Dense NDVI Time Series by Fusion of Optical and SAR-Derived Data

Abstract

Gaps in normalized difference vegetation index (NDVI) time series resulting from frequent cloud cover pose significant challenges in remote sensing for various applications, such as agricultural monitoring or forest disturbance detection. This study introduces a novel method to generate dense NDVI time series without these gaps, enhancing the reliability and application range of NDVI time series. We combine Sentinel-2 NDVI time series containing cloud-induced gaps with NDVI time series derived from the Sentinel-1 synthetic aperture radar sensor using a gated recurrent unit, a variant of recurrent neural networks. To train and evaluate the model, we use data from 1206 regions around the world, comprising approximately 283 000 Sentinel-1 and Sentinel-2 images, collected between September 2019 and April 2021. The proposed approach demonstrates excellent performance with a very low mean absolute error of 0.0478, effectively filling even long-lasting gaps while being applicable globally. Thus, our method holds significant promise for improving the efficiency of numerous downstream applications previously limited by cloud-induced gaps.

Keywords

• Cloud removal
• data fusion
• deep learning
• gap filling
• recurrent neural network (RNN)
• vegetation monitoring