Research On Hyperspectral Inversion Of Soil Iron Oxide Based On Spectral Transformatio

Iron is one of the main elements in soil minerals and the most crucial transition element.Most iron in the soil exists in the form of a combined state.Because the accumulation and migration of iron oxide are affected by chemical weathering,biological cycling and other factors,it reflects the leaching process of soil,the degree of weathering development and the zonal distribution characteristics of the soil.It is often used as one of the indicators to describe soil development and classification.Therefore,mastering the iron oxide content is of great significance to reveal the soil environmental status,guide crop production and efficient land use.The conventional iron oxide content determination technology has high cost and long cycle.The use of hyperspectral technology can quickly and efficiently determine the content of soil iron oxide.Hyperspectral technology has the characteristics of high spectral resolution and strong wavelength continuity,which has become a new means of quantitative estimation of soil iron oxide.However,when soil types and land uses are diverse,the original spectra of soil often become more complex,especially when other substances such as water and organic matter in the soil can mask the spectral characteristics of iron oxide.Therefore,it is necessary to adopt an appropriate spectral transformation method to decompose the overlapping spectra in order to improve the prediction accuracy of iron oxide.This study collected 135 0-20 cm surface soil samples from the southern edge of Dinosaur Valley,Lufeng City,Yunnan Province.The samples’spectral reflectance and iron oxide content were measured in the laboratory.The soil spectral curve is the original spectral curve（OS）after edge band elimination,jump point correction and Savizky-Golay smoothing.Three spectral transformations are used to process the original spectrum,trying to find the best spectral transformation method.In this paper,conventional spectral transformations,including continuum removal（CR）,standard normal variable transformation（SNV）,absorbance transformation（Log（1/R））,continuous wavelet transform（CWT）of scale 1 to 10,and fractional differential transform（FOD）of order 0 to 2 are used to deal with OS.Then the characteristic wavelength was extracted by the correlation coefficient method（CC）,competitive adaptive reweighting algorithm（CARS）and Boruta algorithm.Finally,the inversion models of soil iron oxide in this region were constructed based on partial least squares regression（PLSR）,support vector machine regression（SVR）,random forest regression（RFR）and back propagation neural network（BPNN）,respectively,with the full band and characteristic wavelength as input variables.The results show that:（1）In this study,the iron oxide content and spectral characteristics of the soils in the study area were measured and analysed.After the data sets were divided using the K-S algorithm,the total,calibration and validation sets did not satisfy the normal distribution and showed a right skew.The shapes of the spectral reflectance curves were approximately the same for all samples.Also,the iron oxide content is inversely proportional to the spectral reflectance.The CST,CWT and FOD treatments of the OS showed that CST can highlight spectral features and attenuate noise;CWT can help highlight the features of the spectrum and thoroughly explore the weak features;and FOD can show the changes in spectral details and improve the resolution of the spectral curve,thus capturing spectral features that cannot be captured by integer order differentiation.（2）Characteristic wavelength extraction using CC,CARS and Boruta algorithms.It was found that both OS and Log（1/R）passed the 0.01 significance test for the whole band,while CR and SNV had relatively few significant wavelengths.The determination coefficients of CWT with high iron oxide content were mainly distributed in the visible band of the 3rd to 6th scales.The absolute values of the correlation coefficients between wavelet coefficients and iron oxide content at each scale tended to increase and decrease,and the number of significant wavelengths tended to increase gradually.In FOD,many positive and negative correlation peaks appeared gradually with the gradual increase in order,and positive and negative correlations appeared at adjacent wavelengths.The number of wavelengths passing the0.01 significance test decreases with increasing order,with the maximum correlation reaching a maximum at 0.75-order（-0.620）.The number and distribution of characteristic wavelengths extracted by the CARS and Boruta algorithms are generally similar across the different spectral transformations.The distribution of the selected characteristic wavelengths is mainly concentrated at 400,500,900,1400,1800,1900,2200 and 2400 nm.（3）There is no uniqueness in the combination of spectral transform,feature extraction and modelling methods,and cross-combinations are needed to verify the effectiveness of their methods.The prediction results of BPNN are all better,followed by SVR,and PLSR and RFR are comparable.The FOD obtains the best model prediction accuracy,and the CST and CWT have comparable model prediction accuracy;The CARS and Boruta algorithms can not only effectively extract the characteristic wavelengths,but also outperform the full-band and CC in terms of the number of input variables and model prediction performance.From the point of view of combining spectral transformation and characteristic wavelength selection method,the combination of CST or CWT with Boruta works better,while the combination of FOD and CARS works better.The 0.5-order-CARS-BPNN has the highest prediction accuracy and the best fit by combining the prediction results of all models with solid model stability and prediction ability.Its calibration set R² and RMSE are 0.917 and3.069,respectively,while the validation set R²,RMSE and RPIQ are 0.851,5.186g·kg^-1 and 3.907,respectively,with very excellent model prediction ability.Therefore,the 0.5-order-CARS-BPNN is the best prediction model for soil iron oxide.This paper explores three kinds of spectral transformation in decomposing overlapping spectra and improving the prediction ability of soil iron oxide,compares the influence of different characteristic wavelength selection methods and modelling methods on the prediction performance,finds the best combination of spectral transformation method,characteristic wavelength selection method and modelling method,and provides specific technical guidance for the rapid and accurate prediction of soil iron oxide content in this area.