S2 and L8 Exploitation Synergy 1

Free and open access to satellite imagery, such as Sentinel-2 and the Landsat series among others, have revolutionized the role of remote sensing in Earth system science. Changes in the global environment pose challenges to the science community that are increasingly difficult to address using data from single satellite sensors or platforms due to the underlying limitations of data availability and trade-offs that govern the design and implementation of currently existing sensors. Virtual constellations of planned and existing satellite sensors can help overcome this limitation by combining existing observations to reduce limitations of any one particular sensor. While multi-sensor applications are not new, the integration and harmonization of multi-sensor data is still challenging, requiring concerted efforts of science and operational user communities. For instance, the quality calibration and the spectral and spatial matching, effectively make the Landsat archival record a shared history for new Sentinel-2 measures. 

Defined by the Committee on Earth Observation Satellites (CEOS) as a “set of space and ground segment capabilities that operate in a coordinated manner to meet a combined and common set of Earth Observation requirements”, virtual constellations can be used to combine sensors with similar spatial, spectral, temporal, and radiometric characteristics. In this presentation, we articulate the potential and possible limitations to be overcome regarding virtual constellations for terrestrial science applications, discuss potentials and limitations of various candidate sensors, and provide context on integration of sensors. Thematically, we focus on land-cover and land-use change (LCLUC), with emphasis given to medium spatial resolution (i.e., pixels sided 10 to 100 m) sensors, specifically as a complement to those onboard the Landsat series of satellites. We conclude that virtual constellations have the potential to notably improve observation capacity and thereby Earth science and monitoring programs in general. Various national and international parties have made notable and valuable progress related to virtual constellations. There is, however, inertia inherent to Earth observation programs, largely related to their complexity, as well as national interests, observation aims, and high system costs. Herein we define and describe virtual constellations, offer the science and applications information needs to offer context, provide the scientific support for a range of virtual constellation levels based upon applications readiness, capped by a discussion of issues and opportunities toward facilitating implementation of virtual constellations (in their various forms).

This project is aimed at producing a harmonized Surface Reflectance product from the Landsat and Sentinel-2 missions to ultimately achieve high-temporal coverage (2 to 4 days repeat cycle, depending on the latitude) at high spatial resolution (20-60m). The goal is to achieve a seamless/consistent stream of surface reflectance data from the different sensors. This would provide an improved moderate resolution data set to address science and applications for rapidly changing phenomenon (e.g. agriculture, vegetation phenology, fire, flooding).

The first part of this presentation discusses the basic requirements for such a product and the necessary processing steps: mainly calibration, atmospheric corrections, BRDF effect corrections, spectral band pass adjustments and gridding. We demonstrate the performance of those different corrections by using MODIS and VIIRS (Climate Modeling Grid at 0.05deg) data globally. It should be noted that all the corrections and adjustments are generic in nature and could be applied to any sensor by taking into account specific sensor characteristics.

The second part is devoted to the analysis of the individual performance of the Landsat 8 and Sentinel 2 directional surface reflectance product for AERONET sites.  The protocol developed originally for MODIS for surface reflectance validation is used. This protocol has been used to assess the performance of MODIS over the past 15 years and gives consistent results with the independent analysis of the residual noise in the time series. This protocol has also been used for VIIRS and AVHRR and has also demonstrated its value for testing atmospheric correction algorithm improvements.

The third part is devoted to the derivation of a detailed and traceable error budget of the performance of the merged L8/S2 surface reflectance product based on: 1) the advertised calibration performance of both sensors 2) the theoretical error budget of the directional surface reflectance, revised based on the performance observed over the AERONET sites 3) the accuracy of the correction for BRDF and the residual noise in the spectral adjustments, derived from the analysis of the VIIRS/MODIS time series 4) the registration error between L8 and S2 (scene/land cover dependent).

In the fourth part, the merged L8/S2 surface reflectance products are derived for a series of test sites and the performance of the L8/S2 is evaluated in term of the agricultural monitoring capabilities within the framework of the G-20 GEOGLAM (Group on Earth Observations for Global Agricultural Monitoring) recommendations. The benefits of having both high spatial /temporal resolution in that context are discussed in the conclusion.

Sentinel-2 has tremendous potential for providing detailed information on land vegetation, not the least vegetation dynamics and phenology. This information will be useful in ecosystem research, research on land-atmosphere dynamics, and for generation of improved land cover products. Providing smooth seasonal image data and reliable phenological information from these data is the aim of our research. We base the research on the TIMESAT processing methodology, which is well tested and has been extensively applied to coarse-resolution data (http://www.nateko.lu.se/timesat). Mostly, TIMESAT and similar time-series methodology has been applied to coarse-spatial resolution data, however, new data sources will provide detailed information on vegetation dynamics relevant for a range of applications.

In order to modify the processing to work with high-spatial resolution data we compare a number of existing algorithms in TIMESAT (Savitzky-Golay filter, logistic and Gaussian fits) with newer smoothing algorithms (LOESS smoothing, and splines), so as to develop accurate and computer efficient solutions. We develop these methods to handle irregular data in time and for handling cloudiness. This includes utilizing quality information through weights, and upper-envelope fitting of vegetation indices to reconstruct temporal vegetation trajectories in the presence of noise. We also study methods for processing data with extensive data gaps, e.g. caused by winter conditions or long cloudy periods. Handling these data will require access to long time-series for building statistically robust algorithms. The Landsat archive today contains a very long record of data that is invaluable for providing this basis. Apart from Landsat we use data from the SPOT 5 Take-5 experiment for our algorithm development. To evaluate the accuracy of the methods we utilize a network of fixed multispectral sensors measuring time-series of reflectance in a variety of environments (tundra, forest, open land, wetlands), across a geographical gradient from north to south. This information can help understanding differences due to the different satellite platforms. We also utilize phenological cameras and UAV technology to better characterize vegetation in the test areas. The UAVs carry multispectral and RGB cameras and enable vegetation mapping at very high resolution.

