Soil Texture Changes And Hyperspectral Estimation In A Lakeside Oasis On The Western Shore Of Bosten Lake,China

Soil texture was one of the most important natural properties of soil.Its composition reflects the origin and condition of the parent material,and affects the soil’s resistance to wind erosion,water retention capacity,soil nutrients and soil capacity.Soil texture and soil spectral reflectance were correlated and can influence the hyperspectral reflectance of soil.The study of the hyperspectral response band of soil texture provides theoretical support for the rapid acquisition of soil texture information.In order to effectively use the spectral reflectance data,enhance the correlation between spectral reflectance and soil texture,and find the response spectra of soil texture and reflectance,the lakeside oasis on the western shore of Lake Constance was used as the study area,and the correlation between spectral data and soil texture was enhanced by performing five mathematical transformations and 10 scales of continuous wavelet transform decomposition on the spectral reflectance R.Sensitive wavebands were screened,and statistical analysis models（partial least squares regression）and capability of the models（random forest and support vector machine）were determined using the coefficient of determination R~2,root mean square error RMSE and relative analytical error RPD,and suitable combinations of data processing methods and models were found for estimating the clay,powder and sand contents of soil particles.The model was also used to extract the response bands of soil texture and soil hyperspectral reflectance by combining the mathematical equation of the relationship between spectral reflectance and soil texture.The main findings were as follows.（1）The soil texture at 0～50cm depth in the study area was mainly powder loam,followed by powder and sandy loam,and finally loamy sandy loam,and the soil particles were mainly powder particles,with powder particles（72.69%）> sand particles（21.68）> clay particles（5.63%）.The spatial variation of soil particles at depths of 0～50cm was characterized by a high percentage of clay particles in the south-eastern part of the study area,a high percentage of powder particles in the south-western and northeastern parts of the study area,and the opposite spatial variation of sand particles and powder particles.（2）After five mathematical transformations of the original spectral reflectance R,the correlation coefficient of the first-order differential R’ was higher than other mathematical transformations,and the number of bands that pass the significance test（p=0.01）was the largest,powder grain（628）> sand grain（573）> sticky grain（294）,and the band with the highest correlation coefficient was 1194 nm,with the absolute value of the correlation coefficient being 0.47 The highest correlation coefficient was1194 nm,with an absolute value of 0.47.The sensitive bands were concentrated in 750,1021,1283～1284,1789～1792,2173～2175,2234～2235nm and 2359 nm.After the continuous wavelet transform（CWT）and the decomposition of 10 scales,the 5th scale correlation was the most obvious,and the sensitive bands were concentrated in1194～1195,1663～1664,1745～1748nm and 2359 nm,the highest correlation coefficient r~2 value ranging from 0.24 to 0.25 in from 1745 nm to 1748 nm.（3）Using mathematical equations,it was possible to enhance the response relationship between soil texture and soil hyperspectral reflectance.A total of nine bands of 750,1021,1194,1283,1284,1608,1746,2243 and 2359 nm of R’ both passed the significance test with higher correlation coefficients than other bands and the best fit was 0.87.After CWT transformation,the bands of 1746 and 2359 nm of scale 2,the bands of After CWT transformation,the correlation coefficients of the bands 1194,1282 and 2359 nm at scale 3,1663,1745 and 1746 nm at scale 4,and the four bands from 1745 to 1748 nm at scale 5 were high and the fit was better.（4）The combined CWT5-SVM model is the best estimator,providing a good estimate of the soil fines content and a rough estimate of the sand content.Three models,PLSR,RF and SVM,were used to estimate the clay,powder and sand grains of soil particles in the study area.The RF model was constructed with the sensitive band modelling screened by R’,and the modelling set and validation set with powder particles worked best,the modelling and validation set determination coefficients R~2 value were 0.74 and 0.82,RMSE value were 10.19 and 7.37,respectively,and RPD value was 2.04.The SVM model was constructed with the sensitive band modelling screened by R’ The first-order differential R’ has a good response relationship to both RF and SVM models,and can better estimate the information of soil powder and sand content.After the CWT transformation,the SVM models established by the 5th scale with powder and sand particles had good response relationships,and validation sets R~2 value were 0.81 and 0.80,RMSE value were 7.63 and 8.37,and RPD value were 2.03 and 1.81,respectively.Overall,the first-order differential transformation R’ was better than the other four mathematical transformations,the 5th scale modelling of the CWT transformation was optimal.The SVM model was slightly better than the RF model.The estimation of powder particles was poor due to the sand particles and the estimation of sticky particles.