Analysis of bioclimatic time series and their neural network-based classification to characterise drought risk patterns in South Italy.

A new approach to characterise geographical areas with a drought risk index (DRI) is suggested, by applying an artificial neural network (ANN) classifier to bioclimatic time series for which operational temporal units (OtUs) are defined. A climatic database, corresponding to a grid of 8 km x 8 km cells covering the Italian peninsula, was considered. Each cell is described by the time series of seven variables recorded from 1989 to 2000. Sixteen cells were selected according to land cover homogeneity and completeness of the time series data. The periodic components of the time series were calculated by means of the fast Fourier transform (FFT) method. Temporal units corresponding to the period of the sinusoidal functions most related to the data were used as OtUs. The ANN for each OtU calculates a DRI value ranging between -1 and 1. The value is interpretable as the proximity of the OtUs to one of two situations corresponding to minimum and maximum drought risk, respectively. The former set (DRI = -1) is represented by an ideal OtU with minimum values of temperatures and evapo-transpiration, and maximum values of rainfall, normalised difference vegetation index (NDVI) and soil water content. The second set (DRI = 1) is represented by the reciprocal OtU to the former one. The classification of the cells based on DRI time profiles showed that, at the scale used in this work, DRI has no dependence on land cover class, but is related to the location of the cells. The methodology was integrated with GIS (geographic information system) software, and used to show the geographic pattern of DRI in a given area at different periods.