Advancing Streamflow Prediction by Integrating Diverse Data and AI Models
Our AI model integrating remote sensing images, meteorological forcing sequences, and static catchment attributes accurately predicts streamflow at 531 basins across the CONUS.
Dan Lu
Accurate streamflow prediction is crucial to understanding climate impacts on water resources and developing effective adaption strategies. A global Long Short-Term Memory (LSTM) model, using data from multiple basins, can enhance streamflow prediction particularly at ungauged basins. However, obtaining detailed basin attributes poses a significant challenge. This study introduces a novel Geo-RS-LSTM method that utilizes vision transformer model to extract basin attributes from widely available remote sensing images. These spatial attributes are then integrated with hydrometeorological data and static attributes into the LSTM model to advance streamflow prediction. The method has demonstrated superior accuracy in predicting streamflow at “ungauged” basins under “future” climate conditions, outperforming several baseline models.
This innovative method significantly advances land surface modeling by effectively synthesizing temporal meteorological data with static, spatial catchment attributes to enhance streamflow prediction. By integrating diverse datasets and leveraging their insights, it provides a more comprehensive and effective tool for understanding environmental responses to climate change. This approach represents a major step forward in improving our ability to predict water resources in a changing climate.
Accurate streamflow prediction is vital for understanding the impacts of climate change on water resources and for developing effective adaptation strategies. Recent studies have demonstrated that a global Long Short-Term Memory (LSTM) network excelled in predicting streamflow at ungauged basins and modeling hydrological extremes. However, the effectiveness of the global LSTM model largely relies on its ability to assimilate diverse attribute data from multiple basins. To address the challenges of unavailable, missing, or homogeneous attributes from the current database, we developed a vision transformer model that extracts spatially informed catchment attributes from widely available remote sensing (RS) data. Moreover, we introduced the Geo-RS-LSTM method, which combines meteorological sequences, static attributes, and spatial attributes derived from remote sensing images to enhance streamflow prediction. Applied to 531 basins across the CONUS, our innovative method outperformed state-of-the-art models in both temporal and spatiotemporal extrapolation scenarios. This approach holds promise for widespread application in various environmental science fields, particularly where static catchment attributes are missing or where a deeper analysis of attributes could improve predictions.