Land use change in the tropics is changing our climate through the release of greenhouse gases, changes in surface roughness, impacts on hydrological cycling, and changes in the surface albedo. However, detecting changes in land cover, let alone land use, in the tropics has become more difficult with the rise of small-scale deforestation and the creation of complex landscapes where intact forests, recovering forests, tree crops, agroforestry, pasture, and crops all exist in small patches in close proximity.
Some countries with large areas of tropical forests are impacted heavily by this challenge: the average size of agricultural holdings has declined to less than 2 ha in Indonesia, Vietnam, and India (Lowder, Skoet, & Raney, 2016); small farms also dominate the Demographic Republic of Congo with 1.3 to 2 ha per household (Otchia, 2014); while in Brazil, where the average farm size is much larger, the number of small clearings (< 1 ha) has nonetheless increased by 34% while large clearings with more than 50 ha has decreased by 46% between 2001-2007 and 2008-2014 (Kalamandeen et al., 2018).
Our research takes on the challenging task of detecting small changes using free satellite data with high accuracy. Our method utilises the publically available Sentinel-2 data (10 m resolution) , and a free analysis platform, Google Earth Engine (Nomura & Mitchard, 2018). Sentinel-2 has great potential for mapping vegetation types in complex landscapes as it is a multispectral instrument with 13 bands, some of which (for example, the ‘red edge’ bands) cover very narrow portions of the spectrum, less than 20 nm wide, giving it some of the advantages in classification that were traditionally only available to a true hyperspectral sensor. It’s temporal resolution (covering this area every few days) and spatial resolution (10/20/60 m resolution, depending on the bands) is also considerably higher than other available free datasets.
Our first experiment was conducted in Southern Myanmar, covering about 13,000 ha containing oil palm (Elaeis guineensis), rubber (Hevea brasiliensis), betel nut (Areca catechu) plantations and forests. Using an extensive training dataset and a Random Forest classifier, we achieved 95-96% accuracy rates in classifying the study sites (Nomura & Mitchard, 2018). The resulting land use change showed a fast-paced and an intensive competition among different crops, providing crucial information on the trajectory of land use change and drivers of deforestation.
We explore the potential of much wider application of this method by testing in other regions with different crop types as well as scaling up. Our study so far provided evidence that the high spatial resolution of Sentinel-2, at 10 to 20 m, should be sufficient to classify even very small plantations, without needing to resort to object-based classification of higher resolution data, making it the ideal tool for mapping fragmented landscapes.

References
Kalamandeen, M., Gloor, E., Mitchard, E., Quincey, D., Ziv, G., Spracklen, D., … Galbraith, D. (2018). Pervasive Rise of Small-scale Deforestation in Amazonia. Scientific Reports, 8(1), 1600. https://doi.org/10.1038/s41598-018-19358-2
Lowder, S. K., Skoet, J., & Raney, T. (2016). The Number, Size, and Distribution of Farms, Smallholder Farms, and Family Farms Worldwide. World Development, 87, 16–29. https://doi.org/10.1016/j.worlddev.2015.10.041
Nomura, K., & Mitchard, E. T. A. (2018). More Than Meets the Eye: Using Sentinel-2 to Map Small Plantations in Complex Forest Landscapes. Remote Sensing, 10(11), 1693. https://doi.org/10.3390/rs10111693
Otchia, C. S. (2014). Agricultural Modernization, Structural Change and Pro-poor Growth: Policy Options for the Democratic Republic of Congo. Journal of Economic Structures, 3(1), 8. https://doi.org/10.1186/s40008-014-0008-x

Area A - Board 13

Forest managers have had detailed inventory of forests since the 19th century in Hungary. The inventory is made on forest compartment level, which covers 2-3 hectares of homogenous growth conditions and tree species. The inventory also stores attributes about the forest parameters which are generally collected with field sampling methods. Therefore, there is a limited opportunity to get spatial information about species distribution on sub-compartment level in this system. The renewal period of the forest inventory is 10 years now. During this long period minor and major changes could take place inside the compartments, like species mixture ratio change after silvicultural operations or occurrence of fast-growing invasive tree species. Such cases are difficult to handle in the current forest inventory system. However, in most cases the 10-year renewal period is satisfactory for the management, it does not compete with the fast information flow standard of the 21st Century.
Optical satellite remote sensing can derive important parameters for the forest canopy, including the tree species. Despite of the limitations with multi layered forest, this tool can provide useful information on high spatial and temporal resolution about the forest types. Previous studies derived forest type maps based on a single date by using only the spectral features of the satellite images. In the era of open Earth observation data, multiple studies have made use of spectral-temporal features of time series data which significantly improved the mapping accuracy. By extending the spectral-temporal features with spatial features, even more detail could be extracted from the data.
In this study the mapping was done in the Börzsöny mountains, Hungary on 60000 hectares to improve the forest inventory with spatial extension of forest types. The forests were classified into 11 major forest type. The imagery of Sentinel-2 satellite mission was used from the vegetation period of 2017. A cloud and gap free time series was constructed for the analyse. The cloud masking was done with a Kalman filter based multi temporal method, while the gap filling was done based on the reflectance change of the neighboring valid pixels. The multi temporal image stack was segmented into homogenous parts, where each represented a forest type patch. To make use of the spatial features of the homogeneous regions a deep learning type convolutional neural network (CNN) was used for classification.
The created forest type map was validated against field data and compared with a map, which was made from only spectral-temporal features of the imagery. The used method slightly over performed the previous methods for forest type mapping. However, it has a significantly higher computational cost. The map provides reliable information for young and mature forests. Over mature forests the imagery has more spatial (textural) features, where the CNN classifier could perform better. The proposed method for preprocessing and classifying the time series of Sentinel-2 data can be reused on similar sized study areas to improve forest inventory, where enough training site could be established for image classification.

