A Comparative Study of Deep Learning and CA-Markov Methods for Land Use/Land Cover Change Prediction

The term "land cover" indicates the physical characteristics of a land (vegetation, water...), while the term "land use" is the way in which people use a land (agriculture, conservation...). Land use and land cover change (LULCC) refers to the conversion of land from one use to another. The analysis of LULCC is a real challenge and can be very useful in various fields. Forecasting deforestation, urban growth, or soil moisture, to name a few, are some of the commonly studied applications. For a very long time, predictions of LULCC were mostly made by Cellular Automata and Markov Chain (CAMarkov) methods. However, with the success of deep learning (DL) algorithms in predictive analysis, and thanks to the wide availability of satellite images, DL algorithms are increasingly being used to predict land cover changes. This paper aims at comparing these two aforementioned methods for LULCC prediction. Models based on CAMarkov, ConvLSTM and CNN-LSTM are applied in turn on Sentinel-1A images from the Bouba-Ndjida National Park (Cameroon), in order to predict the next image of the input sequence. Metrics such as the peak signal to noise ratio (PSNR), the structural similarity index (SSIM), the coefficient of correlation (r), the mean squared error (MSE) and the number of pixels close to zero (NPCZ) are used to assess the models. The experimental results indicate that models based on DL algorithms, and specifically on ConvLSTM networks, are better suited for LULCC prediction.