Information on water depth in river channels is important for a number of applications in water resource management but can be difficult to obtain via conventional field methods, particularly over large spatial extents and with the kind of frequency and regularity required to support monitoring programs. Remote sensing methods could provide a viable alternative means of mapping river bathymetry (i.e., water depth). The purpose of this study was to develop and test new, spectrally based techniques for estimating water depth from satellite image data. More specifically, a neural network-based temporal ensembling approach was evaluated in comparison to several other neural network depth retrieval (NNDR) algorithms. These methods are described in a manuscript titled "Neural Network-Based Temporal Ensembling of Water Depth Estimates Derived from SuperDove Images" and the purpose of this data release is to make available the depth maps produced using these techniques. The images used as input were acquired by the SuperDove cubesats comprising the PlanetScope constellation, but the original images cannot be redistributed due to licensing restrictions; the end products derived from these images are provided instead. The large number of cubesats in the PlanetScope constellation allows for frequent temporal coverage and the neural network-based approach takes advantage of this high density time series of information by estimating depth via one of four NNDR methods described in the manuscript: 1. Mean-spec: the images are averaged over time and the resulting mean image is used as input to the NNDR. 2. Mean-depth: a separate NNDR is applied independently to each image in the time series and the resulting time series of depth estimates is averaged to obtain the final depth map. 3. NN-depth: a separate NNDR is applied independently to each image in the time series and the resulting time series of depth estimates is then used as input to a second, ensembling neural network that essentially weights the depth estimates from the individual images so as to optimize the agreement between the image-derived depth estimates and field measurements of water depth used for training; the output from the ensembling neural network serves as the final depth map. 4. Optimal single image: a separate NNDR is applied independently to each image in the time series and only the image that yields the strongest agreement between the image-derived depth estimates and the field measurements of water depth used for training is used as the final depth map. MATLAB (Version 24.1, including the Deep Learning Toolbox) source code for performing this analysis is provided in the function NN_depth_ensembling.m available on the main landing page for the data release of which this is a child item. To develop and test this new NNDR approach, the method was applied to satellite images from the Potomac River near Brunswick, MD, acquired in July and August of 2021. Field measurements of water depth available through another data release (Duda, J.M., Greise, A.J., and Young, J.A., 2020, Potomac River ADCP Bathymetric Survey, October 2019: U.S. Geological Survey data release, https://doi.org/10.5066/P9GOZZYX) were used for training and validation. The depth maps produced via each of the four methods described above are provided as GeoTIFF files, with file name suffixes that indicate the method employed: Potomac_mean-spec.tif, Potomac_mean-depth.tif, Potomac_NN-depth.tif, and Potomac-single-image.tif. The spatial resolution of the depth maps is 3 meters and the pixel values within each map are water depth estimates in units of meters.