南京林业大学学报(自然科学版) ›› 2024, Vol. 48 ›› Issue (5): 255-266.doi: 10.12302/j.issn.1000-2006.202306008

• 研究论文 • 上一篇    下一篇

基于多分类器集成的土地覆盖分类及预测研究

蒋路凡1(), 苏慧毅1, 张银1, 刘琴琴1, 张小伟2,*(), 李明诗1   

  1. 1.南京林业大学林草学院,南方现代林业协同创新中心,江苏 南京 210037
    2.浙江省森林资源监测中心,浙江省林业勘测规划设计有限公司,浙江 杭州 310020
  • 收稿日期:2023-06-12 修回日期:2023-08-04 出版日期:2024-09-30 发布日期:2024-10-03
  • 通讯作者: * 张小伟(xiaoweiz099618@gmail.com),高级工程师。
  • 作者简介:

    蒋路凡(lf_jiang@yeah.net)。

  • 基金资助:
    国家自然科学基金项目(31971577);江苏高校优势学科建设工程资助项目(PAPD)

Land cover classification and prediction based on multiple classifier ensemble learning

JIANG Lufan1(), SU Huiyi1, ZHANG Yin1, LIU Qinqin1, ZHANG Xiaowei2,*(), LI Mingshi1   

  1. 1. Co-Innovation Center for Sustainable Forestry in Southern China, College of Forestry and Grassland, Nanjing Forestry University, Nanjing 210037, China
    2. Zhejiang Forest Resources Monitoring Centre, Zhejiang Forestry Survey Planning and Design Company Limited, Hangzhou 310020, China
  • Received:2023-06-12 Revised:2023-08-04 Online:2024-09-30 Published:2024-10-03

摘要:

【目的】为发挥不同单分类器各自的优势进而采用集成学习方式提高土地覆盖分类精度,据此比较不同土地覆盖变化模拟模型性能从而执行最优的土地覆盖变化预测,为土地资源合理开发与利用决策制定提供参考。【方法】基于南京市江宁区2000、2010和2020年的Landsat TM/OLI影像,结合研究区实际定义了水体、建筑、林地、草地、耕地和未利用地等6种土地覆盖分类体系,在测试了最大似然法、马氏距离法、最小距离法、神经网络和支持向量机等基分类器性能基础上,采用随机森林和证据理论2种不同的集成学习方法对5种基分类器的分类结果进行集成,比较了集成性能后构建了最终的土地覆盖分类结果。然后,基于2000和2010年的最优集成土地覆盖分类图,运用CA-Markov、PLUS和ANN-CA模型分别对2020年研究区的土地覆盖格局进行模拟,并将不同的模拟结果与2020年真实集成分类结果进行了空间一致性检验,以此确定土地覆盖变化预测的最佳模型并用其预测2030年江宁区的土地覆盖模式。【结果】在单分类器分类结果中,2000年支持向量机算法取得了最佳分类效果,总体精度达到了88.75%,Kappa系数为0.77;2010年神经网络方法表现最佳,总体精度为88.75%,Kappa系数为0.83;2020年最大似然法取得了最佳分类效果,总体精度为82.75%,Kappa系数为0.74。在2种集成方法中,随机森林在2000年取得了最佳集成分类效果,总体精度和Kappa系数分别为91.25%和0.85;证据理论在2010年取得了最佳集成效果,总体精度和Kappa系数分别为90.80%和0.86;随机森林在2020年取得了最佳集成效果,总体精度和Kappa系数分别为93.75%和0.91。就土地覆盖预测而言,PLUS模型获得了98.54%的空间一致性。根据PLUS模型预测2030年土地覆盖结果可知,江宁区各土地覆盖类型变化较小,建设用地略有扩张但范围有限,耕地稍减少但在可控范围内。林地、草地和未利用地变化较小,水体面积相对减少。【结论】恰当的集成策略能明显改进土地覆盖分类精度,合适的、适应局部土地利用状态和政策调控的预测模型可有效预测区域未来土地覆盖分布模式,它们能为区域土地利用、城市管理决策等提供更可靠的方法和数据支持。

关键词: Landsat 影像, 地表特征, 土地覆盖分类, 集成学习, 土地覆盖变化模拟

Abstract:

【Objective】This study aims to exert the respective advantages of individual base classifiers, different ensemble learning strategies were tested in the present study to improve the final land cover classification accuracy. On this basis, different land cover change simulation models were implemented and compared to obtain more accurate land cover projection results to provide references for formulating decisions of rational development and utilization of land resources in the future.【Method】Based on the Landsat Thematic Mapper (TM)/Operational Land Imager (OLI) images acquired in 2000, 2010 and 2020 covering Jiangning District of Nanjing City, the classification scheme, including water body, built-up land, forest land, grass land, crop land and unused land, was defined by referring to relevant industry standards and the actual conditions of the study area. The land cover classification performances of five base classifiers, including the Maximum Likelihood Classifier, Mahalanobis Distance Classifier, Minimum Distance Classifier, Neural Network, and Support Vector Machine(SVM) Classifiers, were implemented and quantitatively evaluated, followed by an integrated combination of the five individual classification results using the random forest algorithm and Dempster-Shafer (D-S) evidence theory. The final integrated classifications in 2000, 2010 and 2020 with higher overall accuracy were created after comparing the respective ensemble performances of the two algorithms. Based on the optimal integrated land cover classification maps in 2000 and 2010, the cellular automata (CA)-Markov model, PLUS model, and artificial neural network (ANN)-CA model were used to predict the land cover pattern of 2020 in the study area, and the prediction results of different models were compared with the real integrated classification results in 2020 to derive a spatial agreement index, determine the best model for land cover change simulation, and generate the projected land cover pattern of 2030 in Jiangning District.【Result】An independent validation showed that for a single base classifier, SVM algorithm achieved the best classification effect in 2000, with an overall accuracy at 88.75% and a Kappa coefficient of 0.77. In 2010, the neural network algorithm obtained the best classification effect, with an overall accuracy of 88.75% and a Kappa coefficient of 0.83. In 2020, the maximum likelihood algorithm had the best classification effect, with an overall accuracy of 82.75% and a Kappa coefficient of 0.74. Regarding the two ensemble methods, the random forest algorithm achieved the best integrated classification effect in 2000, with an overall accuracy of 91.25% and a Kappa coefficient of 0.85. The evidence theory process achieved the best integrated classification effect in 2010, with an overall accuracy and Kappa coefficient of 90.80% and 0.86, respectively. Moreover, the random forest algorithm achieved the best integrated classification effect in 2020, with an overall accuracy and Kappa coefficient of 93.75% and 0.91, respectively. Regarding land cover prediction, the PLUS model obtained a spatial agreement measure of 98.54%, outperforming the CA-Markov and ANN-CA models.According to the optimal prediction results in 2030 made by the PLUS model, there was a slight expansion of built-up land, but the extent was limited. The areal reduction in cropland was observed but within a controllable range. The changes in forest, grassland, and unused land were slight, while the water body was decreased.【Conclusion】Overall, the present study demonstrated that proper assemble learning strategies produce more accurate land cover classifications in contrast to a single base classifier. In addition, a suitable land cover change prediction or simulation model that adapts to local land use status and policy regulations effectively predicts the future land cover classification pattern. Thus, these strategies provide more reliable methods and data support for the development of land use and urban management decisions.

Key words: Landsat imagery, surface features, land cover classification, assemble learning, land cover change simulation

中图分类号: