| In this study, the forest volume of the key forestry city, Longquan in Zhejiang province of China, was predicted dynamically. First, the related factor set was established, including topography, climate, soil, forest structure, and spectral characteristics of forest etc. It was based on integration of satellite images, elevation, forest inventory data, fixed sample survey data and other data sources. Then, the membership of each variable was empirically fitted by polynomials, and the forest volume was estimated via an improved BP neural network model. The study can provide references for dynamic update of forest volume in the state of natural growth, and facilitate the decisions of forestry management departments. The main results of this study are as follows:(1) A comprehensive environmental factor set were established to estimate forest volume cost-effectively, including17factors:the depth of soil layer, the depth of soil A layer, altitude, slope, aspect, surface curvature, solar radiance, ground wetness, tree ages, canopy closure, normalized differential vegetation index (NDVI), and spectral of seven bands. To be specific, first, the satellite images of Longquan in the years of2003,2007and2010were radiometric and geometric corrected and cropped and overlayed with forest inventory data and fixed field sampling data in ENVI5.0and ArcGIS9.3to extract seven spectral factors:NDVI, and the spectral characteristics of the seven bands. Second, Altitude, slope, aspect, surface curvature, solar radiance, ground wetness were derived from Digital Elevation Model (DEM) in ArcGIS9.3. Third, the depth of soil layer, the depth of soil A layer, tree ages, canopy closure, forest volume per unit were calculated from forest inventory data.(2) The membership of each variable factor was obtained by polynomial fitting empirical data in Matlab R2011b, and an improved BP neural network model was established, which was based on Levenberg-Marquardt optimization algorithm and dominant species (fir, pine, hardwood category, Taiwanensis). On this basis, carry block training and simulation were carried out for2007forest resource inventory data in the study area. The average relative errors of the individual (IARE) were from27.00%to41.69%, with an average of33.73%; Groups relative errors (GRE) were from4.94%to7.55%, with an average of6.14%, this means that groups estimation was more than85%which is the overall sampling accuracy standard about volume of forest resource inventory, can be used to guide the production practice.(3) In Matlab R2011b, the average unit stock volume of fixed sample of subcompartment in2004was inversed via the model above by dominant species. The point pairs by the composition of predicted and measured values uniformly distributed on both sides of the diagonal, only with the exception of Taiwanensis. Group relative errors were from0.91%to10.48%, with the average of3.67%, which means that groups estimation reaches a very high accuracy (more than95%) and was more than the overall sampling accuracy standard about volume of forest resource inventory. Individual average relative errors were from17.28%to39.04%, with an average of27.31%. The most ideal result was on hardwood Masson class, the worst one was Taiwanensis.(4) In Matlab R2011b, the average unit stock volumes of the year2010fixed sample subcompartments were predicted via the model above. It is seen from the scatter plot and the error curve, the estimated and measured values of point pair mostly concentrated within a narrow range along the diagonal center line. The group relative error of prediction was6.73%, and the groups estimation accuracy was93.27%which was much more than the overall sampling accuracy standard about volume of forest resource inventory, can be used to guide the production practice. The individual average relative error was24.14%that there are61.84%of them were less than20%. These further demonstrate the established estimation model is of strong generalizability. |
PDF Full Text Request