USGS Flow Photo Explorer is Still Going and Growing!
The Flow Photo Explorer (FPE) platform continues to grow rapidly as a national resource for using imagery to monitor environmental conditions. As of early December 2025, FPE now supports more than 350 users, operating across more than 600 monitoring sites. The database has expanded to over 12 million images, 800,000 annotations, and approximately 160 trained models, reflecting accelerating engagement from federal, state, tribal, academic and nonprofit partners.
Please see two critical updates below. Thank you for your continued support and contributions, we’re looking forward to many exciting improvements in the year to come!
When will we train a model and produce a predicted relative streamflow hydrograph for your station?
- We have plans to operationalize a data release pipeline and pilot an automated modeling pipeline this coming year, but in the meantime, we are still running model batches every 4 to 6 weeks.
- We recommend that you collect image data at a station for at least 300 days, and then, once all the imagery is uploaded, annotate at least 1000 pairs per station before we train a model and produce a predicted hydrograph for your site.
- These guidelines ensure that the imagery and annotations capture a full range of seasonal conditions in at least one year. This will optimize ranking model performance and deliver the best results.
- These guidelines will be a moving target as the platform grows, we run more models, and learn more about how to deliver the best possible results. We will work on updating the website more frequently as our knowledge evolves over the coming year.
New Publication: Technical note: A low-cost approach to monitoring relative streamflow dynamics in small headwater streams using time lapse imagery and a deep learning model
Our recent Hydrology and Earth System Sciences paper describes the modeling framework that underpins Flow Photo Explorer and evaluates model performance compared to observed streamflow data. Highlights:
- We evaluated performance of the deep learning model at eight streamflow sites in western Massachusetts where low-cost game cameras are co-located with traditional stream gages. Drainage areas at the study sites range from 1.3 to 32.6 km2, and monitoring periods range from approximately 3 to 5 years.
- Kendall’s Tau rank correlation for model performance during the annotation period ranged from 0.6-0.83 with a median of 0.75, where 1 is perfect performance.
- We investigated the relationship between deep learning model performance and three factors likely to affect performance: observed streamflow variability, annotation accuracy, and camera stability. We found that sites where images captured a wider range of streamflow were easier to annotate accurately, resulting in better model performance, and that camera stability had a surprisingly small effect on performance.
- These patterns suggest that this method is more effective for monitoring streamflow dynamics in small streams with highly variable runoff response to precipitation and snowmelt.
- A scaling simulation determined that model performance improvements are limited after 1,000 annotation pairs, suggesting a target annotation requirement necessary for acceptable model performance.
📄 Read the full paper: https://hess.copernicus.org/articles/29/6445/2025/
|
