Retrieval Of Vegetation Leaf And Canopy Water Content From Hyperspectral Data Based On Spectral-shape Information

The quantitative remote sensing research on vegetation water content has considerable significance for rapid recognition of the drought and forest fire monitoring.It is an important input parameter to ecosystem models,and the accurate vegetation water content helps reduce the uncertainty in the results from these models.This study was aimed at establishing high-precision remote sensing models for leaf water content and canopy water content retrieval.The abilities of the models were tested using two in situ datasets(LOPEX93 and PANAMA)and one simulation dataset(PROSPECT)at the leaf level and one in situ dataset(Yunnan experiment)and one simulation dataset(PROSPECT+4-Scale)at the canopy level.Furthermore,the sensitivity analysis was conducted for the proposed models to the biophysical and biochemical parameters of plants as well as other parameters that can influence the spectral signals.Besides,generality and limitations of the proposed models were also discussed.Following conculusions can be drawn:(1)This study describes a methodology used to create spectral similarity water indices(SWIs)to accurately retrieve leaf water content based on the similarity between the leaf reflectance spectra and the specific water absorption spectrum.The abilities of six common distance metrics to capture spectral features were tested using two in situ datasets(LOPEX93 and PANAMA)and one simulated dataset(PROSPECT).These three datasets were also used to determine the most appropriate spectral intervals and to verify the performance of SWIs against six frequently used spectral indices that were specifically designed to estimate vegetation water content.Our results demonstrate that the Spectral Angle Cosine(SAC)is the most effective metric to capture useful spectral information pertinent to the equivalent water thickness(EWT),and three spectral intervals(970-1150,1330-1350 and 1584-1760 nm)are suitable for the retrieval of leaf water content.The SWIs were then created based on the SAC distances in these three spectral intervals respectively.The results demonstrate that SWI indices are better indicators of leaf water content and more tolerant to different species than the six spectral indices,including the Shortwave Angle Normalized Index(SANI),Shortwave Angle Slope Index(SASI),Moisture Stress Index(MSI),Normalized Difference Infrared Index(NDII),Normalized Difference Water Index(NDWI)and Maximum Difference Water Index(MDWI).In addition,the most accurate estimates of EWT were achieved from a single SAC distance with nRMSE of 4.08%((?)=0.98),3.63%((?)=0.95)and 8.11%((?)=0.80)for PROSPECT,LOPEX93 and PANAMA,respectively.SWI models adjusted for Hyperion bands produced a good performance as well,which provided a foundation for the retrieval of canopy water content.(2)The spectral downscaling model and leaf water content inversion model were combined to estimate mean leaf EWT at canopy level((?))and canopy water content per unit ground area(EWTcanopy)from the remote sensing image.Simulation analysis shows that the angle indices(AIs)which are the angles formed at vertex NIR by the Red-NIR-SWIR reflectance were negatively correlated to the probabilities of viewing sunlit foliage(PT)and were positively correlated to the probabilities of viewing sunlit background(PG).The AI(645,858,1640)was the best indicator of PT and PG for the Hyperion image,with high R2 of 0.837 and 0.800,respectively.The spectral downscaling model was obtained to estimate leaf reflectance spectra from hyperspectral remote sensing image through rational simplifying of the geometrical-optical model 4Scale by using AI.Good agreements were obtained between the ground measured and leaf reflectance spectra retrieved from the Hyperion image by using the spectral downscaling model in the Yunnan experiment.The (?) was estimated from leaf reflectance spectra retrieved from the Hyperion image,and EWTcanopy was then estimated based on the spatial distributions of the (?) and the LAI.The results demonstrate that the proposed EWTcanopy model((?)=0.45,nRMSE=20.82%)was better than the models based on the direct inversion method and the six spectral indices.(3)Analysis shows that the proposed leaf water content inversion models were not sensitive to the variation of leaf structure parameter(N)and dry matter content(Cm)as well as leaf thickness(T)and specific leaf area(SLA).The transferability(from simulation to in situ data)and generality(with different measured data)of the models were generally better than the six spectral index-based models.The results indicated that the proposed models have the advantages of strong anti-disturbance capacity and more universal applicability.In the application,parameters of the models should be adjusted according to different bandwidths of input spectrum,and the models based on SWI(1584、1760)and SWI(970,1150)require bandwidth of input spectrum to be lower than 30 nm and 13 nm,respectively.AI was not sensitive to the variation of leaf reflectance and forest background reflectance,and it can be used to retrieve the PT and PG from Hyperion and MODIS images with different spatial resolutions.The results indicate that the spectral downscaling model can work on different ecosystems and different spatial resolution images.The spectral downscaling model and canopy water content inversion model are sensitive to the variation of LAI,and the application of the models are limited when the vegetation coverage is much too low.