Accurate estimation of soil moisture is essential for the successful simulations of boundary layer evolution, mesoscale circulations and convection. However, since soil moisture observations are insufficient for direct real-time initialization, we are adapting an Artificial Neural Network (ANN) model to this problem. Our goal is to develop the new methodology capable of estimate soil moisture with sufficient spatial and temporal resolution.

The ANN model was originally developed for precipitation estimation based on infrared (IR) satellite data. The ANN model is trained and tested for its performance in estimate soil moisture from combined LDAS skin temperature data and USGS composite NDVI data. Soil moisture information from LDAS model output is used as the target data to adjust the ANN parameters.

Preliminary training results show that soil moisture estimated from the ANN differs the most from the LDAS output in the dryer and moister regions. The correlation coefficient between the ANN estimate and LDAS output in soil moisture is 0.558. The closer the correlation coefficient value is to 1.0, the closer the ANN model estimate matches the trend of the LDAS data.

Data for a different date is selected to test (sequential training) the ANN model. The testing results show that the difference between the LDAS and ANN has been reduced and the correlation coefficient is improved to be 0.78. More testing with different days and different input variables are being conducted.