High Temporal Resolution Rainfall Information Retrieval from Tipping-bucket Rain Gauge Measurements

Abstract

Disdrometer can play a vital role in restoring detailed rainfall process by providing rainfall at a high temporal resolution. Rainfall rate derived from the widely used “Tipping-Bucket rain gauge” usually neglects its temporal variation especially during the low rainfall intensity periods. This study explores a heuristic artificial neural networks (ANN) approach along with the conventional Cubic Spline Algorithm (CSA) and Multivariate Linear Regression method (MLR) for high temporal resolution rainfall rate retrieval for the period of 2007 to 2009 at Chilbolton, U.K. The Supervised Levenberg-Marquardt backpropagation algorithm and the K-folds cross-validation method are integrated in a feed-forward neural network as to implicitly detect complex nonlinear relationships and to avoid model overfitting. Results indicate ANN is performing equivalently well with CSA after training, however, with poor generalisation in test due to low correlation between input and target data, as well as the curse of dimensionality in optimum model complexity selection. MLR can be an alternative approach in rainfall rate estimation but it highly depends on the data quality.