A study on land use change simulation based on PLUS model and the U-net structure: A case study of Jilin Province

Abstract

Deep learning and spatial models are typical paradigms for land cover change simulation, each exhibiting certain differences in representing land use change mechanisms. This study investigates land use changes in Jilin Province from 2000 to 2020, independently constructing land cover change simulation frameworks using the PLUS model and U-Net neural network, and simulates land use for the year 2030. The simulation differences were quantified through spatial consistency analysis. By comparing the performance of the two methods in land use change prediction, this study reveals the divergent characteristics of the two approaches in capturing complex transformation patterns. The findings indicate that: 1) Cross-validation based on actual 2020 land use data shows that the overall Kappa index for the PLUS model and U-Net neural network are 0.802 and 0.810, respectively, with spatial consistency values of 87.88 % and 88.99 %, and the MSE value for U-Net is 0.43. These results demonstrate that both models possess reliable simulation performance. 2) The two methods exhibit significant differences in predictive performance. The U-Net model, which utilizes convolutional neural networks to extract multi-scale spatial features and addresses the class imbalance issue with the OHEM-Dice composite function, significantly enhances the prediction accuracy of nonlinear dynamics. It is more sensitive to short-term land use changes, but its generalization ability is constrained by sample size and balance. The PLUS model, on the other hand, is more suited for long-term trend prediction and shows significant advantages in simulating land types with fewer pixels, particularly maintaining high stability even under conditions of missing data or sample imbalance. 3) In terms of land use transformation patterns, the PLUS model primarily predicts unidirectional transformations. Arable land transitions to construction land and forest land by 326.73 km 2 and 301.82 km 2 , respectively, while wetland transitions to forest land by 0.28 km 2 , reflecting an orderly transformation driven by policies. Conversely, the U-Net model predicts more complex bidirectional flow features, with arable land transitioning to construction land and forest land by 396.43 km 2 and 498.06 km 2 , respectively, and significant bidirectional transformations between arable land and grassland. This suggests that deep learning can capture nonlinear coupling processes of “human decisions-ecological responses” that traditional models struggle to quantify, providing a new approach for land sustainability assessment.

Topics & Concepts