Soil Salinity Monitoring Model On UAV Multi Spectral Remote Sensing Under Vegetation Cover

Soil salinization is a soil degradation caused by natural and human factors,which affects the normal growth of crops and has a negative impact on agricultural development.Remote sensing monitoring of salinization provides a reliable basis for salinization control.High resolution is a significant feature of unmanned aerial vehicle(UAV)remote sensing,which can carry out fine research on spectral characteristics of different ground objects,providing a new idea for efficient monitoring of regional small-scale salinization.In this study,an UAV multi-spectral platform equipped with a small multi-spectral remote sensing camera is used to collect multi-spectral images during the vegetation coverage period of the Shaao Channel Irrigation test site in Hetao Irrigation District,Inner Mongolia.The threshold method is used to classify the multi-spectral images.Based on the classification results,the average reflectance of soil and vegetation canopy was extracted respectively,and the spectral index and image texture characteristic parameters were calculated to analyze the spectral characteristics and image texture of soil and vegetation spectrum under different salt conditions.The differences of soil salt content inversion models based on partial least squares(PLS),support vector machine(SVM)and extreme learning machine(ELM)is compared,and the model with the highest accuracy was selected.To extract the sample point single pixel reflectivity and calculate the spectral index,vegetation coverage pixel dichotomy model is used to calculate the pixel,analysis of bare soil(T1),low vegetation coverage(T2),neutral vegetation coverage(T3),high vegetation coverage(T4)of the four depths of 0-10cm,10-20cm,20-40cm,40-60cm of soil salt content and the relation between spectral index,and compare the coverage based on whole subset selection under different depth of sensitive spectral variables of PLSR and ELM differences in soil salinity inversion model and screen the model with the highest accuracy.The research results are as follows:(1)Establish soil salinity inversion model based on spectral reflectance of vegetation canopy spectrum,soil background spectrum and mixed spectrum of planting soil.By comparing the image classification accuracy of different vegetation indices,the results show that NDVI has the best image classification effect,with an average of 96.9%overall accuracy and 0.92 Kappa coefficient,which is the maximum of all vegetation indices.The spectral characteristics of vegetation and soil are analyzed and it was found that there are significant differences in spectral characteristics under different salinization levels.The spectral reflectance of vegetation in Band4,Band5 and Band6 bands is greater than that of soil spectrum,while that of vegetation in Band1,Band2 and Band3 bands is smaller than that of soil spectrum.With the increase of soil salinity,the reflectance of soil spectral bands increased.The reflectance of vegetation spectrum decreases in Band4,Band5 and Band6bands,and increases in Band1,Band2 and Band3 bands.PLSR,SVM and ELM regression algorithms is used to construct models under different treatments.It is found that ELM algorithm had the best modeling effect,and the R~2 of soil salinity inversion model is up to0.529.Comparing the model accuracy of the three treatments,the results show that the model based on soil spectrum has the highest accuracy,while the model based on soil mixture spectrum has the lowest accuracy.(2)Establish a soil salinity inversion model based on the full subset screening spectral index and soil texture characteristic parameters before and after removing soil background.By comparing the texture characteristics of images before and after removing soil background,it is found that compared with the image texture before removing soil background,the texture grooves of image after removing soil background become shallow,the texture increases and the texture becomes coarse,but the image clarity decreases.SVM and ELM soil salinity inversion models were established.The results showe that the ELM model have higher accuracy,and the R~2 of the ELM model reach 0.683.The models based on vegetation index and vegetation index plus texture feature parameters were established respectively,and it is found that adding texture feature parameters could effectively improve the model accuracy.It is found that removing soil background does not always improve the accuracy of soil salinity inversion model.In the soil salinity inversion model based on vegetation index and image texture characteristics,removing soil background reduces the accuracy of the model.(3)An inversion model of soil salinity with four coverage and four depths is constructed based on the combination of full subset screening sensitive variables.The total subset screening method was used to screen out the optimal sensitive spectral indices and sensitive band combinations with four coverage degrees and four depths.It is found that the salt index is more sensitive than the vegetation index under T1 coverage degree.Under the coverage of T2,T3 and T4,the number of vegetation index is greater than that of salt index,and the vegetation index is more sensitive.PLSR and ELM algorithms is used to construct four soil salinity inversion models under T1,T2,T3 and T4 coverage,respectively.It is found that the accuracy of the inversion model under bare soil and high vegetation coverage was higher than that under low vegetation coverage and medium vegetation coverage.By comparing the accuracy of PLSR and ELM SSC inversion models,the inversion accuracy of ELM model is higher than that of PLSR model.