Mapping Growing Stem Volume Of Chinese Fir Forest Based On Spectral Saturation Of Features In Huangfengqiao Forest Farm

Forest Growing stem volume(GSV)is the main parameter for evaluating the quality of forest resources.The accuracy of remote sensing estimation of GSV depends on the extracted remote sensing features and the selected models,especially the feature evaluation criteria,which are highly correlated with the accuracy of GSV estimation.Currently,commonly used feature evaluation standards only evaluate the sensitivity of remote sensing features and GSV from a single perspective of correlation or importance.In addition,the problem of spectral saturation frequently encountered in remote sensing estimation of GSV further leads to the phenomenon of underestimating high GSV areas,which seriously hinders the accuracy of GSV remote sensing estimation.Therefore,improving and delaying the phenomenon of spectral saturation can help to improve the accuracy of GSV estimation.This paper focuses on the artificial Chinese fir forest in Huangfengqiao Forestry Farm,Hunan Province,and employs three optical remote sensing datasets,namely GF1,Sentinel-2,and Landsat 8,to extract various remote sensing features,including band values,vegetation indices,and texture factors,after image preprocessing.The spectral saturation of different features is quantified using the Kriging spherical model and quadratic model,and remote sensing feature evaluation criteria that consider spectral saturation and correlation coefficient criteria(Pearson correlation coefficient and distance correlation coefficient)are constructed,namely the PS method(Pearson&Saturation)and DS method(DC&Saturation).Moreover,the optimal feature sets for the three datasets are obtained using the remote sensing feature evaluation criteria that consider spectral saturation and the stepwise regression algorithm,and four models,namely multiple linear regression(MLR),k-nearest neighbors(KNN),support vector machine(SVM),and random forest(RF),are employed to estimate GSV.Based on this,a GSV estimation model based on multi-source remote sensing data is established by combining features from multiple optical data sources.The main conclusions drawn from this experiment are as follows:(1)Both spherical and quadratic models have the capability of quantifying spectral saturation,with minimal differences observed in saturation quantification for the same feature across different saturation quantification models,while larger differences in spectral saturation were observed for different types of remote sensing features.Among the top ten features with the highest correlation from different data sources,the saturation of remote sensing features extracted from GF-1 ranged from 201.52 to 474.54 m3/hm2,from 152.21 to 474.54 m3/hm2 for Sentinel-2,and from 169.72 to 474.54 m3/hm2 for Landsat 8.(2)The remote sensing feature evaluation criteria based on saturation and correlation coefficients contribute to the sensitivity assessment of remote sensing features and GSV through both spectral saturation and correlation.The results indicate that the feature evaluation criteria based on saturation and correlation perform well in GSV estimation,with a significant improvement over correlation criteria.Specifically,compared to the Pearson correlation coefficient method,the rRMSE reduction ranges of the four GSV models(MLR,KNN,S VR,and RF)using the PS method are between 1.11%and 6.78%,while those of the DS method compared to the distance correlation coefficient method are between 1.38%and 5.37%.The best model is the Sentinel-2 SVR model in the PS method(R2=0.52,rRMSE=26.65%).(3)Compared to single optical data,the combination of features from multiple optical data sources achieves complementary advantages and disadvantages between data sources,delays spectral saturation,and further improves the accuracy of GSV estimation.Among the three feature combinations in the multi-source data,GFSentinel-Landsat(G-S-L)shows the highest accuracy,and the best model is the RF model in the PS method(R2=0.63,rRMSE=23.72%)which reduces the rRMSE by 4.95%,6.46%,and 6.34%compared to the single data source PS method’s RF models of GF-1,Sentinel-2,and Landsat 8,respectively.Furthermore,the method based on multi-source optical data can significantly improve the overestimation of low GSV,especially improving the accuracy of GSV estimation in the range of 100-300 m3/hm2.The results demonstrate the significant advantages of multi-source optical data that consider spectral saturation in GSV estimation.