Modelling land cover change in a Mediterranean environment using Random Forests and a multi-layer neural network model

Abstract

The present study seeks to identify the changes that have taken place in the Mediterranean island of Lesvos (Greece) between 1995 and 2007 in the seven main land cover types of the island. We also attempt to predict the changes that will occur by the year 2019. Three Landsat 5 TM summer scenes were used spanning 12 years. A combination of Random Forests (RF) classification with expert rules was then applied for achieving high overall classification accuracies (95%, 94% and 91%, respectively). The 1995 and 2001 classified data were then used to train a multi-layer perceptron neural network (MLPNN) model and predict land cover for the year 2007. Seven possible transitions were included in the MLPNN model which was trained with the 1995 and 2001 classified data successfully: accuracy rate of 93% after 5000 iterations. The quantity of change in each transition was modelled through Markov chain analysis. The modelling results for 2019 provide an anticipated prediction for the end of the decade: economic activity will remain centred to the agricultural sector, as crops and olive groves will expand. A rather unanticipated prediction is the significant increase in the area of forests.