Estimation Of Forest Stock Based On Multi-Source Data With Different Canopy Cover

Forest stock is a factor reflecting the richness of forest resources and an important parameter for forest quality evaluation.Accurate estimation and research of forest stock based on remote sensing means is helpful to grasp the status of high-precision forest resources in real time,which is of great significance to the sustainable management of forests and the realization of the goal of"carbon peak and carbon neutrality".In this study,the forest of Huangshan District,Anhui Province was used as the research object,and GF remote sensing image,Sentinel-2 image,airborne laser scanner（ALS）point cloud data and sample survey data were used as data sources.Firstly,five data sources（GF,Sentinel-2,ALS,GF+Sentinel-2,GF+Sentinel-2+ALS）were divided to explore the Principal Component Analysis（PCA）,Pearson Correlation Coefficient（PC）,Combining Pearson Correlation Coefficient and Principal Component Analysis（PCA-P）three characteristic variable screening methods in volume estimation to find the optimal characteristic variable screening methods for different data sources.Secondly,Multiple Linear Regression models（MLR）,Decision Tree Regression（Dtree）,BP-Artificial Neural Network（BP-ANN）and Random Forest models（RF）were constructed to explore the optimal model suitable for forest stock estimation.Finally,the correspondence between Canopy Cover（CC）and plot volume was explored,and the applicability of five remote sensing data sources under different CC levels were evaluated from the aspects of accuracy and data acquisition cost according to the optimal volume estimation model.The main conclusions reached are as follows:（1）Among the three screening methods,PCA method is best when using three single remote sensing data sources,GF,Sentinel-2 and ALS.Among the two joint data sources,passive remote sensing combination（GF+Sentinel-2）and active and passive remote sensing combination（GF+Sentinel-2+ALS）,PCA-P method is the best.The PC method does not achieve the best results among the five data sources in this study.The PCA method and PCA-P method can reduce the number of independent variables,ensure the accuracy of the model,improve the modeling speed,and are effective variable screening methods.（2）Compared with the other three forest stock estimation models,the RF model has achieved the best estimation results in five different remote sensing data sources.The coefficient of determination（R²）of the model is between 0.59-0.82,the mean absolute error（MAE）is between16.04 m³·hm^-2-24.47 m³·hm^-2,and the root mean square error（RMSE）is between 23.92 m³·hm^-2-36.05 m³·hm^-2.（3）The three single remote sensing data of GF,Sentinel-2 and ALS have certain potential in forest stock estimation,and the estimation capacity order is ALS(R² is 0.79,RMSE is 25.58m³·hm^-2),>GF(R² is 0.61,RMSE is 35.19 m³·hm^-2)>Sentinel-2(R² is 0.59,RMSE is 36.05m³·hm^-2).The passive remote sensing joint data source can slightly better than the forest stock estimation results of the single GF and Sentinel-2 data source,with an R²of 0.63.The combined data source of active and passive remote sensing can fully consider the characteristic information such as spectrum,texture and altitude to obtain the optimal forest stock estimation result,and the R² is 0.82.（4）The results show that there is a good correspondence between CC and forest stock.Among them,low-level CC（≤30%）corresponds to low volume(≤70 m³·hm^-2),medium grade CC（30%-80%）,medium-volume(70-170 m³·hm^-2),and high-level CC（CC≥80%）corresponds to high volume(≥170 m³·hm^-2).（5）The optimal data sources for forest stock estimation under different levels of CC are different,and the optimal data sources for volume estimation at low,medium and high CC levels are Sentinel-2（R² is 0.62）,active and passive remote sensing（R² is 0.69）and ALS（R² is 0.72）.Without considering the acquisition cost of ALS data,the texture and spectral characteristics provided by Sentinel-2 data（10 m resolution）can be well fitted with low volume for forests with low CC level.For forests with medium CC level,fully combining the characteristics of different remote sensing data sources can effectively improve the estimation accuracy of medium volume,especially the highly correlated features provided by ALS data.For forests with high CC level,the combination of multi-source remote sensing data can reduce the accuracy of volume estimation,and the use of a single ALS data can provide a better fitting effect.If the cost of ALS data acquisition is considered,a single GF remote sensing data（2 m resolution）is used to estimate forest stock at medium and high CC levels,and R² can reach 0.60 and 0.56,respectively.