Hyperspectral remote sensing has the proven ability to provide bathymetric estimates over relatively clear water. High turbidity caused by suspended sediments scatters light, decreasing light penetration to depth and is thus regarded as the most confounding factor in remote sensing-based bathymetry. Therefore, the effect of varying sediment-laden flow conditions on remote sensing-based bathymetry measurements was investigated in this study to assess the applicability of hyperspectral imagery and a robust method was proposed to account for such conditions. To acquire spectral data under various suspended sediment and bottom conditions, experiments involving the injection of sand and yellow loess were performed in outdoor field-scale channels with three bottom conditions (coarse sand, vegetation, and fine sand) under shallow stream conditions (< 1 m). Using the acquired data, the random forest with recursive feature elimination (RF-RFE) method was used to select and learn the optimal spectral band combination for each bottom type. The results were then compared with those of the optimal band ratio analysis (OBRA) method, which has been most widely used for bathymetry measurements via remote sensing imagery. The reflectance spectra differed depending on the bottom type although they were similar regardless of the bottom type when high reflectance was distributed at high concentrations. Given such optically complex data, OBRA exhibited very low accuracy because it was substantially affected by suspended sediment. However, RF-RFE showed an accuracy of R2 ≥ 0.95 and mapped a stable bathymetry distribution under sediment-laden flow conditions by utilizing 6–12 spectral bands as the optimal combination, depending on the bottom type. This result demonstrated that although the contribution of bottom reflectance was weak when the suspended sediment concentration was high, hyperspectral remote sensing is still applicable if spectral bands of various wavelengths can be incorporated to account for nonlinearity. While the effectiveness of this method could be limited to the specific conditions of suspended sediment and bottom properties, it is expected to be more robust to various confounding factors after being trained under diverse spectral conditions.