We aim to deliver fast algorithms for smooth and cloud-free image production at user-defined time-steps. These algorithms will be useful in research on vegetation phenology, vegetation carbon cycle, plant ecophysiology, and for improved land cover classification by combining spatial and spectral information.  Thereby they will be relevant for research as well as for a range of applied fields, e.g. forestry, agriculture, societal planning, etc. The methods are general in structure and will also be useful with coarser-resolution data, e.g. from Sentinel-3/OLCI.

While the number and capacity of current Earth Observation systems is unprecedented, also information needs are increasingly complex. In the context of food security, for example, a differentiation of food and energy crops is needed, while carbon sequestration assessments would benefit from the reliable separation of natural forests from plantation forests. Systematically generated value added products from high to medium resolution satellites provide an important step towards addressing these issues, while operational products are currently only produced from coarse resolution sensor data. Sentinel-2 offers several relevant improvements over other medium to high resolution multi-spectral sensors. Besides improved spectral and spatial characteristics, the increased temporal repeat frequency is most important to a wide range of applications in the land remote sensing domain. In its operational constellation with two satellites, Sentinel-2 will achieve global coverage every 5 days. The effective number of cloud free observations will be considerably lower for many regions around the globe, though, due to persistent cloud cover. Fortunately, as the Sentinel-2 mission was designed to provide continuity to Satellite Pourl’Observation de la Terre (SPOT) and Landsat data. The combined and synergetic use of such data therefore presents an unique opportunity to improve effective observation frequency and thus to enable a variety of innovative analysis approaches.

Here we present the first components of an image processing and analysis framework developed within the Living Planet Fellowship project Island2Vap (Integrating Sentinel-2 and Landsat-8 data to systematicallygenerate value-added products at high resolution). We demonstrate results obtained from data acquiredduring 2015 over a Berlin-Brandenburg test site. This includes Spot-5 imagery from the Take-5 experiment, all available Landsat-8 and Landsat-7 data as well as early Sentinel-2 imagery.

In a first step, gridding and compositing of multi-sensor imagery into temporally equidistant composites is performed on atmospherically corrected reflectance data. Individual sensor reflectance observations are radiometrically normalized using a stable reflectance reference. Compositing is performed for different temporal intervals (i.e. 10, 14, 21 days) and spatial resolutions (10m, 20m and 30m). We assess the quality of the resulting composite time series by analyzing spectral homogeneity, synopticity and quantify the surplus value of multi-sensor integration versus individual sensor data. Two sets of value added products are subsequently derived from the time series of reflectance composites: phenology metrics and land cover maps. Phenology metrics are derived by transforming the multi-sensor composites into vegetation indices, cleaning and fitting the resulting time series and estimating key phenological dates (e.g. green-up, length of season, end of season) from the fitted time series. We compare our estimates derived from multi-sensor time series to those obtained from near-nadir MODIS vegetation index time series. This is done for selected pixels within large, homogenous landscape features. Land cover information is derived from the equidistant composite time series by unmixing class-characteristic temporal signals. We perform unmixing on different sets of composites and vegetation indices, targeting a relatively complex class catalogue including several specific crop classes and urban structure types. We compare the achieved results against those obtained from a more conventional Random Forest based time series classification. We focus on an evaluation of potentials and challenges of these new temporal products and based on this we will present strategies for making best use of the synergistic products. Overall our results underline the value of combining data from both sensors and suggest that deriving high resolution phenology and reliable land cover information from a synergistic use of Sentinel-2 and Landsat data is feasible.

A long-standing goal of the land imaging community has been to harness the full range of international sensing systems to monitor land dynamics.  Many applications require either more frequent observations than possible through individual national satellite programs, or require multi-modal observations (e.g. radar and optical).  As one example, the GEO Global Agricultural Monitoring (GEOGLAM) initiative has called for ~weekly data on crop condition, type, and area at spatial resolution finer than 50m.  Increased access to national satellite archives, combined with rapid development of high-end computing, has opened the door to “virtual constellations” in order to meet these needs.

The NASA Land Cover / Land Use Change Program recently selected projects for a Multi-Source Land Imaging (MuSLI) science team.  The goal of MuSLI is to prototype generation of standard, higher-level land products from international, moderate-resolution data streams.  While a high priority is the merging of data from the USGS/NASA Landsat and ESA Sentinel-2 missions, MuSLI also solicited for projects combining optical data with radar (e.g. Sentinel-1, PALSAR).  Selected projects include development of products for agricultural monitoring (crop type, area), vegetation phenology, urban mapping, wetlands and rice cultivation, burned area mapping, and forest dynamics.  The successful proposals demonstrated a clear need for high temporal resolution or multi-modal data, as well a clear community desire for the information.  In addition, a key component of MuSLI is collaboration with non-US researchers in order to establish “best practices” for using the international data sets.  

Supporting elements for the MuSLI activity include high-level computing provided by the NASA Earth Exchange (NEX) at NASA Ames Research Center, and development of a harmonized Landsat/Sentinel-2 surface reflectance product at NASA GSFC.  NASA is collaborating with USGS and ESA to ensure data provisioning for the research teams, and to characterize the Landsat-8 and Sentinel-2 systems.  This talk will provide an overview of the MuSLI program, and discuss evolving opportunities to generate new standardized products from international sources.