The soil moisture information comes from a Puerto Rico soil moisture network, and from radar and satellite information. An empirical model was developed based on studying the behavior of the soil moisture response. Soil moisture exhibits spatial and temporal variability. It has been noted that the temporal variability for a given area exhibits long- and short-term variations that can be expressed by an empirical model. The major components of the empirical model are the long-and short-term memory. The long-term variability is mostly associated to long-term climatology patterns and modeled by an artificial neural network. The variables for every grid (1 km) used to train a neural network are: monthly soil moisture (in-situ observations), monthly accumulated rainfall (NEXRAD), monthly vegetation index (MODIS), monthly air temperature (MODIS), soil texture, elevation, and slope.

The short-term memory model is a nonlinear stochastic transfer function that estimates the soil moisture response in hourly basis for every grid. Estimation of soil moisture in an hourly response requires using the cumulative rainfall during the last week (NEXRAD), the hourly rainfall (NEXRAD), the hourly temperature estimated from MODIS and the initial level of soil moisture estimated from the neural network. The instantaneous soil moisture response is modulated by exponential factor that is applied to the impulse response function.

The empirical model was designed to estimate soil moisture at 2, 4, 8, 20 and 40 inches depth and was applied to Puerto Rico climate conditions. The proposed method may be implemented to a similar tropical region. The proposed method can be used to create the initial conditions of soil moisture for running atmospheric and hydrological models such as: the regional atmospheric modeling system (RAMS), the mesoscale model (MM5), the MIKE SHE and VFLOW hydrological models, etc. Cross-validation techniques show that the proposed algorithm is a potential tool to estimate soil moisture over densely vegetated areas.