Research On The Estimation Of Farmland Soil Salinity Based On UAV Hyperspectral Data

Soil salinization is an important resource and ecological problem in the world.As one of the major forms of soil degradation,the increase of soil salinization poses a serious threat to crop growth,agricultural product quality,biodiversity and regional sustainable development.Rapid and accurate information on soil salinization on a large scale is particularly important for understanding the current situation of soil salinization and effectively restoring the soil environment.In recent years,remote sensing technology has been widely used for soil salinization monitoring due to its high spatial and temporal resolution and large amount of information.The Yellow River Delta is rich in land resources and is one of the largest and most promising river deltas in China.Due to the serious seawater intrusion,high salinity of groundwater,high salinity of soil-forming parent material,and high evaporation,soil salinization is a serious problem,which has become an important factor limiting the sustainable development of agriculture and ecosystem in the region.In order to achieve high precision quantitative estimation of soil salinity,this study explores the soil salinity estimation method based on UAV hyperspectral data using the farmland in Nanzhang Village,Huanghekou Town,Yellow River Delta as the study area.First,the response law of UAV hyperspectral soil salinity spectra was explored by analyzing the response characteristics and consistency analysis of the measured soil spectra and UAV hyperspectral.Combined with the soil salinity data collected in the field,one-dimensional feature spectra and two-dimensional feature spectra sets were constructed.Four different methods of Partial Least Squares Regression（PLSR）,Random Forests（RF）,Support Vector Machines（SVM）and Back Propagation Neural Network（BPNN）were used to develop a quantitative estimation model of soil salinity.The estimation accuracy of soil salinity by different dimensional feature spectra,preprocessing methods and quantitative models were explored separately,and the best combination of methods for UAV hyperspectral soil salinity estimation was selected and determined.The main research contents and conclusions are as follows:（1）The measured spectra were compared with the UAV hyperspectra to clarify the pattern of the UAV hyperspectral response of soil salinity.The quality of the S185 UAV hyperspectral data was evaluated using the measured spectra of ASD.The results showed that the soil spectral data of S185 had high quality in the wavelength range of 450～850 nm,and the change trend of hyperspectral reflectance of ASD was basically consistent with that of the hyperspectral reflectance of S185 UAV.The hyperspectral images obtained by UAV were denoised based on Multiple Scattering Correction（MSC）,Standard normalized variate（SNV）and Savitzky-Golay filtering techniques.The results show that the signal-to-noise ratio of the UAV hyperspectral images after MSC processing has been greatly improved,and the signal-to-noise ratio in four bands,including 622 nm and 674 nm,has been significantly improved.Then the spectral reflectance before and after denoising was compared,and it was found that the reflectance distribution of each sample was more concentrated after MSC treatment,which indicated that the baseline shift and nonlinear shift caused by the scattering effect due to the uneven distribution of particles among the spectral data of each sample were well corrected.Finally,the correlation between the spectral reflectance and soil salinity after the denoising treatment was analyzed,and it was found that the correlation between the spectral reflectance and soil salinity was significantly enhanced after the MSC treatment.After a comprehensive comparison,MSC has the best denoising effect.（2）Fully exploit the hyperspectral soil salinity sensitive information of UAV to construct one-dimensional feature spectra and two-dimensional feature spectra.The original spectral reflectance was subjected to nine transformations,such as inverse,logarithmic,first-order differential and second-order differential,to obtain the one-dimensional spectrum.The one-dimensional spectra and soil salinity were analyzed by correlation analysis and gray correlation analysis,respectively.The results showed that after correlation analysis,the bands that passed the significance test the most were 1/R,log（1/R）and log R transformations.After 1/R conversion,it is negatively correlated with soil salt content in the wavelength range of 450～494 nm and 658～850 nm.The maximum correlation coefficient between log（1/R）and soil salt content is 0.578 at 554nm.After log R transformation,it is negatively correlated mainly in the wavelength range of 498～654 nm and862～850 nm.By analyzing the gray correlation,the log R transform had the highest correlation with soil salinity,log（1/R）,（log R）’and 1/R in that order.combining the results of correlation analysis and gray correlation analysis,the final one-dimensional feature spectral sets were selected as 1/R,log（1/R）,（log R）’and log R.Two-dimensional spectra were obtained by arbitrarily combining all the bands of nine one-dimensional spectra into pairs using a two-band algorithm.The correlation analysis and gray correlation analysis of the two-dimensional spectra with soil salinity were carried out separately,and the final two-dimensional characteristic spectral groups were obtained as follows:1/R-SI₂,log R-SI₂,log（1/R）-SI₂,and 1/R-DSI.（3）To reveal the effects of nine spectral transformations,multidimensional feature spectra（1D and 2D）and four estimation models on the accuracy of UAV hyperspectral soil salinity estimation,select the optimal combination of methods and construct the best estimation model.Soil salinity estimation models were established by using PLSR,RF,SVM,BPNN with 1D and2D feature spectra as independent variables and soil salinity content as dependent variables,respectively,and finally,the best estimation model was selected by comparing the accuracy of multiple models.The results show that compared with the one-dimensional spectra,the two-dimensional spectral indices can fully amplify the spectral response characteristics and effectively enhance the correlation between the spectra and soil salinity content,so that the accuracy,stability and applicability of the model can be significantly improved.For the one-dimensional feature spectral model,the RF model constructed based on log（1/R）has the highest accuracy and better model stability;for the two-dimensional feature spectral model,the RF model constructed based on 1/R-DSI has the highest accuracy and best model stability.For different modeling methods,overall the modeling accuracy and validation accuracy of RF and BPNN models were higher than other models,and the model stability was better than other models.By comparing and analyzing each soil salinity estimation model,the 1/R-DSI-RF model had the highest accuracy with modeling set R² of 0.829 and RMSE of 0.716,validation set R² of 0.815,RMSE of 0.619,and RPD of 1.921.