1. Introduction
In recent years, the concept of essential variables (EVs) has been introduced to assess progress towards Sustainable Development Goals across policy domains [
1]. EVs that capture the progress of the agricultural domain are currently defined by the Group on Earth Observations Global Agricultural Monitoring (GEOGLAM) [
2]. In this catalogue, phenology—more specifically, the current crop stage—is listed as an EV. It has been established that phenology provides crucial information for crop management because it strongly relates to plant productivity and growth. Furthermore, certain stages of the crop lifecycle are highly sensitive to meteorological conditions [
3,
4,
5,
6]. In the context of an increasing frequency of and more extreme weather events, as well as complex topics such as climate adaption and resilience, this kind of information is in high demand [
7,
8,
9]. Consequently, Earth observation data and products have been widely researched in the field of agriculture as a potential source of such information. This field has been dominated by multispectral optical sensors such as Landsat or the Moderate Resolution Imaging Spectroradiometer (MODIS) for a long time. However, spaceborne synthetic aperture radar (SAR) data have been researched increasingly during the last decade, either as an additional or an alternative data source, as SAR provides textural or structural instead of spectral information and it is independent of weather and illumination conditions [
8,
10]. More specifically, the start of the European Sentinel-1 mission in 2015 caused an increase in research activities using SAR data for various applications [
9,
11,
12]. In the context of tracking crop phenology, three major forms of analysis have been established: machine learning classifiers, such as random forest or deep learning [
13,
14], stochastic or statistical modeling [
15,
16] and time series metrics (TSMs) [
17,
18,
19].
The study at hand focuses on TSMs, more specifically extreme value and break point analysis. For description, a crop phenology classification scheme has been established by Biologische Bundesanstalt für Land und Forstwirtschaft, Bundessortenamt und CHemische Industrie (BBCH) The BBCH scale categorizes plant growth according to micro and macro stadia [
20]. In this case the phenological development of the following crops is tracked: wheat, sugar beet, canola, and potatoes. These crops cover a wide range of physiognomic properties and management requirements (e.g., irrigation of potatoes); thus, a more general assertion can be framed. Hence, this study offers a wider perspective of the signal–target interaction during the crop lifecycle than studies that focused only a specific crop type or crop family [
15,
17,
21]. Moreover, the addition of InSAR coherence (hereafter referred as “coherence”) to polarimetric features, as suggested by Lobert et al. [
14], might lead to improvements in monitoring crop phenology, since coherence might be more sensitive to early stages of the plant lifecycle due to the temporal decorrelation of the signal caused by the emergence of crops. In the context of a multiannual approach, we also address the dilemma outlined by Harfenmeister et al. [
22] that the chronological occurrence of TSMs, especially break points, cannot be reliably used to allocate phenological development by SAR time series. Hence, we introduce agrometeorological data, namely, growing degree days (GDDs) [
23], as an artificial baseline for calibrating and validating the occurrence of TSMs and their associated progress of plant growth. GDDs contribute to the assessment of reliability by adding information on thermal growth potential to the occurrence of TSMs. Apart from these gaps, the mission ending malfunction of Sentinel-1B produced a new aspect: integrating multiple orbits to conserve a comparatively dense time series. As of January 2022, it is no longer possible to rely on the six-day repetition rate, provided by the twin constellation [
24] over our study area. Hence, we used data from the existing archive to track phenological development simultaneously across different orbits. Recently, such an idea was addressed over wheat fields and sunflower plantations [
25,
26,
27], but the focus was either on deriving biophysical parameters or investigating the intensity of the backscatter only. Hence, this study aims at generating information regarding if and how these orbits differ in their response towards phenological development of multiple crops and S1 InSAR and PolSAR features. The current assumption is a negligible difference as soon as there is a certain volume of biomass on the fields [
13,
25,
26,
27].
Since all of the abovementioned issues are highly dependent on how the respective time series are generated, and especially on the usage of smoothing algorithms and their parametrization, a density-based framework across multiple parametrizations of a single smoothing algorithm is used in this study. Such stacking of different degrees of smoothing helps to mitigate the inevitable loss of information of highly smoothed time series. Thus, entire periods, instead of points, which cause a phenologically induced signal change, are revealed. The fundamental hypothesis of such an approach is as follows: extrema and break points, which originate from events that drastically shape a crop-specific time series, remain visible throughout various intensities of smoothing.
