Estimating Time Series Leaf Area Index Based on Recurrent Neural Networks

Abstract

Leaf Area Index is a key parameter of many meteorological,environmental and agricultural models.At present,global LAI products of several sensors have been released.However,due to the limitations of the retrieval methods and the qualities of the reflectance products,the released LAI products,generally retrieved from a single sensor data,have been found maybe not continuous in time series and can not characterize the natural growing vegetation well.Many research results in different domain have found that the neural network has nearly perfect performance in recognizing patterns in complex,nonlinear data,e.g.,a three-layered neural network could be used to approximate any continuous function if nonlinear increasing function and linear function are used respectively in its hidden layers and output layer.As to a span of the LAI values of an individual vegetation type in a same area,if there were no natural disasters or human destructions,they should present as a nonlinear continuous time series.In this study,by fusing the MODIS and VEGETATION products,time series LAI were used to construct recurrent neural networks,namely the NARX neural network,for six typical vegetation types.Meanwhile,time series reflectances in red,near infrared and shortwave infrared bands and the corresponding sun-viewing angles were introduced into the NARX neural networks as the exogenous inputs to estimate time series LAI.The validation results show that the NARX neural network is competent to estimate time series LAI and the predicted LAI of it is more continuous and smoother than that of the original MODIS LAI in time series.Thus the proposed method may be helpful to improve the quality of LAI products of the typical vegetation types.