Bi-LSTM Model for Time Series Leaf Area Index Estimation Using Multiple Satellite Products

Abstract

Time series leaf area index (LAI) is essential to studying vegetation dynamics and climate changes. The LAI at current status can be regarded as the accumulative consequence of the counterpart at prior times. Although the deep learning algorithm - Long short-term memory (LSTM) can capture long-time dependencies from sequential satellite data for time series LAI estimation, it only uses the information at prior statuses, and neglects the backward propagation of current vegetation change information. Thus, the LSTM-based LAI quality might be limited. In this letter, the bidirectional LSTM (Bi-LSTM) approach was proposed to integrate the information of multiple satellite products from both the past and future for temporal LAI retrieval. The fused values from GLASS, MODIS, and VIIRS LAI products, as well as MODIS reflectance in 2014-2015, serve as the output response and input for the Bi-LSTM training. Then, we compared the Bi-LSTM predictions with the counterparts from the LSTM, the fused LAI and three products using independent validation datasets in 2016. Results illustrated that our proposed Bi-LSTM method achieved better performance with higher accuracy (R²=0.84, RMSE=0.76) when compared to the LSTM estimation (R²=0.83, RMSE=0.82) and LAI products (R²<0.68, RMSE>1). Furthermore, our proposed method provided smoother and more continuous temporal profiles of LAI than other retrieval approaches.