Here, we employed the frequently used “locally estimated scatterplot smoothing” (LOESS). This decision was based on two factors. Firstly, Cai et al. (2017) [
28] stated that there is no mayor difference between locally weighted smoothing techniques. Secondly, commonly applied smoothing techniques in studies related to phenology are locally weighted (e.g., Whittaker [
29], Savitzky-Golay [
30], LOESS [
17,
18]). Considering the stacking of different smoothing degrees and the small differences between the individual algorithms, it is assumed that our approach produces representative crop signatures, despite employing only LOESS.
In sum, this study addresses the following issues centered around crop monitoring via SAR time series: (i) combining information obtained from time series of different smoothing intensities [
28] (ii) as well as different S1 viewing geometries, ergo, relative orbits in a singular conceptual framework without angle normalization. Furthermore, (iii) the variance in chronological occurrences of break points and extreme values [
21] is investigated. This study addresses these issues by (i) estimating the density of TSM occurrences by stacking time series of different smoothing intensities at the field level and aggregating the findings to derive landscape-wide (the extended DEMMIN site covering an area of 25 km by 25 km) patterns. Here, the time series analyses are conducted separately for each orbit generating insights into (ii) orbit-related discrepancies within the patterns at landscape level. In regard to the issue of chronological occurrence, a GDD baseline is introduced for calibration and validation. This provides a temporal coordinate system consisting of day of year and thermal growth potential. Because this approach aims at generating systematic insights into the relationship of crop phenology and the inherent randomness of SAR signals [
31], the study encompasses seven S1 features, three relative orbits, and an observation period of five years [
31].
4. Discussion
4.1. Discussing Patterns of Major Signal Change
The results of the investigation into patterns of major signal changes revealed and reaffirmed serveral key aspects of SAR-based time series analyses. This is, to the extent the authors are aware of, the first systematic study to analyze the relation between C-Band SAR signals and phenological development of crops in an investigative approach encompassing three relative orbits, five years, seven S1 features, two types of TSMs, and four different crop types, as well as their BBCH stages. In addition to that, a novel approach of detecting windows of phenologically induced change was applied by stacking time series of different smoothing intensity to estimate the density of TSM occurrences.
Because the coverage of the targeted BBCH stages also implies insights into reliability and offset, this is discussed in detail in
Section 4.2. But apart from the targeted stages, the potential to cover additional time windows was demonstrated. Furthermore, the integration of a GDD baseline provided a second temporal coordinate to evaluate the tracking reliability of such events. Judging by their DOY values, the example of wheat signatures contains likely freezing and thawing events around DOY 55 to 67 as well as the transition into the catch crop stage alongside the reemergence of subsequent winter crops. Similar observations were made for the other three crop types as well. Potato and sugar beet also displayed additional TSM occurrences during their development up to the likely period of harvesting. Such patterns were better discernable for break points as extrema exhibited larger time windows. This reiterates the need to filter phenologically nonrelevant extrema [
30] or to limit the tracking to specific stages such as flowering of canola [
29].
By conducting this analysis in a multiannual framework encompassing various agro-meteorological conditions, phenological events that shape a crop-specific time series regardless of weather conditions and signal processing were revealed in relation to their viewing geometry. Hereby, highly active and inactive periods in terms of signal change were placed into the context of plant lifecycles. In addition, this demonstrated the potential to reveal parts of said plant lifecycles that are reliably traceable and not covered by the DWD monitoring framework. This would suggest an added value of Earth-observation-based monitoring setups for phenological monitoring, because data gaps would be filled. The multiannual framework also highlighted the clear borders of the calibration periods required for such an approach. The first local extreme values, which are not at the start and end of the time series, were found as early as 20 to 25 days into the yearly observation period break points provided only insights into yearly developments between DOY 50 and 300. This quantification of the calibration periods constitutes a contribution to the discussion of information availability in time series [
21] within the context of civil years.
4.2. Discussing Reliability and Systematic Offsets
As this study uses the DWD statistics on a federal level, the fact that major signal changes are not necessarily linked to the onset of that particular micro stadium cannot be fully discarded. Moreover, the results of the temporal thresholding (see
Figure 8,
Figure 9 and
Figure 10) showed that signal changes were not necessarily related to the onset of BBCH stages represented by in situ data. It is considered likely that adjacent micro stadia, in the case of wheat and stem elongation BBCH 31 or 35, rather than 30 were depicted. This corroborates the findings by [
18,
22].