Area A - Board 14

The Reduced Emissions from Deforestation and Degradation (REDD+) framework has triggered the need of accurate maps of tree cover and tree cover change. While global initiatives of forest change mapping exist (Hansen et al., 2013), national institutions often look for methods upon which they can have a complete control and understanding. The current quality, quantity and free availability of remote sensing data is helping to achieve that goal. In Africa however, alternative methods have still to be found to overcome limited technical capacities, computing power and data availability (Badjana et al., 2017). In the case of East Africa, those limitations are adding up to the challenge of mapping or assessing tree cover in the dry domain (Bastin et al., 2017; Bouvet et al., 2018; McNicol et al 2018). This is mainly due to the tree cover density, ranging from dense forest with tall trees, down to sparsely distributed forests with trees not higher than a few meters. In addition, many tree species are deciduous, with a phenological activity related to the precipitations, which requires a monitoring throughout the whole season.

The launch and free data policies of Sentinel-1 (S1) and Sentinel-2 (S2) platforms could change the paradigm. The combination of both sources of data opens new avenues for thematic applications such as demonstrated by a previous study for detecting and monitoring forest fires in the Congo Basin (Verhegghen et al, 2016). The data have an increase spatial and temporal resolution that has to be investigated for overcoming the challenge linked to seasonality and sparse tree cover of dry forests monitoring.

In the framework of the ReCaREDD project, the JRC is conducting capacity building activities with East African countries. The objective is to develop a toolkit for (i) facilitating the access to pre-processed S1 and S2 data, (ii) producing meaningful layers for dry forest mapping and monitoring and (iii) allowing the incorporation of available national or local training datasets. A study case has been carried out over Tanzania, where the potential of S1 and S2 data to better capture the tree and shrub cover is assessed. The aim is to produce a map for the year 2017 as a benchmark for future or past change detections.

The first step is the production of a dataset from S1 and S2 amongst which a choice can be done for the classification. Vegetation indices and spectral composites are produced over the whole country from the S2 5 days time series for the year 2017. A cloud mask is applied to single image of Level 1C. Then, monthly composites of three different types are generated. First, monthly composites of NDVI are obtained from the maximum NDVI value. Second, shadow index monthly composites are calculated as the percentile 90 of the shadow index values in a month. Third, monthly composites of the SWIR band (B11) are computed as the median value in a month. In addition, three months composites are created taking the median reflectance values for the period of January-March, April-June, July-September and October-December. Regarding S1, the full time series for the year 2017 is processed into monthly composites for each polarization, VV and VH. The S1 data is resampled to a 10-meter resolution.

For the classification, 23 layers derived from the above dataset are selected. For S2, the 12 monthly NDVI, the annual cumulative NDVI, the annual minimum shadow index value, the annual median SWIR value, the July-September and October-December composite for the Red (B4), NIR (B8) and SWIR (B12). For S1, the annual VV minimum and VH minimum values.
These input layers are used into two different thematic mapping approaches: the first approach classifies the pixels into tree or shrub cover according to their intrinsic spectral values (radar and vegetation indices), while the second approach uses a contextual spectral approach to classify pixels into forest types considering their neighbouring pixels.

We collected training data for sample plots (with 2 sizes: one pixel and 0.25 ha) via the Collect Earth interface of Google Earth (Bey et al, 2016). Sample plots were pre-selected with a stratified random sampling scheme and only plots where very high-resolution imagery was available close to year 2017 have been interpreted. A first interpretation was done for the central pixel (10m x 10m or 0.01 ha) based on the woody cover proportion (tree or shrub) within the pixel. A minimum of 5 m crown diameter was set for interpreting a tree versus a shrub. Second, a window of 0.25 ha size around the central pixel (corresponding to 5 x 5 pixels) was interpreted in different types of forest or shrubland cover classes, from open to closed canopy cover. A specific map was created for each legend by applying the Random Forest algorithm on the combined S1 and S2 derived composite layers.

The random forest algorithm (with 100 trees) was used for the pixel-based classification approach to produce a first map of tree and shrub cover (figure 1). Other deep learning algorithms able to directly take into account the neighbourhood will be tested for the forest type classification. The maps and method will be validated with the National Carbon Monitoring Center of Tanzania. With this approach, the expert users can focus on the choice of most relevant layers and the preparation of the training dataset. The best technical options for running the classification remotely and sharing the results with the National Carbon Monitoring Center will be explored.

References

Badjana, H. M. et al. (2017). Mapping and estimating land change between 2001 and 2013 in a heterogeneous landscape in West Africa: Loss of forestlands and capacity building opportunities. International Journal of Applied Earth Observation and Geoinformation, 63, 15–23.

Bastin, J.-F., et al. (2017). The extent of forest in dryland biomes. Science, 356, 635–638.

Bey A. et al. (2016) Collect Earth: Land Use and Land Cover Assessment through Augmented Visual Interpretation Remote Sens. 8, 807.

