| Soil salinization restricts the sustainable development of irrigated agriculture.For soil salinity with strong temporal and spatial variability,a variety of scale conversion methods can realize remote sensing data scaling up and facilitate remote sensing inversion of surface salinity information at different scales,so as to improve the monitoring accuracy of soil salinity by remote sensing technology.In this research,the Shahao Canal Irrigation Area in the Hetao Irrigation District of Inner Mongolia is taken as the research area,and four plots with different salinization degrees in the irrigation area are selected as the experimental area.The soil samples collected from the field in April(before spring irrigation)and June(after spring irrigation)are used as the research objects.In addition,the spectral parameter information is obtained from UAV images and satellite images that are consistent with the ground data collection time,and the UAV upscaling model is constructed based on the combination of different model input variables and the spatial scale conversion method,and the satellite data is corrected,so as to realize the integrated air-space monitoring of soil salinity;In the thesis,fractal theory is used to find suitable observation scales in UAV scale space.On this basis,the relationship between remote sensing spectral index and soil salinity is analyzed by grayscale correlation,and a corresponding soil salinity monitoring model is constructed to explore scale conversion,so that the inversion accuracy of soil salinity by different models before and after scaling conversion is explored;This thesis also uses the image spectrum analysis theory to analyze the differences of remote sensing spectral parameters under different conversion paths from the perspective of energy,and compares and evaluates their influence on the monitoring accuracy of soil salinization.The main results of this research are as follows:(1)The upscaling model constructed based on the dominant variation weight method can correct the GF-1 satellite remote sensing data and improve its inversion accuracy of soil salinity.Through statistical indicators such as mean,standard deviation,information entropy and average gradient,the three scale conversion methods are used to evaluate the promotion effect of the three scale conversion methods on the four-band image scale of the UAV in the test area,and the indicator system is analyzed and compared with the original image.It is found that the dominant variation weight method has the best conversion effect.According to the construction of soil salinity monitoring model under the combination of three model inputs for different data sources,it is found that in the UAV upscaling model,the dominant variation weight method has the best monitoring effect,the local average method is the second,and the nearest neighbor method has the worst effect.By screening out the two models with the best modeling and verification effects under the multiple linear regression and BP neural network methods,and combining the band ratio mean method to revise the GF-1 satellite data,the results show that the best monitoring model is based on multiple linear regression models for groups of mixed variables.Its R_v~2is 0.420 and RMSE_vis 0.219,which is 0.217 higher than R_v~2and 0.013 lower than RMSE_vof the multiple linear regression model of the mixed variable group obtained directly from the GF-1 satellite data.It can be seen that the upscaling model based on the dominant variation weight method can effectively improve the accuracy of the GF-1 satellite remote sensing inversion of soil salinity.(2)The optimal transformation scale in the UAV scale space is determined through fractal theory,so as to construct a soil salinity monitoring model and compare it with the original scale.A continuous spatial scale conversion model for the normalized vegetation index NDVI is established,and the most reasonable scale level is determined to be level 14through the evaluation parameter system,and the spatial resolution of the image corresponds to 1.4m.Under the original scale(0.1m)and the optimal converted scale(1.4m),a soil salinity monitoring model for two sampling depths is constructed by combining mathematical modeling methods and grayscale correlation analysis.The results obtained are that the models with the best monitoring effect at the depths of 0-10cm and10-20cm are the SVM models with the scale of 1.4m,and the corresponding R_v~2are 0.743and 0.763 respectively,and the RMSE_vare 0.172 and 0.136 respectively.Compared with the 0.1m scale model,the R_v~2of the modeling set and the validation set at 10-20cm is improved by at least 0.8,and the RMSE is also significantly reduced.(3)The monitoring effects of remote sensing spectral parameters under different scale conversion paths are evaluated.Taking the ratio vegetation index RVI-2 as an example,RVI-2 is deduced to different scales by local average method and three variation weight methods.It is found that there are differences in the spectral statistical curves of RVI-2features in different orders,that is,aggregation before inversion and inversion before aggregation.The peaks of the two spectrum curves change obviously with the scale,and the specific performance is that the peak of the spectrum decreases with the increase of the scale.The angular spectral energy is concentrated atθ=158°,and the radial spectral energy is concentrated at low frequencies(r=2).At the same scale,the radial and angular spectral peaks of the RVI-2 image obtained by the dominant variation weight method in aggregation before inversion transformation method are the highest,and its model accuracy is the best among all regression models constructed according to the aggregation before inversion transformation method;The radial and angular spectral peaks of the RVI-2 image obtained by the local averaging method in aggregation after inversion transformation method are the highest,and the model accuracy of arithmetic mean variation weight method is the best among all regression models constructed according to the transformation mode of aggregation after inversion. |
