Abstract: Urban space structure and its simulation are important prerequisites for urban scientific management and planning. Based on the extreme learning machine, this paper proposes an urban extended cellular automaton model (ELM-CA) that takes into account the differences and intensities of different non-urban land conversions into urban land use. The experimental results show that the urban simulation accuracy of ELM-CA model reaches 70.30%, which is 2.21% and 1.54% higher than logistic regression and neural network respectively. The FoM coefficient is increased by 0.025 9 and 0.017 9 respectively, and the Kappa coefficient is improved by 0.024 7 and 0.016 9 respectively. And the Moran I index is close to the actual value, which shows that the extreme learning machine model can simulate and predict the spatial shape and change of urban expansion more effectively than logistic regression and neural network; the training time of ELM model is only about 1/3 of the neural network, it reflects the advantage of ELM learning speed; In the small sample case, both logistic regression and neural network are significantly affected, and the extreme learning machine can maintain good performance, which makes it have obvious advantages when the sample is difficult to obtain. The comparison between urban expansion simulation and real data of two phases shows that the urban extended cellular automata model (ELM-CA) based on the extreme learning machine simplifies the complexity of the CA model and can effectively improve simulation accuracy under small sample conditions. The proposed model is suitable for urban expansion simulation and prediction under complex land use conditions.

Key words: urban space expansion, complex land use, land conversion differences, cellular automata, extreme learning machine