基于参数化和非参数化法的棉花生物量高光谱遥感估算

【目的】利用高光谱遥感数据快速、无损地估算棉花生物量,评估参数化与非参数化方法在棉花上的表现差异。【方法】本研究以4个棉花品种在2个年份(2004和2005年)的试验资料为基础,将2年数据分别进行建模和验证,采用参数化算法(植被指数法、连续小波变换)与非参数化算法(偏最小二乘回归、随机森林、人工神经网络、回归树、袋装树和增强树、支持向量机和高斯过程回归)分别构建吐絮前和吐絮后的生物量估算模型。【结果】近红外与红边波段仍然是棉花生物量遥感监测中最有效的波段区间。参数化方法运算简单,效率高,其中,CI_{red edge}证明是棉花生物量估算上表现最好的植被指数,具有较高的独立验证结果(吐絮前:RMSE=27.23 g·m^-2;吐絮后:RMSE=48.81 g·m^-2)。基于连续小波变换的方法缓解了植被指数的低估现象,尤其是吐絮后(吐絮前:RMSE=31.54 g·m^-2;吐絮后:RMSE=37.57 g·m^-2);在非参数化法中,随机森林是棉花生物量估算的最优算法(吐絮前:RMSE=2...

Remote Sensing Estimation of Cotton Biomass Based on Parametric and Nonparametric Methods by Using Hyperspectral Reflectance

【Objective】The aim of this experiment was to use hyperspectral remote sensing data to estimate cotton biomass quickly and non-destructively, and to assess the performance differences between parameterized and non-parametric algorithms on cotton. 【Method】This experiment was based on the dataset of four cotton varieties in two years (2004 and 2005), and the two-year data were modeled and verified respectively. The biomass estimation models in different periods (before and after boll opening) were built by utilizing parameterized algorithms, including vegetation index method and continuous wavelet transform, and non-parameterized algorithms, including partial least squares regression, random forest, artificial neural network, regression tree, bag tree and enhanced tree, support vector machine and Gaussian process regression, respectively. 【Result】Near-infrared and red edge bands were still the most effective bands in monitoring cotton biomass of remote sensing. The parameterized method was simple, efficient and accurate. Among the parametric methods, CI_{red edge} was proved to be the best vegetation index for cotton biomass estimation with high independent verification results (before boll opening: RMSE=27.23 g·m^-2; after boll opening: RMSE=48.81 g·m ^-2). The result based on continuous wavelet transform alleviated the underestimation phenomenon of vegetation index, especially after boll opening (before boll opening: RMSE=31.54 g·m ^-2; after boll opening: RMSE=37.57 g·m ^-2). Among the non-parametric methods, the random forest was the best algorithm for cotton biomass estimation (before boll opening: RMSE=20.48 g·m ^-2; after boll opening: RMSE=30.28 g·m ^-2). The estimation accuracy of the two types of algorithms was affected by cotton wool, and the estimation accuracy after boll opening was significantly lower than before boll opening. 【Conclusion】In this study, the cotton biomass estimation models based on parameterized and non-parameterized algorithms were evaluated, and it was proved that the non-parameterized algorithm had high inversion accuracy and could be used as an important method for non-destructive monitoring of cotton biomass.