Remote Sensing Estimation Of Forest Crown Closure Based On ZY-3 Imagery

Crown closure(CC)is a crucial parameter for indicating forest ecosystem and growth conditions.Estimating CC with high precision is important for calculating soil erosion,evaluating forest productivity benefits,improving the accuracy and quality of forest resources survey,and is also of practical significance to the construction of ecological civilization.In this study,the Fuerhe Basin,located in the southwest of Dunhua city,was taken as the study area.High spatial resolution image(ZY-3 image),digital elevation model(DEM,spatial resolution of 6m) and field data were used for the extraction of spectrum and textural features,topographical parameters,and measured CC.The stepwise linear regression(SLR),back-propagation neural network(BPNN)and linear spectral mixture model(LSMM)were chosen to develop the inversion model of CC.Those prediction models of CC were validated with field data and compared by coefficient of determination R~2,root mean square error(RMSE),relative root mean square error(rRMSE)and estimation accuracy(EA).The best model was applied to estimate the CC of the study area and the spatial characteristics of CC were analyzed.The basic conclusions are drawn as follows:(1)Establishment of CC inversion modelsCC inversion model based on SLR.17 independent variables were selected from 50 original variables using correlation analysis and principal component analysis for the establishment of the model,including B2_DIS,B4_HOM,B3_HOM,B2_HOM and Red*NIR/Green(RNG).The R~2,RMSE and EA of the model were 0.845,0.035 and 94.95%.CC inversion model based on BPNN.19 variables were selected as the input layer,and CC was defined as output variable.Learning rate,training function,output function,transferring function and number of neurons in the hidden layer were adjusted,and finally the 19-9-1 network structure was confirmed.The R~2,RMSE and EA of the model were 0.883,0.038 and 97.20%.CC inversion model based on LSMM.Three types of end-members were selected from ZY-3 image including vegetation,low reflectivity features,and bare soil or sand.Fully Constrained Least Square was sued to un-mixing pixel and four sequent images were obtained.The mean value of RMSE was 0.01,and the R~2,RMSE and EA of the model were 0.571,0.12,and 82.07%.(2)Comparison of CC inversion modelsThe accuracy of three models was compared based on statistical indices.The BPNN had outperformed the SLR and LSMM,whereas SLR was better than LSMM.Through comparison of inversion results,we found that LSMM had a large degree of underestimation on high CC value.BPNN and SLR had a higher spatial agreement.The SLR overestimated when CC was greater than 0.7,and the BPNN was more reasonable in the estimation of lower and higher CC.The BPNN had a higher precision than SLR and LSMM.The main reasons included:First,the multiple scattering effects on ground objects cannot be ignored in forest areas with complex terrain,and the nonlinear model was superior to the liner model.Second,all of the input variables were comprehensively considered when BPNN was used to predict forest CC,while deep learning method tends to fully explore the internal relationship between independent variables and CC,which can provide more abundant information for a high-precision model.Therefore,the BPNN was selected to estimate CC of the study area.(3)Estimation and analysis of spatial characteristics of CCThe predicted values of CC ranged from 0.2 to 0.92,mostly concentrated between 0.46 and 0.73,which indicated that there was little extremely dense and sparse forest.Timely harvesting of forests with high CC can not only help young trees grow fast,but also help to reduce the surface fuel load and reduce the potential risk of fire.In the southwest and non-forested area,low CC below 0.46 were observed,while high CC distributed in the north and southeast area.The relationships among spatial distribution of CC with elevation,slope,dominant tree species,RNG and B4-HOM were analyzed.The results showed that:(1)Forest was distributed in the area that the elevation was below 834.7 meters,slope was below 17.1 degree,and CC value had positively correlations with elevation and slope.(2)The dominant tree species had a great influence on CC,the mean value of CC was higher of the Tilia amurensis and betula platyphylla,and lower of Juglans mandshurica,Acer mono,Fraxinus mandshurica and Phellodendron amurense.(3)More than 60% of the RNG values in the study area ranged from 0.2 to 0.3,and with the increase of CC,the proportion of RNG between 0.2 and 0.26 decreased,meanwhile the proportion of RNG between 0.3 and 0.98 increased.(4)The mean value of B4-HOM in forests was 0.28,which indicated the grayscale of forest interior texture was quite different.CC had similar spatial pattern with B4-HOM.As the values of CC increased,the proportion of B4-HOM greater than 0.32 increased,while the proportion of B4-HOM less than 0.2 decreased.