Countries face a large challenge to regularly report on 232 Sustainable Development Goal (SDG) indicators, but this is essential to monitor and report annually on progress towards the goals. Approximately 20% of the indicators can be interpreted and measured either directly or indirectly through geospatial data; in some cases in combination with other statistical data. Global datasets can potentially fill national data gaps, while identifying priorities to improve national data, thus reducing costs for annual updating. Global datasets which are available for the entire reporting timeframe (2015-2030, for example those derived from Sentinel data), can also be used to stimulate transparency across countries, and communicate comparable performance to build confidence or compliment national reports. Among the available remote-sensing and geospatial data sources and products, land cover and land use (LCLU) data products in particular are essential for monitoring 4 of the 17 goals, and can be complimentary to at least 3 others. Taking a closer look at the targets and indicators, reveals that LULC datasets can be used to monitor at least 33 indicators. Since a number of LULC datasets exist, selection of the most appropriate datasets is crucial. However, there are no internationally established methods or standards for 57 of the indicators (which are classed as tier III), and those which do have methods can also benefit from the integration of additional datasets. There is therefore, a need for guidance on the use of geospatial data for monitoring, or for baseline estimation of SDG indicators.

The Task Stream on Application of satellite Earth Observation Data for the SDG indicators within the Working Group on Geospatial Information of the Inter-Agency and Expert Group on the Sustainable Development Goal Indicators (https://unstats.un.org/sdgs/iaeg-sdgs) has a mandate to provide a critical analysis of the high-level methodological guidelines on SDG, as well as recommendations on how to better include Earth Observations for SDG monitoring. This group aims to leverage and enhance partnerships with space agencies active in the Group on Earth Observations (GEO) Earth Observations for the Sustainable Development Goals (EO4SDG) initiative, and this presentation will build on the outputs from this working group, in particular those led by GOFC-GOLD, and will:
- Describe the user needs related to the use of LCLU data for SDG monitoring, including the link to national data availability and individual countries’ capacities. Criteria for the utilization of complimentary data will include data quality; potential to augment national data (interoperability, standardization applied); temporal, spatial, and thematic resolution; access to, ease of download and use of data including organization and management aspects such as data ownership .
- Provide an overview of the available LCLU data, an assessment of the core datasets which are applicable for SDG monitoring. The use of the land cover products within the Copernicus Global Land Service https://land.copernicus.eu/global/ in this context will be highlighted.
- Present case studies of the use of LCLU data for specific indicators, including requirements for the integration of complimentary data. These case studies may include:
o Indicator 2.4.1 Proportion of agricultural area under productive and sustainable agriculture (tier III). FAO is the custodian agency, and statistical data from FAO can be used, however land cover data which has thematic detail to separate sustainable and productive agriculture from other agricultural systems is required.
o Indicator 9.1.1: Proportion of rural population who live within 2 km of an all-season road (Tier III). Along with combining global road network data and geospatial population datasets, land cover maps which can separate urban and rural classes can be used.
o Indicator 15.3.1: Proportion of land that is degraded over total land area (Tier II). This indicator has 3 sub-indicators, and land cover data is central to the proposed methodologies, for 2 of the 3 indicators. These sub-indicators are Land Cover, where the data is used directly, and Soil Organic Carbon, where land cover changes can be used along with carbon conversion factors for each class changes to determine whether degradation has occurred for that sub indicator.
- Building on the last points, describe the use of LCLU data for underpinning interventions which aim to support the achievement of SDGs.
- Finally, give recommendations for the promotion of geospatial datasets to support SDGs, including the need for an active dialogue between data providers and countries to identify priorities, and the needs for a consistent long-term service vision for geospatial data in this context.

Area A - Board 115

Land cover monitoring is an integral part of land management. Many stakeholders, such as land owners, governments, NGOs and international organisations are interested in keeping track of changes on the ground, so that these changes could be managed. This can be done using land cover maps, however, most land cover maps are only produced for a single date. For example, the GLC2000 product was produced with the intention of representing the global land cover in the year 2000. Yet, the target of land cover monitoring is the land cover change over time, which is an indicator of ongoing processes on the ground.

One way to express land cover change is by producing multiple single-date maps, and analysing the differences between them. However, if the probability that the pixel belongs to a certain class is close between two classes, then even a small change in model input data may cause the prediction to shift to a different land cover class.

In this study, we propose to update land cover maps by performing change detection on time series of satellite imagery. If there is an abrupt change in seasonal pattern in the time series of a pixel, then a land cover class transition is likely to have occurred at that location. Only areas with detected change need to be updated for producing the next iteration of the map, making it more consistent over time, and taking less computational effort to reclassify the map. In addition, performing change detection allows us to avoid extracting temporal features for classification from periods of change. If no land cover change happened, then it is possible to use a longer history period for extracting temporal features, therefore increasing the robustness and accuracy of the classification model (see Figure 1). However, a limitation of using change detection for map updating is that we assume that the original land cover map is accurate and can be used as a basemap for updating, which may not always be the case.

To see if this method is viable and scalable, we compared the performance of three time series break detection algorithms: strucchange::breakpoints, BFAST Monitor and annual t-test. The algorithms were run over the time series of vegetation indices derived from MODIS 250 m data over the entire continent of Africa. A computer cluster hosted by VITO was used for this big data processing task. This was done in the context of the Copernicus Global Land System: Dynamic Land Cover (CGLS-LC100) land cover project.

