Early Career Researcher Presentation 1: Machine learning in the physical sciences: applications in hydrology
Katherine Breen (Institute of Pure and Applied Mathematics (IPAM))
In the geosciences, it is oen necessary to simulate physical processes that are impractical or impossible to observe directly such as subsurface water ux. e classic approach to simulation and prediction for many years has been development, calibration, and validation of physics-based models; however these approaches have notable limitations including introduction of user and model bias, sparse observational datasets, not to mention computational burden. Over the past few decades applications of articial intelligence, most notably deep neural networks (DNNs), have grown in popularity due to advances in computer hardware in conjunction with the availability of remotely sensed data sets. A DNN is a layered network of densely interconnected information-processing nodes trained to recognize the relationship between model inputs and desired outputs to make predictions as a weighted linear combination of inputs.
Here, a DNN was used to map input parameters (soil/land-use characteristics, weather) from the Soil & Water Assessment Tool (SWAT) to remotely sensed soil moisture from NASA’s Soil Moisture Active Passive (SMAP) satellite. e objective of this research was to accurately predict regional soil moisture corresponding to temporally synchronous weather observations/forecasts using a DNN. Predicted soil moisture may then be used as a parameter for risk assessments (e. g. ooding and crop viability), providing near-real-time, high-resolution results.