled with water during wet seasons. While wells next to banana holes were tidal, pools in the banana holes were not, indicating pools were perched. Our observations, supported by hydrological models, suggest banana holes may form in the vadose zone in transient, perched aquifers on exposure zones. Runoff from ridges infiltrates through vadose fast-flow routes until encountering low permeability exposure zones, where flow is directed laterally. Dissolution by perching on exposure zones would create thin, laterallyextensive chambers radiating from injection points. Subsequent roof collapse results in their surface expression. Because dissolution occurs when waters become perched, chambers could form throughout low stands and may therefore not reflect rapid phreatic dissolution.