Nevertheless, crop- and stage-specific sets of S1 features that produced close TSM occurrences were identified. Overall, more stages could be closely tracked by break points rather than extrema. This is most likely related to the wider range of uncertainty of extrema and their respective time windows of phenological changes (see
Figure A2 and
Figure A4).
With regard to break points, the targeted stages of winter wheat were best tracked by VV intensity and coherence of orbits 95 and 146, reflecting the plant’s vertical structure. The listing of coherence for stem elongation and harvest (BBCH 30 and 99) demonstrated the features sensitivity to significant changes in volume of biomass [
17,
18]. The absence of orbit 168 suggests that tracking wheat phenology might be susceptible to moisture content when using moderate incidence angles. Otherwise, a clear preference of steeper incidence angles was only found at heading (BBCH 50). Such a clear overall preference in viewing geometry was not found for canola. Only stage-specific preferences for the start and end of flowering (BBCH 60 and 69) were found regarding moderate incidence angles. Here, h2a_entropy and h2a_alpha, as well as intensity_cr, featured the closest occurrences of break points. The preference of h2a and CR is most likely related to the complex structure of canola plants, as they reflect a change in dominant scattering type. Furthermore, VV intensity being the favorable S1 feature for tracking onset of flowering (BBCH 60) fits the increased sensitivity of VV polarization to superficial changes of the canopy [
29,
31]. Some of these findings are also replicable for sugar beet signatures. First of all, significant changes at the land surface such as emergence (BBCH 0) or leaf development were captured by coherence. Secondly, h2a and intensity_cr could be linked to canopy closure, which coincides with the time when the plant’s physiognomy is comparatively complex. However, sugar beet favors incidence angles of moderate range, which is most likely related to its low height in comparison to wheat or canola. The break points within signatures of potato exhibit a similar behavior, similar to wheat orbit 95, and to some degree, 146 was favored. A tendency towards steeper incidence angles was only found at leaf development (BBCH 10).
Results of the extreme value analyses overlap with these observations; however, there are some notable differences. Heading (BBCH 50) of wheat produced local maxima in coherence signatures, suggesting that the crop volume was stable enough to increase the temporal correlation above noise level, regardless of the incidence angle. This particular occurrence has not been observed by comparable studies [
17,
19]. A similar observation was made for inflorescence (BBCH 50) of canola in VV coherence; however, this was only related to moderate incidence angles. Another, more evident difference was discovered for sugar beet. Here, extreme values did not show any preference of orbit, but displayed a similar performance across all viewing geometries. This is most likely related to the aforementioned finding, that extrema can be produced over larger time windows.
While unequivocal indications for a complementary tracking potential of different S1 features was only found for BBCH 69 of canola and BBCH 0 of potatoes, the increased number of listings around strong changes in biomass (harvest, stem elongation) or disturbances of the soil agree with findings of previous studies [
14,
17,
18,
59], that InSAR coherence increases the trackability of certain phenological stages. On the other hand, h2a_entropy and h2a_alpha seem to provide less added value and can be replaced by intensity, especially CR, in a dual-pol C-Band framework.
Considering the results of break points and extrema in relation to time of overpass, moisture content within the canopy is rated as a secondary effect. But because the results of the orbit-wise comparisons of wheat, canola, and sugar beet differ mostly by comparatively small margins, it is also likely that the increased closeness to in situ observations originates from a favorable revisiting schedule that is temporally closer to the onsite phenological developments, especially because this study uses the DWD statistics at the level of federal state. On top of these issues, the analysis of potato signatures was additionally hampered by the comparatively small number of fields (twelve fields of potato vs. 500 fields of wheat), which increases the impact of outliers on a generalized pattern at the landscape level. Moreover, fields of potato in Demmin are characterized by rather heterogeneous conditions due to the cultivation of different sorts of potato as well as the application of sprinkle irrigation. Therefore, the statements made about phenological developments of potatoes are considered less robust and serve more as a first impression.
