Land Cover 1

A detailed and accurate knowledge of the land cover is crucial for
many scientific and operational applications, and as such, it has been
identified as an Essential Climate Variable. This accurate knowledge
needs a frequent update of the information.

The reference land cover mapping product at the European scale is
Corine Land Cover (CLC), which offers a rich nomenclature going
beyond land cover and addressing land use. The main drawback of
CLC is its lack of timeliness: CLC 2012 was made available in 2015.

For most applications, timeliness is more important than a detailed
nomenclature. Timeliness can only be achieved with automatic methods
which are robust and reliable. Also, new kinds of image data with
respect to SPOT and Landsat are needed to obtain the required
accuracy.

Decametric or metric spatial resolution imagery is needed in order to
produce detailed maps, but many land cover classes can only be
recognised by their temporal dynamics, and therefore, high temporal
resolution is needed. The availability of Sentinel-2 imagery, with its
unique characteristics (290 km swath, 10 to 60 m spatial resolution,
5-day revisit cycle with 2 satellites, 13 spectral bands) will make
possible the implementation of land cover map production systems able
to deliver up to date and accurate information with the appropriate
timeliness.

The state of the art in land cover mapping uses image classification.
In the case of global mapping systems using medium to low resolution
imagery, as for instance ESA's GlobCover, rule or expert based
approaches have been implemented. This was possible due to the
definition of the classes and the coarse target resolution. At finer
scales, rule sets would have to be highly local and therefore the
implementation of global mapping systems seems unfeasible.

Supervised classification is known to be superior to unsupervised
approaches. Its main drawback is the need of training data (ground
truth or other reference data). This has often been criticised and
seen as a blocking point for the implementation of global scale
operational systems.

However, the impact of the scarcity and spatial distribution of
reference data for the mapping of large areas using supervised
classification has not been assessed in the literature. Unsupervised
systems also need reference data for quality assessment and sometimes
for class recognition. This has led to the development of
classification approaches combining the strengths of both approaches
and terms like semi-supervised, transductive or domain adaptation have
emerged in the machine learning literature in the recent years.
However, no real world operational use of these approaches has been
made public.

One of the main issues involves achieving spatial continuity and
consistency of the final product. Indeed, covering a whole country may
need several orbit tracks which will be observed at different dates.
Cloud cover may also imply that not all areas are observed with the
same frequency. These differences in temporal sampling may cause
spatial inconsistencies in the final product.

The French Theia Land Data Centre has set up a Scientific Expertise
Centre whose aim is to implement an operational fully automatic land
cover map production system using mostly Sentinel-2 image time series.
The product will be updated once a year and will contain 20 thematic
classes mapped at 10 m. resolution. The system will be operational
in 2017.

In this contribution, we will present the design of the current
pre-operational land-cover processor, its performances using real data
as well as the foreseen evolutions. In particular, we will detail:
- how the full Sentinel-2 time series is used without needing date selection;
- how spatial consistency is ensured across adjacent orbit tracks;
- which is the impact of scarcity and spatial inhomogeneity of the
  reference data used for training and validation;
- how synergy between Sentinel-2 and Landsat-8 time series is implemented;
- how to overcome the lack of any reference data for the current
  period to be mapped by using outdated land cover maps or other
  reference data sets.

Background and given situation

The diverse applications of LCLU data have led to the development of many classification systems. Most of these contain a mixture of land cover and land use information. Each application emphasises particular aspects of land cover and land use, related to specific requirements and purposes. Furthermore, data collection methods, scales, tailored-to-purpose definitions and lack of completeness hamper the data transfer from one application to another. The concept and nomenclature of CORINE Land Cover (CLC) has established itself as the quasi-standard for LCLU mapping in Europe. On the one hand, the technical circumstances including quality of affordable satellite imagery, data computing and storage capacities and methodologies have developed rapidly, which open a whole new analytical approach regarding the extraction of land information from remote sensing data. On the other hand, thematic requirements and political reporting obligations have also evolved and changed. Consequently, the need for a revision of the CLC concept has become evident.

The EAGLE concept

The EAGLE data model behind the concept has been design according to a number of key criteria. It clearly separates between land cover and land use information, is scale independent and also tackles the temporal aspects of transient or altering phenomena. Its structure is flexible enough to react on user needs arising from different fields of application and institutional levels. The EAGLE concept is technically represented as a UML (Unified Modelling Language) chart and a cross-table (referred to as the EAGLE matrix).

Figure 1 shows the simplified structure of the object-oriented EAGLE data model, consisting of three main segments of Land Cover Components (LCC) distinguished by their (bio‑) physical appearance: Abiotic (non-vegetated) artificial and natural surfaces, Biotic (vegetated) surfaces and Water surfaces. Each segment contains a number of LCCs hierarchically ordered in sub-branches. The LCCs are further described with additional landscape characteristics (CH) that express more specific details about their properties.

