Detection of land cover change using an Artificial Neural Network on a time-series of MODIS satellite data

Abstract

An Artificial Neural Network (ANN) is proposed to detect human-induced land cover change using a sliding window through a time-series of Moderate Resolution Imaging Spectroradiometer (MODIS) satellite surface reflectance pixel values. Training of the ANN is performed on data from two pairs of different, but adjacent areas: (i) degraded vs. non-degraded and (ii) urban settlements vs. natural grasslands. The close proximity of the sites limited natural variability in rainfall, soils and vegetation type. It was therefore assumed that the ANN based its classification decisions on human modifications of the land cover, specifically in the form of land degradation and urban expansion. Numerical results are presented for locations in the Limpopo and Mpumalanga provinces, where the non-degraded class was located inside the Kruger National Park. It was found that some 80% of the pixels were correctly classified, and simulations demonstrated that change from non-degraded to degraded could be detected reliably. In Gauteng 87% of pixels were correctly classified as either urban settlements and natural grasslands and the ANN would be able to accurately detect urban expansion