Plant phenology plays a critical role in structuring ecosystem processes such as carbon assimilation and energy flux, and is a sensitive indicator of climate change. Land surface phenology (LSP) measurements from orbital platforms have been available for decades, but typically only at coarse resolution because of the perceived importance of high image temporal density. The combination of Landsat 8 and Sentinel-2 satellites now offer 3-day repeat coverage of nearly the entire Earth at moderate resolution, offering an unprecedented opportunity to collect large-scale LSP measurements at a spatial scale that better comports with ecosystem processes. Because LSP measurements are self-consistent, they are invaluable indicators of intrannual change in ecosystem function, and their climate drivers. However, LSP products have often been criticized as lacking in ecological relevance. For instance, it is not well-known what particular plant developmental stages correspond to phenometrics derived from time series of spectral vegetation indices. Moreover, the ecological meaning of phenometrics is likely to vary across different biomes and plant functional types. Previous efforts that have attempted to link LSP phenometrics to ground observations of plant developmental stages have been frustrated by the heterogeneity of the land surface at coarse resolution. Here, we use LSP phenometrics derived from multisource, moderate resolution imagery along with ground observations of plant seasonal development collected by citizen scientists across the United States to reveal the ecological meaning of moderate resolution phenometrics. We show that these LSP measurements have clear, often strong, correspondence to ecologically-relevant plant developmental stages observed on the ground, but the nature and strength of the relationships varies across plant functional types. Our results demonstrate that LSP observations are not just consistent indicators of intrannaual change, but also contain valuable ecological information about the nature of those changes. Additionally, we demonstrate an effective strategy for utilizing large amounts of citizen-science observations of variable quality that may be useful to similar investigations. Broadly, our results support the use of moderate resolution phenometrics derived from Landsat 8 and Sentinel-2 time series in investigations of ecosystem response to changing climatic drivers, and in environmental modeling activities.

Phenology assessment by satellite at coarse spatial scale is today more or less operational, with a number of established phenology data products and methods. Most commonly, data from the MODIS sensor has been used for developing phenology information from NDVI or EVI at 250-500 m resolution. The information is useful in ecosystem research, research on land-atmosphere dynamics, and for generation of improved land cover products. However, obtained parameters are sometimes inconsistent, and phenology metrics do not always accurately match ground data observations. Based on a novel vegetation index called the Plant Phenology Index (PPI) that uses BRDF-correct MODIS data and that tackles saturation in dense forest and disturbance due to winter-time snow, we have generated phenology trends across Europe for the period 2000-2016. Significant trends in growing season parameters occur over this short time period, in particular for start of the growing season. These match observed trends based on phenological ground data. End of the growing season parameters are more difficult to verify due to limited accuracy in comparison to validation data sets.

The coarse resolution of MODIS data means that phenology information from different vegetation types may get intermixed in heterogeneous landscapes, leading to difficulties in interpreting and verifying the signals. In these situations, data from Landsat and Sentinel-2A and 2B sensors have tremendous potential to provide detailed information on land vegetation, enabling phenology mapping at individual forest parcels or cropping fields. In order to process the data, which has lower temporal frequency than MODIS, we have developed robust methodology for smoothing and gap-filling the data. One of our methods is based on box-constrained fits to logistic model functions, and provides robustness against data gaps by including shape prior information. We evaluate the derived seasonal curves and phenology metrics against near-ground measurements. These measurements are from phenocams and fixed multispectral sensors, located at flux-tower sites across Scandinavia. The spectral sensors measure reflected radiation in visible, near-infrared, and short-wave infrared wavelength bands at 10-minute frequency. The sensors provide accurate verification data for satellite time-series products, free from atmospheric influences. The phenological estimates from satellite are also verified by studying vegetation transects along topographical gradients, and by comparing against flux tower measurements. The evaluation shows that the two Sentinel-2 satellites provide phenological information with higher accuracy and with better description of spatial variability in phenology than that from MODIS. The dynamic vegetation information derived from Sentinel-2 is useful and relevant for operational applications within forestry, agriculture, and for biogeochemical and ecological modeling.

