<p>The shear-wave velocity structure of the subsurface plays a critical role in evaluating seismic site amplification. One established technique for estimating velocity structure is the spatial autocorrelation (SPAC) method, which involves array-based observations of ambient seismic noise. The classical SPAC method assumes a single-mode Rayleigh-wave field, although the assumption does not always hold. To account for higher Rayleigh modes, previous studies have extended the SPAC method by introducing an effective phase velocity. However, inverting for subsurface structure using effective phase velocities requires strong assumptions regarding the distribution of seismic sources and propagation paths. To our knowledge, there is no widely adopted SPAC-based methods for reliably treating wavefields that include higher modes. Here, we propose an approach for identifying frequency ranges where higher modes are present, based on the standard deviation of the effective phase velocities estimated from multiple array sizes and analysis methods at each frequency. Unlike conventional approaches, we use the effective phase velocity solely to identify the presence of higher-mode contributions, not for inversion. We develop a neural-network-assisted method that stably estimates phase velocities for multiple array sizes and analysis methods, and calculate the standard deviation. The method is first verified on a representative model, then validated with synthetic microtremor data, and finally tested on a broad set of subsurface-structure models from across Japan. The proposed method contributes to improving the accuracy of velocity-structure estimation for multimodal wavefields analysed using SPAC-based methods.</p> Graphical Abstract <p></p>