基于LSTM-CA模型的土地利用动态模拟

Dynamic simulation of land use based on the LSTM-CA model

及时准确获取土地利用空间格局演变规律,能够有效为城市生态环境保护和科学管理提供依据。文章利用卷积神经网络提取多个时期土地利用空间特征,结合多种空间驱动因子建立长短时记忆网络(long short term memory network,LSTM)的元胞自动机(cellular automata,CA)模型(LSTM-CA)。以张家口市中心城区1995年、2000年、2005年、2010年及2015年5期时序土地利用分类、地形及城市交通等数据为基础,开展2020年城市土地利用模拟方法研究。通过与多层感知机(multi-layer perceptron,MLP)的CA模型(MLP-CA)进行精度对比分析,结果表明所提方法Kappa系数达到0.90,FoM指标为0.39,各项指标均优于MLP-CA模型,LSTM-CA更能充分挖掘历史土地利用变化之间的内在关系,可以有效提升模拟精度。

Timely and accurate acquisition of the spatial pattern evolution of land use can effectively support urban ecological environment protection and scientific management.In this study,the spatial characteristics of land use in multiple periods were extracted using a convolutional neural network.Then,they were combined with multiple spatial driving factors to build the long short-term memory network-cellular automata(LSTM-CA)model.Based on the data of land use classification,terrain,and urban traffic of the Zhangjiakou central urban area in 1995,2000,2005,2010,and 2015,this study investigated the simulation methods for urban land use in 2020.By comparison with the precision of the multi-layer perceptron-cellular automata(MLP-CA)model,the proposed method has a Kappa coefficient of over 0.90 and FoM of 0.39.All indices are better than those of the MLP-CA model.Therefore,the LSTM-CA model can fully explore the internal relationships between the changes in historical land use and effectively improve the simulation precision.