干旱区研究

干旱区研究 ›› 2023, Vol. 40 ›› Issue (1): 69-77.doi: 10.13866/j.azr.2023.01.08

• 水土资源 • 上一篇    下一篇

基于深度学习的玛纳斯土地利用时空格局变化与预测

王娇娇1,2(),尹小君1,2(),刘陕南1,王帝盟1,2   

  1. 1.石河子大学信息科学与技术学院，新疆 石河子 832000
    2.兵团空间信息工程技术研究中心，新疆 石河子 832000
  • 收稿日期:2022-07-04 修回日期:2022-08-19 出版日期:2023-01-15 发布日期:2023-02-24
  • 通讯作者: 尹小君. E-mail: yinxiaojun2018@163.com
  • 作者简介:王娇娇（1996-），女，硕士研究生，主要从事空间遥感土地利用研究. E-mail： 2272745959@qq.com
  • 基金资助:
    国家重点研发计划项目(2017YFB0504203);石河子大学国际合作项目(GJHZ201905);兵团社科基金项目(20YB23)

Study on the change and prediction of spatiotemporal pattern of land use in Manasi region based on deep learning

WANG Jiaojiao1,2(),YIN Xiaojun1,2(),LIU Shannan1,WANG Dimeng1,2   

  1. 1. College of Information Science and Technology, Shihezi University, Shihezi 832000, Xinjiang, China
    2. Geospatial Information Engineering Research Center, Xinjiang, Shihezi 832000, China
  • Received:2022-07-04 Revised:2022-08-19 Online:2023-01-15 Published:2023-02-24

RichHTML

28

PDF (PC)

212

赞

摘要：

土地利用时空格局变化与预测对土地资源管理与优化至关重要。本文基于遥感时空序列数据，协同景观指数与深度学习的长短时记忆网络（Long-Short Term Memory，LSTM）模型，对玛纳斯进行长时间序列土地利用时空格局演变特征分析和预测。结果表明：（1） 1992—2020年耕地、草地和建设用地增加，林地、水域和未利用地减少。（2） 耕地破碎化程度降低，林地和水域的景观指数轻微波动；草地破碎化程度降低，形状趋于规则化；建设用地处于持续扩张状态，破碎化程度加深，形状趋于复杂；未利用地破碎化程度增加，但形状趋于规则化。（3） 比较了LSTM模型、多层感知人工神经网络(Multi-Layer Perception Artifical Neural Network，MLP-ANN)模型、逻辑回归(Logistic Regression，LR)模型和CA-Markov模型的预测精度。LSTM模型的Kappa系数为95.31%，较其他模型准确度高，符合实际土地利用格局分布。LSTM模型表明2025年土地利用类型可能仍以耕地、草地和未利用地为主。

关键词: 时空格局变化, 土地利用预测, 深度学习, LSTM模型, 景观指数

Abstract:

Land use change and prediction are crucial for land resource management and optimization. In this paper, based on remote sensing spatial and temporal series data, a synergistic landscape index and long short-term memory (LSTM) model were used to characterize and predict the evolution of spatial and temporal patterns of land use in the Manasi region over a long time series. Results showed that (1) from 1992 to 2020, the cropland, grassland, and building land increased, and tree cover, water bodies, and unused land decreased. (2) The degree of fragmentation of arable land gradually decreased. The landscape indices of tree cover and water bodies were in a state of slight fluctuation. Grassland aggregation increased, fragmentation decreased, and the shape exhibited regularization. The landscape index of building land showed the most dramatic state of continuous expansion, with a deepening fragmentation and a tendency for complex shapes. The fragmentation of unused land gradually increased, but the shape tended to be regular. (3) Different models for predicting land use change, including the LSTM, multi-layer perception artificial neural network, logistic regression, and CA-Markov models, were compared. The Kappa coefficient of the LSTM model was 95.31%, which is more accurate than that of other models and consistent with the actual land use pattern. The LSTM model suggests that in 2025, land use types will still be dominated by cropland, grassland, and unused land.

Key words: spatial and temporal pattern change, land use prediction, deep learning, LSTM model, landscape index