Improving the Efficiency of Neural Network Retrieval of Soil Moisture from Active Microwave Data

Soil moisture mapping by conventional ground based point sampling techniques, is a long and tedious process as soil moisture varies in space and in time. Furthermore, the dynamic of soil moisture is generally affected by the variation in soil surface characteristics such as: land cover, vegetation density, soil texture, and soil material. In the last two decades, various remote sensing techniques have been evaluated and proven to be a valuable source of information for different hydrological applications. Particularly, microwave remote sensing have proven to be an efficient alternative to traditional methods for estimating spatial soil moisture based on the large contrast between the dielectric properties of wet and dry soil.

The main purpose of this project is to find the optimal neural network configuration for soil moisture retrieval from active microwave data. An attempt has also been made to examine the effect of sub-pixels variability of land cover on the estimation of soil moisture. Indeed, a successful application of neural networks in satellite data classification requires a good understanding of the effect of some internal parameters related to the neural network structure and training process. Artificial neural networks have been widely applied to image processing, and have shown a great potential in the classification of a wide range of remote sensing data.

In this study, a back-propagation neural network has been used to estimate the surface soil moisture from Synthetic Aperture Radar (SAR) data from study area located in Oklahoma (97d35'W, 36d15'N). The soil moisture data measured by ESTAR Instrument (Electronically Scanned Thinned Array Radiometer) during the SGP97 campaign (operated by NASA) were used as truth data in the training and the validation processes.

Confusion matrices and Kappa Coefficients calculated from independent datasets have been used to evaluate the accuracy of the retrieved soil moisture as well as to select the best combinations of input data. Further, all the training and validation pixels (800 m resolution) have been labeled as either homogeneous or heterogeneous based on land cover type and number of sub-pixels of 25m resolution. The results showed the importance of neural network configuration parameters on classification accuracy and especially these related to the training process. Furthermore, the same results showed that homogeneous pixels are more likely to have better accuracy than heterogeneous pixels in soil moisture classification.