Following the structure of the UML model, one or several LCCs with their attached characteristics build a Land Cover Unit (LCU) that also can have its own specific characteristics. The LCU is completed by the additional information of LUA originating from the Hierarchical INSPIRE Land Use Classification System (HILUCS) and extended by EAGLE-specific sub-types. LCCs are mutually exclusive. Several LCCs can occur inside a LCU, but they cannot overlap. LUAs, however, can occur in an overlapping manner.

Application of the EAGLE concept

The following two main applications are proposed for the EAGLE concept:

As a tool for semantic comparison and translation between class definitions within one or between several different classification systems. Abstract class definitions can be analysed and decomposed with the model elements in a descriptive and diagnostic manner (without looking at any real landscape scenario).

As guidance for mapping activities to describe real landscapes by using generic descriptive elements to characterise single land cover units which can then be classified according to an appropriately chosen application purpose and nomenclature.

Once a land monitoring system based in the EAGLE concept will be established and is working operational, the field of work of statistics can benefit from such a harmonised approach to describe landscape component-wise instead of classifying it with inflexible pre-defined classes. Then, more robust figures can be calculated based on a more object-oriented and parameterized perspective on landscape rather than losing information due to cut-off-effects by threshold.

 Some use cases have already emerged to assess the applicability and usefulness of the EAGLE concept in the field of habitat monitoring and citizen science.

Key messages

1) The EAGLE matrix aims at being a tool for analytic decomposition of class definitions and for semantic translation between recent or future nomenclatures.

2) The EAGLE model offers a conceptual basis for a future harmonised European land monitoring system and is open to be implemented as an object-oriented guideline for mapping and monitoring initiatives.

3) The EAGLE concept does not represent another classification system but instead is a descriptive vehicle for harmonisation of LC/LU information supporting both top-down and bottom-up approaches.

State of play and outlook

The long-term target is to implement a future European land monitoring framework upon the principles of the EAGLE concept. As the core role, the concept shall help to harmonize LC/LU data between countries and also to integrate national data into pan-European datasets in a bottom-up approach and vice versa (top-down).

Under service contract with EEA, several tasks and work packages have been worked out so far by the EAGLE group. An explanatory documentation has been published as well as the EAGLE matrix and the UML data model. To ease the use of the matrix/model to semantically analyse and comparison nomenclatures and theirs classes, a web-based matrix population and comparison tool has been developed in a pre-beta version. Further, a physical database has been developed and tested to store test samples of spatial data which is expressed in terms of the data model. For a coherent streamlining of the countries´ bottom-up approaches, the question of common generalisation and aggregation rules was outlined and addressed.

On a new web-platform under the Copernicus ‘Land’ domain of EEA, the documentation and methodology of the concept is available, as well as the tools will be made accessible. Further, a forum has been installed to facilitate the communication and exchange of experiences with the user community.

The 5^th IPCC Assessment Report into the impacts of climate change highlighted land surface models as one of the key contributors to uncertainty in climate change impacts projections. Here, we present a study that uses state of the art satellite-derived land cover maps to identify locations for and drivers of uncertainty in the representation of the vegetation fractions in land surface models.

Current generation Land Surface Models (LSMs) use the concept of Plant Functional Types (PFTs) to group different vegetation types and species according to similar physiological, biochemical and structural characteristics. LSMs define PFT-dependent parameters (fixed values) in order to solve equations that describe how different types of plants interact with the biosphere. Therefore, it is critical to have accurate information on the amount, and which, PFTs occur in each model grid cell, and to understand the sensitivity of models to known uncertainties in the spatial distribution of PFTs.

Land cover classes are converted to fractions of PFTs using a cross-walking procedure that assigns a fractional coverage of each PFT for each land cover class. In this process, two key uncertainties exist. Firstly, there is uncertainty in the land cover classification algorithm, which assigns the most likely land cover class depending on the satellite-derived surface reflectance. Alternative land cover classes to the most likely class may have a lower probability, but may still be plausible given the surface reflectance. Secondly, there is uncertainty in the land cover to PFT cross-walking procedure. The fraction of each PFT within a land cover class is determined using the Land Cover Classification Scheme (LCCS) definition of a land cover class, which is further refined using expert judgement. However, alternative cross-walking fractions would be equally plausible, within the range specified by the LCCS. This is especially the case for mosaic or mixed classes and open canopies.

In this study, we quantified the sensitivity of 3 LSMs (JSBACH, JULES, and ORHCIDEE) to land cover class uncertainty, and cross-walking uncertainty with regard to key indicators of processes in the carbon, hydrological and energy cycles. In order to achieve this, we express uncertainty in the context of either minimising or maximising biomass. We used this approach to identify locations where the uncertainty in PFT fraction is greatest, and identify which PFTs are most affected by the different sources of uncertainty. Furthermore we identified the impact of this uncertainty on LSM estimation of carbon, moisture and energy fluxes.

This work will enable us to both advise the land cover mapping community about the accuracy requirements for land cover maps, and to provide insights to the earth system modelling community on the implications of decisions taken when converting from land cover to PFTs.

