Accurately predicting river flow is critical for managing water resources, especially in dry regions like Arizona. Computer models that forecast river flow often rely on simplified information, using a single average value for an entire river basin, which can miss important details on the ground.

In this study, we tested newer, more advanced computer models to see if we could improve predictions for over 400 rivers across Arizona. We compared a standard model with a more powerful one called a Transformer. We also experimented with giving the models detailed satellite maps of the landscape to see if they could use that rich information more effectively than a simple average.

We found that the newer Transformer model was consistently much more accurate, representing a significant improvement for forecasting. While using satellite data helped, we discovered that getting the models to take full advantage of the detailed, map-like information is still a key challenge. Our work helps create better forecasting tools for Arizona and guides future research toward making these tools even smarter.