Estimating forest aboveground biomass in tropical forests with Landsat time-series data and recurrent neural network

The insensitivity and saturation of sensor signals significantly limit the application of Landsat data on estimating biomass in tropical forests with a high level of biomass. The previous studies with single-date Landsat data have lower accuracies in tropical forests compared to boreal and temperate forests. Landsat time-series data provide a promising opportunity to improve the accuracy by enhancing the relationship between Landsat spectral reflectance and forest aboveground biomass with disturbance and recovery dynamics. Compared to the single-date image, Landsat time-series data can capture abrupt spectral changes (e.g. harvesting and fire) and show the regrowth process in forested pixels. However, very limited studies take advantage of Landsat time-series data to estimate aboveground biomass in tropical forests. Recurrent neural networks (RNNs) are powerful deep learning techniques to capture time dependencies in sequence data. However, the application of RNNs in estimating forest biomass has not been explored yet. Therefore, we integrate the long short-term memory network (LSTM) and the fully connected neuron network (FNN) to establish an RNN-FNN model for estimating forest biomass with Landsat time-series imagery and airborne LiDAR data. We compared the proposed model with the commonly used Random Forest model and linear regression model which are implemented with single-date data. The results show that the RNN-FNN model can deal with Landsat time-series sequence data to enhance the relationship between Landsat spectral reflectance and forest aboveground biomass. The proposed model achieves the R2 of 0.63 and RMSE of 25.5 Mg/ha, which significantly outperformed the Random Forest model and linear regression model with Landsat single-date data. This study demonstrates the value of RNN and Landsat time-series imagery in estimating forest biomass for tropical forests.