基于机器学习算法的喀斯特峰丛洼地石漠化程度评估  

Evaluation of rocky desertification degree in karst peak cluster depression based on machine learning

作　　者：张亚丽 田义超 王栋华 ZHANG Yali;TIAN Yichao;WANG Donghua(School of Resources and Environment,Beibu Gulf University,535011,Qinzhou,Guangxi,China;Key Laboratory of Marine Geographic Information Resources Development and Utilization in the Beibu Gulf,Beibu Gulf University,535011,Qinzhou,Guangxi,China;College of Environmental Science and Engineering,Guilin University of Technology,541004,Guilin,Guangxi,China)

机构地区：[1]北部湾大学资源与环境学院,广西钦州535011 [2]北部湾大学北部湾海洋地理信息资源开发与利用重点实验室,广西钦州535011 [3]桂林理工大学环境科学与工程学院,广西桂林541004

出　　处：《中国水土保持科学》2023年第5期51-61,共11页Science of Soil and Water Conservation

基　　金：国家自然科学基金“桂西南峰丛洼地流域生态系统服务权衡关系及其驱动机制”(42061020);广西高校中青年教师(科研)基础研究能力提升项目“桂西南峰丛洼地流域生物多样性情景模拟及其驱动力机制”(2021KY0431)。

摘　　要：喀斯特石漠化区域水土流失严重,生态环境脆弱。准确定量评估喀斯特区域石漠化程度,对水土流失防治和生态环境治理具有重要的意义。基于遥感数据、气象数据和野外调查等数据,结合特征选择方法和优化算法对支持向量机参数优化,采用最佳模型评估桂西南峰丛洼地流域的石漠化程度。结果表明:Boruta构建的特征集具有最佳的降维效果和最高的拟合精度,裸岩率和植被覆盖度在石漠化程度评估中贡献率较高,其次是坡度;优化算法可以有效辅助支持向量机的超参数调优,混合评估模型的总体精度均达到93%,Kappa系数均达到0.90。其中,萤火虫优化算法、粒子群和万有引力搜索优化的支持向量机模型具有较高的评估精度和运行效率,适合应用于大区域的石漠化程度评估;研究初期(2001—2010年)重度石漠化、极重度石漠化呈现出大面积连片分布,后期(2010—2020年)由于国家石漠化治理效果显著,极重度石漠化呈局部零星分布,该流域内石漠化呈改善的趋势。本研究有效地提高了石漠化程度分类准确性,为大范围的石漠化程度监测提供新的思路和方法。[Background]Rocky desertification(RD)has become one of the most serious ecological and environmental problems in karst areas.The ecological environmental security problems such as soil erosion caused by rocky desertification have seriously affected people s living environment and sustainable development.Accurately evaluating rocky desertification is the key to implementing soil and water conservation projects,and ecological projects in karst areas.The study is aimed to compare the performance of seven different optimization algorithms machine-learning models so as to assess the degree of rocky desertification and then invert the spatial and temporal distribution of rocky desertification in a typical peak cluster depression basin in Southwest Guangxi by selecting the optimal model.[Methods]Based on Boruta,the Maximal Information Coefficient(MIC)and Extreme Learning Machine(ELM)feature selection method,Cuckoo Search(CS)algorithm,Whale Optimization Algorithm(WOA),Firefly Algorithm(FA),Artificial Bee Colony(ABC)optimization algorithm,Difference Evolution(DE)algorithm,Particle Swarm Optimization(PSO)algorithm,and Gravitational Search Algorithm(GSA)were used to adjust the super parameters of Support Vector Machine(SVM)model for assessing the extent of rocky desertification,using Moderate-resolution Imaging Spectroradiometer(MODIS)remote sensing data,topographic data,and meteorological data from 2001 to 2020 as well as the field survey data in 2020.[Results]1)The bare rock rate and fractional vegetation coverage played an important role in assessing the degree of rocky desertification,followed by slope.2)As shown by comparative analyses of three feature selection methods,the feature set constructed by Boruta had the best dimension reduction effect and the highest accuracy.3)Seven intelligent optimization algorithms could effectively assist in the super parameter optimization of SVM.In addition,the accuracy of the optimization models,in descending order,were PSO SVM,FA SVM,GSA SVM,CS SVM,ABC SVM,WOA SVM,and DE SVM.The co

关 键 词：石漠化 支持向量机 优化算法 特征选择 机器学习 峰丛洼地 

分 类 号：P931.5[天文地球—自然地理学]