In order to validate the algorithm performance, the CGLS-LC100 project is collecting reference points from Africa. Land cover change is identified through image interpretation of very high resolution imagery, as well as Sentinel-2 images and NDVI profiles. Once the reference data is collected, algorithm performance will be assessed using confusion matrix statistics, such as sensitivity, specificity, false positive rate, likelihood ratio for positive tests and positive predictive value.

Preliminary results from West African drylands show that land cover change is an uncommon phenomenon, as only 3.4% of the 1010 reference points show change. All of the tested algorithms tended to overestimate the detected change in this region. The annual t-test algorithm was the fastest, but it does not provide any temporal metrics, only whether there was a break in time series over a chosen year. In contrast, strucchange::breakpoints provides the estimated day of year when a break in the time series is detected. BFAST Monitor is in between, in that it gives a rough estimate of when the break occurred within a particular year.

Once additional reference data about land cover change is collected, we will investigate options to further tune the algorithms to detect land cover change, so that it would improve the classification accuracy of the final land cover map for a given year. This will also reduce the data amounts that need reprocessing each year to produce an updated map. Limiting the amount of data to process is becoming more important, as land cover maps move into higher spatial resolution, e.g. 20 m resolution based on Sentinel-2 imagery.

The framework set in this study is independent of a particular satellite sensor, and so the findings are applicable to all land cover maps that are in an operational phase, i.e. are regularly updated. Due to the scalability of the approach, the methods have the potential to improve the accuracy of both local-scale and global-scale land cover classification products, as well as make them easier to update.

Area A - Board 116

People have increasingly been shaping the surface of our planet. Land cover/land use change (LULC) and land management dynamics are highly interlinked and occur with the same order of magnitude, commonly even at the same time and location. The continuous expansion of agricultural land at the expense of forests rank among the most distinctive traces humanity has left on earth. Such land use transitions are one of the main contributors to greenhouse gas (GHG) emissions and biodiversity loss. Simultaneously, they are essential for providing food and goods for a rapidly growing world population with changing consumption patterns. Thus, land use change assessment is key for tackling pressing sustainability challenges such as climate change, environmental protection and food security.
For better understanding its underlying processes and associated environmental impacts, more accurate, spatially explicit and global-scale land change reconstructions are needed, particularly for earth system models. However, data sets on LULC still hold considerable limitations for land change reconstruction at large scales, such as fragmented content, varying formats and scales, or lacking spatial explicitness. Recent and ongoing developments in earth observation are an increased spatial, spectral, and temporal resolutions, analysis-ready land cover data at large-scales (e.g. Globeland30, ESA CCI-Land Cover) as well as a shift towards free and open-access data policies (e.g. Landsat archive, ESA Sentinel missions). We attempt to overcome the still existent uncertainties regarding the extent and environmental effects of human-induced land change by exploiting this growing potential of available remote sensing products.
In particular, we develop a more comprehensive reconstruction of global land use change from 1950-2015. The proposed model allocates land change based on a harmonised integration of multiple data sets, including satellite-based maps and statistical inventories, quantifying both net and gross LULC transitions at 1 km spatial resolution across the globe. Remote sensing-based LULC products are ranked according to multiple criteria such as timeliness, resolution, accuracy and spatial extent. Based on derived likelihood maps for each LULC class, land change statistics are spatially allocated. As a result, distribution and transition maps covering the major LULC classes - settlement, cropland, pasture, forest, natural grass-/shrubland and other land (barren, ice/snow cover) - are generated for each year. This reconstruction aims to mirror past and ongoing LULC change patterns, providing valuable information on the state and dynamics of human activity on the land surface.
As a follow-up and in still ongoing work, we link the observed land transitions with information from land management inventories for croplands, pastures and managed forests across different world regions. Cropland expansion and intensification patterns as well as the underlying land change legacy will be analysed at the same time, gaining new insights into location and extent of land change and its induced GHG emissions. Further, afforestation patterns and trends in Europe and associated carbon sequestration are estimated and put into a global context. Using crop trade statistics, land change displacements can be located and quantified more accurately and finally be re-allocated to the importing country. By feeding the reconstruction model with this updated change matrices, hypothetical land change maps for Europe can be derived. They visualise how European land area might look like if displaced LULC transitions, which are involved in European food production but occur in remote world regions (e.g. conversion of natural forests to cropland used for fodder production), were led back to the consumer’s land surface.
With this purely data-driven approach, we address current research needs of the earth system modelling community by providing new layers of land use change and land management with unprecedented level of detail. Learning from the recent past, understanding how management and land cover dynamics interactively affect the climate and how land change patterns are highly globalised is essential for implementing measures of mitigation and sustainable land use policies. In this context, a solid information base can support informed decision-making.

