受污染胁迫玉米叶绿素含量微小变化的高光谱反演模型

Retrieval Model for Subtle Variation of Contamination Stressed Maize Chlorophyll Using Hyperspectral Data

通过野外实验测试和室内样品化验,获得3个不同污染状况农田样地自然环境下玉米的高光谱反射率、叶片的叶绿素含量、叶片和土壤的重金属含量等数据。对高光谱数据的可见光波段(400～800 nm)进行导数光谱计算和连续统去除处理,得到吸收谷位置、吸收深度、绿峰位置、绿峰处归一化反射值、红边位置、红边处归一化反射率、红肩位置、吸收宽度、光谱不对称度等光谱特征参数。分析上述参数的物理含义并将其和玉米叶绿素含量变化进行相关分析,选择并确定与玉米污染胁迫叶绿素微小变化有一定关系的参数,作为输入因子,建立BP神经网络模型,逐步增强并提取农田污染胁迫状态下玉米叶绿素含量的微小变化信息。

Chlorophyll content is an important indicator of photosynthesis activity, stress and nutritional state. In the present paper, the hyperspectral data, foliar chlorophyll content and heavy metal contents in foliar and soil were measured for the maize growing in three natural fields. In most previous research, the contamination stress was controlled artificially in laboratory by adding chromium, zinc or copper pollutant etc. to the soil, and the pollutant concentration added was much higher than that in natural environment. The three sample fields were under different heavy mental contamination level, but all located at the Changchun region, Northeast China, where is called Golden Maize Belts in the world. After continuum removal （400-800 nm）, ten spectral indices were computed including max absorption position, normalized reflectance at max absorption position, absorption depth, green peak, normalized reflectance at green peak, red edge, normalized reflectance at red edge, red peak, absorption width, and asymmetry degree. The physics meaning of the above indices and their correlation with maize foliar chlorophyll content were analyzed. It was found that there were close relationships between these indices and foliar chlorophyll content except max absorption position, green edge and asymmetry degree. Besides the asymmetry degree, five indices were selected in the stepwise multiple linear regression for estimating chlorophyll content and its determination coefficient （R^2） is 0. 702 7. Furthermore, in order to measure the weak change information of foliar chlorophyll content under the contamination stress, the BP artificial neural network （ANN-BP） was used. Several ANN-BP models were built and tried with different structure, namely five nodes, seven nodes or ten nodes in input layer, one hidden layer or two hidden layer, and different nodes number in hidden layers. It was found that the highest accuracy of estimates was obtained by the model with two hidden layers, ten nodes in input layer, seven nodes in first hidden layer and 4 nodes in second hidden layer （R^2 =0. 975 8）.