Hyperspectral imagery is a promising tool for bathymetric retrieval because that might be more robust than traditional imagery. However, in fluvial sciences, this approach has only been applied to short reaches for which sparse bathymetric information is available and studies extrapolating such models to longer reaches are lacking. In this paper, we use linear regression and random forest (RF) regression to derive bathymetric estimates from hyperspectral data acquired under two different flow conditions: an unmanned aerial vehicle (UAV) acquisition at 27 m3.s−1 (base flow) and an airplane acquisition at 127 m3.s−1 (mean flow) for reaches of 0.35 km and 20 km, respectively. Using a 2D bathymetric model based on a green LiDAR acquisition, we generate calibration and validation data for the 20 km reach and the two flow conditions. First, we show that the airplane dataset is able to retrieve base flow bathymetry with a similar accuracy to the UAV acquisition (5–10 cm) despite having been acquired at a higher discharge. Then, we show that we can extrapolate bathymetric models calibrated on a smaller reach to the 20 km study reach with a loss in accuracy compared to an RF model calibrated on the study reach (20–30 cm vs. 5–10 cm). By analysing the errors from our models, we show that the accuracy of the extrapolated models dropped due to an underprediction of deep pool areas and changes in the optical properties of the water column and substratum across the study reach, with some spectral bands performing more poorly. The better performance of the random forest calibrated on the study reach led us to suggest using multi-band machine learning models when attempting bathymetric predictions over long river corridors, and highlights the need to cover the range of environmental conditions in the target reach when acquiring field data.