Land surface phenology (LSP), defined as recurrent biological events in the annual profile of vegetated land surface at ecosystem scale as observed from remote sensing, is the focus of our work in the frame of the GlobDiversity project. LSP is not only widely used as indicator for terrestrial ecosystems’ responses to changes of environmental conditions, but can also serve as indicator for characterising species composition and biodiversity of an ecosystem.
We developed a new approach to observe LSP on a large spatial scale within the frame of a project funded by the European Space Agency (ESA) with the aim to contribute to the use of remote sensing data in the frame of the Essential Biodiversity Variables EBVs. Within the EBV-framework, biodiversity shall be observed and monitored in high-resolution and on a global scale, similar to the concept of Essential Climate Variables ECVs.
Monitoring LSP requires a dense time series of vegetation activity in the region of interest in order to study the yearly vegetation profile. This data is available from space in high-resolution and sufficient dense time series when combining Sentinel-2A and 2B observations with Landsat 8 images. In our study, we investigated time series of the years 2017 and 2018 and extract LSP from visual and near visual bands at pilot sites located in different biomes (temperate forest, arctic tundra, Mediterranean wetland, sub-tropical savannah and shrublands). Our project demonstrates the feasibility of calculating LSP on global scale and high resolution while using the Normalized Difference Vegetation Index (NDVI) and a double logistic model. We also developed a quality control flagging system in order to allow a reliable and fully automatic processing and quality measures.
The spatial phenological maps are then compared to the vegetation profile of observed on ground by phenocams installed in the pilot sites. We propose an approach how the satellite data can be matched to the satellite-derived LSP maps based on the cameras imaging the vegetation in the visible color channels. Main challenges in the development of the algorithm is the matching of the field of view of the phenocam with the satellite pixels as well as combining the green chromatic coordinates (GCC) from the phenocam with LSP derived from NDVI.
In the interpretation, we focus on ecological processes and how these can be used for biodiversity assessment. This method is a crucial step for narrowing the scale gap between plant phenology and land-surface phenology and shows the potential for application at large – even global – spatial extent.

Autumn phenology in broadleaf forests remains understudied relative to spring phenology, despite the importance of autumn events for determining season length, carbon uptake, and interactions between migratory species and their environment. The difficulty of defining autumn transition dates complicates the study of fall phenology, as spring phenological events are often discrete in time (e.g., budburst), while autumn events such as leaf coloring and leaf drop are continuous processes. Further, remote sensing predictions of spring phenology are consistently more accurate than predictions of autumn events, due to the clear and abrupt signal of budburst in the spring compared to gradual leaf senescence in autumn. Here, we aim to improve understanding and characterization of autumn phenology using remotely sensed data and explore the drivers of autumn phenology using a combination of in situ observations, PhenoCam data, and a dense time-series of Landsat and Sentinel-2 imagery over the Hubbard Brook Experimental Forest in the Northeast United States. First, we use ground observations and PhenoCam data to describe the timing of autumn events, such as the start of leaf coloring, peak of leaf coloring, and the timing of leaf drop from 1990 to 2017, and then investigate how these autumn events are manifested in Landsat and Sentinel-2 data. Specifically, we assess dynamics in spectral indices that are sensitive to changes in leaf area index and greenness (i.e., Enhanced Vegetation Index (EVI)), as well as indices that are sensitive to changes in leaf pigments (i.e., red chromatic coordinate (RCC)), and allow for a multi-index approach to monitoring autumn phenology. Using this improved interpretation of autumn events in Landsat and Sentinel-2 data, we then provide a regional assessment of variability in autumn phenology, and investigate the drivers of this variability using climate data and information on spring phenology. Initial results suggest that the date of 50% greendown in EVI corresponds more closely in time to the onset of leaf coloring (RMSE = 6.8 days, R2=0.13 across 9 sites and 3 years), but is more strongly correlated to the timing of leaf drop (R2=0.50, RMSE=14.5 days). Through continued analysis to improve retrievals of both leaf coloring and leaf drop, we aim to describe both the timing and duration of autumn events, leading to an improved characterization of autumn phenology in temperate broadleaf forests.