Bouvet, A. et al. (2018). An above-ground biomass map of African savannahs and woodlands at 25 m resolution derived from ALOS PALSAR. Remote Sensing of Environment, 206, 156–173.

Hansen, M. C. et al. (2013). High-Resolution Global Maps of 21st-Century Forest Cover Change. Science, 342, 850–853.
McNicol I.M., Ryan C.M., Mitchard E.T.A. (2018) Carbon losses from deforestation and widespread degradation offset by extensive growth in African woodlands. Nature Communications 9, 3045.

Verhegghen A. et al (2016) The Potential of Sentinel Satellites for Burnt Area Mapping and Monitoring in the Congo Basin Forests. Remote Sens. 8, 986

Area A - Board 15

The planning of wind energy in remote areas creates a need for easily accessible information of the land surface. Wind turbines provide more than 5% of the annual global electricity consumption. The wind global power capacity by the end of 2017 was 539 GW. In 2017, 52.6 GW new wind power capacity was installed. Installation is expected to increase over the coming years.

Prior to each new wind farm project, a calculation of the wind resource is made. This is based on inputs of terrain height, land surface roughness and the observed or modelled wind distribution. The project InnoWind aims to increase the reliability and increase the speed of these calculations using novel Earth Observation data. The Copernicus archive of land cover types and vegetation, forest type and density are used in combination with digital elevation models to derive the surface drag force. The surface drag is the combination of roughness and form drag exerted by the surface to the wind flow. This is a function of the tree height, spacing, density and other features in the landscape. Forests have the strongest impact on the wind flow and represent 75% of new land based wind energy developments. Thus mapping the forest is of key importance as input to the flow modelling for new wind energy installation on land.

The satellite-based input is compared with aerial lidar scan (cloud point data analysis) for verification at the site Østerild in Denmark. This site has open terrain and forest. The novel satellite-based data layers are used as new input to a flow model. Flow model results on wind speed at several heights are compared to observations of winds at the tall meteorological mast at Østerild.

The new approach in flow modelling for wind energy using Copernicus data is expected to eliminate current indirect and subjective assessment of surface properties and reduce the uncertainty on the wind resource. Current practice is to perform field visits by a wind engineering expert that estimates the roughness of the nearby terrain. The new approach is expected to lead to more accurate, automatic and consistent assessments at different sites worldwide.

The InnoWind project has research focus on forested areas at high resolution (around 10 m) and land cover information at the mesoscale (around 1 km). There is a linking from mesoscale flow modelling to microscale flow modelling. The linking (modelling chain) of models for resource assessment is being adapted to utilize novel satellite-based input layers. This part is not trivial. The presentation focusses on the different Earth Observation data layers tested at various sites in both microscale and mesoscale, to demonstrate the methodology in different types of terrain.

Area A - Board 16

Living Wales is a unique and novel world-first concept that aims to capture the state and dynamics of Wales’s landscape, in near real time, historically and into the future (over the long term) through integration of Earth observation data, supportive ground measurements and process models. Living Wales is working with and supporting existing initiatives to collate, correctly process, analyze and make publicly available the vast archive of optical and radar satellite observations over Wales (at least since the 1980s) and derived spatial outputs that directly address key issues relevant to the Welsh economy, policy and populations.

The research component is focusing on:

 The advancement and implementation of algorithms that target the interrogation of dense time series’ of Earth observation data (including from the European Copernicus Sentinels and NASA’s Landsat) data to routinely inform on, for example, agricultural, forestry, water and urban use and change.
 The provision of quantitative biophysical data (e.g., on vegetation productivity, structure and carbon dynamics, hydrology and flood sequences) for assessing trends attributed to climate or other environmental change.
 The routine production of detailed and high-resolution land cover and habitat classifications across Wales for multiple points in time based on relevant taxonomies.
 Identification of options for optimizing land use, conserving and restoring landscapes and ecosystems, and monitoring progress.

Living Wales seeks to utilize and capitalize on existing and new geographical information (e.g., on agricultural use), models (e.g., carbon, hydrology) and scientific knowledge (e.g., habitat links to biodiversity) and is contributing to the coordinated collection and collation of field, drone and airborne datasets acquired over Wales. The Living Wales project is providing openly available software and capability that will also benefit countries internationally and is working closely with organisations in Europe and Australia to maximize benefit and cooperation. Living Wales, through its use of Copernicus data and products is anticipated to lead to significant benefits in terms of society, economics, policy and environment in Wales but also internationally.