Assessing another anticipated shortcoming, the references of DOYs corresponding to BBCH 0 and BBCH 10 for winter crops as general points of orientation within phenological progress of the civil year worked well in the chapter on major signal changes, and the deliberate focus on the civil year ensured the comparability between the progress of all four crop types. The GDDsim baseline is also affected by this decision, resulting in larger counts of relevant GDDsim accumulations at the start of this season. Therefore, further studies on that subject should account for this aspect. Nevertheless, the GDDsim baseline enabled an initial quantification of relevant TSM patterns which were not covered by the in situ observations. In addition, each of the TSM occurrences, as well as the in situ observations, are now associated with a GDDsim value; thus, progress towards a certain BBCH stage or the next occurrence of relevant TSM patterns is easily quantifiable in subsequent studies on this subject.
Needless to say, this study relies on the availability of information on crop types and field boundaries, but these are also issues that the Earth observation community has investigated. Therefore, there are solutions to overcome the lack of information on field boundaries [
60,
61] or crop types [
62].
4.3. Outlook
This successful investigation into phenologically induced pattern of S1 time series at the landscape level opens up two major directions for future research.
On the one side, an allocation of field-level developments within the general pattern of the landscape is possible. This information can be used to check whether a field is early or late in its phenological development when compared to the general trend at landscape level. This would result in a more spatially explicit information than the interpolated 1 km grid of phenological progression provided by DWD. Additionally, a strong deviation from the landscape level patterns may serve as an indicator for field-specific crop stress such as lodging or pest infestation. On the other side, an archive of recorded phenological events and their relation to a growing degree baseline has been established. As pointed out by Harfenmeister et al. [
22], and as was made evident during this study, back-looking approaches such as break point and extreme value analyses are not easily converted to a near-real-time application due to their calibration phases that may stretch across multiple weeks. However, the satellite- and GDDsim-based records of said archive could be leveraged in a comparative scenario analysis to assess the performance of an ongoing season. This could be accomplished by a comparison between its S1-based time series and the accumulation of GDDs with the successfully tracked phenological development of past seasons. Furthermore, by including weather forecasts, even an outlook may be provided, as the distance to the next phenological event on record also contains a respective GDDsim value. Hence, the projected GDD accumulation from the weather forecast can be employed to estimate the phenological progress of the ongoing season in comparison to the seasons on record.
Finally, the transferability of the approach in space and time, especially with regard to the added value of a GDD baseline, as well as the inclusion of spectral information are deemed sensible objectives for further investigation.
5. Conclusions
This study produced comprehensive insights into how phenology of wheat, canola, sugar beet, and potato shape their respective S1 signatures in different weather conditions, their inherent viewing geometry, and regardless of how strongly the time series is smoothed or processed otherwise. This leads to the following conclusions about TSM analyses:
- (i)
Break points constitute a better tool for entire-season monitoring, while extrema are mainly suitable for specific stages, because of the great variations detected in patterns of extreme values.
- (ii)
It is therefore crucial to optimize time signal processing for targeted stages or focus on specific parts of the crop lifecycle when employing extreme value analysis.
- (iii)
This study demonstrates its suitability for entire-season coverage of the crop lifecycle, as a single optimized smoothing of time series will inevitably hamper the capabilities of a monitoring framework to detect certain macro or micro stadia of crop development.
This approach also generated a designation of reliable S1 features by crop type and relative orbit for the area of Demmin. An overall great suitability of backscatter intensities was confirmed, and orbits favorable for tracking specific crop types were discovered. The most prominent findings are as follows:
- (iv)
A moderate range of incidence angles between 31°and 41° (orbit 146 and 168) is better suited than a range between 41° and 45° (orbit 95) for tracking crop phenology of sugar beet and wheat, whereas potatoes favor steeper angles. There are, however, BBCH- and S1-feature-specific exceptions to these tendencies.
- (v)
Relevant signal changes often corresponded to surrounding micro stadia of the onset of a macro stadium, instead of the onset itself.
- (vi)
With respect to times of overpass, only wheat displayed visible tendencies to favor the late afternoon.
- (vii)
VV and VH intensity outperformed the other features in terms of overall reliability. There are, however, specific combinations of crop type and BBCH stadia, where a complementary dataset of PolSAR and coherence provides added value and increased robustness.
By adding an artificial GDD baseline, the quantification of progress towards targeted stages by their thermal growth potential was achieved. This enables the relation of the established patterns at the landscape level to phenological developments of individual fields, because every TSM and BBCH stage at the landscape and at field levels were assigned a corresponding GDD value. This, in turn, sets up objectives of further studies such as the analysis of spatial patterns in phenological developments within the landscape and the potential to highlight field-specific anomalies in the crop lifecycle.