Naïve Bayes integration with Markov chain and cellular automata for modelling urban land use change

Away from the conventional Logistic Regression (LR) technique, this study presents a novel Naïve Bayes (NB) integration with Markov Chain (MC) and Cellular Automata (CA) for modelling urban land use change. Although NB is a simple parametric algorithm similar to the Logistic Regression (LR), previous studies have shown that it can outperform the LR when the features are independent, and the dataset is small and highly dimensional, such as land use datasets. In this experiment, NB was used to classify land use into urban and non-urban classes. The training data, test data, and validation data were extracted from the land use data. The training and test datasets were used to train and evaluate the NB; while the validation data was used to validate the CA prediction. The NB probabilities, transition probabilities from the MC model, neighbourhood interaction, and transition rule from the CA model were combined to determine the evolution of each cell in the prediction. Land use drivers were derived from Landsat-5 Thematic Mapper (TM, 1984), Landsat-7 Enhanced Thematic Mapper Plus (ETM+, 2000), and Landsat-8 Operational Land Imager/Thermal Infrared Sensor (OLI/TIRS, 2020) imageries. Binary urban scenarios (urban and nonurban cells) were modelled using twelve salient urban land use drivers. Urban transitions from 1984–2000 to 2000–2020 were used to predict the urban state in 2060. The Model validation techniques employed were McNemar’s test and Receiver Operating Characteristic (ROC) curves. The computed mean Area Under the Curve (AUC) for 1984–2000 and 2000–2020 were 0.7969 ± 0.0842 and 0.8357 ± 0.0613 respectively; which indicated that the predicted maps aligned closely with the validation data. We discussed the advantages and limitations of the NB algorithm relative to nonparametric algorithms for predicting urban land use change. We compared the results from NB and LR, and NB emerged a slightly better algorithm for the prediction. The resulting maps demonstrated the robust predictive power of the NB model.