Area A - Board 17

Olive production in the EU member states increased by 8% within the last decade. As a result, the complex agricultural landscapes in the Mediterranean region are increasingly characterized by olive groves of heterogeneous parcel size and structure, which are located in flat terrain, or occur terraced across steep topographies and large elevation gradients. The structural attributes of olive groves further relate to their productivity. For instance, an optimized tree spacing is desired, as per area canopy cover determines farm productivity. Olive trees are evergreen and thus do not underlie a strong seasonality. However, layers of herbaceous vegetation commonly prevail in the understorey. Precipitation and productivity-enhancing management practices such as irrigation or ploughing affect the presence of understorey vegetation, causing additional local heterogeneity. Mapping of olive tree cover thus requires methods that enable the separation of ephemeral variations in herbaceous understory vegetation from perennial vegetation signal.
The structural and land surface phenological traits of olive groves pose specific requirements to remote-sensing based characterization, which relate to spatial resolution and revisit frequency. The 10 to 20 m spatial resolution and 5-day revisit frequency of the current Sentinel-2A and 2B satellite constellation offer new opportunities for characterizing land surface phenology (LSP) at high spatial detail and are thus suited for improving current characterization of complex managed landscapes. This study aims at assessing the value of dense Sentinel-2A and 2B time series for identifying olive groves and characterizing tree cover on heterogeneous olive groves in a Mediterranean setting. The Greek island of Crete, which is among the major European olive producing regions, serves as an ideal case to test these methods, due to the strong variability in topography, precipitation, and management regimes, coupled with low cloudiness and thus high rates of clear sky observations. Recent market dynamics led to rapid increases in the extent of Cretan olive groves, which are characterized by an outstanding heterogeneity and below-average farm sizes compared to other important European olive producing regions.
We produced topographically normalized and atmospherically corrected Sentinel-2A and 2B time series for 2017 and 2018 at 10m spatial resolution. We extracted dense time series of vegetation indices, and derived LSP metrics that decompose the time series into permanent, and seasonal LSP components. While the first mostly relate to the standing tree biomass, the latter relate to the strong seasonality of the understorey. Olive grove reference points were compiled from ancillary EUROSTAT datasets, and very high resolution imagery. We amended the reference samples with information on tree cover fractions at the pixel- and parcel-level. The integrated Sentinel 2A/B time series were used for classifying olive groves at 10 m spatial resolution. Next, the perennial component of the LSP was related to olive tree cover fractions using machine learning regression techniques, resulting in a spatially explicit representation of tree cover for Cretan olive groves.
This study explores the applicability of employing intra-annual time series and LSP metrics for mapping olive tree plantations and their tree cover for the case of Crete, Greece. Given the current constellation of Sentinel-2A and 2B, we discuss the potential of the LSP time series concerning their applicability in a Circum-Mediterranean setting. This could for the first time allow for a synoptic characterization of Mediterranean olive production system at 10m spatial resolution.

Area A - Board 18

Fruit-trees farming is one of the most important agricultural activity from the Romanian Subcarpathian area, having economic significance as well as social and environmental benefits for both rural and urban communities. In this respect, accurate mapping and monitoring of the fruit-trees plantations are required as a prior condition for developing efficient and sustainable agricultural management policies.
The large archive of free optical remote sensing data at different spatial and spectral resolutions as well as the newly artificial intelligence algorithms creates new premises for the delineation of the fruit-trees plantations at local, regional or global scale. But doing this activity is challenging because of the spatial and spectral characteristics of the orchards. Firstly, in the Romanian Subcarpathian area, the spatial pattern of the orchards (composed mainly by plum, apple and pear trees) is very fragmented and heterogeneous, alternating from well-organised large farming systems to irregular smallholder’s patches with a randomly distribution of the trees. Secondly, the spectral response of these orchards, is quite similar to other woody vegetation like deciduous forests, shrubs, dispersed groups of woody vegetation or hedges, generating confusions in the thematic classifications. From another perspective, the phonological characteristics of orchards analysed on a set of optical images regularly acquired within a temporal window of interest represent a useful approach to discriminate them from different types of vegetation.
The purpose of our paper is to test several machine learning (ML) algorithms, Suport Vector Machine and Random Forest, on different satellite time series images in order to map the interface between orchards and other land cover types in an effective way, within a selected area from Romanian Subcarpathians. Our work focuses on the comparison and evaluation of the potential offered by ML algorithms in respect with hyperspectral and multispectral time-series data at medium spatial resolution in the field of orchards crop classification.
The approach integrates intra-annual time series hyperspectral - Hyperion (30 m) and multispectral data - Landsat 8 OLI (30 m) and Sentinel-2 MSI (20 m) and it is organised between multiple levels: 1) the calibration and the processing of the selected images for the analyses in order to produce spectral reflectance information; 2) the creation of the intra-annual time series sets for each type of satellite data, covering the temporal window from April to October, necessary to capture the temporal and spectral variability of the orchards within the vegetation season; 3) the definition of the training data for both feature classes: orchards and non-orchards and for each set of time series images; 4) the testing of the ML algorithms on different configurations of the time series datasets, referring to the use of the full spectral resolution of the time series or just parts of them; 5) the validation of the results using independent data; 6) the comparative evaluation of the results.
The highest classification accuracy of the orchards is recorded with Hyperion hyperspectral images, followed by Sentinel-2 and Landsat data, and it is explained by the high number of bands in NIR and SWIR. Both ML algorithms performed very good, with a small increase in accuracy for Random Forest.

Area A . Board 19

Earth Observation data has been extensively used in country-wide national mapping and monitoring activities in Hungary since the beginning of the 1990’s. BFKH FTFF and its predecessor, FÖMI (Institute of Geodesy, Cartography and Remote Sensing) has been active in this area ever since then. Moreover, a number of in situ geospatial data sources (topography, orthoimagery, cadastral maps, Land Parcel Identification System - LPIS) are also maintained by BFKH FTFF in its function of National Mapping and Cadastral Authority, facilitating the operational use of EO data through data fusion, machine learning and GIS analysis in various application areas.

The presentation focuses on three different application scenarios, with each providing specific land cover products for well-defined user groups.

