| Research focusing on land use change analysis is of tremendous importance in global change studies. Land use change modeling, which is a prerequisite to understanding the complexity of land use dynamics, is an effective way to describe the change patterns and delve into the causes for the changes. Despite the development of many models in the past, several important issues still remain to be addressed such as spatio-temporal non-stationarity, spatio-temporal correlation, and individual effect. The primary objective of this research is to make improvements on.the traditional logistic models to suit the characteristics and requirements of land use change modeling. Specifically, three enhancements have been made. The first enhancement deals with spatio-temporal non-stationarity, the second improvement aims to incorporate spatio-temporal autocorrelation, and the third includes individual effect.;Three spatio-temporal logit models for land use change analysis, namely, geographically and temporally weighted logit model (GTWLM), spatio-temporal panel logit model (ST-PLM) and generalized spatio-temporal logit model (GSTLM), are proposed accordingly to deal with the aforementioned issues. GTWLM, which considers spatio-temporal non-stationarity, includes temporal data in a spatio-temproal framework by proposing a spatio-temporal distance. ST-PLM incorporates the spatio-temporal correlation and individual effect in one model, where spatio-temporal correlation is considered in the random individual effect with an assumption that the correlation between such components is inversely proportional to the spatio-temporal distance. By integrating GTWLM and ST-PLM, the GSTLM explores spatio-temporal non-stationarity and correlations simultaneously, whilst considering their individual effects to construct an integrated model.;Based on the models, a case study is performed on multi-temporal land use change analysis in the Special Economic Zone (SEZ), Shenzhen. The results show that all the proposed models outperform the traditional logistic regression model: multinomial logit model (MNLM), which overlooks the aforementioned issues. Compared with MNLM, GTWLM and ST-PLM increased the percentage of correctly predicted (PCP) values from 74.1% to 82.3% and 79.4%, respectively. McNamara's test shows that the differences between those models are significant. The kappa coefficients reveal that the GTWLM and ST-PLM are better than MNLM. In particular, the GSTLM yields a considerably higher PCP of 85.9%. The Kappa coefficients also indicate that the GSTLM is the most optimal model. Furthermore, the GTWLM allows the model parameters to vary across space and time, which provides deep insights into the spatio-temporal variations of the land use pattern. Assisted with the visual results, the spatio-temporal land use distribution patterns in Shenzhen are analyzed and the results presented thereafter. |
