基于人工神经网络的大豆叶面积高光谱反演研究

Soybean LAI Estimation with in-situ Collected Hyperspectral Data Based on BP-Neural Networks

目的探索不同高光谱模型监测大豆叶面积指数LAI的精度。方法实测不同水肥耦合作用下,大豆冠层的高光谱反射率与叶面积指数(LeafAreaIndex)数据,对二者进行相关分析;采用敏感波段(801nm,670nm)构建RVI,NDVI,SAVI,OSAVI和MTVI2植被指数,建立大豆LAI估算模型;最后采用相关系数较大的波段作为神经网络模型的输入变量进行大豆LAI的估算。结果大豆LAI与光谱反射率在可见光波段呈负相关、近红外波段呈正相关、红边处相关系数由负变正;微分光谱在三边处与大豆LAI关系密切,在红边处取得最大回归确定性系数(R2=0.86)。植被指数可以较为精确反演大豆LAI,确定性系数R2>0.84。人工神经网络模型可以大大提高大豆LAI的估算水平,当隐藏层节点数为2时,R2为0.92,随着隐藏层节点数的增加,R2可高达0.96;在没有黄熟期数据干扰的情况下,神经网络可以进一步提高大豆LAI的反演精度,R2可高达0.99。结论与基于植被指数建立的模型相比,神经网络模型可以有效避免因LAI过高而出现的过饱和现象,大大提高了LAI的反演精度。

[Objective] An experiment was carried out to evaluate the precision of hyperspectral reflectance model for monitoring soybean leaf area index （LAI）. [Method] Soybean canopy reflectance data collected with ASD spectroradiometers （350- 1 050nm）, which were cultivated in water-fertilizer coupled control conditions, and soybean LAI were collected simultaneously with LI-COR LAI-2000. Firstly, correlation between reflectance, derivative reflectance against soybean LAI were conducied; secondly, five vegetation indices with reflectance at bands 801nm and 670nm were applied to regress against soybean LAI; finally, ANN-BP was established for soybean LAI estimation with changeable nodes in hidden layers. [Result] It was found that soybean canopy reflectance showed a negative correlation with soybean LAI, while it showed a positive correlation with soybean LAI in near infrared region. Reflectance derivative had an intimate Co relation with soybean LAI in blue, green and red edge spectral region, and got maximum correlation coefficient in red edge region. All five vegetation indices had an intimate correlation with soybean LAI, with regression determination coefficient R2 ranged from 0.84 to 0.88. ANN-BP model could greatly improve soybean LAI estimation accuracy. Determination coefficient （R^2= 0.92） obtained with 2 nodes in hidden layers, however, R^2 still can be improved with nodes in hidden layers increasing, and R^2 = 0.96 with 8 nodes in hidden layers. Still, it should be noticed that without indecent phonological soybean data participate model establishing, ANN-BP model could improve estimation accuracy with large room, and Determination coefficient （R^2= 0.99） could be obtained with 8 nodes in hidden layers. [Conclusion] By above analysis, it is concluded that ANN-BP model could be applied to in-situ collected hyperspectral data for vegetation LA1 estimation with quite accurate prediction, and in the future, ANN-BP model still should be applied to hyperspectral data for other vegetation biophysical and biochemical parameters estimation.