Study On Wetland Information Extraction And Above-ground Biomass Estimation Supported By Worldview-2Images

Currently wetland resource is considered to be important the same as forest and marineecosystems and thus its dynamics and sustainable protection and utilization is of a majorconcern. Using remote sensing technology to monitor wetlands has become a more and morepowerful tool, which can solve some scientific problems we are facing in wetland research,such as wetland classification, wetland landscape mapping, wetland change detection and so on.Without doubt, the wetland study will promote the development of wetland ecological resourceprotection and therefore provide useful information for government’s scientific and correctdecision making in wetlant protection and reservation development.This research selected a core area of the wetland nature reserve of Dongting Lake inHunan provice as the study area. In order to accurately obtain wetland resource information ofthe core wetland region, accurate and detailed wetland type and vegetation classification wasconducted using high resolution image WORLDVIEW-2images. Moreover, the methods toextract wetland vegetation biophysical parameters were explored. An empirical model and avegetation structure parameter based above-ground wetland biomass estimation model weredeveloped to obtain the information of the wetland resources. The important finding were asfollows:(1)Wetland lassificationIn this study, a refined wetland classification system was first developed, consisting offour classes at the first level and eight classes at the second level. At the first level, thewetlands were classified into river, lake, grassland, and mudflat. At the secodn level, thegrassland was divided into Reed, Sedge, Polygonum hydropiper l and Mud artemisia. Afteratmospheric and geometric corrections, the spectral and wavelength reflectance characteristicsof WORLDVIEW-2images were then analyzed and four modified spectral indices includingmodified NDWI, modified NDVI, modified NDSI and NHFD were developed. Using theindices, the wetland classification was made. The obtained accuracy of the wetland classification was92.24%with Kappa coefficient of0.902. Moreover, object-orientedclassification for the detailed wetland vegetation types was carried out based on segmentationof spectral angle threshold and the obtained overal classification accuracy was higher than85%.(2)Wetland vegetation structure parameter modelingPearson product moment correlation between vegetation indices and leaf area index (LAI)was first analyzed. A total of seven vegetation indices including NDVI, RVI, DVI, SAVI,MSAVI, EVI and RDVI were selected as predictor variables of LAI. Various regression modelsincluding univariate linear regression model, polynomial (quadratic and cubic) models,exponential model, logarithmic model, and power regression model were used to fit the LAIdata and develop VI-LAI models. A sample consisting of23observations were employed forvalidating the models. The obtained optimal model to predict LAI was the NDVI basedexponential model with an accuracy of74.34%.After analyzing eight bands of WORLDVIEW-2images, a modified NDVI*basedvegetation fraction cover (VFC), NDVI*-VFC, was developed and this model was validatedusing the collected field observations. The obtained accuracy was87.8%. The overallvegetation fraction cover of the wetland region was64.3%. The VFC map was displayed withfive classes:0-20%,20-40%,40-60%,60-80%,80-100%. The results showed that the VFC ofthe study area could be accurately predicted using the modified NDVI*.In order to scientifically and reasonably predict vegetation biomass of the wetland region,the concept of plant structure parameters that have close relationship with vegetation biomasswas introduced, including average density PD, average height PHand average radius PLCRof theplants. In addition, the correlation analysis between each of eighteen image derived spectralvariables and each of the vegetation structure parameters were conducted and the first sixspectral variables having the highest correlation with vegetation structure parameters of sedge,red-knees herb, reed and artemisia selengensis were selected to develop linear stepwiseregression. The spectral variables were selected at a significant level of0.05. Whenheteroscedasticity appeared in a model and the quality of the regression model was dtramatically degraded, the spectral variable was removed. A total of9groups of multipleregression models were obtained. The results showed that the accuracy of all the models wereover75%. The obtained R2values and average relative errors of the estimates were reasonable.(3)Estimation of above-ground biomass for the wetland regionThe correlation between LAI and above-ground biomass was first analyzed. A total of35above-group biomass field observatons were then used to develop a relative growth model(Y=aXb) that accounts for the relationship of the above-ground biomass with LAI. The resultsshowed that the obtained:AGBiomass48.018LAI1.0278had a high correlation withabove-ground biomass. The relationship was stable and robust and can be used to accuratelyestimate above-ground biomass of the wetland vegetation.Moreover, vegetation structure parameters that were highly correlated with above-groundbiomass, including LAI, VFC, PD, PHand PLCR, were introduced into the above-groundbiomass regression models. Through correlation analysis of these predictor variables withabove-ground biomass, the optimal estimation models of biomass were obtained for typicalsedge, red-knees herb, reed and artemisia selengensis,respectively, with the accuracy of morethan70%. The reed above-ground biomass estimating model was the best, in which thecoefficint of determination was0.7288and root mean-square error was0.5121, averagerelative error was22.54%, and the overall accuracy was77.46%.In addition, the above vegetation structure parameter based biomass model was comparedwith the relative growth model and the widely used NDVI-based biomass model obtainedpreviously. The results show that the results from the vegetation structure parameter basedabove-ground biomass model were closer to the measurements and the estimate was12440.5294tons for the study region.Overall, this study led to the following innovations:(1) modifying four spectral indices(NDWI, NDVI, NDSI and NHFD) and greatly incraesing the accuracy of wetland classificationcompared with traditional classification methods;(2) putting forward a new idea with whhichvegetation structure parameters (PD、PHand PLCR) were introduced into estimation of wetland above-ground biomass and further using multivariate stepwise regression analysis method toseek the most important variables that were sensitive to the variation of wetland biomass and toobtain the optimal models for typical plant types (sedge, red-knees herb, reed and artemisiaselengensis); and (3) developing a relative growing model that is related to biophysicalmechanism and introducing LAI, VFC, PD, PHand PLCRwhich had close relationship withbiomass into biomass estimation to develop the vegetation structure parameter basedabove-ground biomass model. In a word, this study showed that using WORLDVIEW-2highspatial resolution imagery to perform wetland classification and information extraction,vegetation structure parameter inversion and biomass estimation was practical and the obtainedresults were accurate. The obtained methods through this research greatly enhanced the contentof wetland remote sensing research.