Effect Of Sampling Scale On Soil Moisture Content Monitored By UAV Multi-spectral Remote Sensing

Soil moisture is essential for crop growth and development.Timely and accurate access to soil moisture is fundamental and critical in guiding irrigation decisions.Currently,UAV remote sensing has become a popular research direction for monitoring soil water content due to its advantages such as high mobility and timeliness.However,in addition to soil water content,there are a number of factors that can affect the imaging of UAS remote sensing systems,the individual image elements of UAS remote sensing images are not very representative(spatial heterogeneity within features),and there is currently no relatively standardized size sampling window for spectral information and its selection method when using UAS remote sensing to monitor soil water content.To address this issue,this paper uses three typical fields within the river-loop irrigation area as the study area,and acquires soil water content data and UAV multispectral remote sensing images of the study area during the bare soil and vegetation cover periods respectively.Different window sizes were used to extract the spectral information of the sampling points during the bare soil period,to construct their monitoring models for soil moisture content respectively,to test the significance of differences in the accuracy of different models using ANOVA,and to explore the choice of the optimal spectral sampling window size with a local variogram.Four soil moisture content monitoring models were developed for the vegetation cover period,depending on whether or not feature classes(variability in feature classes)and spectral information sampling windows(spatial heterogeneity within features)were taken into account,and finally the accuracy of the models was evaluated and the specific effects of feature classes and spectral information sampling windows on the model results were analysed.The main findings of the study are as follows:(1)The variance of the local variogram of remote sensing images in the monitoring area is 11*11-15*15,which is basically the same as the size of the optimal sampling window for UAV remote sensing information,and it is feasible to use the local variogram to select the optimal sampling window size.As the sampling window of spectral information increases,R~2tends to increase first and then decrease,reaching a maximum value of 0.2605 at a window size of 13*13.RMSE tends to decrease first and then increase,reaching a minimum value of 0.0167 at a window size of 7*7.The different size of the window for extracting spectral information from UAV remote sensing images affects the accuracy of the soil moisture monitoring model,and the difference in model accuracy tends to increase and then decrease as the difference in window size increases.(2)Spatial heterogeneity within features and variability in feature classes can have a negative impact on UAV multispectral remote sensing monitoring of soil water content.When the only feature classes are soil and vegetation,a reasonable filter window size can be selected based on the vegetation cover.Both filtering the original image and classifying the features for discussion can improve the accuracy of UAV multispectral remote sensing monitoring of soil water content.The highest accuracy of the soil water content monitoring model was constructed by combining the two,with a coefficient of determination R~2of 0.279and a root mean square error RMSE of 0.014.(3)The AVG model(which takes into account the spectral information sampling window but not the classification)shows an increasing and then decreasing trend with increasing vegetation cover,and is optimal in the 0.6-0.8 cover range.The monitoring effectiveness of the OG-C model(which takes into account the classification but not the sampling window of the spectral information)improves as the local variance increases and does not affect the monitoring effectiveness of the model when the local variance is greater than 0.1.Spatial heterogeneity within the same feature and differences in the spatial distribution of different features interact in influencing the monitoring effects of the OG model(which does not take into account classification and also does not take into account sampling windows for spectral information).