Optimising Phenological Metrics Extraction for Different Crop Types in Germany Using the Moderate Resolution Imaging Spectrometer (MODIS)

Xingmei Xu, Christopher Conrad, Daniel Doktor
2017 Remote Sensing  
Phenological metrics extracted from satellite data (phenometrics) have been increasingly used to access timely, spatially explicit information on crop phenology, but have rarely been calibrated and validated with field observations. In this study, we developed a calibration procedure to make phenometrics more comparable to ground-based phenological stages by optimising the settings of Best Index Slope Extraction (BISE) and smoothing algorithms together with thresholds. We used a six-year daily
more » ... d a six-year daily Moderate Resolution Imaging Spectrometer (MODIS) Normalized Difference Vegetation Index (NDVI) time series and 211 ground-observation records from four major crop species (winter wheat/barley, oilseed rape, and sugar beet) in central Germany. Results showed the superiority of the Savitzky-Golay algorithm in combination with BISE. The satellite-derived senescence dates matched ripeness stages of winter crops and the dates with maximum NDVI were closely related to the field-observed heading stage of winter cereals. We showed that the emergence of winter crops corresponded to the dates extracted with a threshold of 0.1, which translated into 8.89 days of root-mean-square error (RMSE) improvement compared to the standard threshold of 0.5. The method with optimised settings and thresholds can be easily transferred and applied to areas with similar growing conditions. Altogether, the results improve our understanding of how satellite-derived phenometrics can explain in situ phenological observations. Spectrometer (MODIS), provide daily vegetation index (VI) observations for evaluating phenological characteristics, among others in agricultural areas [6, 7] . VIs are indicators of the abundance of different leaf components (pigments, water, nitrogen etc.) that are responsible for the absorption of radiation [8] . Therefore, changes in canopy biophysical properties are reflected in VI evolution over the growing season of crops [7] . The increase and decrease of VI during crop growth and senescent periods are mainly due to changes of leaf area index, the relationship between chlorophyll and other pigments together with water content [9] . For cereals, VIs usually reach their maximum at the late booting to the beginning of heading stage after the last leaf (e.g., flag leaf) is fully unfolded [10] [11] [12] . During this period of time, the development of all shoots (main stem and tillers) on the same plant also gradually synchronises [13] . As VI time series are usually affected by noise caused by clouds, atmospheric constituents and varying viewing geometries of the sensor, several compositing algorithms have been developed for noise reduction. The maximum value composite (MVC) method is widely used to minimize negatively biased noise by selecting the highest VI value to represent a certain period (8-16 days) [14] . However, the use of MVC methods can result in unnecessary loss of data when there are several noise-free VI values within one compositing period. The Best Index Slope Extraction (BISE) algorithm [15] has advantages over MVC as it potentially retains more observations which contain true vegetation signals. BISE could also eliminate data transmission errors which cause abnormally high VI values. The performance of this method can be optimised by varying a so-called sliding period. The sliding period defines a time window within which changes in VI time series, i.e., phenological variability, should be accepted. This time window also determines how many true VI measurements (i.e., not hampered by clouds) can be expected. Despite of its importance, we did not find a systematic assessment on the sensitivity of sliding periods for crop phenology monitoring in the literature. Many different smoothing and model fitting methods have been applied on composited products to construct high-quality VI time series [16] [17] [18] . Each method has its own advantages as well as limitations. In contrast to model fitting, filtering algorithms tend to require fewer parameters and can better preserve irregular intra-seasonal VI curves, which are typically observed in agricultural fields with more than one growing season in a year [19] . More recently, Zhao et al. [20] compared double logistic functions [21] and the Savitzky-Golay filter (SG) [17] to detect phenological stages (phenostages) of summer crops in the United States. Nevertheless, to our knowledge, comprehensive comparisons of smoothing methods for crop phenology mapping over agricultural areas and multiple years, i.e., including crop rotations, have not been published yet. Calibration and validation of satellite-derived crop phenology, i.e., investigations of deviations between phenometrics and field-observed phenostages, are not straightforward and are rarely addressed. This situation is in large part due to a lack of dense and evenly distributed in situ data, and the difficulties of matching ground observations to the satellite-derived phenological metrics (phenometrics). Typically, thresholds, such as the half maximum or maximum slope, have been used to retrieve general vegetation seasonal metrics [7, 21] . However, some of these metrics, for example for the onset of greening and senescence, are unable to represent specific agronomic stages. Some limited their studies to only one stage of one crop due to lack of available data [22] [23] [24] . Only a few studies have detected specific agronomic stages from satellite-derived time-series data [10, 25, 26] . Although high estimation accuracies were achieved, their two-step filtering model was more suitable for areas where deviations of phenostages were relatively small between different years and different test sites. Besides, the settings of the models needed to be re-adjusted for other study areas and for different crops (maize and soybean). Therefore, it is necessary to have a more general and transferrable approach to define crop phenometrics associated with the specific agronomic stages. In Germany, crop rotation is dominated by winter cereals, oilseed rape, sugar beet, maize and potato, which results in unevenly distributed growing seasons. Phenological data of these species recorded by the German Meteorological Service (DWD) have been applied to investigate crop-weather
doi:10.3390/rs9030254 fatcat:ihtff4mqbvdvvpce7ckh3e5ar4