Vegetation phenology has a great impact on land-atmosphere interactions like carbon cycling, albedo, or water and energy exchanges. To understand and predict these critical land-atmosphere feedbacks, it is crucial to measure and quantify phenological responses to climate variability, and ultimately climate change. While a wide range of mechanistic phenological models have been applied to estimate key phenological parameters (i.e. leaf unfolding), it has been shown that these models fail to predict interannual variability in land surface phenology measured from global satellite archives. It is therefore necessary to improve our understanding of how phenological models relate to land surface phenology measured from various spaceborne remote sensing sources.
In this study, we analyze the potential of multi-sensor, medium-resolution time series for estimating spring phenology of broadleaf forests across Germany. We combine dense time series of Sentinel-2, Landsat 7, and Landsat 8 acquired in 2017 and 2018. To build multi-sensor reflectance time series, Sentinel-2 and Landsat data are atmospherically corrected using the Framework for Operational Radiometric Correction for Environmental monitoring (FORCE). We fit two phenological models, logistic regression and generalized additive models, using NDVI and EVI time series at 12,767 sampling sites, and compare their suitability for representing phenology across our study area. We further evaluate our estimates against state-of-the-art mechanistic models calibrated using meteorological data from 426 meteorological stations and literature-based species parameters. This comparison also allows to characterize the spatial variability in the drivers of spring phenology.
Our results show that estimates of spring phenology are feasible for single years by integrating multi-sensor time series. It also shows, though, that data density, choice of model, and the respective vegetation index can have a strong impact on predictability. EVI time series outperformed NDVI time series in logistic regression and generalized additive models across the study area. Preliminary results also indicate that the generalized additive models are more robust than simple logistic regression fits. Logistic models converged for 81.5% of all sampling sites, whereas almost all (99.7%) of the generalized additive models converged. Start of season dates varied from early April to end of May in Germany, with local patterns overruling general gradients. The study shows that dense multi-sensor time series from integrated Landsat and Sentinel-2 data hold the potential to advance our knowledge of phenological dynamics and their drivers across broader spatial scales.

The start of the growing season in temperate forests is signalled by the emergence of leaves on deciduous trees. The timing of the leaf-out affects essential ecosystem processes such as carbon sequestration and plant-animal interactions [1]. Climatic changes such as earlier increases in spring temperature are the main controls for an earlier onset of leaf-out in temperate deciduous forests [2]. Satellite remote sensing has become highly valuable to monitor changes in land surface phenology such as changes in the start of growing season (SOS). In this study, we propose the estimation of leaf-out events from microwave satellite observations that are sensitive to vegetation structure and water content. It has been shown that vegetation optical depth derived from the Advanced Scatterometer (ASCAT) on-board the Metop-A satellite is sensitive to seasonal changes in the water content and structure of crops [3][4]. However, little is known about the sensitivity of ASCAT to the phenology of deciduous trees.

Here we aim 1) to assess if leaf-out events of temperate deciduous broadleaved forests (DBF) can be estimated from ASCAT backscatter observations, 2) to investigate to which extent the spring DBF signal can dominate grid cells with a main land cover other than DBF, and 3) to explain year-to-year differences in the timing of leaf-out.

Since 2007, ASCAT backscatter measurements are available globally every 1-3 days with a spatial resolution of approximately 25 km [5]. The observed backscatter depends on the incidence angle, soil moisture, surface roughness and vegetation state; changes in the vegetation state, e.g. water content or structure, lead to a change in the slope of the backscatter-incidence angle relationship. In this study, we investigate if leaf-out events of DBF can be detected in the slope time series, with a focus on temperate DBF in Europe and North America. We compare the slope time series to SOS derived from data of the PhenoCam network [6][7] and to Sentinel-1 backscatter, which has been found to also show a relative maximum during spring leaf-out [8]. ESA CCI land cover [9] is used for locating DBF and calculating DBF fractions in the grid cells.

