Study On Stand Structural Parameters Estimation And Site Quality Evaluation Of Eucalyptus And Chinese Fir Plantation Forests

With the rapid development of China’s economy and society,facing the national timber security,ecological security and green sustainable development and other major strategic needs,effectively promote the cultivation of planted forests to improve the quality and efficiency and forest quality precision,for the protection of national ecological civilization,to promote sustainable economic and social development is of great significance.Real-time and accurate grasp of stand structure and site quality of plantation forests is an important basis for precision silviculture of plantation forests,which is the basis for realizing the principle of suitable land and suitable trees for plantation forests and provides important information support for effectively improving the quality of plantation forest cultivation and productivity of forest land.Traditionally,the acquisition of stand structure and site quality mainly relied on ground measurement,which has problems such as high cost of data acquisition,poor timeliness and limited coverage area.Multi-source and multi-scale remote sensing data can provide the possibility to characterize stand structural parameters and site quality with high accuracy,low cost and large scale.This study took two typical tree species（Eucalyptus and Chinese fir）of southern plantation forests in China as the research objects,combined ground survey data,cooperated with various remote sensing data such as UAV,airborne,satellite,and cloud computing platform,and used various algorithms and models such as machine learning,deep learning and theoretical growth equation to characterize key stand structural parameters（diameter at breast height,stand height,volume,stem density）and forest age of southern plantation forests.We extracted key stand structural parameters and forest age with high accuracy,and finally applied the status index approach to quantitatively evaluate site quality of southern plantation forests.This study aimed to provide an effective theoretical basis and methodological support for the high-precision evaluation of sited quality of typical plantation forests in southern China,so as to promote land and tree suitability,precision silviculture and forest quality improvement of plantation forests.The main study results were as follows:（1）A technical framework for estimating structural parameters of southern plantation forests based on multi-source remote sensing and advanced modeling algorithms was created to effectively improve model estimation accuracy,achieve high-precision,low-cost,and wide-range inversion of stand structural parameters and their spatial distribution,and quantitatively analyze model uncertainty caused by error accumulation.A deep learning regression algorithm（Deep-RBN）for high accuracy estimation of stand structural parameters under limited sample numbers was developed based on airborne Li DAR point clouds and sample plot data for Eucalyptus and Chinese fir.The algorithm was based on quadratic asymptotic fitting of optimized radial basis networks and deep neural networks for small-sample deep learning regression.The efficiency and capability of the algorithm were evaluated by comparing the estimation accuracy and mapping results of the traditional multiple linear regression algorithm（MLR）and fully connected network algorithm（FCN）for key stand structural parameters（mean diameter at breast height,mean stand height,volume and stem density）.The results showed that the Deep-RBN algorithm achieved the highest prediction accuracy among all modeling algorithms（R²=0.67-0.86,r RMSE=6.95-20.34%）and the FCN algorithm（R²=0.58-0.81,r RMSE=8.27-30.99%）had slightly lower modeling accuracy than Deep-RBN,and the MLR algorithm had the lowest modeling accuracy（R²=0.52-0.76,r RMSE=8.32-31.03%）.In terms of forest structural parameters mapping results,Deep-RBN characterized more variance and detailed information.Through the parameter tuning process,the Deep-RBN model has the lowest error and its model training process was more stable when the learning rate is 0.001.By testing the sensitivity analysis of sample numbers on the modeling accuracy of Deep-RBN algorithm,it was found that the model estimation errors of DBH,mean stand height,and stocking volume were lower than 20%when the number of training samples was greater than 70.By testing the effects of terrain and tree species factors on the Deep-RBN algorithm,the differences between the two factors on model accuracy were found to be r RMSE=0.04-2.36%and r RMSE=1.04-3.13%,respectively.Second,an improved two-stage scale extrapolation method based on random forest algorithm was constructed to realize the scale extrapolation and spatial distribution mapping of stand structural parameters in collaboration with ground data,UAV-Li DAR strip point clouds and domestic GF-6 satellite imagery,and the model uncertainty caused by error accumulation in the extrapolation process was quantitatively analyzed,and the influence of UAV-Li DAR point cloud density and sampling intensity on the scale extrapolation modeling accuracy was also evaluated.The results showed that the algorithm