Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

2.1 Introduction

Over the past 50 years, anthropogenic ecosystem changes were more rapid and extensive than in any comparable period in history (MEA 2005) and nowadays no ecosystem is free of pervasive human influence (Vitousek 1997). Changes in land surface characteristics mirror a multitude of processes induced by human alteration of the Earth system as a whole. Earth observation (EO) allows for repeated, synoptic and consistent measurement of the Earth surface and has long been used for environmental assessments (Lambin and Strahler 1994). The process of monitoring changes and modifications of land surface characteristics by means of a series of EO data is often referred to as time series analysis.

Approaches for time series analysis in remote sensing have long been restricted to the domain of wide swath, coarse spatial resolution sensors. Such systems commonly achieve complete global coverage on a (near-) daily basis. Image products can be generated based on a defined temporal interval, such as the 8- or 16-day Normalized Difference Vegetation Index (NDVI) composites derived from AVHRR data. This allows the direct utilization of time series methods, many of which require equidistant observations and were more commonly used in econometric or meteorological sciences. The temporal repeat frequency of higher spatial resolution sensors, however, does not readily allow applying these types of methods. On the one hand, orbital and engineering characteristics drive repeat acquisitions on the order of 7–20 days. On the other hand, acquisition strategies and cloud coverage govern the actual availability of scenes acquired with low or no cloud cover contamination.

Recently, great advances have been made in the use of remote sensing data that allow analyses at landscape to regional scales, such as Landsat’s Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), and Operational Land Imager (OLI). Generally, the availability of optical spaceborne imagery has tremendously increased. This was achieved by numerous new sensors in space, but the sea change is due to open access policies of governmental agencies. While coarse spatial resolution data have traditionally been available free of charge, this was not the case for higher resolution imagery. During 2004, the Brazilian Space Agency (INPE) pioneered free and open access to medium resolution satellite imagery by first providing China-Brazil Earth Resources Satellite (CBERS) data for free and subsequently making the Brazilian Landsat data holdings available at no cost. Most importantly, however, in 2008 the USGS adopted a free data policy for the U.S. Landsat holdings, the largest archive of Landsat data (Woodcock et al. 2008). The global Landsat archive constitutes a unique record of the way humans modify the land surfaces (Roy et al. 2014), and Landsat’s spatial resolution enables chronicling of anthropogenic and natural change at all scales (e.g. Gutman et al. 2008). The USGS Landsat archive alone currently contains more than four million scenes and Landsat 8 is currently contributing an unprecedented 650 scenes daily. Additionally, the ongoing Landsat Global Archive Consolidation effort is repatriating unique scenes from data holdings of the network of international receiving stations (Wulder et al. 2012). Next to data availability, providing ready-to-use data is the second prerequisite to promote new time series analysis applications. Landsat data are distributed in precision terrain corrected (L1T) format (Landsat Project Science Office 2010; USGS 2012), providing the radiometric quality and spatial precision needed to directly employ the imagery in end-user workflows. The new L1T processing of Landsat data has set a ‘gold standard’ for the quality of absolute and relative geometric accuracy. Careful radiometric calibration over the series of Landsat sensors has made consistent conversion of digital counts to radiance and reflectance possible (Chander et al. 2009). Substantial user input and interaction requirements have until recently prevented the correction of atmospheric effects and the subsequent conversion to surface reflectance for large amounts of Landsat data. Automated atmospheric correction algorithms and reliable cloud screening algorithms are available today that allow bulk processing for creating time series of reflectance data (Ju et al. 2012; Masek et al. 2006; Zhu and Woodcock 2012). This increasing availability of Landsat data as well as improved data quality and newly emergent pre-processing algorithms has spurred considerable methodological innovation for time series analyses with Landsat data.

The open data model that was spearheaded with Landsat has since been adapted for several sensor systems, including coarse resolution sensors such as MODIS or MERIS, and higher resolution sensors such as SPOT (i.e. SPOT World Heritage programme, CNES 2014) or the Sentinel programme (ESA 2014). Similarly, the development of important pre-processing (Zhu and Woodcock 2012) and analysis (e.g. Huang et al. 2010; Kennedy et al. 2010; Verbesselt et al. 2010) tools has embarked on a trend towards open source and open access, fostered for example by code sharing platforms and open source programming languages. Distributed and hosted processing services have appeared recently by commercial or scientific institutes (e.g. NASA’s Giovanni system, Berrick et al. 2009), several more have been announced in order to address the big data challenges in remote sensing. Simultaneously. Concern regarding observational continuity of long running programs (Wulder et al. 2011) has spurred the discussion regarding suitable imaging systems as a replacement or for different systems to function in a virtual constellation. Time series analytic capabilities can be considerably enhanced in terms of temporal extent or observational density by taking advantage of tools for data fusion (e.g. for merging Landsat and MODIS reflectances, Gao et al. 2006; Zhu et al. 2010) or long term time series models (Jonsson and Eklundh 2004). While these recent developments hold true for a wider range of sensor systems, in the following we focus on Landsat data and time series analyses in the context of land systems.

