A cellular automata model coupled with partitioning CNN-LSTM and PLUS models for urban land change simulation

Abstract

Urbanisation is a key aspect of land use change (LUC), and accurately modelling of urban LUC is crucial for sustainable development. Cellular automata (CA) are widely used in LUC research. However, previous studies have overlooked the significant temporal dependence and spatial heterogeneity associated with LUC. To address these gaps, this study proposes a novel model called KCLP-CA, which integrates k-means, a convolutional neural network (CNN), a long and short-term memory neural network (LSTM), and the popular patch-generation land use model (PLUS). Initially, k-means and CNN are utilised to address spatial heterogeneity, while LSTM tackles temporal dependence. The LSTM and land expansion analysis strategy (LEAS) models of PLUS are employed to obtain land use conversion probability maps. Finally, a simulation of land use dynamic change was conducted using a linear weighted fusion conversion probability map that accounts for random factors. To validate the KCLP-CA model, land use data collected from Hangzhou between 1995 and 2000 were employed. The results showed that the KCLP-CA model outperformed traditional methods, including artificial neural networks and random forest model, with the figure of merit (FoM) index increasing from 2.12% to 4.19%. Random forest analysis of drivers impacting LUC revealed that distance to water and road network density exerted the greatest influence on urban land development in Hangzhou. Incorporation of various policy planning factors affecting urban development yielded simulation results aligning more closely with reality, resulting in a FoM index increase of 1.64–1.76%. In summary, the model developed in this study combines the strengths of two sub models to deliver an accurate and effective simulation of future land use.