The first scenario fits within the context of agricultural mapping and monitoring for the control of agricultural subsidies. The main objective is the delineation of permanent grasslands (i.e. confirmed presence through a period of 5 years) based on EO imagery. The established method is based on machine learning with a combined use of optical and radar time series, enriched with spectral indices and polarimetric descriptors integrated over selected time periods. Reference data used for learning and validation are integrated from various data sources including farmers’ claims for subsidies, Copernicus and national HR land cover data. The outcome is used for the targeted, optimized updating of the permanent grassland layer of the LPIS. Besides this, a highly reliable country-wide crop map is produced, that is currently under evaluation for being included in agricultural monitoring activities.

The second scenario is related to the mapping and assessment of ecosystem services (MAES). The 2nd target of the EU Biodiversity Strategy requires the member states to assess and map the most important ecosystem services (ES) and integrate these results into policy decisions. Led by the Ministry of Agriculture, a project entitled „Strategic Investigations on the long-term preservation and development of natural heritage of Community Importance and on the implementation of the EU Biodiversity Strategy 2020 objective”, co-financed by the EU, has started in Hungary in 2016. The main objective is to build up spatial databases of ecosystems and ES in Hungary, and assess those using biophysical, economic and social indicators. The first version of the ecosystem base map is now available and is undergoing expert validation. It was created by integrating a large variety of in-situ datasets along with newly created maps based on machine learning on EO data. After validation, it will constitute the basis for the assessment of ecosystem conditions and services. It is also evident that regular updates will be necessary to monitor relevant changes and fulfil policy objectives.

The third scenario focuses on the creation and regular update of a general-purpose national high resolution land cover map (NHRL). Currently in experimental phase, it is based on the principle of Copernicus High-Resolution Layers (HRLs), extended and refined in a national context. It provides a country-wide, wall-to-wall general thematic land cover raster with a resolution of 10 m. Mapping is done by machine learning methods based on seasonal time series of Sentinel-1 and Sentinel-2 imagery and an integrated set of reference data coming from ancillary in-situ data bases. Validation of the first results is currently being undertaken, with the perspective of extensive evaluation by different user groups. Subject to a positive evaluation, the provision of regularly updated NHRL enabling monitoring can be initiated.

Area A - Board 20

Detailed thematic descriptions of the Earth’s terrestrial land cover (LC) are essential for natural resource management and ecosystem change assessment. The increased volume of freely available Earth observation data from multiple sensors together with powerful machine learning algorithms and advances in computing power have opened up new possibilities for repeated production of regional to global LC and LC change maps at 20-30 m resolution. Temporal metrics from seasonal time-series have been shown to enable more thematically detailed LC assessments and to improve the discrimination of classes that are spectrally similar in broadband multispectral data.

Most ecosystems are inherently diverse and composed of a heterogeneous LC mosaic with gradual spatial-temporal transitions between LC types. Assessments adopting discrete classifiers aim at mapping the dominant LC present in a pixel, whilst not accounting for gradual transitions between natural surfaces. In contrast, subpixel mapping, i.e., unmixing, helps estimating fractions of various LC types present in a pixel, and thus allows a more realistic representation of the heterogeneous composition of natural surfaces while preserving the information on the dominant LC.

In this study, we present a flexible analysis framework for producing large area vegetation type fraction maps across a range of California’s ecoregions. These include natural grasslands, shrublands, as well as broadleaf and coniferous woodlands. We make use of a dense seasonal time-series from intergrated Landsat/Sentinel-2 data and machine learning regression-based unmixing. The regression model training relies on a seasonal endmember spectral library and synthetic mixtures generated thereof. Model generalization for large area fraction mapping is made possible by including endmembers from a variety of different ecoregions.

Our results show that this analysis framework leads to accurate fraction maps depicting the heterogeneous composition of vegetation communities and spatial transitions between vegetation types and ecoregions. The hierarchically structured class labels attached to the library entail a flexible and user-oriented steering of the regression models, allowing the production of broader vegetation fraction maps, e.g., woody vs. non-woody vegetation, to more detailed fraction maps presenting different vegetation life forms, e.g., coniferous trees, broadleaf trees, shrubs, or natural grassland. For more detailed vegetation type descriptions in particular we found great value in seasonal endmembers. Findings from this study illustrate the step from discrete land cover classification to quantitative land cover fraction maps and this way contribute to possible pathways for generating future global LC products.

