Research On Muti-scale Spectral Mixture Analysis Method For Sparse Photosynthetic/Non-photosynthetic Vegetation In Arid Area

Arid area accounts for one third of the whole land area in the world.Desert vegetation plays an important role in ensuring the ecological integrity of the arid area ecosystem.Photo synthetic vegetation(PV)and non-photo synthetic vegetation(NPV)are the important factors for describing plant survival and monitoring plant productivity in arid area,and also important indicators to indicate the change of ecological environment.Therefore,timely monitoring of photosynthesis/non-photosynthesis desert vegetation coverage is a necessary condition to guide management practice on land desertification and research on the mechanism driving vegetation recession.There are considerable potential in accurate estimation of photosynthetic vegetation fractional coverage(fPV)and non-photosynthetic vegetation fractional coverage(fNPV)by remote sensing images with the diversity of spatial resolution,spectral resolution and temporal resolution characteristics.Due to the special environment in arid area,there are still many difficulties in estimating fPV and fNPV accurately by remote sensing technology.Therefore,it is a major problem for how to improve the accuracy of the sparse fPV and fNPV estimation in arid area by remote sensing technology.In view of this,Minqin oasis-desert transition zone of Gansu province in western of China was taken as the research area in the paper,and we took sparse desert vegetation in arid area as the study object.Then we systematically carried out the research on the spectral mixture analysis for fPV and fNPV estimation in arid area,and linear spectral mixture model(LSMM)and nonlinear spectral mixture model(NSMM)and their corresponding unmixing algorithms were selected to quantitatively estimate fPV and fNPV by remote sensing images with the different spatial resolution and spectral resolution.And we analysized the spectral mixing process between vegetation and bare soil,then it was applyed to monitor fPV and fNPV with long time series in Minqin.The specific research results are summarized as follows:(1)In this paper,an improved FCLS algorithm(named p-FCLS)based on Fully Constrained Least Squares algorithm(FCLS)is proposed,which is suitable for the sparse vegetation endmember abundance estimation in arid area.The efficiency and accuracy of the algorithm is compared with FCLS,Non-negative Matrix Factorization algorithm(NMF),Least Squares Orthogonal Subspace Projection(LSOSP)and its extended kernel-based nonlinear unmxing algorithms for KFCLS,KNMF,KLSOSP and p-KFCLS.The results show that FCLS algorithm is suitable for abundance estimation of LSMM/NSMM with small data,while p-FCLS is more suitable for LSMM with the remote sensing data to estimate fPV and fNPV.(2)By comparing the performance of LSMM and NSMM in retrieving fPV and fNPV from two typical desert vegetation(Nitraria canopy and Haloxylon canopy),the effects of spectral nonlinearity on fPV and fNPV estimation were studied.The main results were as follows:The correct selection of endmembers is important for improving the accuracy of vegetation coverage estimates,and in particular,shadow endmembers cannot be neglected.Secondly,for desert vegetation canopy,the Kernel-based Nonlinear Spectral Mixture Model(KNSMM)with nonlinear parameters was the best fit unmixing model,but the nonlinear spectral mixing process is not obvious at 10-60m spatial scale.In consideration of the computational complexity and accuracy requirements,LSMM is still the best model choice in the feasibility and performance of estimating photosynthetic and non-photosynthetic vegetation fractional coverage on the spatial resolution scale from 1m vegetation canopy to 60m remote sensing image.Furthernore,the vegetation canopy structure(planophile or erectophile)determines the strength of the nonlinear spectral mixture effects.(3)By comparing the performances of Sentinel-2A Multi-Spectral Instrument(S2 MSI)with Landsat-8 Operational Land Imager(L8 OLI)and GF-1 Wide Field View(GF1 WFV)sensors in distinguishing PV-NPV-BS in arid region.Results demonstrate that:With the higher spatial and spectral resolution,Sentinel-2A Multi-Spectral Instrument(S2 MSI)sensor shows the clear advantage for retrieving PV and NPV fractions compared to Landsat-8 Operational Land Imager(L8 OLI)and GF-1 Wide Field View(GF1 WFV)sensors,which solved a difficult problem of hindering the inversion of fNPV by multispectral satellite data.Secondly,Through incorporating more red-edge and NIR bands,the accuracy of NPV fractional coverage estimation could be improved greatly,but it is important to improve the accuracy of fPV and fNPV estimation simultaneously with more VIS and NIR bands(especially red-edge band)or with narrow bands,and the increasing of NIR bands proportion can reduce the sensitivity of NSMM to noise.Last but not least,high spatial resolution is especially important for inversion of fPV and fNPV in arid region.(4)Auto-Monte Carlo spectral Unmixing model(AutoMCU),it was concluded that AutoMCU is more suitable model to solve the problem of endmember variability in arid area by comparison.At the same time,by Google Earth Engine geographic cloud computing platform,a set of PV and NPV fractional coverage products with 30m resolution were generated(once every one year)by annual NDVI maximum composite Landsat series images from 1987 to 2017 in Minqin,Gansu province.Through the analysis of long-time series fPV and fNPV maps,it was found that:The PV and NPV fractional coverage increased significantly during the past 30 years,and the period 2007-2008 was the time node of PV conversion to NPV.The spatial distribution of the PV fractional coverage increased was mainly near the water source in the core area of Shiyang river and its southern area,while the area of NPV fractional coverage increased was distributed in the desert area of Minqin county outside the cultivated area.According to the above conclusions,the causes and root problems of PV and NPV fractional coverage changed in the study area were analyzed,and the construction measures of photosynthetic/non-photosynthetic vegetation were formulated.In general,spectral mixing analysis is considered to be the best method to distinguish sparse vegetation PV and NPV from bare soil in arid area.And linear spectral mixing model has been applied to many research fields,however,nonlinear spectral mixing model will be applied to a larger prospect in the future.In the study,we realized high-precision quantitative inversion of sparse PV and NPV coverage in arid area based on hyperspectral and multispectral remote sensing data,and fully exploited the potential of multispectral remote sensing,and studied the spectral mixing process of different vegetation types.In brief,we successfully expanded the application field of vegetation remote sensing.