In this master's thesis we study conditional generation of missing data in time series. We focused on the problem of gappy NDVI time series. This index is in our case based on the data from the optical sensors of the Sentinel-2 mission, and is therefore highly sensitive to atmospheric conditions. We developed a method for imputing missing values, leveraging the radar imagery of the Sentinel-1 mission to address the limitations of NDVI time series. We employed a conditional diffusion model. The model was trained on data for agricultural parcels spanning over one vegetation cycle. The approach proved to be promising. The results demonstrate that the model outperforms the basic interpolation methods and demonstrates actual trends that are left undetected by basic interpolation methods. Despite that, a number of cases exhibit non-intuitive values, possibly due to the sensitivity of radar data to other influeces.
|