Area A - Board 21

High-resolution land cover database are necessary for land use planning and natural resources management. Despite the increasing performances of machine learning methods to extract valuable information from very high-resolution images, combining several sources of information has a greater potential than the optimization of the classifier. Indeed, some land cover classes are better discriminated when information about their phenology or specific spectral bands are used. On the other hand, small landscape structure and objects are only observable with images at submetric spatial resolution. While a dataset combining a high spatial, temporal and spectral resolution is possible in theory, this is in practice not available over large areas.
Our study area is the Walloon region, covering about 16.5 thousand square kilometers. A single date mosaic of aerial images at 25 cm was available for the year 2016. This dataset provided spectral information in red, green, blue and near-infrared together with height information computed by photogrammetry. A study of users’ requirements identified nine classes of interest including impervious surfaces, buildings, water, broadleaved trees, coniferous trees, grasslands, herbaceous crop, shrub and bare soils. Many of those classes can be distinguished solely by photointerpretation of the single date aerial images, but in some cases, this is not possible with sufficient accuracy. We therefore added information of higher spectral and temporal resolution by using Sentinel-1 and 2 images at 10 (or 20) meter resolution. Those images mainly contributed to the discrimination between forest types and between grassland and herbaceous crops, however their resolution was too coarse to efficiently map small landscape elements like isolated buildings, roads, hedgerows or small water bodies.
Ortho-images were processed with object-based and pixel-based supervised machine learning methods considering the height information. The training was performed based on existing (outdated) database and the classification parameters were locally optimized. On the other hand, Sentinel-2 time series were combined with i) the proportion of high feature in each 10 m resolution pixel with and ii) multitemporal filtering of gamm0 processed Sentinel-1 data for two key time periods with respect to vegetation phenology in the study area. At this stage, two wall-to-wall land cover maps have been produced for the whole study area: one at 25 cm and the second one at 10 meters.
Using these different classification results and their confidence maps, new maps will be derived by context-dependent Bayesian data fusion, knowledge-based approach and Dempster-Schafer theory. Validation of preliminary results with an independent dataset attest for a high overall accuracy of above 90% with most methods. A new confidence map will also be computed to highlight the areas of uncertainty in the final maps and hence identify further areas of improvement.

Area A - Board 22

Land cover (LC) and land use classification is one of the core applications of satellite images. It provides information necessary for monitoring a surface of the Earth in order to assess a state of the environment and detect environmental changes caused by human activity, natural disasters and climate changes. Additionally, information on LC helps, among other things, to determine indicators for assessment of sustainable development. However, large amount of satellite data currently available can be effectively analysed and process only by the use of automatic methods. To address this issue, a project group consisting of Space Research Centre of Polish Academy of Science (CBK PAN), Industrieanlagen-Betriebsgesellschaft mbH (IABG), EOXPLORE UG and Friedrich-Schiller-Universität Jena carried out an ESA funded project the S2GLC (Sentinel-2 Global Land Cover). The goal of the project was to develop LC classification approach applicable to global mapping with the use of Sentinel-2 imagery. The prepared methodology was fully automatic and implemented in prototype software.

The developed pixel-based classification approach utilizes the Random Forest classifier. It is characterized by two basic features: (i) existing global land cover (GLC) databases are used for training, (ii) the final result of classification is calculated by aggregation of single classifications of multi-temporal datasets. The aggregation is performed pixel by pixel with the use of information on the number of occurrences of a certain LC class within a time series and the prediction score of particular classes resulting from the Random Forest. The legend of classified LC types was constructed after reviewing existing GLC databases. The S2GLC approach has been tested thoroughly by classification of the whole areas of Germany and Italy and large test polygons located in China, Namibia and Columbia. The testing area exceeded over 1 300 000 km2 in total.

In spring of 2018 the S2GLC project was extended. The extension aims at implementation of the S2GLC classification approach for classification of the whole Europe. As a source of training data two existing LC databases are used: Corine LC and High Resolution Layers.

Due to the characteristics of the European landscape and LC databases used the global way of proceeding has been modified. Intensive tests were carried out in selected test areas spread across Europe in deferent climate zones. Applications of different sets of spectral channels, classification features and sources of training data were tasted as well as the method for selecting images for classification from a time series and possibility for detecting new LC classes.

The S2GLC legend originally developed for global mapping has been extended by three classes. The methodology for training data selection was refined by filtering (cleaning) all layers of the source databases used for training. Additionally, a new method for image selection for a time series has been applied. The aggregation procedure that is the core part of the classification approach and the post-processing method, which addresses only pixels classified with low prediction score values, were applied in the same way as in the initial part of the S2GLC project.
Due to the amount of data the whole classification process will be performed using the CREODIAS infrastructure. Over 15 000 Sentinel-2 images from the year 2017 will be classified fully automatically with the use of software developed by CBK PAN. The final classification will be validated using approx. 45 000 points spread over the whole Europe.

Area A - Board 23

European directives INSPIRE (2007/2/EC) and PSI (2003/98/EC, revised by 2013/37/UE) induced major changes to EU member states geomatics ecosystems. In Wallonia, INSPIRE and PSI were respectively transposed into the INFRASIG and the Open Data decrees in 2010 and 2017. INFRASIG proposed the development of an official Georeferential. This Georeferential aims to provide free and open-access reference geodatasets to all Walloon stakeholders. The Georeferential should integrate basic geographical data which cover consistently the entire territory, refer to a general interest and have a guaranteed updating strategy. The current list of geodata includes administrative boundaries, mailing addresses, road networks, buildings, aerial images and digital elevation models. Other geodatasets are envisaged to be added in the future, one of them is land cover and land use.
The current land cover and land use map of Wallonia was published in 2007. Its nomenclature was designed in compliance with the european Corine Land Cover map. It does not distinguish the land cover from the land use. Despite offering finer minimum mapping unit than Corine Land Cover, the map does not address the current users’ needs in terms of positional accuracy, nomenclature and updating strategy. In the frame of the Georeferential, the Walloon authorities supports a pilot project called Walous for a collaborative design of a new land cover and land use database. Walous combines very high spatial with very high temporal resolution datasets. Aerial images and normalized digital surface models are combined with Sentinel-1 and 2 satellite data. Walloon thematic ancillary databases are used for training the classifiers, postprocessing and validation steps. A user survey gathered the interests of Walloon end-users in land cover and land use and integrated these people in a collaborative database development process. Walous conducted this consultation process for clarifying data requirements such as the definition of a potential common classification scheme, the resolution and quality needed as well as input data expected. During the consultation process, the end-users’ technical expectations were compared to other national cases and to the most recent European land monitoring initiatives looking for transregional standardization.
The poster introduces this users’ consultation methodology and illustrates how the results impacted the priorities and the specifications of the future Walloon land cover and land use database.

