Estimation of maximum water storage in reservoirs using satellite data and machine learning methods

Abstract

Water reservoirs fulfill numerous functions for ecosystems and society. They are crucial for flood control, hydrological cycles, and biodiversity. Knowledge about the storage capacity is useful for adaptive water management, risk assessment, and hydrological modeling. In-situ measurements of reservoir storage capacity are challenging due to the high financial and temporal costs of field survey, while Earth observation sensors are incapable to penetrating deep or turbid water. If in-situ measurements exist, the public access to data is often limited. Although reservoir storage capacities are typically determined during reservoir planning and construction, globally accessible and consistent storage records remain incomplete or unavailable for many reservoirs. Therefore, modeling approaches are required to support adaptive water management where direct measurements or reliable public records are lacking.
This study proposes a novel deep learning approach to estimate the maximum reservoir storage capacity directly from spatial raster inputs, without requiring simulated data or explicit bathymetry reconstruction. The proposed model is based on the ConvNeXt Tiny architecture with a regression head and uses publicly available global digital elevation model (DEM) data, surface water frequency information, and reservoir information from the Global Reservoir and Dam (GRanD) database. The developed Convolutional Neural Network (CNN) learns spatial terrain and maximum water occurrence patterns associated with maximum reservoir storage capacity and predicts continuous storage values directly. The final model was developed through staged hyperparameter tuning and testing different oversampling strategies. The model was evaluated on 1,050 reservoirs from the GRanD database, spanning storage capacities from 0.09 to 1,750 million cubic meters (mcm). It achieved a mean absolute error (MAE) of 39.78 mcm and a mean absolute percentage error (MAPE) of 76.83%. On an independent dataset of 16 reservoirs with in-situ measurements in New York, the model achieved an MAE of 32.44 mcm and an MAPE of 52.04%. Prediction accuracy varied substantially between reservoirs and may be affected by uncertainties in the reference storage capacities used for model training and evaluation. Nevertheless, the results indicate that reservoir storage capacity can be estimated directly from globally available raster data. Compared with existing machine learning approaches, the proposed model is globally applicable, computationally lightweight, requires only a limited amount of publicly accessible input data, is fast to train and apply, and can be continuously expanded as new, more accurate data becomes available.
In addition, the study examined the relationship between predicted error and five reservoir characteristics: surface area, catchment area, maximum storage volume, average water depth, and coverage area ratio. No clear relationship between these characteristics and either the absolute error or the logarithmic error ratio was identified. A comparison of 59 reservoirs further showed that the CNN-based estimates deviated less from the reported GRanD storage capacities than estimates from an ellipsoid-based approach. Overall, the study demonstrates the feasibility of CNN-based methods for large-scale reservoir storage estimation and provides an alternative to traditional geometric approximations and bathymetry reconstruction methods for hydrological and water resource applications.

Popular Abstract

Estimating maximum reservoir volume with satellite data and machine learning*
Determining how much water a reservoir can store is challenging because direct measurements are difficult and expensive to obtain. Reservoirs are surface water bodies that are often associated with dams but can also include lakes. They play a crucial role in society and the environment by providing recreational opportunities for people and supporting biodiversity. Moreover, knowing the maximum reservoir volume is important for hydrological modeling, adaptive water management, e.g., for agricultural irrigation and power generation, and for risk assessment, e.g., when a dam is in danger of breaking.
Direct in-field measurements can measure water depth in a regular grid, but are related to high time and financial costs. These data are often lacking or not publicly available. Instead, over the years, different approaches were developed to estimate the reservoir volume. Traditional approaches are based on geometrical approximation, for instance, fitting the reservoir to a bowl shape (ellipsoid method) and calculating the reservoir volume by applying a geometrical formula. More recently, machine learning models were developed. Machine learning methods can learn from examples. By analyzing thousands of reservoirs, they can identify patterns that help estimate the maximum volume of other reservoirs. Unlike traditional methods, they do not require assumptions about the reservoir's shape and can learn various kinds of relationships.
In this study, the aim was to estimate the maximum reservoir volume as a single value by creating a machine learning model. The model used images containing information about the terrain surrounding reservoirs and the maximum extent of the water body. The model was created using around 5,200 reservoirs worldwide.
Throughout the process, various model settings were tested to support optimal learning. At the end, the model with the best performance was selected and applied to 1,050 reservoirs from a reference dataset and 16 reservoirs in New York with field measurements. The model produced reasonable estimates for reservoirs worldwide and showed similar performance when tested on independent reservoirs with field measurements. The results indicate that the model successfully learned relationships between landscape characteristics and reservoir volume, although uncertainties remain. Some of the uncertainty may originate from the training data itself, as many of the reservoir volumes used to build the model were estimates rather than direct field measurements. It can be expected that the model accuracy could be improved with a higher quality of the volume data on which the model was created.

Moreover, it was analyzed whether the model performance varies with reservoir characteristics. The deviation between the estimate and the reference volume in the dataset tended to increase with increasing volume. Apart from that, no clear relationship was visible.
Lastly, the machine learning model was compared to a model based on the geometrical approximation of a bowl shape. The machine learning model was closer to the reference volumes and showed, particularly for larger reservoirs, more accurate volume estimates.
Overall, the results demonstrate that satellite data and machine learning can be used to estimate maximum reservoir volume worldwide, providing a practical alternative where direct measurements are unavailable or difficult to obtain.
*Original title: Estimation of maximum water storage in reservoirs using satellite data and machine learning methods

• author: Höller, Anne Kathrin
• supervisor: Zheng Duan; Connor Chewning
• organization: Department of Earth and Environmental Sciences (MGeo)
• course: NGEM21 20261
• year: 2026
• type: H2 - Master's Degree (Two Years)
• subject: Earth and Environmental Sciences
• keywords: Geographical Information Science, Physical Geography, Deep Learning, Convolutional Neural Networks (CNN), ConvNeXt, Reservoir storage capacity estimation, Remote Sensing, Water reservoir, Digital elevation model (DEM)
• publication/series: Student thesis series INES
• report number: 758
• language: English
• id: 9241739
• date added to LUP: 2026-06-22 10:33:04
• date last changed: 2026-07-29 13:57:51