NDVI time series filling over wheat fields using the Sentinel-1 data

The Normalized Vegetation Different Index (NDVI) has proved its relevance to describe crop dynamics with a high spatial resolution. Therefore, it’s widely used to monitor and describe vegetation in different agronomic applications such as land use classification, yield forecasting, and biophysical variables’ estimation. As an optical index, the NDVI is limited by weather conditions where the cloud presence impacts the pixel information. In the present study, we retrieve the NDVI values according to wheat growth stages using the normalized polarization ratio (IN) and the estimated coherence in Vertical-Vertical polarization from C-band Sentinel-1 radar data. To estimate the NDVI values, we used empirical equations and the machine learning algorithms such as the support vector regressor (SVR) and random forest (RF). During the two years from 2018 to 2020, we divided the wheat cycle into two periods. The first period is extended between the sowing and heading event. The second one covers physiological maturity which is subdivided into two sub-periods where the NDVI values are lower or higher than 0.4. For the first period, the NDVI estimation is characterized by root mean square of error (RMSE) values varying between 0.07 and 0.1. When the NDVI values are lower than 0.4 through the senescence phase, the RMSE values are lower than 0.06. Throughout the grain maturity (NDVI ≥ 0.4), the RMSE values exceed 0.19 using the calibrated empirical equation as a function of IN against a moderate performance characterizing the use of machine algorithms with the IN and as features. The developed approach to estimate the NDVI according to the wheat development stage was tested on several fields. The overall RMSE values vary between 0.12 and 0.19 with a correlation coefficient fluctuating between 0.64 and 0.87 and a bias value ranging between -0.06 and -0.02. The combination of the radar variables improved the NDVI estimations during the wheat cycle. The developed approach can be tested on other crops and climatic contexts.