Welcome to the Flow Photo Explorer
The Flow Photo Explorer (FPE) is an integrated database, machine learning, and data visualization platform for monitoring streamflow and other hydrologic conditions using timelapse images.
The goal of this project is to develop new approaches for collecting hydrologic data in streams, lakes, and other waterbodies, especially in places where traditional monitoring methods and technologies are not feasible or cost-prohibitive.
Want to add your photos or help annotate? Request an account to upload your photos or help annotate photos uploaded by other users.
Questions or Feedback? You can reach us at ecosheds@usgs.gov.
Video produced by the USGS MD-DE-DC Water Science Center
FPE is a collaboration between U.S. Geological Survey (USGS), U.S. Environmental Protection Agency (USEPA), Walker Environmental Research, Microsoft Research, and many contributing partners. Funding was provided by USGS, USEPA, and National Geographic Society. See About for more information.
- April 16, 2025: Preprint, FPE Growth, Annotator Thank You, Model Development Report
- April 18, 2024: Interface Updates, Model Results
A preprint is available for a new publication entitled A low-cost approach to monitoring streamflow dynamics in small, headwater streams using timelapse imagery and a deep learning model. This paper presents an evaluation of FPE model performance at 11 camera sites colocated with USGS reference gages in western Massachusetts.
Model predictions are now available at 84 public stations on the Photo Explorer. While most of these stations focus on streamflow predictions, there is one station in Alaska where the model is being used to track a seabird colony and a few stations in the midwest for monitoring harmful algal blooms (HABs).
As part of the USEPA-funded ROAR project, we completed an in-depth exploration of the model methodology and performance, and evaluated the potential for (1) adding weather data as another input to the model and (2) applying pre-trained models between sites. You can read more about this in the Year 1 ROAR Project Report.
Lastly, we released our second newsletter where you can find additional information about all these updates and more. To sign up for the our email newsletter, click here.
FPE uses an artificial intelligence/machine learning (AI/ML) deep learning model to estimate relative streamflow using timelapse imagery. The model is trained using pairs of images for which a person (a.k.a. an annotator) has selected which of the two images in each pair appears to have more flow. From this, the model learns how to sort the images from lowest to highest apparent flow. The rankings of the sorted images then serve as indicators of the relative amount of streamflow.
See the following publications to read more about our methodology.
Goodling, P., Fair, J., Gupta, A., Walker, J., Dubreuil, T., Hayden, M., and Letcher, B. (2025). A low-cost approach to monitoring streamflow dynamics in small, headwater streams using timelapse imagery and a deep learning model. EGUsphere [preprint]. https://doi.org/10.5194/egusphere-2025-1186
Walker, J.D. (2025). Low-cost Streamflow Monitoring using Timelapse Imagery and Machine Learning Models. Year 1 Final Report, USEPA Regional-ORD Applied Research Program (ROAR) Project #2554. Prepared for US Environmental Protection Agency Office of Research and Development. April 3, 2025. https://doi.org/10.5281/zenodo.15133342
Gupta, A., Chang, T., Walker, J., and B. Letcher (2022). Towards Continuous Streamflow Monitoring with Time-Lapse Cameras and Deep Learning. In ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies (COMPASS) (COMPASS '22). Association for Computing Machinery, New York, NY, USA, 353–363. https://doi.org/10.1145/3530190.3534805
Phase I (2020-2022): a database and cloud-based data pipeline was developed for storing, managing, and accessing timelapse imagery of streams and rivers along with associated flow and stage data. The system allows registered users to upload and manage their own photos and (optionally) flow data at multiple locations. The images and flow data are accessible through the Photo Explorer, which provides an interactive and exploratory interface for viewing the timelapse imagery coupled with observed flow data. The images and flow data in the FPE database will serve as the primary data source for developing and training the machine learning models in Phase II. All images are screened for the presence of Personal Identifying Information (PII) using the MegaDetector object detection model.
Phase II (2023-2024): the photos and data uploaded to FPE were used to develop deep learning models for estimating relative flow and other hydrologic parameters using timelapse imagery. See the Deep Learning Model section on the right or Gupta et al., 2022 for more details.
During this phase, models were developed for:
- USGS West Brook Study Area: the initial set of models will be trained on stations located in the West Brook study area in central MA. This area has been the focus of a long-term research project on fish population dynamics, and is where the idea for FPE originated. Most of these stations also have streamflow gauges collecting observed flow data that can be used to directly evaluate model performance.
- Other USGS Stations: the next batch of stations to receive models will include USGS stations outside the West Brook, beginning with those having a co-located streamflow gauge.
- USEPA and Other Collaborators: following roll out across the USGS, the model will be trained at stations for the USEPA and its collaborators through the ROAR project, which has provided core funding for FPE.
- All Stations: lastly, once the model framework and training methodology is sufficiently robust, the model will be made available for training at any FPE station.
Phase III (ongoing): as FPE continues to grow and expand, we are adding many new features and capabilities including:
- Support for training models to predict other hydrologic parameters such as water level, ice cover, snow depth, algal biomass, and more.
- Support for real-time image transmission and model predictions using cell-enabled cameras.
- Packaging the model for local training and prediction, outside of the FPE platform.
- Development of a statistical methodology for converting the model output, which is currently a relative measure of streamflow to actual streamflow in cubic feet per second.
- Development of classification models for identifying flow/no flow or ice/no ice conditions.
- In-the-field deployment of the FPE model using edge computing and satellite data transmission.