BP-ANN在荒漠草地高光谱分类研究中的应用

Application of BP-ANN to classification of hyperspectral grassland in desert

利用高分辨率光谱仪在实地测得的光谱数据来识别新疆阜康地区的7种典型荒漠草种，对原始高光谱数据作预处理（微分和平滑），选取典型荒漠植被的光谱特征（红边、绿峰、红谷、RVI等）作为输入数据，植被类型作为输出数据，构建基于BP神经网络模型的典型荒漠草地分类器，进行了三组基于高光谱特征的草地类型分类实验，结果表明：（1）红边特征较其余吸收特征更能获得精确的分类结果；（2）波段550～790nnl间的窄波段光谱分类间隔中，20nm优于10nm的间隔；（3）草地分类器中BP网络模型的输入层、隐藏层神经元个数与BP网络训练时间、精度具有复杂的耦合关系，不可一概而论。

In order to identify the seven typical desert grasses of Xinjiang Fukang area,high-resolution spectroscopy is used to obtain the hyper-spectral data.After the preprocessing of the original hyper-spectral data,such as differentiation and smoothing,the typical desert grass classifier based on BP neural network is constructed, with the input data of typical desert grasses＇ spectral characters（red-edge, green peak, red valley, RVI, etc.） and the output data of vegetation types.Three groups of grass classification experiments based on hyper-spectral features demonstrate that： （1）Red-edge characteristics perform better than the other absorption features to obtain accurate classification results.（2）Between the narrow-band spectral classification interval 550～790 nm, interval 20 nm performs better than interval 10nm.（3）There are complex relationships between the input,output layers of BP neural network and the training time,precision of BP network.