With an ever-changing environment the need for accurate, timely and high-resolution information on land use/land cover and its changes has increased tremendously over the past years. Until now however, regional or continental land cover maps were solely based on high resolution optical earth observation data such as Sentinel-2 or Landsat while the use of SAR data such as Sentinel-1 in the production of large area land cover maps is still in its infancy.

For this purpose, the European Space Agency (ESA) initiated the WorldCover project which has released in October 2021 a freely accessible global land cover product at 10 m resolution for 2020, based on both Sentinel-1 and Sentinel-2 data. WorldCover contains 11 classes and has been independently validated with an overall accuracy of 74.4%.

A crucial aspect for WorldCover was the involvement of several end users such as WRI, UNCCD, FAO, CIFOR & OECD active in different domains who provided primary input for all engineering aspects and followed the whole project workflow from design up to validation and uptake. Consequently, WorldCover intends to provide a substantial benefit to various user communities and expands the established global land cover base of users and the development of novel services.

In this presentation we will present you the WorldCover product, illustrate the complementary power of Sentinel-1 and 2 for global land cover mapping, discuss tradeoffs made and lessons learned in the production of the WorldCover product, zoom in on the user feedback on the product and show how the product can be improved towards the future.

Decision making at regional, national and international scales can be greatly improved with the availability of regular, consistent, and reliable maps of the land cover and how it changes over time and space. With modern improvements in data accessibility and the advancement of computational resources, operationalizing the production of these products at a large scale is now achievable. The next challenge comes with building systems which are not only just meeting today’s needs, but also have the ability to easily incorporate future anticipated improvements.

Geoscience Australia’s Digital Earth Australia (DEA) in collaboration with Aberystwyth University (Wales, UK) and Plymouth Marine Laboratory (PML) have built a globally applicable method for generating consistent, large-scale land cover maps from satellite imagery. The approach builds on the Earth Observation Data for Ecosystem Monitoring (EODESM) system (Lucas and Mitchell, 2017), which constructs and describes land cover classifications based on environmental descriptors derived from Earth observation (EO) data. The system’s land cover structure is based on the globally applicable United Nations Food and Agriculture Organisation Land Cover Classification System (UN FAO LCCS) taxonomy). This new land cover classification system has been built to be adaptable and modular, allowing for its application to a range of different landscapes, at different spatial and temporal resolutions and with a variety of data sources. One major feature of this new system is the inclusion of native support for the Open Data Cube environment. The system is open source, meaning that the algorithms and code are openly available to researchers anywhere in the world. The structure of the code enables researchers to utilise their own, regionally specific methods to build land cover products tailored to their own needs, while still adhering to the overall UN FAO LCCS framework. It can be run using data from a range of sensors including multispectral optical, radar, LIDAR, as well as custom georeferenced datasets.

This method is currently being used operationally by both DEA and Aberystwyth University to create national land cover products. For Australia, DEA has generated DEA Land Cover, a high resolution (25 m) continental, annual land cover map for each year from 1988 to 2020 by utilising over 30 years of Landsat sensor data. This data is being utilised in Australia’s environmental economic accounting, and is providing valuable insights to researcher and decision-makers. Similarly, Aberystwyth University has worked with DEA and Welsh Government through the Living Wales project to generate national land cover maps for Wales over for four years (https://wales.livingearth.online/) using multiple sensors including Sentinel 1 and Sentinel 2 and is currently extending the time-series back to the mid 1980s using Landsat sensor data (Lucas et al., 2018). In addition, several research bodies across the globe have begun a community of practice to share ideas, algorithms and support each other in implementing this land cover methodology in their own Open Data Cube environments.

References:

Lucas, R.M.; Mitchell, A. Integrated Land Cover and Change Classifications. In The Roles of Remote Sensing in Nature Conservation: A Practical Guide and Case Studies; Díaz-Delgado, Lucas, R., Hurford, C., Eds.; Springer: Cham, Switzerland, 2017; pp. 295–308.

