基于地面高光谱数据的典型作物类型识别方法--以青海省湟水流域为例

Identifying Typical Crop Types from Ground Hyper-spectral Data:A Case Study in the Huangshui River Basin,Qinghai Province

高光谱技术运用于农作物识别与分类目前已成为农业遥感应用领域前沿课题之一。使用ASD FieldSpec4地物光谱仪实测青海省湟水流域大豆、青稞、土豆、小麦和油菜5种典型作物冠层光谱,经数据预处理,利用1/R、d(R)、N(R)、log(R)、d(log(R))、d(N(R))6种光谱数据变换形式和在"绿峰"、"红谷"、"红边"、"光谱吸收特征区"提取的16种光谱特征变量的6种选取结果,分别构建基于BP神经网络的典型作物类型识别模型,通过模型精度比较以寻求用于高光谱农作物分类的有效光谱数据形式和光谱特征变量。结果表明:1/R、d(R)、log(R)、d(log(R))及d(N(R))5种数据变换形式能显著提高模型识别精度,以d(N(R))变换数据构建BPNN模型其辨识精度最高,总体分类精度达88%;在提取的16种光谱特征变量中,以变量数分别为16、14、12的3种选取方案构建BPNN模型其辨识精度较优,总体分类精度分别为88%、86%、84%;BPNN模型能较好地识别5种作物光谱,且采用选取光谱特征变量方法构建BPNN模型其网络训练效率和模型稳定性优于光谱数据变换方法构建BPNN模型。

The hyperspectral technology used in crop identification and classification has increasingly become one of the frontier issues in agricultural remote sensing applications at currently.In this study,using ASD FieldSpec4 spectrometer,canopy spectrum from five selected typical crops including soybean,barley,potato,wheat and rape was measured in the open air in the Huangshui River Basin,Qinghai Province,and then data preprocessing was finished.Six spectral transformations for original reflectance spectrum（R）such as 1/R,d（R）,N（R）,log（R）,d（log（R））,d（N（R））were conducted and six kinds of selection results of 16 spectral characteristic variables selected from ＂Green Peak＂,＂Red Valley＂,＂Red Edge＂and ＂Spectral Absorption Feature＂to construct a typical crop types identification model which based on neural network of BP.Through the comparison of the accuracy of the models to find the effective spectral data form and spectral characteristic variables for crop classification with hyperspectral data.The results showed that 5kinds of transformation methods such as 1/R,d（R）,log（R）,d（log（R））,d（N（R）））could improve identification accuracy of model significantly,especially the BPNN model which created by the data of d（N（R））has the highest overall accuracy,which reached to 88%.The number of spectral characteristic variables of 3projects were 16,14,12,respectively,and the identification accuracy of BPNN from 3projects was much better,which were 88%,86%,84%.The BPNN model can better identify the five crops spectra,the training efficiency and the model stability of BPNN model which created by the spectral characteristic variable selections was better than that created by the transformation forms of the spectrum.