Inversion Study On Optimized KNN Method Of Vegetation Leaf Area Index

Leaf area index(LAI)is an important indicator for evaluating plant growth,development and health.Quickly and accurately obtaining the vegetation leaf area index is an important prerequisite for assessing the vegetation growth status and carbon sequestration capacity in desertified areas.The desert area has sparse vegetation and a vast area.The traditional manual survey method to obtain the leaf area index is time-consuming and labor-intensive.The remote sensing image combined with the ground survey plots to construct a leaf area index inversion model can make up for the deficiencies of the traditional survey methods.The selection of variable selection method and inversion model are two key factors to remote sensing inversion of leaf area index.Therefore,it is of great significance for remote sensing inversion of leaf area index in desertification area to select appropriate feature variable selection method and improve the existing remote sensing inversion method to improve the estimation accuracy and work efficiency of the model.In the study,Ganzhou District of Zhangye City was used as the study area,and 455 sample plots with a size of 30m × 30m were obtained by stratified random sampling.Using the ground area data of leaf area index,combined with Landsat 8 OLI remote sensing images to extract feature variables,linear stepwise regression method and random forest method were used to select the feature variables.In this paper,the optimized kNN based on random forest is proposed to improve the traditional inversion method of leaf area index.Two sets of characteristic variable were used to construct the inversion model and map the spatial distribution of leaf area index.In order to verify the effectiveness of the optimized method,multiple linear stepwise regression(MLSR),Support Vector Machine(SVM),Random Forest(RF),common kNN(k-Nearest Neighbors)and distance-weighted kNN model are compared with the optimized method for accuracy.A method suitable for inversion of vegetation leaf area index in desertification areas was discussed.The main research conclusions include:(1)The characteristic variables obtained from the combination of red,near-infrared and short-wave infrared bands are significantly related to the leaf area index,and have made important contributions to improving the accuracy of the inversion model.Using Landsat 8 OLI remote sensing image combined with leaf area index sampled field data to extract feature variables,a total of 259 feature variables were significantly correlated with leaf area index.There are 128 feature variables with absolute correlation coefficients above 0.500.The first five feature variables with absolute correlation coefficients are SR435,SR425,ARVI,SR415,and NDVI,and their correlations are-0.745,-0.744,0.743,-0.742,and 0.739 respectively.(2)There was no significant difference between the two sets of feature variables selected by the linear stepwise regression method and the random forest method.The feature variables screened by the two methods are SR435,SR51,B5 and SR537,SR527,SR517,SR546,SR547,SR415,SR534,SR536,SR315,NDVI,Elevation,SR56 respectively.The above variables are mainly calculated from a combination of near-infrared,red,green,and short-wave infrared bands,and are basically derived from high-correlation coefficients and high-importance characteristic variables,indicating that these bands are more sensitive to LAI inversion.Taking into account the characteristic variables of different wavelength combinations can better perform LAI inversion.(3)Among the two selection methods of feature variables,the linear stepwise regression method is better than the random forest method as a whole.For all the non-parametric models,optimized non-parametric models,and multiple linear stepwise regressions constructed by the linear stepwise regression method and the random forest method,the performance of ordinary kNN,distance weighted kNN and optimized kNN based on random forest in linear stepwise regression method is better than those of random forest method.However,for the random forest model,the random forest method is a selection variable based on the contribution of the feature variables to the random forest regression,so the modeling effect is significantly better than the linear stepwise regression method.The applicability of different feature variable combinations to the inversion model is inconsistent.During the inversion process,it is necessary to select an appropriate feature variable screening method according to the sample and model characteristics.(4)The inversion effect of the optimized kNN based on random forest is significantly better than other models.In all the established inversion models,the optimized kNN based on random forest has R2 of 0.783(RMSE=0.409,rRMSE=42.2%,MAE=0.246)and 0.762(RMSE=0.441,rRMSE=45.3%,MAE=0.252).Moreover,the RMSE,rRMSE and MAE are lower than other models,and the estimation accuracy is significantly better than other models.The rRMSE of the optimized kNN based on random forest is reduced by 24.8%and 28.4%respectively compared with the ordinary kNN model with the worst estimation effect,and the improvement effect is significant.At the same time,the residuals of the model are randomly distributed on both sides of the X axis,and most of the absolute values of the residuals are within 1,which indicates that the model results are reliable and the prediction effect is good.(5)The spatial distribution of leaf area index of the optimized kNN based on random forest is basically consistent with the actual leaf area index distribution,resulting in the best mapping effect in all models.Combining the Landsat 8 data with the measured LAI of the sample plots,a variety of inversion models were used to map the spatial distribution of the leaf area index in Ganzhou with a resolution of 30m.Both multiple linear stepwise regression and support vector machines produced unreasonable estimates,that is,there were estimates less than zero.The spatial distribution of LAI obtained by the optimized kNN based on random forest constructed by linear stepwise regression method is closer to the actual situation.The results showed that most of the vegetation LAI values in the central and western regions were greater than 1,and the northern regions were Gobi,with lower LAI values.The sparse grasslands are mainly distributed in the southern region,and the LAI value ranges from 0 to 1.The mapping effect of the optimized kNN based on random forest is better than other models,and it can provide a reference for the inversion of leaf area index in desertified areas.The research results can be used to simulate vegetation productivity,evapotranspiration,and net primary productivity,which can help to better understand the general process of ecosystems.