| Forest’s influence on the earth’s eco-system is particularly of importance,which has gained an increasing attention from academia and governments.Forest carbon density has been recognized as one of the effective indicators of forest health and growth conditions.Due to its excellent characteristics including wide coverage,proper spectral and spatial resolutions and frequent revisits,remote sensing technology has been now an effective and reliable means for retrieving large-scale forest structural parameters and forest carbon density.Using theLandsat TM imagery acquired in 1988,1992,1997,2002,2007 and 2012,with a WRS2 Path/Row p122r043 in tandem with the national forest inventory data acquired in 1988,1992,1997,2002,2007 and 2012 the spatio-temporal patterns of forest carbon density in this study area were modeled and mapped in the current analysis.The major objective of this study was to extract the spatio-temporal dynamics of forest carbon density to inform the local forest management agencies.First,based on the original Landsat TM imagery,a bunch of remote sensing based features including the principal component analysis and K-T transform,and vegetation indices like NDVI,SAVI,texturalmeasures derived fromGray Level Co-occurrence Matrixand fourier transform etc were developed as a portion of predictor variables for modeling forest carbon density.And other geographical derivatives such as elevation,slope and aspect were also developed to correlate with the sample plot-level forest carbon density.To identify those optimal features as the inputs of modeling,the importance analysis process of random forests algorithm was implemented.The modeling methods involved in the current analysis included the random forest(RF),the BP neural networks,the support vector machines(SVM),the co-kriging and the geographically weighted regression.Results showed that the sample-plot level forest carbon densitymodelingR2 of RF model was about 0.871-0.912,with a validation R2 value ranging from 0.487 to 0.589.The modelingR2 of SVM model was about0.632-0.792,the validation R2 was at about 0.296-0.451;the modelingR2 of BP network model was about0.701-0.803,the validation R2 was at about 0.312-0.498.The modelingR2 of Co-kriging model was about0.695-0.908,the validation R2 was at about 0.336-0.676;the modelingR2 of GWR model was about0.789-0.911,the validation R2 was at about 0.398-0.702.Through comparing the advantages and disadvantages of each model in terms of accuracy,error,stability and computational efficiency theRF model was ultimately determined to retrieve and map forest carbon density in the entire study area.After a ten-folded cross validation process,the mapped spatio-temporal pattern of forest carbon density was deemed to be reliable to some extentby a created confidence map of forest carbon density,providing a reference and basis for local forest management practices.The spatio-temporal pattern of forest carbon density in a span of 24 years were analyzed and the major findings can be summarized as follows:(1)forest carbon density in the mountainous areaswasmuch higher than that in the plains.(2)There was an increasing trend of forest carbon density in the study area over time,the average of forest carbon density increased from 6.31t/ha in 1988 to 33.65t/ha in 2012.Particularly,between 1992 and 1997,the average growth rate of forest carbon density in the study area increased from 6.45 t / ha to 16.03 t / ha,which reached the peak.(3)The growth of forest carbon storage and carbon density is closely related to policy,afforestation and greening development practice,and ecological public welfare forest construction engineering,which are contributing to improve forest quality,making forest a sustainable increase in forest carbon density possible in the study area. |
PDF Full Text Request