Area A - Board 24

WHY?
In spite of the existence of the EU Biodoversity Strategy towards 2020, biodiversity (nature-for-nature) and the delivery of ecosystem services (nature-for-humans) in Europe have continued to degrade since the 2010 baseline (EC 2015). This degradation is in large part due to losses in the wider environment, i.e. outside dedicated protected areas (Tucker et al. 2018). Much of this can be related to land use changes and the intensive exploitation of natural resources (EEA 2015).

Our study area is Flanders, the northern region of Belgium located in the Atlantic zone of Western Europe. For Flanders, the biological composition of the entire land-system has been assessed in a spatially explicit way, through the creation of the ‘Biological Valuation Map’ based on field surveys in the period 1978-1996 (De Saeger et al. 2017). However, due to labor constraints, this map has been kept up-to-date only for the ‘Special Areas of Conservation (SACs)’ of the EU Natura 2000 network. Outside these dedicated protection areas, information is largely outdated. This is problematic as it hinders policy makers in targeting their efforts in the framework of the EU Biodiversity Strategy.

WHAT?
The aim of this study is to automatically detect land-system changes in Flanders from Sentinel 2 imagery and to assign them to the correct type of change. The final objective is not to deliver a finalized up-to-date land-system composition map, but to deliver a change probability map with an indication of the type of change. This map can be refined by targeted field surveys in those areas with a large probability of change but also uncertainty about the type of change.

The European Sentinel 2 constellation was chosen as the source of input data because it offers free-to-use data with good spatial and temporal resolutions. However, long time series are not available yet so this study has to be considered as a pilot study, designed to explore the possibilities of this relatively new satellite constellation.

HOW?
Our method considers five broad land cover classes that can be converted between them: (i) impervious land, (ii) cropland, (iii) agricultural grassland, (iv) forested land and (v) semi-natural vegetation. The latter consist primarily of open soils; semi-natural grasslands; shrub- and heathlands; and wetlands.
It also takes into account changes in intensity or composition within a single land cover class. For these within-class changes, the method is restricted to the forest and semi-natural vegetation classes. There, a distinction is made between (i) intensity changes (harvesting regimes, grazing patterns,...) and (ii) composition changes (conifer to broadleaved forest, shrubland to grassland, grass encroachment, tree encroachment,...)

From a pair of Sentinel 2 images acquired in different years, one pre-change and one post-change, but at the same phenological stage (time in the growing season), four change indicators are calculated: (i) the Change Vector (CV); (ii) the Relative Change Vector Maximum (RCVMAX); (iii) the differenced Normalized Burn Ratio (dNBR); and (iv) the differenced Normalized Difference Vegetation Index (dNDVI). For more information on these indicators, see Jin et al. (2013).

For each indicator, a kernel-based stratified ranking of the change indicator is performed. Stratification is based on the pixels' post-change phenology, as derived from a seasonal time series of PlanetScope images from 2018. The kernel size is inversely proportional to the local variability in indicator values. In a first tier, the thresholds for between-class change detection are applied. The pixels that are identified as stable within the forest and natural vegetation classes are then subjected to a second tier, where the change in intensity/composition is assessed from the same change indicators, but using other thresholds.
To quantify commission errors a two-step verification is carried out, consisting of a visual interpretation from orthophotos which, if uncertainty remains, is followed by an actual field check. Omission errors are evaluated based on a comparison to existing databases of known land conversion projects.
Depending on the impact of the acquisition date on the change detection success, the method can be adapted to use multiple Sentinel 2 pairs distributed over the growing season.

References:
De Saeger S., Oosterlynck P. and Paelinckx D. “The Biological Valuation Map (BVM): a field-driven survey of land cover and vegetation in the Flemish Region of Belgium”. Documents phytosociologiques - Actes du colloque de Saint-Mandé 2012 - Prodrome et cartographie des végétations de France. 2017. pp. 372-382.

European Environment Agency (EEA), “The European Environment—State and Outlook”. 2015.

European Commission (EC). “The mid-term review of the EU biodiversity strategy to 2020”. 2015, 478p.

Jin S., Yang L., Danielson P., Homer C., Fry J. and Xian G. "A comprehensive change detection method for updating the National Land Cover Database to circa 2011”. In: Remote Sensing of Environment, Elsevier, 2013, pp. 159-175.

Tucker G., Darbi M., Wende W. , Quétier F. and Rayment M. “Conclusions: Lessons from Biodiversity Offsetting Experiences in Europe”. In: Biodiversity Offsets, Springer International Publishing, 2018, pp. 245–252.

