Fractional Vegetation Cover Estimation Algorithm Based on Recurrent Neural Network for MODIS 250 m Reflectance Data

Abstract

Fractional vegetation cover (FVC) is a critical land surface parameter, and several large-scale FVC products have been generated based on remote sensing data. Among these existing products, the global land surface satellite (GLASS) FVC product, derived from moderate resolution imaging spectroradiometer (MODIS) 500 m reflectance data (MOD09A1), has achieved complete spatial-temporal continuity and satisfying accuracy. To further improve the spatial resolution of GLASS FVC product, this study developed a novel FVC estimation algorithm for MODIS 250 m reflectance data based on a recurrent neural network with the long short-term memory unit (RNN-LSTM). The RNN-LSTM was established using sequence training samples derived from the MODIS 250 m reflectance and GLASS FVC products, which were conducted over three vegetation types in mid-West China. Additionally, two machine learning methods, including the back propagation neural network (BPNN) and multivariate adaptive regression splines (MARS), were used to compare with the proposed method. The evaluation results showed that RNN-LSTM derived FVC had reliable spatial-temporal continuity and good consistency with the GLASS FVC product. Furthermore, the smooth temporal profiles of the RNN-LSTM FVC estimation indicated that the proposed method was capable of capturing the temporal characteristics of vegetation growth and reducing the uncertainties from the atmosphere and radiation. Finally, an independent validation case in the Heihe area indicated that the RNN-LSTM algorithm achieved the best accuracy (R2 = 0.8081, rmse = 0.0951) compared with the BPNN (R2 = 0.7320, rmse = 0.1127) and MARS (R2 = 0.7361, rmse = 0.1117). This study provides a new approach by showing the potential of the RNN-LSTM method for land surface parameter estimation and related research.