Hydrographic Predictors of Small Pelagic Fish Acoustic NASC in Jinhae Bay, Korea: A Two-Part Random Forest Model with Spatially Blocked Validation

Nautical area scattering coefficient (NASC) data from small coastal acoustic surveys are strongly zero-inflated and spatially autocorrelated. We analysed 3324 integration intervals from 19 surveys in Jinhae Bay, Korea: May–September 2023 and March–September 2024–2025. Signals meeting a 38–120 kHz backscattering-strength-difference mask were treated as small pelagic fish and integrated at 38 kHz. Concurrent set-net and eDNA sampling showed a multi-species assemblage. Four predictor sets combining hydrographic, seasonal, spatial and topographic variables were compared using a two-part random forest under nested cross-validation with 1500 m spatial blocks. Zeros comprised 67.39% of observations. Hydrographic models achieved detection AUCs of 0.874–0.882 and positive-intensity log-scale R2 values of 0.325–0.330, versus 0.701 and 0.062 for topography plus month. Drop-column analysis showed that surface salinity carried the unique hydrographic signal and that surface temperature was redundant with salinity and month; topographic variables added nothing once coordinates were included. Survey identity alone reproduced the full-model AUC (0.882 versus 0.885): spatial-block performance largely reflects discrimination between surveys, not local interpolation. Leave-one-survey-out and leave-one-year-out validation fell to 0.503 ± 0.122 and 0.541 ± 0.123. Detection was more predictable than intensity, but the apparent skill is largely between-survey; the model should not predict unobserved occasions or estimate abundance.