2.2 Process Dynamics and Time Series Analysis Requirements

Before implementing a time series analysis, it is important to understand which land surface processes are better tackled with time series analyses rather than traditional analyses such as multi-temporal image classification. Generally, analysing phenology-driven, either highly dynamic or gradual and long-term, change processes may profit greatly from the use of time series. Highly dynamic processes require using time series with a sufficient observation density so that ‘hot moments’ can be captured, while a long-term gradual change signal can only be separated from image or phenological noise when time series are long enough.

Typical processes inducing changes in or modifications of land surface characteristics are manifold. They are often directly driven by human activity (e.g. urbanisation, extension of agricultural areas, open pit mining, deforestation), result partly or indirectly from human action (e.g. climate change-induced vegetation change or glacial melting) or arise from natural processes (e.g. forest windfall, pests, landslides or geological processes, El Ninõ). A detailed monitoring of such processes—for example, quantifying rates of gradual change or detailed identification of dates of abrupt change—requires different analysis schemes. Time series analyses employ the temporal signal at pixel level to derive metrics for mapping or monitoring. Such metrics may include linear or nonlinear trends, amplitude, phase or break points.

Examples from two major domains of remote sensing-based monitoring illustrate the different needs for analyzing different processes: Monitoring agricultural intensification and monitoring deforestation. If we, for example, wish to gain a better understanding of processes related to agricultural intensification in central Europe, we need to get a remote sensing-based characterization of intra- and inter-annual changes in land surface phenology. We need season-specific multi-date measurements across 1 year to describe the diverse agricultural growth, yellowing and harvesting trajectories and we need the same specific multi-date observations between years to describe crop rotation cycles or change trajectories related to cropland-grassland conversions. To map forest changes such as wildfires and deforestation, annual anniversary-date observations may be sufficient. However, in many tropical regions, multiple observations a year are required to obtain a complete cloud-free, annual snap-shot for a region of interest. Actual needs may vary though, depending on the type of forest and ecosystem dynamics.

The examples above show some of the process-related aspects in different ecosystems that steer the applicability of time series approaches. Basically, the process to be monitored, the respective ecosystem dynamics (temporal trajectory of static and changing land use/land cover) and ecosystem heterogeneity (spatial-spectral domain) define the framework needed. In many cases, an ideal approach additionally needs to be adapted to actual data availability, especially when processes are to be assessed over large areas with variable observation density (i.e. extending over several footprints).

2.2.1 Variables Used in Time Series Analysis

Time series analyses have been applied to a wide variety of remote sensing based variables ranging from individual spectral bands and derived indices to remote-sensing-based predictions of land surface or biophysical parameters (Main-Knorn et al. 2013). The choice of variable often depends on the system and change process under study, and also the extent to which a priori information is available to translate surface reflectance to a desired land surface parameter. In the simplest case, a single spectral band might be input to further processing. But in many cases, simple metrics derived from spectral bands enhance the representation of land surface properties such as vegetation greenness—for example, vegetation indices (Choudhury et al. 1994; Huete et al. 2002), tasselled cap transformed bands (Kauth and Thomas 1976) or the outcome of problem-specific spectral mixture analysis (Hostert et al. 2003). It has also been shown that for problems related to classified data, classification probabilities also offer great potential for time series analysis (Yin et al. 2014).

Variables used in time series analyses are, for example:

  • Single spectral bands, like the 1.6 μm shortwave infrared in Landsat band 5, which are sensitive for various change processes in forest ecosystems.

  • Time series of vegetation indices, commonly used for various vegetation-related analyses.

  • Tasseled Cap (TC) components, a combination of brightness and wetness for urban environments, or wetness for forests.

  • Integrated indices from tasseled cap; for example, the disturbance index (DI), created for tackling the specific dynamics in temperate forest ecosystems.