Area A - Board 26

In this study, we aim to perform basic methods used widely in satellite imagery, in order to unmix and, later on, classify multispectral images. Our data is Sentinel-2 products, depicting the same ground area in two different seasons, August (summer) and October (autumn), respectively, at three different spatial resolutions (10m, 20m and 30m). Our endmembers selection is based on a hybrid method, consisting of the careful inspection of the image and ground truth. First, we perform Least Squares (LS) spectral unmixing techniques, with and without constraints, in order to extract abundance maps. Their comparison, based on SSIM index results, allows us to decide on which one best performs the unmixing of three basic compounds (rough unmixing). Thenceforth, we are able to perform further unmixing within a basic category (fine unmixing). The second technique to perform is classification, using the Nearest Neighbor classifier algorithm, where each pixel is assigned to the class whose representative spectrum is closer to. Its output is a colored classification map, easily compared to our reference map. The last technique to use is the k-means clustering algorithm, where the optimum number of clusters is decided based upon the square error as a function of the number of clusters graph. The last two techniques require the construction of a confusion matrix and the calculation of its success rate, in order to validate our results. As an overall conclusion we shall confirm that the techniques work adequately well, where especially the unmixing was satisfactory enough, given that we exploit multispectral data.
Versatile Land Cover applications such as material discrimination, resources detection and image classification require further processing of the image pixels, through spectral mixture analysis. The challenge faced here, was to implement three main techniques, Unmixing, Supervised Classification and Clustering on multispectral data acquired for a specified region of Southern Pindos. These techniques are mostly exploited in the field of Hyperspectral Imaging [8->1, 10->2], where spectral information is abundant, and due to the poor thematic information of Multispectral data, bibliographic references for the latter are limited. The research was conducted on part of the Northern Pindos National Park, in the northwestern Greece (between 39o 47’ 8.26’’ and 39o 52’ 32.95’’ N latitude and 21o 8’ 52.48” and 21o 1’ 52.19’’ E longitude), which is the largest protected forestry region, with many topographical diversities, including Natura sites and various types of vegetation. The selection of the area was based on its appropriateness concerning the distinction of basic endmember types.
As far as the data are concerned, we select two Sentinel-2 images of the area corresponding to August 2015 and October 2016, in order to identify the LULC changes in different seasons. The images are subsetted for the respective spatial resolutions of 10m, 20m and 30m. Then, simplistic linearity model is assumed, suggesting that each endmember covers a defined region within the pixel area. Therefore, the measured pixel radiance, as mentioned, is a weighted average of the radiances of the pure components present and the abundances represent the percentage of each material on the pixel. Linear Unmixing Techniques require the a-priori knowledge of the material spectral signatures present in site [7->3], deeming the selection of appropriate Regions of Interest (ROIs), per general class, the first and foremost step of the followed methodology.
Appropriate ROIs were selected based on gross classes Water, Vegetation and Soil on the first level of unmixing (rough unmixing) and then endmembers were extracted by averaging the pixel reflectance values per ROI. Analogously, for the refined unmixing process of the same large classes, we have identified 3 endmembers for water (Fig.1a), 4 endmembers for vegetation (Fig.1b) and 3 endmembers for soil (Fig.1c). Moreover, we evaluate the unmixing results in terms of structural similarity with the initial image, compare the unmixing efficiency seasonally (per image) and with ground truth (Landsat reference map)
Furthermore, we perform an NN classifier algorithm with the 11 endmembers/representative for the two images in the three resolutions. The resulting classification map (Fig.2) is compared to the reference map and evaluated in terms of efficiency with calculated success rates, corresponding to different confusion matrices.

Keywords: Multispectral data - Linear spectral unmixing - NN-classifier - k-means clustering - Image structural similarity - Classification map – Confusion matrix
REFERENCES
[1] Keshava, N.; Mustard, J.F. Spectral Unmixing. IEEE SPM, 2002, vol. 19, pg. 44-57, doi: 10.1109/79.974727
[2] Dobigeon N.; Altmann Y.; Brun N.; Moussaoui S. Linear and Nonlinear Unmixing in Hyperspectral Imaging. Chpt. 6, Data Handling in Science and Technology Elsevier, 2016, Vol. 30., http://dx.doi.org/10.1016/B978-0-444-63638-6.00006-1
[3] Branca, S.; Coltelli, M. and Groppelli, G. Geological map of Etna volcano, 1:50,000 scale. Ital.J.Geosci., 2011, Vol. 130, No. 3, pp.265-291, doi: 10.3301/IJG.2011.15
[4] Zhou, W.; Bovik, A. C.; Sheikh, H. R. and Simoncelli, E. P. Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE TIP, 2004, Vol. 13, pp. 600–612, doi: 10.1109/TIP.2003.819861
[5] Buongiorno, M. F. et al. ETNA 2003 field campaign: Calibration and validation of spaceborne and airborne instruments for volcanic applications. ASI Projects: I/R/157/02, I/R/203/02.
[6] Allard, P.; Behncke, B.; D'Amico, S.; Neri, M. and Gambino, S . Mount Etna 1993-2005: Anatomy of an evolving eruptive cycle. Elsevier Earth-Science Reviews, 2006, Vol. 78, pp. 85-114. doi: 10.1016/j.earscirev.2006.04.002
[7] Singh, K.D.; Ramakrishnan, D.A. Comparative Study of signal transformation techniques in automated spectral unmixing of infrared spectra for remote sensing applications. IJRS, 2017, 1235-1257, doi: 10.1080/01431161.2017.1280625

Area A - Board 27