标题: How can SHAP (SHapley Additive exPlanations) interpretations improve deep learning based urban cellular automata model?

作者: Yang, CL (Yang, Changlan); Guan, XF (Guan, Xuefeng); Xu, QY (Xu, Qingyang); Xing, WR (Xing, Weiran); Chen, XY (Chen, Xiaoyu); Chen, JG (Chen, Jinguo); Jia, P (Jia, Peng)

来源出版物: COMPUTERS ENVIRONMENT AND URBAN SYSTEMS  卷: 111  文献号: 102133  DOI: 10.1016/j.compenvurbsys.2024.102133  Published Date: 2024 JUL  

摘要: Interpretations of the urban cellular automata (CA) model aim to ensure that its predictive behaviors are consistent with real -world processes. Current urban CA interpretations have revealed the impacts of driving factors on land development suitability, or neighborhood effects and random perturbation on simulation results. However, three limitations remain unresolved: (1) the interpretations of deep learning (DL) -based urban CA are seldom integrated with the prerequired feature selection, (2) the input features from different urban CA modules are still explained by separate approaches, and (3) the interpretation results are rarely derived at the cell level to uncover spatially varying urban land development patterns. This study proposes a SHapley Additive exPlanations (SHAP)-based urban CA interpretation framework to address these challenges and improve urban CA. This framework uses model-level SHAP importance to identify dominant features from different modules for constructing the final simulation model. Then, cell-level SHAP importance is used to uncover spatially varying driving forces of urban expansion. The framework's effectiveness is rigorously tested and confirmed using a convolution neural network CA (CNN -CA) model for Dongguan City. The experimental results demonstrate that (1) SHAP-based model interpretation improves feature selection for DL -based urban CA. The figure of merit for CNN -CA calibrated using SHAP-based important features improves by 3%, outperforming the tested baseline methods. (2) SHAP measures the impacts of each feature from different CA modules in a whole. In this case, physical factors are much more important at the model level than proximity and accessibility factors, while neighborhood effect is the second most crucial factor. (3) Cell-level SHAP interpretations uncover spatially different urban land development patterns. For example, due to the extensive industrial land development in the northern Songshan Lake Zone, in the CNN -CA model, proximity to major roads within this region is associated with positive SHAP-based contribution share on cell-level urban expansion.

作者关键词: Model interpretability; SHapley Additive exPlanations; Feature selection; Cellular automata; Convolution neural network; Urban growth simulation

KeyWords Plus: LAND-USE-CHANGE; SENSITIVITY-ANALYSIS; EXPANSION

地址: [Yang, Changlan; Jia, Peng] Wuhan Univ, Sch Resource & Environm Sci, Wuhan 430079, Peoples R China.

[Guan, Xuefeng; Xu, Qingyang; Xing, Weiran; Chen, Xiaoyu] Wuhan Univ, State Key Lab Informat Engn Surveying Mapping & Re, Wuhan 430079, Peoples R China.

[Guan, Xuefeng; Jia, Peng] Hubei Luojia Lab, Wuhan 430079, Peoples R China.

[Chen, Jinguo] Third Team Hubei Geol Bur, Huanggang 438000, Peoples R China.

[Jia, Peng] Wuhan Univ, Sch Publ Hlth, Wuhan 430079, Peoples R China.

[Jia, Peng] Wuhan Univ, Int Inst Spatial Lifecourse Hlth ISLE, Wuhan 430079, Peoples R China.

通讯作者地址: Guan, XF (通讯作者)，Wuhan Univ, State Key Lab Informat Engn Surveying Mapping & Re, Wuhan 430079, Peoples R China.

电子邮件地址: guanxuefeng@whu.edu.cn

影响因子：7.1