  • Spectral unmixing fractions, offering a high degree of freedom to focus on question-specific land cover components, such as fractions of concrete in urban environments, soil fractions in agricultural ecosystems or photosynthetic active vegetation for any green vegetation related metrics.

  • Biophysical parameters such as canopy cover, leaf area index (LAI), fraction of photosynthetic active radiation (fPAR), and aboveground biomass.

Vegetation indices are commonly used to characterize vegetation dynamics across different temporal scales and processes. In phenological applications, weekly and bi-weekly time series of vegetation indices are used to infer the timing of vegetation green-up and senescence (Zhang et al. 2003). At annual intervals, time series of vegetation indices such as the Normalized Burn Ratio, of TC wetness, and the DI have been used to detect forest changes resulting from clear-cut harvest and insect and fire disturbances (Kennedy et al. 2010). Although, vegetation indices are not tailored to a specific scene or ecosystem, they have been very effective for capturing abrupt and long-term vegetation trends across large areas and multiple biomes. However, in most cases ground measurements are still needed to relate changes in vegetation index derived from time series data to a specific land surface change.

Time series methods also provide a means to directly characterize the temporal dynamics of ecosystem properties (e.g. leaf area index and aboveground biomass) and to improve the mapping of these properties at individual points in time. Biophysical variables are often predicted using radiative transfer models (Myneni et al. 2002) or by building upon the empirical relationship between remote sensing data and ground measurements (Cohen et al. 2003). By incorporating trend information from adjacent time periods, residual noise from atmospheric path radiance and pixel geolocation error can be minimized (smoothing), and predictions can be obtained for periods where surface observations are hindered by clouds or acquisition gaps (gap-filling). For example, Powell et al. (2010) and Main-Knorn et al. (2013) applied a trajectory-based segmentation algorithm (Kennedy et al. 2010) to annual Landsat-based predictions of forest aboveground biomass to characterize forest disturbance related-biomass dynamics. Comparing time-series fitted versus field based biomass estimates showed that the time series algorithm also improved predictions for individual years. Similarly, time series smoothing and gap-filling has been used to improve seasonal representations of biophysical parameters such as MODIS-derived fPAR estimates (Nightingale et al. 2009). Although, seasonal time series of ecosystem properties have traditionally been limited to coarse resolution sensor data like MODIS, such applications will likely be possible for many regions at the Landsat-resolution within the next decade.

Basically, any continuous measure with sensitivity to the process to be monitored may be used as input for a time series analysis. In some cases, as in forestry-related research on temperate forest ecosystems, there are established measures (e.g. TC wetness, DI, Healey et al. 2005). However, tackling new research questions often requires sensitivity analyses on the usefulness of different indicators for running a time series analysis prior to actually doing so.

2.2.2 Implementing Time Series Analyses

It is obvious that different process regimes ask for different temporal resolutions and appropriate methods to deal with analysing temporally varying information. The density of suitable data can be an issue in many time series approaches, despite the enormous increase of data available today.

Theoretically—disregarding cloud cover or downlink capacities—with two Landsat platforms operating simultaneously, an 8-day revisit provides about 45 acquisitions per footprint and year. This allows for assessing most process regimes within terrestrial ecosystem dynamics. However, global availability of Landsat data within the USGS archive shows pronounced regional differences (Kovalskyy and Roy 2013). This is due to cloud cover but also to the fact that a truly global acquisition strategy has only been established during the later years of the Landsat 7 mission. For areas outside of the U.S., less data has been acquired and much of the archived imagery resides in data holdings of the cooperating international receiving stations. Therefore in many areas around the world the ideal time series approach cannot be pursued due to data constraints, and trade-offs need to be considered regarding the established time series and the choice of analysis methods. Scale considerations might also require trade-offs as certain time series characteristics might be ensured for a smaller area, but providing the same characteristics for larger regions such as entire continents is often hardly feasible.

Conceptually, we may consider different types of time series for optical remote sensing data:

  • Original image acquisitions: For some applications, such as those with a focus on regions with low cloudiness like deserts and semi-deserts, data availability from Landsat overpasses every 16 days may suffice when using original imagery or mosaicked data (‘native data’). For many applications, though, data on a scene-by-scene basis does not suffice and image compositing techniques are required to increase temporal data density.

  • Data spacing: Some analyses require equally spaced data, such as annual peak phenology, while others can deal with the entire acquisition record of usually unequally spaced time series.

  • Intra- or inter-annual analyses: Depending on the process under research, intra- versus inter-annual time series or combinations of the two might be appropriate.