Lucas, R., Bunting, P., Horton, C., 2018. Living WALES — National Level Mapping and Monitoring Though Earth Observations, Ground Data and Models. IGARSS 2018 – 2018 IEEE International Geoscience and Remote Sensing Symposium. Presented at the IGARSS 2018 – 2018 IEEE International Geoscience and Remote Sensing Symposium, pp. 6608–6610. https://doi.org/10.1109/IGARSS.2018.8519452

Owers, CJ, Lucas, RM, Clewley, D, Planque, C, Punalekar, S, Tissott, B, Chua, SMT, Bunting, P, Mueller, N & Metternicht, G 2021, 'Living Earth: Implementing national standardised land cover classification systems for Earth Observation in support of sustainable development', Big Earth Data. https://doi.org/10.1080/20964471.2021.1948179

High quality training data is crucially important for any land cover/ land use mapping. There are different sources of training data available, including on-ground observations and visually interpreted very high resolution images. Additionally, while creating new maps, already existing land cover/land use maps are being used to extract more training samples. Here, we have explored impact of using various sources of training data on quality of the World Cover land cover map at 10m resolution. The key input used for developing the 2020 World Cover map at 10m resolution was the 2015 CGLOPS data set collected through the Geo-Wiki engagement platform. A specific branch of Geo-Wiki (http://geo-wiki.org/) was developed for collecting reference data at the required resolution and grid (PROBA-V UTM 100 m pixels). It showed the pixels to be interpreted on top of Google Earth and Microsoft Bing imagery, where each pixel was further subdivided into 100 sub-pixels of 10 m x 10 m each, in line with the Sentinel 2 grid. Using visual interpretation of the underlying very high resolution imagery, experts (a group of people trained by International Institute for Applied Systems Analysis staff) interpreted each sub-pixel based on the land cover type visible, which includes trees, shrubs, grassland, water objects, arable land, burnt areas, etc. This information could then be translated into different legends using the UN LCCS (United Nations Land Cover Classification System) as a basis. While this data set was a very good input for mapping land cover at 100m resolution, it was a bit noisy at 10m pixel level due to the shifts of underlying images used for visual interpretation as well as due to land cover changes happened between 2015 and 2020. This was critical for mapping highly heterogeneous areas, such as savannas, with mixed woody vegetation, grasslands, and bare land. We’ve worked on optimizing the training data set and we would like to present our approach to improving the training data quality at 10m resolution for mapping heterogeneous areas, and also present our iteration approach for improving maps quality by using spatial accuracy. Finally, we would like to discuss the requirements for training data collection with a particular focus on highly heterogeneous landscapes.

Semantic segmentation with convolutional neural networks (CNN) has proven an effective method for accurate land use/land cover classification with high-resolution satellite imagery in recent years (Carranza-Garcia et al., 2019, Scott et al., 2017). However, CNNs need full coverage of training labels over input features. Generating such training label data is time-consuming and expensive, which has limited this technique’s applicability for continental-scale mapping efforts.

The LUISA Base Map 2018 (Pigaiani and Batista E Silva, 2021) provides a Europe-covering land use/land cover (LULC) database for the year 2018, distinguishing 46 LULC classes in a hierarchical legend at 50m resolution. The LUISA dataset is the result of combining several human-annotated LULC datasets of Europe for 2018, and therefore presents a unique opportunity to train a CNN to classify LULC for the entirety of continental Europe, potentially in other years than the source year of its training data.

We explored the potential of the LUISA Base Map for deep learning-based LULC classification at multiple levels of spatial and thematic resolution. To do this, we trained UNET-based CNNs on 30m Landsat, 10m Sentinel-2, and 3m Planet satellite imagery of multiple regions in Europe, recorded in 2018. We did this for each of the four levels in the LUISA legend hierarchy (5, 14, 41, and 46 classes), resulting in 12 models. After training, each model was used to classify LULC in a left-out set of European sample locations in 2018 and 2015.

Our accuracy assessment consists of a validation on LUISA data and satellite imagery from 2018 for a number of left-out regions, as well as a cross-reference to all observations from the Land use and land cover survey (LUCAS) of both 2018 and 2015. We also calculated the agreement between the LUCAS observations of 2018 and the LUISA Base Map itself to provide an objective estimate of the maximum attainable accuracy of this validation method.

Our objective was to train a model that can reproduce the LUISA basemap on left-out data, and achieve a similar score on LUCAS validation points from 2015 and 2018. Initial results suggest that the unprecedented spatial and thematic resolution of the LUISA basemap can be reproduced for other years without requiring the costly efforts of creating and combining its component datasets for each target year.

