Estimating Forest Leaf Area Index Based On BP-Nueral Networks

Leaf area index(LAI) is an important biophysical parameter of canopy structure that is related to biomass,carbon and energy exchange,and is an impotant input to ecological and climate change models.It is significance that estimating surface LAI for the study of earth ecological systems.How to quickly and efficiently estimate LAI of regional or global scale becomes a hot topic for research. There were five ground-based methods including destructive harvesting method,leaf litter collection, allometric equation,optical measurements and inclined point quadrat.These methods are limited to estimating LAI in small area,and can't meet to monitor the dynamic changes of LAI at the regional and global scale.Remote sensing is increasingly being applied to estimating LAI at the regional and global scale.There are primarily two approaches to estimate LAI from image data in the early times. The first approach is establishes a VI-LAI empirical relation of one or more variables between VI and biophysical variables LAI.Although it is simple and operationally feasible,there is no single VI-LAI equation that can be applied to remote-sensing images of different surface types.The second approach is canopy reflectance models which base on modeling the relationships between canopy characteristics and reflectance.Canopy model is usually non-linear,multi-input parameters, extremely complex.In recent years,artificial neural network technology has been taking into the study of estimating LAI from remote sensing that greatly enhanced the precision and speed of estimating LAI.In this study,ground-base LAI was estimated indirectly using the LAI-2000 plant canopy analyzer,leaf litter collection and allometric equation.Several studies had demonstrated that LAI-2000 has a tendency to underestimate LAI in conifers while the result of estimate LAI drived from leaf litter collection has a fine accordance with true LAI in situ.To calibrate the LAI-2000 PAI estimates,we established the linear correction model between the LAI determined from litter fall collection and optical LAI-2000 measurements.We corrected 300 values of LAI-2000 measurements with this correction model to provide higher accuracy of data modeling and validation for estimating LAI from remote sensing.In this paper,we use two methods estimating LAI from ETM+ data.One is based on the simple linear models between NDVI and LAI.The other one is use BP neural network training eight kinds of widely used vegetation indices to estimating LAI.The proposed strategy of BP model implemented in six sequential steps:(1) Geometric correction.The original image was correct with the ground topographic map 1:50000.(2) Atmospheric correction. Use COST model to correct the atmospheric influence.And the digital values of the ETM+ image are calibrated to spectral reflectance.(3) Eight kinds of vegetation index were established with the spectral reflectance.These VIs include NDVI,RVI,SAVI,ARVI,MAVI,MSAVI,SBL and TVI. (4) BP neural network training.(5) Estimating of LAI use trained BP neural network.(6) Valitdate the estimating result with the LAI in situ.The main results are as follows:(1) The linear calibration model between the LAI determined from litter fall collection and optical LAI-2000 measurements was useful to provide higher accuracy of ground-based LAI data for remote sensing methords.(2) The estimate result of LAI based on the simple linear model between NDVI and LAI was generally higher than the LAI value of in situ.The precision of LAI estimating from linear model was about 70%(at the level of 90%probability).(3) The estimate result of LAI use BP neural network has a well accordance with the in situ LAI data. The precision of LAI estimating from BP nueral network model was about 84%(at the level of 90% probability).