6533b7d6fe1ef96bd1266e98

RESEARCH PRODUCT

An Improved Forecasting Model from Satellite Imagery Based on Optimum Wavelet Bases and Adam Optimized LSTM Methods

subject

description

This paper proposes a new hybrid approach I-WT-LSTM (i.e., Improved Wavelet Long Short-Term Memory (LSTM) Model) for forecasting non-stationary time series (TS) from satellite imagery. The proposed approach consists of two steps: The first step aims at decomposing TS using Multi-Resolution Analysis wavelet (MRA-WT) into inter-and intra-annual components using 18 different mother wavelets (MW). Then, the energy to Shannon entropy ratio criterion is calculated to select the best MW. The second step is based on the LSTM model using Adam optimizer to predict the future. The proposed approach is tested using TS derived from Moderate Resolution Imaging Spectroradiometer (MODIS) images from 2001 to 2017. The results were compared with the predictions of the LSTM models using different optimizers. Additionally, the importance of selecting the optimal MW is analysed. The obtained results prove that the I-WT-LSTM approach outperforms all other methods.