Modeling dynamic land use changes in the eastern portion of the hexi corridor, China by cnn-gru hybrid model

Abstract

Deep learning models have been extensively employed to simulate dynamic Land Use Changes (LUC) at the regional level. However, existing studies were found to be limited by the inadequate extraction of Spatiotemporal Neighborhood Features (SNF) and ignored obvious Long-Term Dependence (LTD) in time series data, which low simulation accuracies. Consequently, it remains to be explored whether deep learning should be employed to initially extract spatial or temporal features in LUC simulations to achieve better results. We propose here a deep learning model of a Convolutional Neural Network – Gated Recurrent Unit (CNN-GRU), coupled with SNF learning to simulate LUC. Land use data for the Eastern Portion of the Hexi Corridor (EPHC) (2000–2020) were utilized to verify the effectiveness of our model, which was compared with four other models (GRU, CNN, CNN-LSTM, and GRU-CNN). We found that CNN-GRU the highest simulation accuracy (OA = 0.9346), whereas the kappa coefficient increased by 1.22%-4.57%. These results indicated that: (1) EPHC urbanization showed a trend of contraction, rapid expansion, and slow growth. (2) The CNN-GRU model effectively extracted SNF and LTD, which improved the accuracy of the simulation. (3) An enhanced effect could be achieved when spatial features were initially extracted during the SNF deep learning extraction process. (4) Variable neighborhood unit sizes influenced the simulation results, where a 3 × 3 window was most suitable for the study area.

Keywords

• deep learning
• land use change
• long-term dependence
• spatiotemporal neighborhood features
• neighborhood unit