Towards optimising the derivation of phenological phases of different crop types over Germany using optical high resolution image time series

Crop phenological phases have traditionally been observed from the ground, which is a labor-intensive and timeconsuming
activity that also lacks spatial variability due to the sparse and limited network of ground data, if
any are available. In view of this, remote sensing can provide a low-cost avenue to systematically monitor and
detect phenological phases from space. The most common approach for retrieving vegetation phenology from
remotely sensed time series is the dynamic threshold method. However, only a few number of studies have
attempted to calibrate and optimize these derived phenological metrics to relate them to actual crop growth
stages. Accordingly, this study attempted to develop a framework to optimize the derivation of phenological
phases for three major crops (winter wheat, corn and sugar beet) in Germany by investigating the optimal
thresholds and comparing the performances with ground-truth observation data. To this end, the Normalized
Difference Vegetation Index (NDVI) time series covering Germany and for two cropping seasons 2019 and 2020
were obtained and derived from a 10 x 10 km tiling grid of Sentinel-2 analysis ready data using a specific
decentralized cloud platform that combines both a set of satellite imagery (petabytes of data) with huge analysis
capabilities on a very large scale. Since cloud contamination is typically the major drawback for estimating
phenology with optical satellite data, the study suggests first a new smoothing and gap filling method (UEWhittaker)
that is based on both envelope detection and the Whittaker filter and that, in the end, constructs
high-quality NDVI time series that are suitable for phenological analysis. Based on these generated time series,
the estimation of various phenological phases of crops as well as threshold optimisation and calibration were
carried out as a second step. In which we traverse the thresholds from 0 to 1 with an increment of 0.01 for
each specific phase and finds the optimal threshold when the lowest error value is obtained between the satellitederived
DOY and the observed DOY in ground data from the year 2019. Later on, these optimum thresholds were
used to derive phenological phases from the next year, and the results of the calculation of the root-mean-square
error (RMSE) and the mean absolute error (MAE) between the ground reported in-situ phenology observations
and those derived from satellite data reveal that they ranged typically between 3 days and 2 weeks for nearly
all the phenological phases. The findings demonstrate how calibrating and optimising the derivation of different
phenological phases of crops using only optical data could produce a timely and accurate information on crop
growth and its condition for a large area which can be used in agricultural management, crop yield estimation,
and several other related applications.