Summarizing, time series analysis is based on one or several indicators from which appropriate temporal metrics are extracted, such as early maxima of a vegetation index (either allowing or not allowing gaps in the time series), the slope of a linear trend along time series values, a sinusoidal function fitted to a phenological trajectory over several years, or simply a breakpoint in the time series indicating change. Recent research indicates that this selection process can be performed most efficiently by machine learning algorithms, e.g. support vector machines or random forests (Senf et al. 2013). This holds specifically true if tens or hundreds of metrics are calculated. Mapping results from the entire time series analysis process may then again be categorical after classifying the created metrics (classes derived from a time series, such as deforestation, cropping systems or cycles) or continuous (gradients of shrub encroachment on abandoned farmland or soil fraction changes during erosion processes) when using metrics as input to a regression analysis.

2.3 Time Series Analysis Examples

In the following we provide two examples to illustrate the conceptualization of time series analyses. Both examples demonstrate the opportunities and challenges of current time series analysis using freely available Landsat data.

2.3.1 Monitoring Tropical Deforestation

Mapping and monitoring tropical deforestation and forest degradation has been a focus of remote sensing research for several decades (Mayaux and Lambin 1995; Skole and Tucker 1993; Woodwell et al. 1987). Estimating the rate of change in tropical forest cover is a key requirement for understanding the impact of human activities on the Earth’s climate, the functioning of ecosystems, and biodiversity (Lambin et al. 2003). While tremendous progress has been made in developing remote sensing-based methods and protocols for estimating tropical deforestation rates from local to pan-tropical scales (GOFC-GOLD 2013), in many regions accurate forest monitoring has only recently become feasible with dense optical time series of high spatial resolution (<50 m). The necessity of high spatial resolution sensor data for monitoring tropical forest changes arises from the complexity of forest change patterns, which range globally from large clearings for mechanized agriculture to fragmented clearings associated with small-holder agriculture. The forest definition, as agreed in the Marrakech accords (UNFCC 2002), establishes a 0.05–1 ha minimum area, a minimum of 10–30 % tree canopy cover, and a potential of 2–5 m tree height for forest. Such small-scale changes cannot be resolved sufficiently with coarse resolution sensor data.

Prior to the opening of the USGS Landsat archive, the Global Land Survey (GLS) epochal dataset (Gutman et al. 2008) provided the first time series of orthorectified, high-resolution images that covered nearly the entire tropical region. However, the epochal resolution has proven to be insufficient for detecting forest change in some regions or for detecting transient forest change. In Indonesia, where the majority of land-use change comes from conversions of tropical forest to tree and oil palm plantations, Broich et al. (2011) showed that time-series approaches based on all good land observations were more accurate in mapping forest cover change than change maps based on epochal image composites. Similarly, shifting cultivation leads to a temporary removal of forest cover followed by a short period of cultivation and subsequent fallow regrowth. In regions of Southeast Asia, where shifting cultivation is the dominant land use, forest changes occur within complex mosaics of primary and secondary forest vegetation (Ziegler et al. 2012). The detection of clearings from shifting cultivation is complicated by the pronounced vegetation phenology in tropical dry and seasonal forest, where the clearing and burning usually occurs towards the end of the dry season, and by the fast recovery of the spectral signal when the land is left fallow (Fig. 2.1).

Fig. 2.1
figure 1

Landsat image chips (bands 5-4-3) showing the spatial and temporal patterns of slash and burn agriculture of a small area (3,000 × 3,000 m) in Houaphan Province, Lao PDR. A single crop of rice is followed by fallow periods of 2–10 years; 5–10 % of the area is used for successive crops of rice for 2–5 years. The time series plot in the lower right shows the Normalized Burn Ratio (NBR) from all cloud-free Landsat imagery between 2000 and 2004. To illustrate the seasonal variations in NBR a simple harmonic function fitted to the entire pixel time series is overlaid

