| The characterization and modeling of the spatial variability of soil moisture is an important problem for various hydrological, ecological, and atmospheric processes. A compact representation of interdependencies among soil moisture distribution, mean soil moisture, soil properties and topography is necessary. This study attempts to provide such a compact representation using two complimentary approaches.; In the first approach, we develop a stochastic framework to evaluate the influence of spatial variability in topography and soil physical properties, and mean soil moisture on the spatial distribution of soil moisture. Topography appears to have dominant control on soil moisture distribution when the area is dominated by coarse-texture soil or by mixed soil with small correlation scale for topography (i.e., small λZ). Second, soil properties is likely to have dominant control on soil moisture distribution for fine-texture soil or for mixed soil with large λ Z. Finally, both topography and soil properties appear to have similar control for medium-texture soil with moderate value of λ Z.; In the second approach, we explore the recent developments in Artificial Neural Network (ANN) to develop nonparametric space-time relationships between soil moisture and readily available remotely sensed surface variables. We have used remotely sensed brightness temperature data in a single drying cycle from Washita '92 Experiment and two different ANN architectures (Feed-Forward Neural Network (FFNN), Self Organizing Map (SOM)) to classify soil types into three categories. The results show that FFNN yield better classification accuracy (about 80%) than SOM (about 70% accuracy). Our attempt to classify soil types into more than three categories resulted in about 50% accuracy when a FFNN was used and even lesser accuracy when a SOM was used. To classify soil into more than three groups and to explore the limits of classification accuracy, this study suggests the use of multiple-drying-cycle brightness temperature data. We have performed several experiments with FFNN models and the results suggest that the maximum achievable classification accuracy through the use of multiple-drying-cycle brightness temperature is about 80%. It appears that the requirement of rapidly changing decision boundary, in the case of space-time evolution of brightness temperature over large areas, will restrict the FFNN model to yield better accuracy. Motivated by these observations, we have used a simple prototype-based classifier, known as 1-NN model, and achieved 86% classification accuracy for six textural groups. A comparison of error regions predicted by both models suggests that, for the given input representation, maximum achievable accuracy for classification into six soil texture types is about 94%. (Abstract shortened by UMI.)... |
