基于分数阶微分光谱指数的冬小麦根域土壤含水率估算模型

为探讨分数阶微分（fractional-order differentiation,FOD）技术联合光谱指数改善高光谱反演冬小麦根域土壤含水率（soil moisture content,SMC）的效果，该研究以冬小麦为研究对象，测取高光谱反射率和土壤含水率数据，将高光谱反射率经Savitzky-Golay(SG)平滑处理后计算典型光谱指数以此构建偏最小二乘回归（partial least squares regression,PLSR）、随机森林（random forest,RF）和BP神经网络（back propagation neural network,BPNN）3种土壤含水率反演模型；将高光谱反射率进行0～2.0阶（步长为0.2）的分数阶微分处理后计算比值指数（ratio index,RI）和归一化指数（normalized difference index,NDI），分析不同阶的RI、NDI与SMC之间的二维相关性，筛选得出敏感光谱指数并分组，以此构建3种反演模型（PLSR、RF和BPNN）。结果表明：不同典型光谱指数与土壤含水率的相关性存在很大差异，相关系数波动范围在0....

Estimation model of soil moisture content in root domain of winter wheat using a fractional-order differential spectral index

Soil moisture content (SMC) is one of the most important conditions to affect crop growth and development in the irrigation of agricultural production. Monitoring SMC of crop roots is much more conducive to guiding precise crop irrigation. This study aims to investigate the effect of fractional-order differentiation (FOD) combined with the spectral index on the hyperspectral inversion of SMC in the winter wheat root domain. The winter wheat was selected as the research object in the Water-saving Irrigation Experimental Station, Key Laboratory of Agricultural Soil and Water Engineering in Arid Areas Subordinated to the Ministry of Education, Northwest A&F University, China. The data was also measured for the hyperspectral reflectance and soil water content. The hyperspectral reflectance was smoothed by Savitzky-Golay (SG). Then the typical spectral index was calculated to construct the partial least squares regression (PLSR), random forest (RF), and back propagation neural network (BPNN) inversion models of soil water content. The hyperspectral reflectance was processed by fractional differentiation of 0-2 order (step of 0.2), in order to calculate the ratio index (RI) and normalized difference index (NDI). A systematic analysis was made to determine the two-dimensional correlation between the RI of different orders, the normalized index and SMC. The sensitive spectral index was screened and grouped to construct three inversion models (PLSR, RF, and BPNN). The results show that: There was a very different correlation between different typical spectral indices and soil moisture content. The correlation coefficient fluctuated between 0.2 and 0.6, the highest of which with the SMC was the near-infrared spectral RI (about −0.6). The best effect was achieved in the SMC inversion model using the typical spectral index, followed by RF and BPNN models. The coefficient of determination of the improved model was 0.55, the root mean square error (RMSE) was 0.027 g/g, and the relative analysis error was 1.64. There was a high correlation coefficient between the sensitive spectral index and SMC after the fractional-order differential processing. The correlation coefficient between the sensitive spectral index and SMC showed a step change at different fractional orders. The correlation coefficient between the sensitive spectral index and SMC was higher at the order 0.2-1.0, and then fluctuated around 0.76, finally gradually decreasing from the order 1.2. The correlation between the sensitivity spectral index and SMC fluctuated around 0.65 after the order of 1.6. The optimal SMC inversion model was the RF model using the normalized sensitivity index after FOD treatment. The determination coefficient of the improved model was 0.75, the RMSE was 0.024 g/g, and the relative analysis error was 2.08. SMC inversion model with the improved fractional differential shared a significantly improved effect (the coefficient of determination from 0.55 to 0.75), compared with the typical spectral index one. The spectral details were mined as much as possible, while the noise was effectively removed more thoroughly without the information variables. As such, an SMC monitoring model was established using the optimal variable set. The findings can provide a reliable way for the hyperspectral monitoring of soil water content in the winter wheat root domain.