Sub-Compartment Canopy Density Estimation Model Based On Multi-source GaoFen Satellite Data

Forest canopy density is an important basic parameter for forest resources monitoring,which plays an important role in forest resources management.In the past,remote sensing inversion of forest canopy density was based on a single remote sensing data source,and the construction and application of remote sensing estimation model combining optical and radar satellite data were few.What's more,the related research on estimation of forest canopy density combining domestic optical and radar satellite data was not reported.Therefore,taking the Picea schrenkiana Forest in the Western Tianshan Mountains as the experimental area and remote sensing quantitative estimation of sub-compartment canopy density of Picea schrenkiana as target,the feasibility of quantitatively estimating the canopy density based on sub-compartment scale multiple stepwise regression,BP neural network and Cubist models are explored by extracting characteristic samples from remote sensing images based on domestic optical and radar satellite data.It is hoped to provide reference for the techniques and methods of forest resources investigation and monitoring,and to promote the application of domestic satellite image data in forest resources monitoring.The main conclusions are as follows:(1)204 feature factors are extracted from Gaofen-2 satellite images,92 feature factors are extracted from Gaofen-3 satellite images,and 3 feature factors are extracted from DEM data.The correlation between canopy density and various characteristic factors is analyzed,and the factors with good correlation with canopy density are: NDVI in vegetation index,variance(VAR),homogeneity(HOM),contrast(CON),dissimilarity(DIS),entropy(ENT),second moment(ASM)and correlation(COR)in 8 texture features of Gaofen-2 and Gaofen-3 satellite images,and altitude and slope in terrain factors.According to the correlation coefficient between each characteristic factor and canopy density and the correlation between each characteristic factor,17 characteristic factors are finally selected as the input parameters of the model.(2)The sub-compartment canopy density estimation models based on optical image characteristics,radar image characteristics and combined optical and radar image characteristics are constructed respectively.By evaluating the accuracy of the model: Judging from the data sources,it is concluded that the accuracy of the sub-compartment canopy density estimation model based on optical image characteristics is generally higher with an average accuracy of 78.50%,followed by the sub-compartment canopy density estimation model combining optical and radar image characteristics with an average accuracy of 78.11%,and the last is the sub-compartment canopy density estimation model based on radar image characteristics with an average accuracy of 76.90%;As far as the three models are concerned,BP neural network model has the highest estimation accuracy with an average accuracy of 79.19%,followed by Cubist model with an average accuracy of 78.08%,and the last is multiple stepwise regression with an average accuracy of 76.25%;Judging from different canopy density grades,the estimation accuracy of medium canopy density is generally higher with an average accuracy of 84.33%,followed by high canopy density estimation with an average accuracy of 81.86%.The estimation accuracy of low canopy density is the lowest with an average accuracy of 55.48%.(3)The BP neural network model based on optical image features has the highest overall accuracy among all models with an accuracy of 80.51%.In the estimation of low canopy density,the BP neural network model based on optical image features has the highest estimation accuracy with an accuracy of 75.37%.In the estimation of medium canopy density,the Cubist model based on radar image features has the highest accuracy,87.46%,and in the estimation of high canopy density,the Cubist model based on the combination of optical and radar image features has the highest accuracy,89.17%.It shows that the satellite data of Gaofen-2 and Gaofen-3 have a certain potential in the estimation of sub-compartment canopy density of Picea schrenkiana Forest in Western Tianshan Mountains,which can provide reference for similar areas.