CAJ | 학술논문

本文利用实测数据集，发展了基于进化建模方法的 HJ-1 CCD 黄海悬浮物（TSM）和叶绿素a浓度（Chla）遥感反演模型，建模过程中有针对性地设计了适合水色反演的端点集和函数集，并利用转基因方法引入水色先验知识。经实测数据检验，TSM 反演的平均相对误差约为31％（相关系数R2为0．96），Chla反演误差约为33％（R2为0．88）。分析了模型对输入误差的敏感性，当输入端引入±5％的误差时，模型误差的波动在大多数情形下都可控制在±10％以内。与神经网络模型相比，本文发展的进化模型具有检验精度高、结构简单等优势。利用不同季节的黄、东海实测数据进行了模型精度的独立检验。本文的研究工作表明，进化建模方法适用于水色遥感反演建模问题，可由程序自动生成多个满足精度要求、结构形式多样的显式模型，为水色反演应用提供了多种选择，对于拥有数百个波段的高光谱数据水色反演具有更大的应用潜力。本文最后探讨了进化建模方法的改进方向。
본문이용실측수거집，발전료기우진화건모방법적 HJ-1 CCD 황해현부물（TSM）화협록소a농도（Chla）요감반연모형，건모과정중유침대성지설계료괄합수색반연적단점집화함수집，병이용전기인방법인입수색선험지식。경실측수거검험，TSM 반연적평균상대오차약위31％（상관계수R2위0．96），Chla반연오차약위33％（R2위0．88）。분석료모형대수입오차적민감성，당수입단인입±5％적오차시，모형오차적파동재대다수정형하도가공제재±10％이내。여신경망락모형상비，본문발전적진화모형구유검험정도고、결구간단등우세。이용불동계절적황、동해실측수거진행료모형정도적독립검험。본문적연구공작표명，진화건모방법괄용우수색요감반연건모문제，가유정서자동생성다개만족정도요구、결구형식다양적현식모형，위수색반연응용제공료다충선택，대우옹유수백개파단적고광보수거수색반연구유경대적응용잠력。본문최후탐토료진화건모방법적개진방향。
By using the in-situ measuring data,this study developed retrieval models of chlorophylla (Chla)and total suspended matter (TSM)for HJ-1 CCD data in the Yellow Sea based on the evolutionary modeling method. The terminal and function set of the evolutionary modeling method were designed to be adapted to retrieval of wa-ter constituents,and the transgene operator was employed to insert and maintain the prior knowledge.The average percentage difference (APD)for TSM was 31% (the correlation coefficientR2 =0.96),and that for Chla was 33%(R2 =0.88).The error sensitivity of the retrieval models was analyzed,and the output errors were generally less than ±10% when introducing ±5% error of remote sensing reflectance.Compared with neural network method, the evolutionary models have higher accuracy and simpler structures.In addition,in-situ data with different sea-sons was employed to validate the accuracy of the retrieval models.This study shows that the evolutionary model-ing method is applicable for retrieval of water constituents from ocean color remote sensed data.Many explicit models with well accuracy and different structures could be obtained automatically,and they are of potential appli-cations for hyperspectral data.Finally,we discussed how to improve the method in the near future.