Area A - Board 117

During the last decades, the global warming has affected both the extension of glaciers and the surrounding landscape. The retreating glaciers were followed by colonization dynamics of the new ice-free areas. Furthermore, in the lower altitudes, land-use change was the major driver of vegetation dynamics. A remote-sensing approach using Landsat imageries could help in revealing changes in the extension and altitude of different land-cover types along a time-series.
Here, we studied land-cover changes in the surrounding landscape of the largest Italian glacier, i.e. the Adamello glacier. A time-series of Landsat (TM, ETM+) imageries covering the last three decades was selected using the criteria of less than 10% cloud cover, taken during the summer season, i.e. June-September, and being without signs of summer snowfall. The study area was limited to a circular plot of 63 km2 centred at the Adamello glacier. Given the heterogeneous morphology of the mountain landscape characterizing the study area, we excluded unreliable pixel values of topographically shaded area, the clouds and their shadows. As the medium resolution of Landsat imageries can be inappropriate to map alpine vegetation, we partially overcame this issue using a fuzzy approach. The fuzzy clustering calculates the probability for each pixel to belong to the different land-cover types, which can be assumed as the proportion of the diverse components of the pixel. Furthermore, to improve the detection of the very different land-cover types (e.g. ice, water bodies, bare soil and vegetation types), we applied the fuzzy clustering to normalized difference indices instead of reflectance bands. Specifically, we used the normalized difference indices for vegetation (NDVI), soil (NDSI), water (NDWI) and the modified version for water (MNDWI). Analyses were conducted in the R environment.
The use of normalized difference indices in the fuzzy classification improved the interpretation of the land-cover types. In fact, different vegetation types characterising different altitudinal ranges were detected on the basis of the proportions of NDVI and NDSI.
We found a slight decrease in the extension of the glacier during the last three decades. Vegetation types did not show clear trends of upwards migration during the last three decades. This can be explained by the time-lag needed by plant communities in proceeding or starting new ecological successions.
The results obtained in this study will be used in the context of the multi-disciplinary project “CALICE: Calibrating Plant Biodiversity in Glacier Ice”. The project aims to reconstruct plant biodiversity and its trend archived in Alpine glacier ice by pollen and eDNA in the past.

Area A - Board 118

Large parts of the Arctic are characterized by diverse surface water systems. Arctic lowland rivers are highly dynamic, often braided and less influenced by anthropogenic activity than most rivers at the mid-latitudes. Future developments driven by climate change are likely to influence river dynamics through changes in water supply, changes in seasonality and changes in the surrounding permafrost underlying the water bodies. The geomorphological structure of rivers can influence the quality and diversity of habitats adjacent to the river and hence affect the biodiversity in a rivers’ ecosystems. Changes in the geomorphology of rivers are therefore a major concern in the conservation of sensitive riparian habitat. In particular, changes in river sinuosity can act as pointers to increased incidences of flooding, drastic effects of climate change and unprecedented anthropogenic activities in the river basin. Lack of long-term and fine-scale spatial and temporal data has hindered detailed analysis of changes in river sinuosity and the link between such changes and ecosystem dynamics. In remote Arctic areas, understanding changes in the properties of rivers and ecosystems have been more difficult due to the inaccessibility of these areas. The emergence of cloud-based archives of big earth observation data provides new opportunities for quantifying the long-term changes in sinuosity of Arctic river systems and linking the changes in the sinuosity to the vitality of Arctic ecosystems. Normalized difference water index (NDWI) from remote sensing imagery has been shown to be a feasible index for the extraction and characterization of water bodies, including river systems. To characterize vegetation dynamics a multitude of indices has been developed over the years, the normalized difference vegetation index (NDVI) being maybe the most prominent. In recent years, the Tasseled Cap index has been used extensively to detect and monitor disturbances and recovery of vegetation after events ranging from flooding to landslides as well as anthropogenic influences. In this study, we extracted rivers from Landsat-8 and Sentinel-2 imagery using Google Earth Engine (GEE) for selected river catchments in Alaska and northern Canada using the NDWI. From the extracted rivers, we estimated changes in river sinuosity. In addition, NDVI, NDWI and the coefficients from tasseled cap transformation as derived from the Sentinel-2 and Landsat-8 imagery were used to estimate the ecosystem dynamics. The results of this study revealed correlations between changes in river sinuosity and vegetation dynamics for the selected study sites. The outcome of this work can be applied to highlight riparian ecosystems that may be facing increased stress from the effects of climate change and human activities, hence allowing for timely implementation of mitigation strategies to conserve the ecosystems in Arctic river catchments.

Area A - Board 119

The development of gap-free and temporally smoothed time series from freely available satellite imagery offers opportunities for improving our understanding of forest disturbance and recovery, as well as a capacity to inform on key sustainability and conservation issues. Wall-to-wall forest disturbance maps have been created from Landsat imagery for the forested area of Canada and can be used to characterize land change across broad areas. By analyzing long time series and extracting spectral information prior to, during, and after disturbance events, we can now create metrics to better characterize current land conditions. For example, recent work in the boreal forest has directly linked spectral recovery of vegetation indices to key stages of forest regeneration, and has shown that recovery patterns differ between natural disturbances (e.g., fires) and anthropogenic disturbances (e.g., forest harvesting). However, it has yet to be shown whether spectral recovery is a meaningful variable in habitat selection modelling. Wildlife populations have declined dramatically over the last 50 years due, in large part, to habitat loss and degradation, making it critical to understand how disturbances impact habitat selection in order to move toward more sustainable land management strategies. In this study, we analyzed how a threatened population of grizzly bears are responding to forest spectral recovery over an 118,000-km2 interior region in Alberta, Canada. Grizzly bears in interior regions are heavily dependent on vegetative foods, which tend to be tightly coupled with forest seral stage following disturbance, making grizzlies especially dependent on disturbance regimes. We analyzed a 33-year Landsat time series in conjunction with over 450,000 GPS telemetry locations from 160 individual grizzly bears and found that spectral recovery was an important variable in habitat selection models. In general, grizzlies tended to select disturbed areas with intermediate levels of spectral recovery relative to pre-disturbance and avoid or be neutral to disturbed areas with high-spectral recovery. In burned areas, grizzlies avoided areas with very low spectral recovery (i.e., very recently disturbed), whereas in harvested areas avoidance was strongest for very high spectral recovery. These contrasting trends likely reflect the different vegetation recovery trends in a naturally regenerating burned area versus a planted cut block. Indeed, analysis of spectral recovery across the study area showed that harvested areas recovered nearly twice as quickly as areas burned by wildfires, which translated into different long-term habitat provisioning between the two disturbance types. Other time-series derived disturbance metrics, such as disturbance intensity and pre-disturbance spectral indices were also important predictors of habitat selection. This research further indicates that spectral recovery is indeed related to seral stage during forest regeneration and demonstrates its utility for habitat modelling. It also highlights the importance of free and consistent satellite time series (e.g., Landsat, Sentinel) to detect disturbances and create novel metrics related to land conditions before, during and after disturbances to better assess land change in relation to sustainability and conservation goals.