References
Carranza-García, M.; García-Gutiérrez, J.; Riquelme, J.C. A Framework for Evaluating Land Use and Land Cover Classification Using Convolutional Neural Networks. Remote Sens. 2019, 11, 274. https://doi.org/10.3390/rs11030274

Pigaiani, C. and Batista E Silva, F., The LUISA Base Map 2018, EUR 30663 EN, Publications Office of the European Union, Luxembourg, 2021, ISBN 978-92-76-34207-6 (online), doi:10.2760/503006 (online), JRC124621.

Scott, G. J. and England, M. R. and Starms, W. A. and Marcum, R. A. and Davis, C. H.,"Training Deep Convolutional Neural Networks for Land–Cover Classification of High-Resolution Imagery, in IEEE Geoscience and Remote Sensing Letters, vol. 14, no. 4, pp. 549-553, April 2017, doi: 10.1109/LGRS.2017.2657778.

About half of Germany's total land area is used for agricultural production, and about one third is forested. Gathering detailed information on the land cover of these landscapes is of great importance for their ecological and economic valuation. This could improve, for example, the estimation of ecosystem services such as pollination, the quantification of nitrate and nutrient inputs into water bodies and the determination of forest conditions in times of climate change. Forests in particular play a central role in environmental impact assessments for large infrastructure projects. However, spatially explicit information on on tree species and the conservation value of forest types is missing.

High temporal and spatial resolution satellite data of the Copernicus mission allow continuous monitoring of plant dynamics on the land surface. Optical remote sensing is suitable for capturing the spectral characteristics of plant species. Using time series analysis, plant species can be distinguished based on their different phenology. However, approaches are needed to deal with cloud contaminated data and to take into account regional biogeographical conditions when classifying plant species at national level.

In this context, we present the classification approach APiC that has been used to produce maps of agricultural crops and tree species. APiC is a highly automated, data-driven machine learning approach to land cover classification that works dynamically at pixel level. The thematic dimension (definition of the land cover classes) of the training data is determined solely by the algorithm, as is its temporal dimension (time periods for which sufficient cloud-free pixel observations are available). For land cover prediction, a large number of classification models are computed to take the individual cloud cover at pixel level into account. In this way, model performance can be specified for each pixel to provide a more detailed insight into the accuracy of the classification. It follows that with APiC neither cloud-free image mosaics are created nor are artificial reflectance values generated for gap filling.

First, APiC was used to classify crops across Germany based on Sentinel-2 data from 2016. LPIS data served as reference data for training the machine learning algorithm ‘random forest’ and for validating the results. The different growing conditions in Germany were taken into account by carrying out the classification independently in six landscape regions. With an overall accuracy of 88%, a total of 19 crop types were classified, namely winter wheat, spelt, winter rye, winter barley, spring wheat, spring barley, spring oat, maize, legumes, rapeseed, leeks, potatoes, sugar beets, strawberries, stone fruits, vines, hops, asparagus and grassland (www.ufz.de/land-cover-classification). Agricultural crops for subsequent years (2017-2020) are currently being mapped using the same routine. Based on time series analyses, the effect of crop rotation and landscape configuration on pollination performance of bees can be better estimated.

Secondly, the main tree species in Germany were classified using forest inventory data and Sentinel-2 data from 2015-2017. Pine, larch, spruce, Douglas fir, oak, beech, hornbeam, alder and willow were classified with an overall accuracy of 76,6 % across three landscape regions. Due to the heterogeneity of forests and the design of the reference data/inventory data collection, the classification of tree species turned out to be more challenging than that of crop plants. We used the tree species classification map together with information on the potential natural vegetation of Germany, the Red List status of forest types and the canopy height (derived from high-resolution LiDAR data) to determine a conservation value of forested areas. Provided by the Federal Agency for Nature Conservation, the potential natural vegetation represents the vegetation of Germany as it would exist under current climate and soil conditions without human influence. This information was essential to assess the tree species classification map from a nature conservation perspective. Our approach does not take into account other important nature conservation aspects such as deadwood occurrence and undergrowth of the forest. However, through the use of remote sensing, a preliminary conservation assessment of forests could be made at the national level.