Full size image

To detect forest clearings in northern Laos, Pflugmacher et al. (2014) constructed annual time series of the Normalized Burn Ratio (NBR) based on all available Landsat imagery with a cloud cover of less than 75 %. The NBR index contrasts Landsat’s 2.2 μm short-wave infrared (SWIR2) band with the near-infrared (NIR) band (NBR = (Band 4 − Band 7) / (Band 4 + Band 7); Key and Benson 2006). NBR can theoretically range from −1 to 1, where low and negative values are associated with sparsely- or un-vegetated areas and high values are associated with dense vegetation. The difference in the NBR between two images (ΔNBR) has been most prominently used to map burned area and burn severity of vegetation fires (Roy et al. 2008). However, due to its sensitivity to forest structure and vegetation moisture, ΔNBR has also been used to map forest disturbance caused by wind-throw, logging, and insects in temperate forests (Kennedy et al. 2012). Further, NBR is less sensitive to atmospheric water vapour and increased aerosols during the burning season, compared to indices that use Landsat’s visible bands. In order to minimize the effects of vegetation phenology on ΔNBR between years, Pflugmacher et al. (2014) used annual, minimum-value composites of NBR for the dry season (Fig. 2.2). The approach requires images that are atmospherically corrected (Masek et al. 2006) and cloud-masked (Zhu and Woodcock 2012).

Fig. 2.2
figure 2

Annual time series of the minimum Normalized Burn Ratio (NBR) for the dry season (October–April) from a single pixel cleared in 2000 (Houaphan, Lao PDR). The red line connects the undisturbed, disturbed and recovered periods via linear segments. The timing of recovery is defined when the NBR reaches 90 % of the pre-disturbance value

Full size image

Time series approaches that incorporate explicit seasonal models (Verbesselt et al. 2010) or more complex temporal segmentation algorithms (Kennedy et al. 2010) have recently become available, but these have not been rigorously tested in the tropics, yet. However, it has already become apparent that the opening of the Landsat archive and the release of the Landsat Climate Data Record in November 2013 have greatly enhanced the monitoring of tropical forest changes.

2.3.2 Mapping Pan-Carpathian Agricultural Land Use Change

The second example is based in a temperate European setting and illustrates how changes in agriculture production systems can be assessed for larger regions using historical Landsat data. The study focused on the extended Carpathian Mountain range in Eastern Europe, a region comprising of approximately 375,000 km2 and currently extending over seven individual countries. Until 1989, almost the entire region was under the political influence of the Soviet Union, which included collectivized agricultural production. Being heavily subsidized and connected to large output markets, basically all cultivatable land was under agricultural production. With the collapse of Eastern European Socialism, agricultural production experienced a drastic shock and most production ceased, leading to extensification through conversion of cropland to grassland, or the abandonment of cropland with subsequent shrub encroachment. Since 2004, most countries have been accessioned to the European Union, which again affected agricultural land use considerably, including re-cultivation efforts or intensification (Griffiths et al. 2013a).

The open Landsat archive provides a unique opportunity for assessing the spatio-temporal patterns of these crop regime changes over the entire region since 1984. For this, at least 32 Landsat footprints need to be included into the analysis scheme. The region comprises an extensive altitudinal gradient, from plains over foothills to high mountain valleys. Cloud coverage is generally profound. The USGS Landsat holdings for the period of 1984 (the first year of Landsat 5 TM data) to 2012 are heterogeneous, with relative data scarcity for the 1980s, basically no data available between 1995 and 1999, but better data availability after 2000 including Landsat 7 imagery that suffered from the defect of the scan line corrector after May 2003. From a time series methods perspective, two spectral-temporal features need to be captured when attempting to assess these changes in cropping regimes: (1) observations from the spring, summer and fall seasons, which are required to reliably separate cropping systems from grassland land use, and (2) a series of such season-specific observations over the different time periods—socialism, transition and post-EU accession—to be able to detect changes among these different agricultural land use regimes. Image compositing algorithms were used to approximate the desired series of season-specific observations over the entire region (Griffiths et al. 2013a, b). Overall, the dataset included cloud-free multi-season composites for 1984, 2000, and 2010 to represent the socialist, transition and post-EU accession periods respectively. In general, multi-year imagery is required to achieve a complete cloud-free coverage for the region.

First, all available Landsat L1T imagery with a metadata-based cloud cover estimate of no more than 70 % was downloaded (approximately 5,000 images). All images were atmospherically corrected and converted to surface reflectance (Masek et al. 2006), and subsequently cloud/shadow masks were developed (Zhu and Woodcock 2012). To ensure the ability to process imagery on a per-pixel basis over several Universal Transverse Mercator (UTM) zones, all data was then transferred into a common projected continental coordinate system. The applied compositing algorithm provides three types of outputs (Fig. 2.3): best observation composites, spectral-temporal variability metrics and metadata layers (Griffiths et al. 2013b). For the first, all available pixel observations for a given time window were evaluated based on the image metadata (e.g. acquisition year and day-of-year (DOY)) and image characteristics (e.g. pixels distance to clouds). A score was produced for each observation based on these different characteristics and the pixel observation with the highest score was transferred into the best observation composite for a certain year and season (for details on the generation of scores, refer to Griffiths et al. (2013b)). Spectral-temporal variability metrics were calculated for defined seasonal windows based on all cloud free observations available for a given yearly range. These metrics were based on the NDVI and included the seasonal mean, range and standard deviation. Metadata layers provided not only the origin of each pixel, but also the evaluated score and number of cloud free observations on a per-pixel level.