Area A - Board 120

Despite the ecological and social importance of arid and semi-arid forests, they have received less attention in the scientific literature than temperate, boreal and tropical forests. Growing evidence depicts that these forests have been dramatically degraded in many regions of the world, causing an urgent need for a vivid understanding of the processes and reasons leading to forest depletion in semi-arid regions. As one prominent example, an alarming decrease in oak forest habitats has been recently reported. Oak decline has also been observed in the Zagros Mountains in western Iran. The area comprises over 42 percent of the country's forests and provides multiple ecosystem services. The reasons for this decline are manifold and include unsustainable use but also an increased number of insect infestations. In this realm, identifying and mapping oak decline is an important step to manage the spread of infestations. Several earlier studies have demonstrated the potential of remote sensing for such purposes, however, there are still open questions concerning why certain patterns of oak decline can be detected with remote sensing, while others not. Moreover, spatially mapping declined areas is just the first step towards understanding the drivers for oak decline. Once the affected areas have been identified, it is necessary to also understand the underlying drivers. In this paper, we will present the outline of a project for a remote sensing study consisting of four steps: Initially, the minimum size of declined forest patches that are detectable with current multi-spectral satellites will be identified. Next, the influence of co-occurring land-cover classes on these results will be assessed by applying a multi-scale remote sensing approach (including very high-resolution drone imagery and multi-spectral satellite data). Then, the derived spatial patterns of declined forest areas will be analyzed in relation to a range of explanatory variables. Ultimately, dendro-chronological time-series of stress events will be compared to spectral information available from historic Landsat datasets to check the potential of reconstructing stress events over time using satellite data.

Key Words: Oak Decline, Multi-scale Remote Sensing, Dendro-chronology, Forest vitality

Area A - Board 121

Since the mid-1990s, North Korea has gone through a hard time of shortage of food and fuel due to the large scale flood and landslide. This became a vicious circle, which has kept accelerating the deforestation in North Korea.
In order to collect informations on the status of deforestation in North Korea, the change of deforestation in North Korea was analyzed using two different seasonal satellite images of Landsat 5-TM and SPOT-5 between 1999 and 2008. In 2018, we are analyzing deforestation trend using RapidEye, and expect that the result is able to be used next year.
We used unsupervised classification since North Korea is inaccessible area. The results of the study showed that the stocked forest area has decreased 1,379,000 ha compared with those in 1999, whereas the deforested and degraded forest area has increased 1,207,000 ha in 2008.
Based on this information of deforestation in North Korea, we suggested a criteria for forest restoration site selection, considering the successful forest greening experience of South Korea in 1970's and the current forest restoration policy of North Korea. Then, we have estimated areas and produced a distribution map concerning forest restoration which can be applied in the future.
When we look into forest greening of South Korea in 1970s, it was set as a principle that the slash and burn fields with more than 20° slope should be fully restored as forest during the slash and burn fields restoration project(1974~1978). In case of erosion control project, three types of erosion control project were applied according to types of deforestation and slope criteria. Consequently, areas with under 25° slope(revegetation of bare soil), 25~40°(general erosion control) and more than 40°(special erosion control) were differently treated amidst the project.
In addition, criteria on agroforestry and slope land management that North Korea has been autonomously implementing were also taken into consideration. It has three different types of slope land management plan depending on degree of slope : Areas with under 15° slope can be used for farm, areas with 15~30° slope which have potentialy soil erosion risk for agroforestry and areas with more than 30° slope for reforestation.
By referring to the successful case of South Korea’s forest greening experience and reflecting North Korea’s forest restoration policy, we produced the criteria for forest restoration site selection in North Korea. All the deforested areas with more than 15° slope can be restored. To put more details, farmland with 15°～30° slope can be chosen for agroforesty site and the rest of the area besides farmland with 15°～30° slope can be afforested. By this criteria, it is estimated that the potential area for reforestation throughout the nation is approximately 1.9 million ha and the area chosen for agroforestry is 0.76 million ha within it. The steep areas with more than 30° slope accounts for 15.8% among the whole potential reforestation area. The largest region is Hamyeongnam-do province, occupying around 20% of the whole potential reforestation area and Heochun-gun of Hamyeongnam-do province turns out to be the region that has the largest potential area across the nation.
The result of this study can help prepare for inter-Koran forest cooperation in the future and can be served as reference data when estimating the quantity of forest restoration and the expense for rehabilitation after Korea’s future unification. It is also expected analyzing the deforested areas of the sub-national scale in North Korea can enhance the effectiveness of implementing a plan for forest restoration in North Korea.

Area A - Board 122

Temperate forests and semi-natural areas harbour high biodiversity and provide important ecosystem services, and are constantly under land use change pressure. Monitoring and mapping changes in the state and condition of temperate forests and semi-natural areas are therefore one of recent, globally critical issues. The accessibility of freely available moderate to high spatial resolution (10–30 m) satellite imagery received a major improvement with the launch of the Sentinel-1 and -2 sensors. Together with Landsat-8, these sensors provide a wide range of spatial, spectral, and temporal properties which allow to advance and develop approaches increasing the accuracy of various types of vegetation mapping among others also in the temperate zone.

