Research On Soil Moisture Content Inversion Method Based On Visible Remote Sensing Data And Neural Network

As an important part of agricultural water management,soil water content can provide data support for irrigation water monitoring,crop water demand estimation,irrigation water planning and so on.Satellite remote sensing inversion is an effective method to obtain remote sensing information of soil moisture in a large scale.Improving the accuracy of soil water content is very important for the simulation,evaluation and optimization of water resources.optical satellite and visible light remote sensing images are currently the most abundant sources of satellite remote sensing data.At this stage,the visible light-based soil water content inversion method is mainly to use satellite data to invert the soil moisture index or drought index,and then use the measured data to establish a regression relationship between the index and the actual soil water content.However,because the spatial and temporal changes of soil water content are affected by a combination of factors,the regression relationship has nonlinear characteristics,and the relationships determined by single scenes,single-temporal images and measured data rates are often not spatially and spatially portable,making the soil The remote sensing inversion of water volume is difficult to establish a more general model model,and can not be separated from the ground measured data.The neural network model has powerful nonlinear mapping capabilities.In this paper,the BP neural network model is selected and Hebei Province is used as the research area.Based on the domestic high-resolution optical satellites GF-1 and GF-6 WFV data sources,the input and structure of the BP neural network model for soil moisture inversion Preferably,improve the accuracy of the model for single-view and single-temporal soil water content inversion;introduce indicators that describe the difference in spectral characteristics of different spatiotemporal images,enhance the adaptability of the soil water content inversion model to spatiotemporal changes,and effectively improve the soil water content inversion The versatility of the model.The main conclusions are as follows:(1)Using the BP neural network to construct a remote sensing inversion model of soil moisture based on GF-1 data.For the selection of input data,among various combinations of DEM,vegetation index,drought index,and spectral reflectance,DEM,modified soil adjustment The combination of vegetation index(MSAVI),vertical drought index(PDI)and the reflectivity of the four bands has the best training and verification effect;the reflectivity of each band also has a significant impact on the accuracy of soil moisture inversion,of which green light and In the red band,the near-infrared band is the least affected.And the soil water content based on neural network inversion is 0.378% lower than the root mean square error of the traditional MPDI method,and the accuracy is significantly improved.(2)Compared with the data of GF-1,the red edge index constructed by using the newly added red edge band has a significant improvement effect on the accuracy of the soil water content inversion model.When GF-6 is used as the data source,the best input combination is the red edge index MTCI,PDI,DEM and the reflectivity of 8 bands,among which the bands that have the greatest influence on the accuracy of soil water content inversion are the red edges Band 1 and red edge band 2,followed by the yellow edge band,the least affected is the purple edge band,and the soil moisture content based on neural network inversion is 0.627%lower than the root mean square error of the traditional MPDI method,and the accuracy is significantly improved.(3)In order to explore the effect of different spatial and temporal satellite image spectral characteristics on soil moisture inversion,three linear equations of dry edge,wet edge and soil baseline were introduced in the spectral space to describe the red-near infrared spectra of different scene images Differences in feature space,and the slope and intercept of the three linear equations are added to the input of the BP neural network model for unified multi-temporal modeling.The results show that after introducing the spatial parameters of the input variables into the spectral space,the root mean square error of the soil moisture inversion based on GF-1 is reduced by 0.277%,and the correlation can reach 0.756;After the model has been trained and enough ground measured data is trained,the soil water content can be retrieved from the ground measured data.