Hyperspectral inversion of total nitrogen content in calcareous soil in karst areas
Nitrogen is a component of many important compounds in plants,such as proteins,nucleic acids and enzymes,and hence is indispensable for the growth of plants. The nitrogen content in soil is one of the key indicators of soil fertility. Calcareous soil is one of the main soil types in karst areas. A rapid and accurate estimation of total nitrogen (TN) content in calcareous soil is an important guarantee for the scientific evaluation of soil environmental quality in karst areas. In recent years,the rapid development of hyperspectral remote sensing technology has brought new opportunities for a quick assessment of soil physical and chemical properties. However,it is still extremely challenging to rapidly assess soil nitrogen content in karst areas by hyperspectral remote sensing due to the weak optical signal of soil nitrogen and the interference of factors such as the complex ecological environment and the strong spatial heterogeneity of soil TN content in karst areas.Karst areas are extensively distributed in China,where calcareous soil is one of the main soil types,exerting a great influence on ecological protection and agricultural development. Therefore,it is of great theoretical and practical significance for us to develop hyperspectral inversion models suitable for the TN content of calcareous soil. Karst landforms are distributed across 70 counties/cities in Guangxi,with an area of 97,700 km2,accounting for 41％ of the total area of Guangxi and 10.8％ of the total area of karst landforms in China. Taking calcareous soil in the karst areas of Guangxi as the research object,this study performed five mathematical transformations on soil spectra to improve the detection ability of spectral signals while eliminating spectral noise. Meanwhile,given the coexistence of linear and nonlinear relationships between soil TN content and spectra,the hyperspectral inversion capability of three models,namely partial least squares regression (PLSR),generalized neural network (GRNN) and PLSR_GRNN (a combined model of PLSR and GRNN),for soil TN content was compared and analyzed to establish a high-precision and rapid inversion model suitable for the TN content of calcareous soils in karst areas.The results showed as follows. (1) The TN content in calcareous soil was significantly correlated with various spectral bands from 400 to 2,500 nm. Among them,the TN content was more sensitive to the reflectance of the spectral bands near 600 nm,1,300 nm,1,600 nm,1,900 nm and 2,300 nm. (2) The first-order differential transform (FDR),second-order differential transform (SDR),reciprocal logarithmic transform (lg(1/R)),reciprocal logarithmic first-order differential transform ((lg(1/R))') and envelope removal transform (CR) of the original soil spectra can improve the capability of inversion of TN content in calcareous soil to some extent. The transformation effects were roughly ordered by (lg(1/R))'>SDR>CR>FDR>lg(1/R). Overall,the spectral differential transform is superior to the envelope transform as well as the reciprocal logarithmic transform,and can better exploit the detection capability of the spectral signal for soil TN. (3) The PLSR algorithm had excellent predictive ability for the variation of TN content in calcareous soil. In the SDR transform case,the model had the highest accuracy and better model robustness without overfitting,with a coefficient of determination (R2) of 0.84 and root mean square error (RMSE) of 0.55 in the modeling set and R2 of 0.82 and RMSE of only 0.64 in the validation set. Compared with the PLSR algorithm,the GRNN model had greater prediction ability. However,the robustness of GRNN model was worse and the overfitting phenomenon was obvious. In the same SDR transformation case,the modeling set R2 of the GRNN model could reach 0.92,but the validation set R2 was only 0.59,so the overall performance was inferior to that of the PLSR model. (4) The PLSR_GRNN model can integrate the advantages of PLSR and GRNN model,maintaining the high predictability of GRNN model and avoiding the overfitting phenomenon. Among them,the best inversion model was established by SDR,with R2 of 0.92 and 0.90 for the modeling set and validation set,and RMSE of 0.43 and 0.51,respectively,which were suitable for hyperspectral inversion of TN content in calcareous soil in karst areas. In addition,the FDR,(lg(1/R))' and CR transformations also had excellent performance,with R2 above 0.80 for the modeling set and R2 above 0.75 for the validation set.Although the prediction accuracy of the GRNN model cannot be improved by combining the PLSR model with the GRNN model,the overfitting phenomenon can be effectively controlled. This modeling approach,which combines linear and nonlinear models,is more widely applicable than the PLSR model or GRNN model alone,and is more adaptable to more heterogeneous soil types,and will be more widely used. Rapid and high-precision prediction of TN content in calcareous soil in karst areas can be performed based on hyperspectral models. The results can provide a basis for regional soil remediation and utilization.
soil total nitrogencalcareous soilhyper-spectrumpartial least squares regressiongeneralized regression neural network
万方数据
维普