In this study, we explored and compared the synergistic use of Landsat-8, Sentinel-1 and Sentinel-2 data in vegetation mapping with a particular focus on tree, shrub and grass cover in regional scale. Our test area is located in the Polish Carpathian Mountains (20,000 km2). One of the biodiversity hotspots in Central Europe characterises by the complex land use mosaics of forests, grasslands, agriculture and built-up areas. We are making here full use of the Landsat-8 and Sentinel-2 (optical) and Sentinel-1 (radar) image archives for years 2017 and 2018. We collected a training dataset vectorised from very-high resolution imagery and combined with LiDAR dataset (from 2013) available for the entire study area. We employ two different algorithms – random forest and stochastic gradient boosting – to classify various data configurations (optical and synthetic aperture radar (SAR) data).

The first results of this study demonstrate that models jointly using optical (Landsat-8 or Sentinel-2) and SAR data performed better than models using single-sensor imagery. As well, the combination of Landsat-8 and Sentinel-2 (optical data) results in overall classification accuracy improvement over that of single-sensor data used, in particular in areas of dense vegetation cover. Our evaluation shows and highlights both the value of the growing Landsat and Sentinel archives and the importance of multi-sensory approaches for vegetation mapping and monitoring across relatively large areas like for example mountains in the temperate zone.

Area A - Board 123

Satellite technologies have widely demonstrated their potential in large scale environmental monitoring, especially in areas in which the implementation of field surveys is more problematic and expensive. The recent launch of the Copernicus Programme is now providing a large amount of free data allowing the remote sensing community for furtherly improving the understanding of our planet, whose resources and ecosystems are increasingly under the pressure of a growing population and the subsequent demand for more goods, food, and energy.
Among the ecosystems most affected by urbanization, there are forests. They currently cover about 40% of Earth’s ice-free land surface although, in the last centuries, a large fraction of forestland has been converted to agricultural and urban uses. Nevertheless, many human activities rely and exploit forests, such as the production of paper, lumber, and fuelwood. Moreover, forests are fundamental for the preservation of the whole Earth ecosystem since they play a fundamental role in the carbon cycle. With a projected increase in human population and the consequently rising in standards of living, the importance of forests is expected to increase. Therefore, the challenge is to set-up protocols and methodologies allowing for the sustainable management and exploitation of this critical resource according to UN sustainable development goal 15 (Life on Land) requiring in few years the implementation of sustainable management of all types of forest, including the mitigation of forest degradation and the increase of global re-forestation.
The role of remote sensing as a proxy for effective forest management is crucial because of its capability to provide synoptic and repetitive information about several geophysical terrain data and biochemical vegetation parameters to be delivered to decision-makers. In this work, a new framework for unsupervised change detection in forest areas using multitemporal Sentinel-1 synthetic aperture radar data is presented. It is based on the exploitation of the change-detection color composites introduced in [1], in which time-series data are effectively combined to deliver a user-friendly representation of the backscattered radar signal. Information about changed areas is retrieved by applying semantic clustering principles [2]. This technique, exploiting self-organizing maps and object-based reasoning, allows for the automatic assignment of a low level semantics to images. This is done drawing from a database of relevant signatures associated with a color within the input RGB temporal composite. These attributes, referring to clusters’ color, can be exploited for unsupervised extraction of image segments representing changes in the imaged scene. As an example, referring to the attached graphics, blue clusters identify areas affected by land cover changes related with vegetation removal (deforestation) [2].
The proposed framework is tested over a forested area in the Antioquia region in Colombia. Tenths of Sentinel-1 scenes are processed to reconstruct the dynamics of the land cover starting from 2014. The obtained results reveal several areas affected by forest disturbances. By the time of conference, land cover changes will be classified based on their first occurrence and labelled with a higher semantic level, thus allowing for the reconstruction of the recent history of the disturbances occurred in the study area.

References
[1] D. Amitrano, G. Di Martino, A. Iodice, D. Riccio, and G. Ruello, “A New Framework for SAR Multitemporal Data RGB Representation: Rationale and Products,” IEEE Trans. Geosci. Remote Sens., vol. 53, no. 1, pp. 117–133, 2015.
[2] D. Amitrano, F. Cecinati, G. Di Martino, A. Iodice, P.-P. Mathieu, D. Riccio, and G. Ruello, “Feature Extraction From Multitemporal SAR Images Using Selforganizing Map Clustering and Object-Based Image Analysis,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 11, no. 5, pp. 1556–1570, 2018.

Area A - Board 124

Unequal land distribution represents an important problem for countries in Latin America and one of the highest impediments for economic growth. During the past decades, important land reforms were carried out in countries like Bolivia, Mexico, Cuba, Nicaragua, Chile or El Salvador, leading to significant land use / land cover changes.

In Bolivia, the land reform was coordinated by the National Institute of Agrarian Reform (INRA). The main purpose of the agrarian reform laws was land redistribution, which granted property titles to approximately 60 indigenous communities in Bolivia, consisting of more than 25% of the country’s population. According to INRA, more than 600 000 km2 of land were titled between 1996 and 2013 under Bolivia’s land regularization laws. Given that a large part of the land being titled is actually forest land, there is an obvious conflict between the objectives of the land reform and the principles of sustainable forest management.

