Key points are not available for this paper at this time.
Over the past decades, especially since 2014, large quantities of Earth Observation (EO) data became available in high spatial and temporal resolution, thanks to ever-developing constellations (e.g.: Sentinel, Landsat) and open data policy. However, in the case of optical images, affected by cloud coverage and the spatially changing overlap of relative satellite orbits, creating temporally generalized and dense time series by using only measured data is challenging, especially when studying larger areas. Several papers investigate the question of spatio-temporal gap filling and show different interpolation methods to calculate missing values corresponding to the measurements. In the past years more products and technologies have been constructed and published in this field, for example Copernicus HR-VPP Seasonal Trajectories (ST) product. These generalized data structures are essential to the comparative analysis of different time periods or areas and improve the reliability of data analyzing methods such as Fourier transform or correlation. Temporally harmonized input data is also necessary in order to improve the results of Machine Learning classification algorithms such as Random Forest or Convolutional Neural Networks (CNN). These are among the most efficient methods to separate land cover categories like arable lands, forests, grasslands and built-up areas, or crop types within the arable category. This study analyzes the efficiency of different interpolation methods on Sentinel-2 multispectral time series in the context of land cover classification with Machine Learning. We compare several types of interpolation e.g. linear, cubic and cubic-spline and also examine and optimize more advanced methods like Inverse Distance Weighted (IDW) and Radial Basis Function (RBF). We quantify the accuracy of each method by calculating mean square error between measured and interpolated data points. The role of interpolation of the input dataset in Deep Learning (CNN) is investigated by comparing Overall, Kappa and categorical accuracies of land cover maps created from only measured and interpolated time series. First results show that interpolation has a relevant positive effect on accuracy statistics. This method is also essential in taking a step towards constructing robust pretrained Deep Learning models, transferable between different time intervals and agro-ecological regions. The research has been implemented with the support provided by the Ministry of Culture and Innovation of Hungary from the National Research, Development and Innovation Fund, financed under the KDP-2021 funding scheme. Keywords:time series analysis, Machine Learning, interpolation, Sentinel
Simon et al. (Mon,) studied this question.