Fig. 2.3
figure 3

Results of the seasonal compositing of Landsat data for the target year 2010. The top row shows the spring and fall best observation composite on the left and right, respectively (bands 4-5-3). The seasonal and annual characteristics of these composites are then summarized, providing the offset to the respective target day-of-year (middle) and the respective acquisition years (bottom) for each pixel. The overview map (top left) shows the approximate location

Full size image

Once the data were composited and all metrics were derived, supervised Random Forest (RF) models were used to capture the changes of interest (Griffiths et al. 2013a). Training data were collected for a total of 15 stable or transition classes, including cropland-grassland-forest transitions as well as stable cropland or forest classes. The RF model was then applied to a combined dataset containing the best observation composites (3 × 6 spectral bands for three reference periods, 54 bands in total), the variability metrics (3 × 3 metrics for the three reference periods, equalling 27 bands) and selected metadata layers (e.g. day-of-year, year, compositing score), resulting in a total of 100 bands. The overall accuracy of the resulting change map was over 90 %. Individual transition classes were subsequently recoded to assess agricultural regime changes. For example, agricultural abandonment included cropland-grassland conversions but also cropland-forest conversions. Details on the processing, mapping and result interpretation are provided in Griffiths et al. (2013a).

2.4 Challenges and Opportunities

The development of time series analysis approaches has grown swiftly during the last years, as the opening of the Landsat archive and streamlined pre-processing and data delivery created a surge of new methods and applications. The ongoing “repatriation” of Landsat imagery from distributed global archives into the USGS archive will create further opportunities; for example, the European archive alone holds 1.5 million unique Landsat scenes not available elsewhere. Yet, the remote sensing community is just at the verge of a new era, where accelerating global change meets increasing capabilities for remote sensing data analysis. Forthcoming challenges and opportunities relate to three intertwined domains.

From an application-focused perspective, the need for new information products is evident. Classic land use / land cover change analyses need to be improved; for example, maps of heterogeneous (e.g. savannah landscapes) or dynamic ecosystems (e.g. shifting cultivation systems in the tropics) are still deficient. Beyond this evident need for enhancements, there is also the need to create more problem-focused products, such as Essential Climate Variables from remote sensing data. Without such seamless global databases, climate impact research will fail to create vital information for climate policy decision makers. Furthermore, information that supports deciphering land use intensification or extensification patterns where land use itself has not changed is urgently needed to understand limitations in our use of natural resources globally. Both will only be possible with spatial high-resolution dense time series analysis capable of capturing subtle and long-term change patterns.

From a methods perspective, time series densification depends on improved image compositing techniques. In a perfect remote sensing setup, data availability would not be an issue. However, useful optical data are not available with every satellite overpass because of varying cloud coverage across the globe, daily variation in illumination and atmospheric conditions, as well as data storage and downlink-related acquisition constraints. Gaining the maximum amount of usable pixels from partially cloud covered datasets is hence key to further increased data density, which in turn is mandatory to successfully monitor processes that are at the same time driven by phenology and land use.

From a sensor-focused perspective, forthcoming improvements in time series analyses will depend on our ability to create time series from multiple sensors, often referred to as “virtual sensor constellations”. Few studies have used multiple sensors to improve the temporal resolution of time series, mostly employing different sensors of the Landsat family (Hostert et al. 2003; Main-Knorn et al. 2013; Pflugmacher et al. 2012). Creating sensor constellations is a non-trivial task, though, and the remote sensing community is just starting to develop operational constellation setups. Future virtual constellations will also focus on the integration of optical and radar data or multi-scale optical data. The latter has been pioneered based on multi-sensor fusion approaches such as spatio-temporal adaptive reflective fusion model (STARFM, Hilker et al. 2009). The most promising approach, however, will certainly be the integration of new and next-generation optical land imaging systems with similar characteristics. Specifically, the USGS Landsat family and ESA-based Sentinel-2 missions will create unprecedented virtual constellation synergies (Drusch et al. 2012), coming close to the temporal density of coarse-resolution monitoring systems.