achieved higher accuracy in estimating the stand structural parameters at the stand scale（R²=0.64-0.85,r RMSE=7.49-26.85%）.The sensitivity analysis of the UAV-Li DAR point cloud density revealed that the prediction model achieved relatively stable and acceptable accuracy when the point cloud density after decimation was greater than 10%of the original point cloud density(34 pts-m^-2).By the sensitivity analysis of the sampling intensity of UAV Li DAR,the prediction models maintained stable estimation accuracy when the sampling intensity was greater than 20%.For different stand structural parameters,the highest estimation accuracy（R²=0.71-0.92,r RMSE=5.93-11.69%）was achieved with stand dominant height,and finally,stand structural parameter mapping and pixel-level uncertainty mapping were generated for the full coverage of forest farm.（2）A framework for rapid extraction of Eucalyptus plantation forest age based on remote sensing cloud computing was created,which combined the characteristics of frequent disturbance and rapid renewal of Eucalyptus,and achieved a high-precision rapid acquisition of disturbance and restoration dynamics and forest age of Eucalyptus plantation forests with one season as the time scale.Using Google Earth Engine（GEE）as the computing platform,we obtained the image stack by season using Landsat long time series remote sensing images,and obtained the spatial distribution of Eucalyptus plantation forest disturbance and restoration years based on the core index IFZ of the vegetation change tracking model VCT,combined with the comprehensive determination conditions of IFZ and NDVI index.The spatial distribution of Eucalyptus plantation forest age was finally characterized.The results showed that the overall accuracy of the algorithm extracted disturbance years was 85.5%,and the overall accuracy of restoration years was 83.5%.This study also compared the time series dynamics of IFZ and NDVI in three typical sample sites,which demonstrated the feasibility of disturbance dynamics detection of Eucalyptus plantation forests based on remote sensing data.This study demonstrated that the season-scale Eucalyptus forest age detection algorithm based on the remote sensing cloud computing platform GEE can effectively improve the data processing efficiency,fit the growth pattern of plantation forests and meet the demand for high-precision acquisition of Eucalyptus plantation forest age at the forest farm level,which provided a feasible technical solution for efficient acquisition of age information of southern plantation forests on a large scale.（3）A Eucalyptus plantation site quality evaluation model based on theoretical growth equation and multi-source remote sensing was constructed,and the Eucalyptus polymorphic status index curve acquisition and status index spatial distribution mapping were realized.Firstly,based on Deep-RBN algorithm and airborne Li DAR,UAV-Li DAR point cloud and ground measurement data,the first and second phase subfarm-level stand dominant height mapping was generated.At the same time,the GEE-based quarter-scale Eucalyptus forest age detection algorithm was used to generate the first and second phase site-level stand age maps.Secondly,each theoretical growth equation was modeled based on the guiding curve method for the dominant height of Li DAR stands in the first and second phase and the stand age in the same phase.The results showed that the model fitting accuracy of Korf’s theoretical equation based on second phase was the highest（R²=0.922）.Based on this theoretical equation and adjustment coefficients for each age class,a table of Eucalyptus status indices in the study area was prepared.Then,six difference equations were established based on the algebraic difference method to construct the Eucalyptus status index model and polymorphic status index curves in the study area using the dominant height and stand age of the two phases of the stand,and using the UAV-Li DAR strips as a range.The results showed that the model based on Korf’s theoretical equation had the highest fitting accuracy（R²=0.876）.Subsequently,spatial distribution maps of the Eucalyptus status index were obtained using the dominant height and age data of Eucalyptus compartments and the Korf difference equation.The results showed that the largest number of compartments with Eucalyptus status indices of 19-22were found in the study area,accounting for 73.04%.The number of compartments with status indices of 12-18 accounted for 15.72%,and the number of compartments with status indices of 23-31 accounted for 11.25%.Finally,by comparing the existing status index models based on ground data,it was found that when the stand age was 5a,the dominant wood mean height range of the two theoretical equations was the same,indicating the same status index.Compared with the theoretical values of stand dominant height at stand age 1a-5a,the difference between the theoretical values of stand dominant height at stand age 6a-10a gradually increased,and the difference between the theoretical values of stand dominant height at stand age 10a was 4.23 m.The source of this error may originate from the errors in the inverse model of stand dominant height and stand age extraction.