Estimating depth of water columns in reservoirs, referred to as bathymetry, is an important task in water resource management. Satellite-derived bathymetry (SDB), where depths of water columns are estimated using satellite data, is an emerging and promising technique, especially for shallow-water scenarios. In this work, our objective is to obtain deep-learning models for SDB of a shallow water reservoir in Western India using supervised learning. For accomplishing this, after a series of pre-processing steps, we generate training data consisting of band reflectance values provided by Sentinel-2 satellite as input features and corresponding depth values collected by acoustic bathymetry as the ground truth data. The main contributions of the work include obtaining deep learning models that perform better than other classical techniques in terms of estimation accuracy, towards which we first obtain one dimensional (1D) convolutional neural network (CNN) and 2D-CNN models, which estimate depth using satellite data in tabular and image formats respectively. We then propose a hybrid model where we first classify each pixel of the image to correspond to high or low depth values and then perform regression on each of them separately. From the study, we show that (i) considering all the available bands of the Sentinel-2 satellites or those having high correlation with the ground truth depths gives significant performance improvement over considering only a subset of the bands having resolution of 10 meters, (ii) the 2D-CNN model trained using the image dataset yields significant performance improvement compared to other models trained on tabular data, (iii) and that the hybrid model which uses 2D-CNN for both the classification and regression tasks, trained on image data, performs best among all the models.