The aim of our study was to assess the dynamics of land use, with a special look on forestry dynamics, within Bolivian indigenous community titled plots, before and after the land reform. The research was based on the Copernicus Global Land Service (CGLS) - CCI Land Cover products, provided at a 300 m spatial resolution and a series of freely available datasets provided by the Bolivian Government through GeoBolivia platform – the geoportal containing the Spatial Data Infrastructure of the Plurinational State of Bolivia (IDE-EPB).

The change detection analysis was performed both by using the 22-class original legend of the ESA-CCI Land Cover products, defined using the UN Land Cover Classification System and also by reclassifying the 24 annual datasets, from 1992 to 2015, for the Bolivian territory into 9 major classes: cropland, tree cover, grassland, wetland, settlements, shrubland, water bodies and areas covered by snow and ice, sparse vegetation and bare areas. The generalization was based on the IPCC (Intergovernmental Panel on Climate Change scheme and allowed a better understanding of the phenomena.

Several parameters and indices were computed, in order to quantify the effects of the land reform:
• land use / land cover (LULC) types of changes for each individual plot,
• LULC change rates for each individual parcel, calculated both in hectares and as a percentage,
• an assessment regarding the moment of deforestation, which was later compared to the moment of granting the property titles.

Also, for each plot there were derived additional parameters such as road density and distance from the closest urban center, with regard to their impact on land cover changes. The analysis of deforestation rates focused both on Bolivian indigenous community titled plots and other titled plots, which were cross-compared in order to establish regional behavioral patterns regarding forest and cropland evolution in the interval analyzed.

Last but not least, a series of forest fragmentation maps, at various levels of detail, were derived based on the ESA-CCI Land Cover 300-m datasets. Forest pattern analysis maps, fragmentation maps, density maps and forest gain / forest loss maps were computed for different key-moments within the time frame analyzed, using a variety of statistical methods, in order to gain a better understanding of the impact of the Bolivian land reform on land cover dynamics. A detailed analysis, based on higher resolution satellite imagery, was later performed for certain areas designated as “hot-spots”, such as the case of Santa Cruz department, which faced some of the most important changes.

Analyzing land-use and land-cover (LULC) is one of the most important applications in remote sensing. Monitoring land use / land cover changes for a certain area or for a specific land cover class can be extremely useful in the effective planning and management of natural resources, therefore global-scale products consisting of long-term time series of land cover maps have a really high potential in the field, as they provide a solid background for the change-detection analysis.

The proposed approach can be replicated and implemented for any area of interest using the ESA-CCI Land Cover dataset at any moment between 1992 and 2015, a feasible time- and cost-saving alternative to the traditional supervised and unsupervised land-cover classifications.

Area A - Board 125

The UNFCCC REDD+ framework increases the need for highly accurate maps of deforestation and degradation in the tropics. Operational forest/non-forest maps are commonly based on optical imagery. However, especially in the tropics optical images are frequently degraded by the presence of clouds. Therefore, we investigated the potential of hyper-temporal Sentinel-1 synthetic aperture radar (SAR) data to derive forest/non-forest and deforestation maps. Feature selection has been used, to decrease the amount of data and to enhance the signal to noise ratio. This is especially relevant for the use of machine learning, because it is one way to deal with the curse of dimensionality. In this study we compared the use of recurrence quantification analysis (RQA) with traditional multi-temporal metrics (e.g., mean, median, standar deviation, percentiles) for feature extraction from dense Sentinel-1 time series. Recurrence quantification analysis (RQA) is a non-linear time series analysis technique. It quantifies the patterns of recurrences in time series. By means of RQA a number of metrics can be calculated (e.g., determinism, recurrence Rate, laminarity), which describe the complex behaviour of dynamic systems. In contrast to traditional multi-temporal metrics, RQA considers the temporal order of the images of the time series. After calculating RQA and traditional multi-temporal metrics from the Sentinel-1 image time stacks, we performed a signature analysis. For this, we selected forested and deforested areas based on visual interpretation of annual very high resolution (1 m) optical imagery over temperate and tropical forests of Mexico. The signature analysis of the traditional and RQA metrics showed promising results for the classification of deforestation. Obviously the consideration of the temporal order of time series provides additional information compared to traditional multi-temporal statistics. Therefore, RQA can enhance the accuracies of forest/non-forest and deforestation maps. In the next steps we plan to combine RQA metrics and traditional multi-temporal metrics in order to further improve the map accuracies.

Area A - Board 126

Starling Pulse is a new product of Airbus Defence & Space dedicated to the historical analysis of land cover in tropical areas. It processes optical imagery from both LANDSAT & Sentinel 2 archives and creates maps that span over 30 years at a resolution of 20m.

This new service aims at monitoring tropical deforestation and growth of several commodities over the years to ensure that supply chains are sustainable. It provides key analytics computed on areas of interest (mill neighborhoods, protected regions, HCV, etc.) to monitor land cover dynamics and changes: forest loss, age of plantations.

The Starling Pulse workflow first involves the use of the Overland processor, an optical image processing suite developed by Airbus that implements top-of-the-art algorithms and retrieves biophysical variables by inversion of scene reflectance models. The Overland processor is used to process all LANDSAT (4, 5, 7 & 8) and Sentinel 2 (A & B) images from 1985 to today. Normalized Difference Fraction Index, LAI, shade of canopy, soil cover, chlorophyl content and fractions of green/brown vegetation cover are the main biophysical variables generated during the inversion.
Overland outputs are then stacked and co-registered to create dense homogeneous time series. Those time series are usually very noisy due to tropical climatic conditions and not very lengthy (from 200 to 500 usable observations). Quality rules are applied to weight each observation of these time series.
A structural change algorithm is then used to detect breakpoints on the NDFI time series of each pixel. This algorithm only allows breakpoints that results in ecologically possible time series decomposition.
Each regime of each variable of the time series is then fitted using various methods depending on its nature (seasonal, biaised, etc.), and phenological features are extracted for further hierarchical classification using machine learning methods.

