Conference Agenda

Temporal consistency of land cover time series and detection of major land cover change are key requirements from the users’ community to describe terrestrial surface over time. Land cover was listed as an Essential Climate Variables (ECVs) by the Global Climate Observing System (GCOS) as critical information to further understand the climate system and support climate feedback modelling. However, only global land cover maps from the same instruments have been produced till now, keeping the time series rather short. In the framework of the Climate Change Initiative (CCI) supported by the European Space Agency (ESA) climate modelling and remote sensing teams joined forces to design, implement and deliver global datasets matching the climate science needs for long-term global products.

Building on the ESA-GlobCover experiences, this research first revisited the land cover definition into a stable component and seasonal component and designed a new approach to produce consistent land cover time series decoupling consistent mapping and change detection. Then, the archives of several satellite missions including ENVISAT Meris FR and RR, SPOT-Vegetation, the more recent PROBA-V, and the 90s’ archive of 1 km AVHRR, were reprocessed using state-of-the art methods to produce weekly surface reflectance composite and quality flags throughout the years. According to a stratification splitting the world into 22 equal-reasoning areas from ecological and remote sensing points of view, different seasonal composites were compiled to enhance the land cover discrimination. A typology of 22 land cover classes was defined based on the UN Land Cover Classification System and its classifiers to support the further conversion into Plant Functional Types distribution required by the Earth System Models (ESM). The classification processes combined first machine learning and unsupervised algorithms at 300 m resolution from the whole MERIS FR archives using most of the MERIS bands to establish the land cover baseline. Based on similar algorithms annual global land cover maps from 1 km AVHRR HRPT, 1 km SPOT-Vegetation data sets and from 300 m PROBA-V time series were then produced to serve as input in the land cover change detection algorithm. Systematic analysis of the temporal trajectory of each pixel allows depicting the main change for a simplified land cover typology matching the IPCC classes. This new land cover change detection method was found quite reliable for SPOT-Vegetation and PROBA-V thanks to their excellent temporal co-registration. In contrast, the poorer radiometric and geometric quality of AVHRR HRPT time series only provided major change in contrasted landscapes. The change detected at 1 km was then disaggregated at 300 m whenever higher resolution imagery was available. Finally, these products were validated from an independent reference dataset built by a network of international experts.

All land cover maps can be visualized from an interactive web interface and downloaded along with an aggregation tool, enabling re-projection and re-sampling as well as the translation from LC classes into Plant Functional Types for different climate models. Three major ESM successfully completed several joint experiments based on the CCI land cover products.

The Copernicus Global Land Service is the component of the European Copernicus service which ensures a global systematic monitoring of the Earth’s land surface. It provides bio-geophysical variables in near real time describing the daily state, and changes in state, of vegetation, land surface processes and is currently preparing the release of a Moderate resolution Dynamic Global Land Cover layer.

In this presentation the methodology and rationale behind the Moderate resolution Dynamic Global Land cover layer is explained. This layer complements several global land cover ‘epoch’ datasets which have been created at medium (and high) spatial resolution during the last decade by providing a yearly dynamic land cover layer at 100m resolution. We will present the sub-product covering continental Africa for the year 2015. To build this global land cover layer, 100m spatial resolution PROBA-V data are used as primary EO data. Data fusion techniques are applied for areas with insufficient 5-daily PROBA-V100 m data and daily 300 m datasets are fused in. Next, time series metrics together with ancillary data sets (e.g. other Copernicus global land service biophysical products) are used in a supervised classification approach. Finally, at a third level, we build upon the success of previous global mapping efforts and focus on the improvement in areas where the thematic accuracy of the respective maps was insufficient to perform the final classification of each pixel. The map uses a hierarchical legend based on the United Nations Land Cover Classification System (LCCS). Compatibility with existing global land cover products is hereby taken into account, and extended by providing several cover layers. Training data have been collected from multiple sources, among others by using existing reference datasets (e.g. GOFC-GOLD) and by collecting reference data through Geo-Wiki (http://geo-wiki.org/). The product has been validated by local experts. The validation sample design has been random stratified where each sample site has been classified by visual interpratation of a high resolution imagery (Google and Bing).

GlobeLand30 is an open-access 30-m resolution global land cover (GLC) data product with 10 major classes for years 2000 and 2010. Since its first release on the 22 Sept, 2014, it has been utilized by users from about 120 countries and found applications in many Societal Benefit Areas. At the same time, the users have put forward new demands, such as providing more land cover classes, up-to-dateness and time-series. This has led to an international validation and the preparation of the updating of GlobeLand30.

The validation of 30-m GLC data products is facing several critical challenges related to the high spatial heterogeneity of land cover in the entire earth land surface, and the lacking of standardized approaches and efficient on-line tools to support collaborative practices. With the support of GEO and UN-GGIM, a technical specification of validation has been formulated and a web-based validation system has been developed. About 40 GEO-UN_GGIM members have participated in the joint validation of GlobeLand30.

The updating of 30-m GlobeLand30 is different than its original creation, and aims to produce a 2015’s version product. From the technical point of view, change detection with remote sensed imagery is the major approach and the rapidly increasing crowdsourcing information provides another valuable resource. Due to the extreme complexity of spectral heterogeneity of land cover classes, no one change detection algorithm could be universally applicable to all kinds of imagery and geographic regions. In order to support efficient updating with the consideration of the existing land cover data sets, a specific on-line system was developed to facilitate the design and execution of suitable change detection workflow with the help of a domain knowledge-based service relation model and dynamic service composition.

In the context of the Climate Change Initiative supported by ESA, the Land Cover team aims to map the whole Africa based on the entire archive of Sentinel-2 mission. To address the requirement of a high spatial resolution LC map expressed by the climate science community, the research team will generate a prototype map at 10 m resolution over the whole Africa with a consistent legend of 10 classes. This pioneer experiment faces several challenges including the big data management issues, the preprocessing chain development for improved and cloud screened Sentinel-2 surface reflectance, the precise definition of a scalable land cover typology, the land cover processing chain development and the design of reference database collection for validation.
The increase in spatial resolution requires indeed significant methodological adjustments and innovations to the processing chains developed for medium spatial resolution imagery at global scale. For the pre-processing, for example, the topography and adjacency effects have 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, the spatial consistency of surface reflectance between few images becomes a critical aspect in the production of high spatial resolution composites.

Sentinel-2 imagery requires several quality control procedures in order to be processed by large scale processing facilities. From a data management point of view, 72 TB of Sentinel-2 data have been downloaded and preprocessed into surface reflectance L3. A surface reflectance product covering one month of acquisition corresponds to 10 TB and continental surface reflectance mosaic reaches 2 TB of data.

For the LC classification, challenges are of a different nature. The decametric resolution captures the landscape elements diversity and distinct evolution through time due to slightly different seasonality and ecological gradients. Specific effort is made to ensure the consistency between this decametric map and the medium resolution global LC maps already developed within the Climate Change Initiative. While the rich literature on LC 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 LC consistently over space and time.

A review of various regional and global mapping efforts completed at different scales from 30 to 300 m resolution, a land cover typology is proposed and currently tested over different regions. The compilation and harmonization of all available land cover maps have been completed in order to support the processing strategy. A set of 10 test sites widely distributed in Africa and representative of different EO conditions and ecoregions allows benchmarking several automated methods. These benchmarking results support the processing chain development in order to optimize the performances of the each step. Finally, results over the whole Africa as well as regional results are presented, highlighting the potential of Sentinel-2 for global land cover mapping and the challenges ahead.