A comparative analysis on the use of a cellular automata Markov chain versus a convolutional LSTM model in forecasting urban growth using sentinel 2A images

Abstract

ABSTRACT Cities are facing many challenges related to urban growth. This phenomenon has prompted decision-makers to adopt innovative approaches for planning based on accurate forecasting of urban growth. Among the most widely used forecasting methods, there are Cellular Automata (CA) based methods and Recurrent Neural Networks (RNN) based methods. The accuracy of these forecasting models is strongly related to data quality, data availability, Model calibration and Model validation. In this paper, a comparative analysis between three forecasting methods is presented based on a temporal sequence of Sentinel 2A images. The main goal of this study is to assess the performance of these models which are of CA-Markov Chain, MLP-Markov and ConvLSTM in terms of accuracy, complexity, and feasibility. The case study is carried out on the city of Casablanca in Morocco. After implementing these three forecasting methods, the obtained results show that the Kappa coefficient of MLP-Markov, CA-Markov and ConvLSTM is, respectively, 89,40%; 97,20%; and 94,50%. In terms of complexity, the ConvLSTM method is more complex due to the number of elementary operations. In terms of feasibility, the ConvLSTM method is more demanding in terms of data volume since it is a Deep Learning model. Accordingly, CA-Markov based methods, in particular MLP-Markov, show a great potential for forecasting urban growth, especially for short term forecasting when there are not enough satellite images available to adopt a Deep Learning approach such as ConvLSTM.