Multi-attention Generative Adversarial Network for multi-step vegetation indices forecasting using multivariate time series

Abstract

Generative Adversarial Networks (GANs) are one of the most significant research directions in the field of Deep Learning (DL). GANs has received wide attention due to their outstanding ability to produce realistic-looking images and can implicitly learn rich distributions over Spatio-Temporal (ST) correlations in Multivariate Time Series (MTS). Forecasting using MTS remains a challenge that needs to be solved to make successful tasks in different ST related applications such as vegetation forecasting. Vegetation is an extremely valuable component of our global ecosystem, as well as an important indicator of land cover change dynamics and productivity. As a result, the Normalized Difference Vegetation Index (NDVI) is one of the most widely used indices in vegetation related studies. The challenges of accurately forecasting vegetation indices come from both the non-stationary ST correlations of the NDVI data and the complex impact of the external context factors, including weather and remote sensing indices. It is further more challenging for most existing one-step forecasting methods, making an accurate forecast over many future time slots. To achieve this goal, we propose a new multi-attention GAN model composed of three main parts. An encoder network first encodes the driving sequence into latent space vectors and store all the information carried by the driving sequence. A generator is then in charge of generating forecasting data by extracting long-term temporal patterns. Eventually, in order to increase forecasting accuracy, we use as a third part an improved discriminator for classification and feedback purposes. The suggested model has been extensively tested with real-world datasets, demonstrating its effectiveness and robustness compared to existing methods. The test results indicate high performance with a Coefficient of Determination (R2) of 0.95, Root Mean Square Error (RMSE) of 0.04, Mean Absolute Error (MAE) of 0.01, and Mean Absolute Percentage Error (MAPE) of 15.35.