6533b7d2fe1ef96bd125ea52

RESEARCH PRODUCT

Neural Models for Rainfall Forecasting

subject

description

This chapter is focused on obtaining an optimal forecast of one month lagged rainfall in Spain. It is assessed by analyzing 22 years of both satellite observations of vegetation activity (e.g. NDVI) and climatic data (precipitation, temperature). The specific influence of non-spatial climatic indices such as NAO and SOI is also addressed. The approaches considered for rainfall forecasting include classical Auto-Regressive Moving-Average with Exogenous Inputs (ARMAX) models and Artificial Neural Networks (ANN), the so-called Multilayer Perceptron (MLP), in particular. The use of neural models is proven to be an adequate mathematical prediction tool in this problem due the non-linearity of the problem. These models enable us to predict, with one month foresight, the general rainfall dynamics, with average errors of 44 mm (RMSE) in a test series of 4 years with a rainfall standard deviation equal to 73 mm. Also, the sensitivity analysis in the neural network models reveals that observations in the status of the vegetation cover in previous months have a predictive power greater than other considered variables. Linear models yield average results of 55 mm (RMSE) although they need a large number of error terms (12) to obtain acceptable models. Nevertheless, they provide means for assessing the seasonal influence of the precipitation regime with the aid of linear dummy regression parameters, thereby offering an immediate interpretation (e.g. coherent maps) of the causality between vegetation cover and rainfall.