The processing flow makes heavy use of cloud-computing technologies that enable processing at such a speed, resolution (20m) and over very large areas (several Mkm²) and long durations (30 years). Recent advances in data format and storage allows storing, exploring and viewing spatiotemporal data more efficiently.
The time series-based approach makes discovering land use patterns easier and the classification is all-the-more accurate as the number of observations increases. Finally, this product provides both a global and local view of the expansion of several types of commodities in tropical areas.

Area A - Board 127

Earth’s tropical forests are being rapidly lost and degraded by agricultural expansion and commercial logging operations, with population growth projected to further increase pressures on forests globally. While satellite monitoring of deforestation has significantly improved in recent years, methods for reliably detecting and monitoring tropical selective logging are at an early stage, with current methods only able to detect more intensive timber harvest. Historically, the spatial resolution of widely available data sets, like Landsat, have been considered too coarse to measure the subtle changes in forests associated with less intensive selective logging, yet most present-day logging is at low intensity. Moreover, frequent cloud cover in many tropical regions limits the availability of data to passive optical sensors like Landsat or Sentinel-2. To overcome these challenges our project seeks to combine Landsat (5, 7, and 8) and Sentinel-1 to monitor tropical selective logging. We are utilizing detailed selective logging data sets from over 25,000 ha of forest in the Brazilian Amazon to develop Random Forest machine-learning algorithms for detecting selective logging from optical and radar data, respectively. We show that texture measures significantly improve detection of selective logging for both optical and radar data. In addition, imagery (both optical and radar) acquired before the cessation of logging activities (i.e. late dry season during logging) was better at detecting selective logging than imagery acquired at the start of the following dry season (i.e. early the next dry season). With optical data, the detection rate of logged pixels was approximately 60% (with roughly 10% commission and 2% omission error rates) and approximately 40% of the area inside selective logging tracts were labelled as logged. Application of this algorithm to 6152 ha of selectively logged forest at a test site in Para, northeast Brazilian Amazon, resulted in the detection of 2316 ha (38%) of selective logging (with 10% commission and 2% omission error rates). In general, the detection of selective logging with Sentinel-1 was poor and suffered from low detection and high commission error (35%, with roughly 65% commission and 15% omission error rates). Optical and radar data sets generally provide complimentary information in classification settings and we are in the process of combining these models into a single predictive framework. Our initial results suggest, at a minimum, Landsat data can be used effectively, in a supervised machine-learning framework, to detect selective logging across large areas of the Amazon. Moreover, further refinement and integration of the Sentinel-1 model with the Landsat model should improve detection further. This work is thus an important step forward in developing systems for detecting selective logging pan-tropically with freely available data sets.

Area A - Board 128

Food production for a growing world population contributes to climate change by triggering increased land use and land cover changes. Agricultural land expanding into natural ecosystems leads to deforestation and land degradation causing greenhouse gas emissions. In the context of global sustainable development, estimating and mapping forest change at large scale is more important than ever before. In the last decade, due to freely available satellite imagery, time-series analysis of satellite data has been established as a tool for detecting and monitoring forest change at small land scale. This methodology allows for intra-annual detection of deforestation, and, with frequent satellite observations, almost continuous monitoring of the forest cover.
So far, most studies concentrated on developing the methodology of the time-series algorithms, and less on their applicability at large scale. This research aims to develop methods that advance time-series forest change detection techniques from local to regional and national scale usage, to the point where they can be used as tools for international reporting such as that required by United Nations Framework Convention on Climate Change (UNFCCC). In this sense, we emphasize the need to make time-series data and approaches open and more accessible to users, in particular in developing countries, and, to undertake this challenge, we collaborate with the Food and Agriculture Organization (FAO) of the United Nations to develop and integrate our methods with their System for Earth Observation Data Access, Processing and Analysis for Land Monitoring (SEPAL).
To begin with, our research addresses the data processing issue of exploiting such tremendously big open global satellite archive. With terabytes to petabytes of data needed in order to cover a large area both spatially and temporally, taking advantage of this data-rich situation has become challenging. First, the storage space and the RAM memory of a regular workstation are not sufficient. Considering this, we are using FAO’s high-performance computing system, SEPAL, as a sandbox for our developments, and we found it to be a suitable solution to tackle the computing power and storage necessity. Secondly, existing approaches are too slow for processing and analysing all the available data for very large areas. One of the most prominent time-series analysis approaches is the break detection for additive season and trend (BFAST) family of methods [1], and specifically the BFAST Monitor algorithm, which has been established as a robust approach for monitoring forest change, being widely applied on different forest ecosystems all over the globe. However, upscaling this method remains challenging, as the data is analyzed sequentially causing the runtime to be a bottleneck. To tackle this issue, we propose to adapt state-of-the-art time-series methods to new technological developments in massively-parallel compute devices such as the new concept of general purpose computing on graphics processing units (GPUs). Our initial developments [2] show great potential for the analysis of large datasets, our many-core implementation of the BFAST Monitor algorithm being up to four orders of magnitudes faster than the commonly used R implementation, given both artificial and real input datasets. Nevertheless, this initial implementation could only work with datasets with no missing values (e.g. from masking clouds), thus it was tested only on a plantation forest in the Atacama Desert, the driest non-polar desert in the world, providing the maximum number of cloud free observations possible over a vegetated area. To overcome this limitation we are currently developing a more general implementation that can effectively handle real-world scenarios with time series potentially containing many missing values. Here we will present a proof of concept, for the forest ecosystems of the African continent.
Next, we focused on tackling the problems associated with larger scale forest loss detection using time series analysis, caused by spatial heterogeneity. Specifically, we replaced the selection of algorithm parametrization based on expert knowledge with artificial intelligence decision making to account for inherent large-scale variation of forest type, data availability, and deforestation drivers. We applied our method on Landsat Surface Reflectance time-series for an area covering multiple countries in East Africa, with multiple forest types, and different cloud coverage, presenting a variety of deforestation drivers. Following good practices and using the stratified random sampling approach, we created the training and validation datasets using both Landsat and the European Union's Copernicus Sentinel 2 sensors for visual interpretation.
Our research consolidates the path towards using time-series satellite data for large scale forest change monitoring, and we hope to show our preliminary results, and discuss the outcomes and implications of our findings.

