Neural network modelling of rainfall interception in four different forest stands

Abstract

The objective of this study is to reveal whether it is possible to predict rainfall, throughfall and stemflow in forest ecosystems with less effort, using several measurements of rainfall interception (hereafter 'interception') and an artificial neural network based linear regression model (ANN model). To this end, the Kerpe Research Forest in the province of Kocaeli, which houses stands of mixed deciduous-broadleaf forest (Castanea sativa Mill., Fagus orientalis Lipsky, Ouercus spp.), black pine (Pinta nigra Arnold), maritime pine (Finns pinaster Aiton) and Monterey pine (Pinus radiata D. Don), was selected study site. Four different forest stands were observed for a period of two years, during which rainfall, throughfall and stemflow measurements were conducted. These measurements were separately calculated for each individual stand, based on interception values and the use of stemflow data in strict accordance with the rainfall data, and the measured throughfall interception values were compared with values estimated by the ANN model. In this comparison, 70% of the total data was used for testing, and 30% was used for estimation and performance evaluation. No significant differences were found between values predicted with the help of the model and the measured values. In other words, interception values predicted by the ANN models were parallel with the measured values. In this study, the most success was achieved with the models of the Monterey pine stand (r(2) = 0.9968; Mean Squared Error MSE = 0.16) and the mixed deciduous forest stand (r(2) = 0.9964; MSE = 0.08), followed by models of the maritime pine stand (r(2) = 0.9405; MSE = 1.27) and the black pine stand (r(2) = 0.843, MSE = 17.36).