Our results show that relative maxima occur in the ASCAT slope time series during leaf-out events. This indicates indeed that ASCAT tracks high vegetation water content at the time of leaf development. However, the ASCAT footprint covers an area of multiple km² and hence we might likely observe only a mixed signal from different vegetation types. To test this, we compared the relative maxima in spring with the fraction of DBF forests in a grid cell, and found similar patterns also in heterogeneous grid cells with smaller fractions of DBF. Hence, DBF can dominate the slope time series in mixed grid cells.
Leaf-out dates derived from ASCAT compare well to SOS dates from the PhenoCam dataset, with a mean absolute difference of less than 3 weeks for 80% of the analyzed sites. The comparison of the ASCAT leaf-out dates to ERA Interim air temperature [10] showed that year-to-year differences in the timing of the relative maxima can to some extent be explained by variations in the winter and spring temperatures.

Thanks to the long life time and the recent extension of the Metop satellite programme, it will become possible to estimate trends in spring phenology from the consistent long-term ASCAT time series in the future. Leaf-out dates from ASCAT provide thus a valuable complementation to the existing methods, especially for continental and global applications.


References
[1] Richardson, A. D., Keenan, T. F., Migliavacca, M., Ryu, Y., Sonnentag, O. and Toomey, M.: Climate change, phenology, and phenological control of vegetation feedbacks to the climate system, Agricultural and Forest Meteorology, 169, 156–173, doi:10.1016/j.agrformet.2012.09.012, 2013.
[2] Buermann, W., Forkel, M., O’Sullivan, M., Sitch, S., Friedlingstein, P., Haverd, V., Jain, A. K., Kato, E., Kautz, M., Lienert, S., Lombardozzi, D., Nabel, J. E. M. S., Tian, H., Wiltshire, A. J., Zhu, D., Smith, W. K. and Richardson, A. D.: Widespread seasonal compensation effects of spring warming on northern plant productivity, Nature, 562(7725), 110–114, doi:10.1038/s41586-018-0555-7, 2018.
[3] Vreugdenhil, M., Hahn, S., Melzer, T., Bauer-Marschallinger, B., Reimer, C., Dorigo, W. A., & Wagner, W. (2017). Assessing vegetation dynamics over mainland Australia with metop ASCAT. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 10(5), 2240-2248.
[4] Pfeil, I., Vreugdenhil, M., Hahn, S., Wagner, W., Strauss, P., & Blöschl, G. (2018). Improving the seasonal representation of ASCAT soil moisture and vegetation dynamics in a temperate climate. Accepted for publication
[5] Naeimi, V., Scipal, K., Bartalis, Z., Hasenauer, S., & Wagner, W. (2009). An improved soil moisture retrieval algorithm for ERS and METOP scatterometer observations. IEEE Transactions on Geoscience and Remote Sensing, 47(7), 1999-2013.
[6] Richardson, A. D., Hufkens, K., Milliman, T., Aubrecht, D. M., Chen, M., Gray, J. M., ... & Melaas, E. K. (2018). Tracking vegetation phenology across diverse North American biomes using PhenoCam imagery. Scientific data, 5, 180028.
[7] Hufkens, K., Basler, D., Milliman, T., Melaas, E. K., & Richardson, A. D. (2018). An integrated phenology modelling framework in R. Methods in Ecology and Evolution, 9(5), 1276-1285.
[8] Dostálová, A., Wagner, W., Milenkovic, M., & Hollaus, M. (2018). Annual seasonality in Sentinel-1 signal for forest mapping and forest type classification. International Journal of Remote Sensing, 1-23.
[9] Bontemps, S., Defourny, P., Radoux, J., Van Bogaert, E., Lamarche, C., Achard, F., ... & Zülkhe, M. (2013, September). Consistent global land cover maps for climate modelling communities: current achievements of the ESA’s land cover CCI. In Proceedings of the ESA Living Planet Symposium (pp. 9-13).
[10] Dee, D. P., Uppala, S. M., Simmons, A. J., Berrisford, P., Poli, P., Kobayashi, S., ... & Bechtold, P. (2011). The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quarterly Journal of the royal meteorological society, 137(656), 553-597.