Snow depth retrieval from passive microwave imagery using different artificial neural networks

The high degree of nonlinearity in the analysis of hydrologic systems demonstrates that artificial neural networks are suitable methods for this purpose. Artificial neural networks and passive microwave imagery have been combined for monitoring snow parameters, particularly in arid and semi-arid areas where the hydrologic process of the water basin is very dependent on the snow conditions (snow depth, snow water equivalent, snow density, snow cover area, snow stratigraphy, the shape of the crystals of snow). The multilayer perceptron is learned by different methods (Levenberg-Marquardt, scaled conjugate gradient, gradient descent with momentum and adaptive learning rate). These multilayer perceptrons are compared with radial basis function and multilayer perceptron–genetic algorithm for evaluating snow depth. Snow depth is estimated using passive microwave brightness temperature of a special sensor microwave/imager sensor. The results indicate that multilayer perceptron–genetic algorithm outperformed other artificial neural networks (multilayer perceptron–Levenberg-Marquardt, multilayer perceptron–scaled conjugate gradient, multilayer perceptron–gradient descent with momentum and adaptive learning rate, multilayer perceptron–genetic algorithm and radial basis function) with r = 0.97, RMSE = 0.18, and NSE = 0.83 for training and r = 0.90, RMSE = 0.19, and NSE = 0.39 for testing.