Advancing the prediction of water and carbon cycles with artificial intelligence and process-based model

The first example is a hybrid precipitation nowcasting model that combines a generative adversarial neural network (GAN) and a numerical weather prediction (WRF) model to generate more accurate rainfall predictions. The GAN-based model uses radar reflectivity data and static characteristics of the target basin to generate rainfall forecasts. In contrast, the physical process-based WRF model generates rainfall forecasts based on atmospheric physics. The two models are blended using a hybrid method to complement each other's limitations and produce improved results. The model was tested using a rainfall event in South Korea in 2018 and showed improved accuracy up to a lead time of 6 hours. 

The second example is a two-step deep learning method called FireDL in conjunction with NCAR CLM-BGC. FireDL sequentially uses Long Short-Term Memory (LSTM) and Artificial Neural Network (ANN) to predict fire duration and burned areas in Alaska. In the first step, LSTM is used to predict the fire occurrence and persistence conditions (i.e., fire duration) using lightning, vegetation, and climate datasets. In the second step, the total burned area is predicted using ANN with fire duration and climate datasets as input. FireDL showed reasonable performance in capturing large fires (>10,000ha). Additionally, the daily-scaled burned areas based on FireDL were implemented into NCAR CLM5-BGC to improve carbon emission prediction. 

In addition, Professor Kim will introduce the AI & Water Informatics Program in Civil and Environmental Engineering, Yonsei University, Seoul, South Korea.

We look forward to seeing you there!

Register by 22 August for the seminar and a light lunch.