Historical reconstruction of floodplain inundation in the Pantanal (Brazil) using neural networks

The relations between hydrology, geomorphology and vegetation provide the basis for understanding the ecological processes in the floodplains of rivers. In this paper, long-term (1969–2009) local flood characteristics (magnitude, duration, frequency, return period) are quantified for the floodplain of the Cuiabá River (Pantanal wetland in Brazil), using a predictive model based on artificial neural networks (ANNs) with input given by the historic river-level record. Morphological features were then described and their relation to vegetation analyzed and used to reconstruct floods. In addition to the river-level data used to train and validate a three-layer ANN, data were also used from 11 floodplain gauges across a 12 km transect lateral to the river, which were monitored during the annual floods of 2004–2007. The resulting neural network gave satisfactory estimates of water depth over the floodplain, with an absolute error of 0.09 m not exceeded in 95% of occurrences. The ANN showed that in almost all years, the level reached by the river was high enough to flood the profile completely: the exceptions were 1971 and 2001 when only 50% and 58% was flooded, to mean depths of 0.34 and 0.48 m, respectively. The highest flood was that of 1995 when the floodplain was flooded to a mean depth of 2.56 m. The median flood event (return period 2 years) produced a mean flood depth 1.80 m and lasted 119 days. Among the most important morphological features is the presence of palaeo-channels which provide hydrological connectivity between the river and the floodplain. The distribution of phytophysiognomic units is significantly influenced by local geomorphology, which determines spatial variation in the magnitude, duration, and frequency of flooding.