Scaling Effect And Scaling Transfer Method Research Of LAI Remote Sensing Inversion Incorporating Land Type Information

The inversion of leaf area indicator（LAI）by remote sensing has been widely concerned by researchers,but its results are affected by scaling effect,and there are significant differences between scales.Although there are many LAI scaling transfer methods,these methods are usually not perfect and cannot meet the practical application requirements of LAI product scaling transfer.Driven by application,this research integrates the information of the land cover types within the large pixel,and conducts in-depth exploration and research on the evolution law of scaling error and the method of scaling up.Integrate information entropy and internal structure parameters to quantitatively express the internal spatial heterogeneity of a single large pixel,and based on this,summarize the evolution and propagation laws of scaling error,and build a bridge between scaling errors and scaling transfer;Based on the law of conservation of matter,the scaling invariant theory of regional LAI true mean is proposed.The scaling error weight of pixels is determined one by one by using the variation coefficient weight method,so as to calculate the correction number of scaling error and establish the scaling transfer model;Optimize the evaluation indicator of image similarity,set the uncertainty threshold to meet the application requirements of LAI through literature review,so as to establish the evaluation indicator of the qualified rate of LAI image pixels after scaling transfer,and carry out objective and effective evaluation of the scaling transfer method in meeting application requirements.The following conclusions are drawn:（1）A coefficient C_sh suitable for quantifying the degree of spatial heterogeneity in LAI scaling is constructed.This coefficient introduces the surface context information,and integrates the three factors:ground category information entropy,ground category variance and ground category contribution,and expresses the influence that the intra-class and inter-class differences of the high-resolution pixels contained in the large pixels contribute to the spatial heterogeneity of the ground.Verified by simulated data and ground-measured data,it is shown that C_sh can more accurately and precisely describe the influence of ground object category information on scaling effects.（2）The influence of nonlinearity of LAI inversion model on scaling error is analyzed and summarized.It is demonstrated that the nonlinearity of the inversion model is a necessary condition of the scaling error,and a shape coefficient k is designed to measure the nonlinearity of the inversion model.The relationship between the four LAI inversion functions（quadratic term function,power function,logarithmic function and exponential function）and the nonlinear degree coefficient k of the curve is tested by experiments.It is found that the relatively specific numerical change of relative scaling error e is directly related to the inversion function,but the overall variation trend is almost positively related to the nonlinearity of the function.The scaling error caused by the reflectivity difference cannot be ignored.The relative scaling error caused by the reflectivity difference is 43.51%at the 240m pixel scale and 44.82%at the 480m pixel scale.（3）The influence of surface spatial heterogeneity on scaling error is analyzed and summarized.It is demonstrated that the spatial heterogeneity of the surface is a necessary condition for scaling errors.Through the simulation data experiment,the influence of the difference in the types of land cover and the proportion of different land cover types on the scaling error is discussed.The study found that when the number of land cover types mixed in the pixel is 2,the second type is gradually combined into the homogeneous surface（pure pixel）.As the ratio of the mixed class increases,the spatial heterogeneity coefficient becomes larger until the ratio reaches 0.5;it then reaches the peak and decreases as the ratio increases,showing a normal distribution.When the ratio of the mixed type is the same,the magnitude of spatial heterogeneity is positively correlated with the absolute value of the difference between the two types of land cover.When the number of land cover types mixed in the pixel is 3,if the pixel value of the combined third type is equal to or similar to the weighted average value of the other two types of initial pixels,the opposite compensation effect will be generated,and the spatial heterogeneity coefficient will decrease as the ratio of the third class increases.When the opposite compensation effect does not occur,the spatial heterogeneity coefficient resembles a normal distribution.If the ratio of mixed new types is the same,the spatial heterogeneity coefficient of the 3 mixed land cover types is larger than that of the mixed 2 types.Different from the 2 mixed land cover types in the pixel,the maximum spatial heterogeneity coefficient does not appear at the average ratio of the three types.（4）When studying the scaling error caused by the difference in spatial resolution,it is found that with the increase of the scale,the scaling error first gradually increases,and then the scaling error gradually decreases with the increase of the scale.Taking Landsat as an example,the 30m spatial resolution increases with the pixel scale,and the scaling error gradually increases.When the scale is increased to 480m,the scaling error reaches a peak value,and then as the pixel scale increases,the scaling error gradually decreases.（5）It is proposed that the truth mean of the LAI in region has scale-invariant properties.Under the framework of the law of conservation of matter,it is demonstrated that the truth mean of the LAI in region is a constant at any scale when the regional and time conditions are determined,that is,the truth mean of the LAI in region has the characteristic of scale invariance,and then the concept of the relative the truth mean of the LAI in region is extended.On this basis,two scaling-up methods,the mean difference removal method and the mean scaling error correction method,were developed,and the scaling transfer application tests were carried out in the Huailai Remote Sensing Experiment Station research area and the Maoershan Experimental Forest Farm research area respectively,and the transfer effect was evaluated according to the four statistical indicatores of MBE,RMSE,MAPE and Corr.The results show that the mean difference removal method and its improvement method,the true mean replacement method,have a certain effect on scaling up,but the effect is not significant.The root cause is that the transformation idea is simple.From the perspective of mathematical statistics,it is just considered that the average difference of scale error is the overall bias of scale error,which is corrected by subtracting the bias from the initial value,without considering the cause of the scaling error.The mean scaling error correction method uses the quantitative expression method of surface spatial heterogeneity and the coefficient of variation to determine the weight,and constructs four types of weights:the mean variation coefficient method,the variance variation coefficient method,the land types information entropy variation coefficient method and the C_sh variation coefficient method.The first two methods do not distinguish the land types,and operate the high-resolution pixels that constitute the large pixels as a whole;the latter two methods introduce land types information.The land types information entropy variation coefficient weighting method and the C_sh variation coefficient weighting method have obvious improvement effects.By observing the change range of each indicator,it is found that the land types information entropy variation coefficient weighting method has a narrower change range than that of the C_sh variation coefficient weighting method,indicating that the information entropy variation coefficient weighting method has better stability and is better than the C_sh variation coefficient weighting method;The mean variation coefficient weighting method has a little effect on high-resolution scaling up and large-scale scaling error correction,but its robustness is not strong;the variance variation coefficient weighting method has the worst performance and there is basically no correction effect.（6）Two evaluation indicatores,the spatial structure similarity based on edge information and the transfer qualification rate that meeting the application requirements of LAI,are proposed.Considering the fuzzy spatial structure when the scale spans large,based on the original GSSIM method,it is proposed to use the Laplacian operator to extract the edge texture information to test the spatial structure similarity between the test image and the reference image;The uncertainty of quantitative remote sensing products is an important indicator to measure the quality of remote sensing products.Based on the LAI uncertainty indicator given by some organizations,we have developed a pass rate indicator that meets the requirements of LAI applications.Through testing in the study area,it is found that the results displayed by the GSSIM evaluation indicator are relatively consistent with the aforementioned error statistical indicators;the results displayed by the LAI product qualification rate evaluation indicator are basically consistent with the GSSIM indicators,but there are also differences.