基于MODIS影像的北极冰水分类方法对比研究  

Comparative study of five methods for Arctic sea ice and water classification based on MODIS images

作　　者：孙绍哲 邝慧妍 叶玉芳 于志同 惠凤鸣 SUN Shaozhe;KUANG Huiyan;YE Yufang;YU Zhitong;HUI Fengming(School of Geospatial Engineering and Science,Sun Yat-sen University,Zhuhai 519082,China;Key Laboratory of Comprehensive Observation of Polar Environment,Ministry of Education,Sun Yat-sen University,Zhuhai 519082,China;Southern Marine Science and Engineering Guangdong Laboratory(Zhuhai),Zhuhai 519082,China;Qian Xuesen Laboratory of Space Technology,China Academy of Space Technology,Beijing 100094,China)

机构地区：[1]中山大学测绘科学与技术学院,广东珠海519082 [2]中山大学极地环境立体观测与应用教育部重点实验室,广东珠海519082 [3]南方海洋科学与工程广东省实验室(珠海),广东珠海519082 [4]中国空间技术研究院钱学森空间技术实验室,北京100094

出　　处：《极地研究》2025年第1期39-54,共16页Chinese Journal of Polar Research

基　　金：国家自然科学基金青年项目(42106225);广东省自然科学基金面上项目(2022A1515011545);南方海洋科学与工程广东省实验室(珠海)创新团队建设项目(311021008);风云卫星应用先行计划项目(FY-APP-2022.0201);五院CAST创新基金资助

摘　　要：北极海冰作为冰冻圈的重要组成部分,对全球气候变化具有重要影响。开展极地冰水分类研究可以获取精细的海冰覆盖范围变化,为船舶航行提供关键的冰况信息。本文基于25景北极地区的MODIS近红外波段光学影像,依据海冰和海水在该波段反射率差异大的特点,利用直方图动态阈值法、迭代阈值法、最大熵阈值法、大津法以及基于大津法的遗传算法5种方法开展冰水分类研究和定量化对比评估。结果表明,对于反射率直方图为双峰分布的影像(影像中无薄冰),5种方法均能得到较为准确的冰水分类结果,总体分类精度均在0.82以上。当反射率直方图为三峰分布时(影像中有薄冰),迭代阈值法、最大熵阈值法和大津法均存在错分现象,总体分类精度均在0.76以下,而直方图动态阈值法和基于大津法的遗传算法能够准确获取冰水分类阈值,其分类结果的总体精度均能达到0.94以上。总体上,迭代阈值法和最大熵阈值法在冰水分类求取阈值上存在明显的高估问题,特别是对于反射率直方图为三峰分布的影像;大津法在面对反射率直方图为双峰或三峰分布的影像时,均能识别出阈值所在的波谷位置,但在分类结果上存在一定偏差;而直方图动态阈值法和基于大津法的遗传算法在基于阈值法的冰水分类方法中具有较强的鲁棒性,更加适用于北极地区不同冰况下的冰水分类研究,其结果可为极区航行以及大尺度海冰观测与数据模拟提供重要参考。Arctic sea ice is an important component of the cryosphere and it has substantial impact on global climate change.Ice water classification is used to obtain detailed information on sea ice coverage changes and ice conditions critical for maritime navigation.In this study,five methods(i.e.,the dynamic thresholding method using histogram(DTM),iterative thresholding method(ITM),maximum entropy thresholding method(MTM),OTSU’s method(OM),and genetic algorithm based on OTSU’s method(GAM))were used to classify Arctic sea ice and water in 25 near-infrared MODIS images.Quantitative comparison and evaluation revealed that for images with a bimodal reflectance histogram(i.e.,no thin ice in the image),all five methods obtained accurate ice water classification results,with overall accuracy(OA)of>0.82.However,when the reflectance histogram had a trimodal distribution(i.e.,with thin ice in the image),the ITM,MTM,and the OM all exhibited misclassification problems,with OA of<0.76,whereas the DTM and GAM accurately obtained ice water classification thresholds,with OA of up to 0.94.Generally,the ITM and MTM both tend to overestimate the threshold for the ice water classification,especially for images with a trimodal distribution in the reflectance histogram.The OM can broadly identify the minimum for the threshold but with nonnegligible biases.The DTM and GAM have strong robustness as threshold-based ice water classification methods and are best suited for ice water classification under different ice conditions in the Arctic region.The results represent an important reference for polar navigation and for large-scale sea ice observation and data simulation.

关 键 词：光学影像 冰水分类 近红外波段 直方图动态阈值法 基于大津法的遗传算法 

分 类 号：P23[天文地球—摄影测量与遥感]