Research On Monitoring Model Of Soil Salinization Based On Remote Sensing Of UAV

Soil salinization is a global ecological and environmental problem,which mainly occurs in arid and semi-arid areas.It will cause soil compaction and fertility decline,which will affect crop yields.The Mongolian Hetao Irrigation Area is an important agricultural production area in China.At the same time,it faces very serious soil salinization problems,which seriously hinders the rapid development of agricultural production in the area.UAV remote sensing technology can quickly obtain remote sensing information of ground objects in a large range,and can effectively monitor the distribution of soil salinization in real time and dynamically.The ultimate goal is to formulate a certain soil salinization prevention plan to reduce the harm of soil salinization and increase crop yield.In this paper,four different degrees of salinization arable lands were used as the test area in Shahaoqu Irrigation Area of Hetao Irrigation District,Inner Mongolia.The soil salinity was obtained in the laboratory by collecting saline soil samples at different times and locations.At the same time,Related remote sensing images were collected by the UAV with cameras.The research explored the correlation between the information including the ground surface spectral reflectance,spectral index,and canopy temperature and soil salinity.Finally,three machine learning methods including support vector machine(SVM),back propagation neural network(BPNN)and Extreme Learning Machine(ELM)and partial least squares regression(PLSR)were used to build different soil salinity inversion models,it also evaluated and analyzed the accuracy of each model.In the end,the following main conclusions can be obtained:(1)The inversion models of soil salinity based on spectral reflectance in bare soil period are constructed.It is found that in the 6-band spectral reflectance,the blue band(490nm),the green band(550nm),the red band(680nm),and the near infrared band(800nm)all show a strong correlation with the soil salinity,the most relevant band is the red band.the model accuracy with different machine learning methods turn out to be significantly different:ELM>SVM>BPNN.Among them,the soil salinity estimation model constructed by ELM has the best prediction effect in June,and its modeling R~2and RMSE are 0.695 and 0.182,respectively,the R~2and RMSE of prediction are 0.717 and 0.171.In addition,the study also find that by removing the film from the remote sensing image,the correlation between the spectral reflectance and the soil salinity and the accuracy of the model inversion can be improved to a certain extent.(2)The inversion models of soil salinity based on spectral index in bare soil period are constructed.The research has analyzed correlation between 13 spectral indces and the salt salinity of the soil.Among them,the 8 spectral indices of S3,S4,S5,S6,SI1,SI2,SI3,and BI are significantly correlated with soil salinity,while the correlations between the 5 spectral indices including S1,S2,SI-T,SR,and NDSI and soil salinity are lower.The study also find that the inversion accuracy of the soil salinity model can be improved by the measure of film removal and constructing spectral index based on spectral reflectance.Among the soil salinity inversion models constructed based on 3 different machine learning methods,most of the models modeling and verification R~2are above 0.6,and the modeling and verification RMSE are below 0.2,showing a good inversion effect.Further comparison and analysis find that the ELM model has the highest inversion accuracy,followed by SVM,and the BPNN model is relatively the worst,which indicates that the ELM is the best model for soil salinity inversion.According to compare and analyze the data modeling effect in different months,it is found that the model inversion effect of the data construction in May is better,and the effect of removing the film in June is more obvious.(3)The inversion models of soil salinity under the condition of sunflower coverage based on optimal spectral index and canopy temperature are constructed,and the models inversion effect are compared and analyzed.The correlation between the soil salinity and the corresponding sunflower canopy temperature in different growth stages and different depths is different,but it all show a positive correlation.The correlation between soil salinity and canopy temperature of sunflowers in two growth stages is highest at the depth of 0-20cm,and the correlation coefficients are 0.422 and 0.404,respectively.In the inversion of the model,the inversion effect of the salt model in the budding stage is better than that in the flowering stage,and the inversion effect of the salt model at 0-20cm and 20-40cm depth is better than that at 40-60cm depth.The inversion effect of the salt model constructed by the salt index and spectral index variable group combined with canopy temperature is better than the salt model based on the vegetation index group.In addition,the accuracy of the salt model constructed by the 4 modeling methods is comprehensively analyzed.It is found that the salt inversion model based on the machine learning method is better than the partial least square regression model.Among the 3 machine models,the inversion effect of the ELM model is the best,followed by the SVM model,and the worst is the BPNN model.