A Globally Applicable Method for NDVI Estimation from Sentinel-1 SAR Backscatter Using a Deep Neural Network and the SEN12TP Dataset

Abstract Vegetation monitoring is important for many applications, e.g., agriculture, food security, or forestry. Optical data from space-borne sensors and spectral indices derived from their data like the normalised difference vegetation index (NDVI) are frequently used in this context because of their simple derivation and interpretation. However, optical sensors have one major drawback: cloud coverage hinders data acquisition, which is especially troublesome for moderate and tropical regions. One solution to this problem is the use of cloud-penetrating synthetic aperture radar (SAR) sensors. Yet, with very different image characteristics of optical and SAR data, an optical sensor cannot be easily replaced by SAR sensors. This paper presents a globally applicable model for the estimation of NDVI values from Sentinel-1 C-band SAR backscatter data. First, the newly created dataset SEN12TP consisting of Sentinel-1 and -2 images is introduced. Its main features are the sophisticated global sampling strategy and that the images of the two sensors are time-paired. Using this dataset, a deep learning model is trained to regress SAR backscatter data to NDVI values. The benefit of auxiliary input information, e.g., digital elevation models, or land-cover maps is evaluated experimentally. After selection of the best model configuration, another experimental evaluation on a carefully selected hold-out test set confirms that high performance, low error, and good level of spatial detail are achieved. Finally, the potential of our approach to create dense NDVI time series of frequently clouded areas is shown. One limit of our approach is the neglect of the temporal characteristics of the SAR and NDVI data, since only data from a single date are used for prediction.