High resolution, reliable quantitative information on land cover and land use dynamics at regional to national scales derived via automated processes are urgently needed. Such services are in demand by numerous national and European stakeholders and decision makers and the global community. While the open USGS Landsat archive provides great opportunities for deriving historic information on the terrestrial land surface at global scale, the current operational Landsat-8 satellite and the Sentinel-2 mission as part of the European Commission’s (EC) and the European Space Agency’s (ESA) Copernicus Programme will ensure data continuity for global land surface monitoring at high to medium spatial resolution.

The TimeTools solutions developed at the Earth Observation Center (EOC) at the German Aerospace Center (DLR) is an extendible, modular framework for the automated pre-processing, processing, and value-adding of optical, high resolution Earth Observation data. The framework provides a set of easy to use algorithms supporting the preprocessing of data beginning with the automatic selection and acquisition of appropriate satellite data for a used defined area of interest, up to the generation of gridded cloud-free surface reflectance composites. Implemented methods related to the generation of land indicators currently support the generation of spectral indices, the exploitation of seasonality through the generation of multi-temporal metrics, as well as an automated method to derive land cover information and related land cover statistics. Currently, implemented functionalities are limited to the processing of satellite data from the Landsat archives. Future efforts, however, will focus on the adaption of algorithms to support the processing and value adding of Sentinel-2 data.

TimeTools is directed towards the provision of operational products such as land-cover maps or land change maps, being of central importance to related land monitoring and reporting services. The TimeTools workflow was already applied for the full-automated generation of land cover maps at 30m resolution from Landsat-5 to Landsat 8 for selected areas in Europe, Asia, and Brazil.

Essential Climate Variables (ECVs) were listed by the Global Climate Observing System (GCOS) as critical information to further understand the climate system and support climate modelling. In response, the European Space Agency (ESA) launched its Climate Change Initiative (CCI) in order to deliver global datasets matching the need for long-term satellite-based products for the climate domain. It focuses, through individual projects, on 14 ECVs selected in the atmospheric, oceanic and terrestrial domains on the basis of ESA priorities.

One project, named the ESA Land Cover CCI (LC_CCI), is dedicated to global Land Cover (LC). This project, built on the ESA-GlobCover experiences, aimed at revisiting all algorithms required for the generation of global LC products from various Earth Observation (EO) instruments that match the needs of key users of the climate modelling community.

At the end of its first phase, the LC_CCI project delivered a new generation of satellite-derived global land cover products consisting in three maps at 300 m spatial resolution for three epochs centered on the years 2010 (2008-2012), 2005 (2003-2007) and 2000 (1998-2002). These maps were obtained from SPOT-Vegetation and ENVISAT-MERIS time series. Other significant outputs were (i) 7-day surface reflectance time series for the whole archive of MERIS Full and Reduced Resolution data (2002-2012), (ii) land surface seasonality products describing the seasonal variability of the land surface for three variables (vegetation greenness, snow cover and burned areas) and (iii) a global map of open permanent water bodies at 300 m spatial resolution derived from Envisat ASAR images and ancillary data.  All products were delivered along with an aggregation tool, enabling re-projection and re-sampling as well as the translation from LC classes into Plant Functional Types for the different climate models. Three major Earth System models already investigated these new products as land surface information.

During the second phase (2014-2016), the project has a twofold objective. On one hand, it aims at generating new global LC products covering the 1990s and 2015 epochs. To this end, new pre-processing and classification methods based on AVHRR, PROBA-V and the forthcoming Sentinel-3 sensors need to be developed. The challenge is to ensure a consistency between the successive global LC maps while they are derived from different sensors. On the other hand, the project aims at addressing the requirement of higher spatial resolution expressed by the climate science community. To reach that goal, the project will generate a first 10-20m prototype map over Africa based on Sentinel-2 and supplemented by Landsat-8 datasets, thus demonstrating the feasibility and performance of such approach. Taking also advantage of the Sentinel era, the potentiality to derive a new water body product based on Sentinel-1 mission is analyzed.

The increase in spatial resolution requires significant methodological adjustments and innovations with respect to the processing chains developed for medium spatial resolution imagery at global scale. For the pre-processing, the neighborhood effect has to be taken into account in the atmospheric correction. In addition, due to the lower revisiting capacity of high spatial resolution sensors such as Sentinel-2 and Landsat-8, the spatial consistency of surface reflectance between few images becomes a critical aspect in the production of high spatial resolution composites. For the land cover classification, the challenge is of a different nature. The decametric resolution captures the diversity of the landscape elements and their temporal shift between each other due to slightly different seasonality and ecological gradients. While the rich litterature on land cover mapping at high resolution supports the processing chain development, the data flow provided by Sentinel-2 forces to revisit the classification strategy to map the land cover consistently over space and time.

A last but important challenge is related to the validation of all these products. Validation is a critical step in the acceptance of products by the users’ communities. Different components are identified: confidence-building, independent statistical accuracy assessment and comparison with other global land cover products. For the statistical accuracy assessment, specific effort is made to build a global validation database containing high spatial resolution interpretations for all epochs.