| Poyang Lake Wetland is China’s largest freshwater lake ecological wetland.It is an important organic carbon pool on land.The density of organic carbon in soil is high.It can store carbon for a longer period of time.The change in the content of organic carbon in the soil reflects the dynamic balance of oxygen and carbon dioxide in the atmosphere.With the rise of the global terrestrial biosphere carbon cycle research,the international community has gradually paid attention to the carbon cycle process and characteristics of this particular type of ecosystem.How to quickly and effectively obtain the organic carbon content in wetland soils has also become an important research topic in assessing environmental carrying capacity.With the widespread application of remote sensing technology,especially hyperspectral remote sensing technology,its advantages of high spectral resolution and multi-band provide an accurate,fast,non-destructive method to obtain soil organic carbon content indexes in a timely and effective manner.However,the soil hyperspectrum reflects comprehensive information about the physical and chemical properties of the soil,and how to effectively extract the characteristic spectrum and sensitive information of soil organic carbon in the spectrum with complex properties has become an urgent task.This has far-reaching theoretical significance for reducing the complexity of soil organic carbon estimation model and improving the robustness of the estimation model,and has important application value for realizing rapid monitoring of soil organic carbon in a region or a plot.This research focuses on the above hotspots and difficulties.Taking the wetland soils of Poyang Lake Basin in Jiangxi Province as the research object,246 sampling points were selected in-situ and statically collected using the ASD FielDSpec Pro FR spectrometer to collect their in-situ vis-NIR spectra.After air-drying,grinding,and sieving,the indoor spectra were collected,and the organic carbon content data of all samples were obtained by chemical methods.The feasibility of predicting soil organic carbon(OC)based on in-situ vis-NIR spectra was studied.Then use Absorptivity conversion,Savtzky-Golay smoothing,Absorptivity conversion + Savitzky-Golaysmoothing,and First derivative analysis + SG.Five methods of continuous removal(Continuous removal)are used to preprocess the spectral data.The processed spectral data is analyzed by linear PLSR and nonlinear LS-SVM respectively.For the same model,different preprocessing methods are compared horizontally Impact on model accuracy;for the same preprocessing method,the prediction effect of different models is compared longitudinally.At the same time,the modeling accuracy of indoor and outdoor in-situ spectra under the same preprocessing method is compared.Based on this,three methods,External parameter orthognolization(EPO),Direct Standardization(DS),Piecewise Direct Standardization(PDS),and Generalized least squares weighting algorithm(GLSW)are used to absorb Environmental influences are removed from the converted in-situ spectra.Finally,the prediction accuracy of in situ spectral organic carbon after three kinds of moisture removal algorithms are compared.The study found:(1)After using the continuum removal algorithm to amplify the spectral characteristics,the main difference between the original and the removed in-situ spectra in the field and the indoor spectra is that the absorption valley of the field spectral curve of wetland soil is much larger in width and depth than that of the indoor spectral curve,which is mainly caused by the long-term immersion of wetland soil.It is also found that the difference between the field in-situ spectrum and the corresponding indoor spectrum is the largest at the water absorption band.Therefore,moisture is the main environmental factor in the process of in-situ spectrometric measurement..(2)Among all the spectral preprocessing methods,the spectral modeling accuracy is the highest after Absorptivity conversion + SG smoothing.For in-situ spectra,the RPD of PLSR model after Absorptivity conversion + SG smoothing was increased from 1.96 to 2.52 and R2 from 0.74 to 0.84 compared with that of the model without treatment;the second largest one was the spectrum after First derivative + SG smoothing,and its RPD was increased from 1.96 to 2.34 and R2 from0.74 to 0.81;both pretreatment methods could significantly improve OC predictionThe accuracy of modeling can achieve the purpose of effective quantitative prediction.At the same time,compared with several spectral preprocessing methods,we also found that the improvement of modeling accuracy by a single processing method is weaker than that by a combination of two preprocessing methods,because the two combined methods not only take into account the amplification of spectral characteristics,but also balance the introduction of noise and invalid information.(3)Whether to indoor or in situ spectroscopy,the data mining algorithm based on nonlinear LS-SVM on the prediction precision of the proposed model is superior to the data mining algorithm based on linear PLSR will originally wild spectrum of the proposed model and found that the results of laboratory spectral modeling under each kind of pretreatment method of indoor spectroscopy originally prediction accuracy is superior to the wild because the spectral data because of its characteristics of high dimensional attribute causes the the relationship between soil organic carbon content is very complex,conventional linear method is often difficult to parse vis-NIR spectrum of soilAnd nonlinear data mining algorithms due to the strong distributed data processing as well as the ability to learn,in the operation of large special space can be fully approximation of complex non-linear relationships at the same time,because indoor spectral eliminates the environmental factors(especially water)to cover of spectral characteristics of soil organic carbon,thereby increasing the effective extraction of soil organic carbon spectrum information,cause indoor spectral prediction accuracy is superior to the in situ spectrum(4)The original spectra processed by the four moisture removal algorithms can be used to establish a high-precision organic carbon prediction model.Among them,the original spectral modeling processed by EPO and GLSW has the greatest improvement in accuracy.In comparison,the spectral modeling accuracy R2 after EPO processing can reach 0.84 and RPD can reach 2.55;the spectral modeling accuracy R2 after GLSW processing can reach 0.81,which is slightly lower than EPO,and the RPD reaches 2.58,which is higher than EPO.Among the four algorithms,EPO,DS,and PDS need to select a representative small sample from the entire sample set and bring it back to the laboratory to obtain the original spectrum in thefield and the indoor spectrum,respectively,and establish a transformation matrix through the correspondence between the two.,While calculating the conversion coefficient.GLSW uses a complete set of samples to establish a filtering matrix.The study found that in the process of establishing the conversion matrix,the singular sharp peaks appeared in the PDS converted spectrum,which further affected the accuracy of the subsequent modeling.The first-order differential processing of the PDS spectrum can alleviate this situation.The ordering of the modeling accuracy of the four algorithms from low to high is DS,PDS,EPO,and GLSW. |
PDF Full Text Request