[1] Verbesselt J, Zeileis A, Herold M. Near real-time disturbance detection using satellite image time series. Remote Sensing of Environment. 2012 Aug 1;123:98-108.
[2] von Mehren M, Gieseke F, Verbesselt J, Rosca S, Horion S, Zeileis A. Massively-parallel break detection for satellite data. In Proceedings of the 30th International Conference on Scientific and Statistical Database Management 2018 Jul 9 (p. 5). ACM

Area A - Board 129

In the last decades increasing greenhouse gas (GHG) emissions accelerated global climate change. Circa 20% of the global GHG emissions are caused by tropical forest cover loss. A key role plays Indonesia showing globally the highest increase of forest cover loss between 2000 and 2012 due to mainly small and large holder agricultural crop expansion and timber production. While fire use for land management is forbidden by law, a wide range of fire-related practices are still used today causing varying GHG emissions. For example, fires are used before logging operations to clear understory, simultaneously to burn the forest directly or after logging to prepare areas for agricultural usage. However, the relationship between different fire-related land management practices, forest cover loss and related GHG emissions remains uncertain.
Remote sensing has proven to be a valuable tool for monitoring forest cover loss and fire activity in tropical environments. Several studies exploited the advantages of combining and utilizing the distinct features of optical and radar satellite data streams for monitoring forest cover loss. Optical sensors are capable of evaluating the photosynthetic activity of vegetation, whereas radar sensors are sensitive to water content and the structure of vegetation. The combination of satellite-derived active fire alerts and optical and radar time series provides the opportunity to investigate how fires and forest cover loss are spatially and temporally related. Furthermore, different fire related forest cover loss events, such as crown fires or understory fires, can be characterized and linked to various land management practices. However, up to now, characterizing the spatial-temporal relationship of fires and forest cover loss events utilizing remote sensing time series was limited due to lacking data availability and insufficient spatial resolution.
The launch of ESA’s Sentinel satellite missions in 2014 provides high resolution and free of charge radar (Sentinel-1) and optical (Sentinel-2) time series data for the first time. Both missions, each consisting of two identical satellites, have about weekly revisit times in the tropics resulting in dense time series. Daily active fire alerts are now provided by numerous satellites (e.g. MODIS, Suomi NPP and Landsat). Here we combine Sentinel-1 and Sentinel-2 time series with satellite derived active fire alerts to create a framework for improving the spatial-temporal characterization of fire-related tropical forest cover loss. Sumatra (Indonesia) is the study area showing wide spread forest cover loss and a variety of fire-related land management practices. The first step of the framework is to individually detected forest cover loss for Sentinel-1 and Sentinel-2 time series based on a probabilistic machine learning approach. Second, detected forest cover loss events are co-located with daily active fire alerts. Third, we separate fires that occur before, simultaneously and after the detected forest cover loss and exploit the response of the optical and radar time series signal. This spatial-temporal relationship is used to characterize different fire-related forest cover loss events and relate them to key land management practices. We will present first results of this study, discuss implications, and highlight the next steps.

Area A - Board 130

Fire is included within the Essential Climate Variables (ECVs) in the Global Climate Observing System (GCOS), due to its great significance in terrestrial and atmospheric systems. Burned Area (BA) information is critical to assess the impacts of biomass burning, both in atmospheric emission models and dynamic vegetation studies. Current global burned area products extend from 1995 to 2017, being only consistent since 2001, with products such as MCD64 (NASA’s official product), or FireCCI51 (ESA CCI’s official product). FireCCI51 is generated from the MODIS sensor, which is the sensor with the most reliable spectral information for the detection of fire on long temporal scales. Longer time series would be desirable to analyse the impacts of fire on climate cycles. This study presents the development of a global BA algorithm adapted to the Long-Term Data Record (LTDR, version 5) based on images acquired by the Advanced Very High Resolution Radiometer (AVHRR 2/3) sensor on board the National Oceanic and Atmospheric Administration (NOAA) satellites from 1981 until the present. This dataset includes atmospheric and geometric corrections of AVHRR data and provides daily thermal and reflectance values at 0.05º (≈ 5 km) spatial resolution.
LTDR daily data is processed in monthly composites using the maximum temperature criterion. Compositing is recommended since daily data has clouds, noise, radiometric instability and distortion or acquisition problems. Unburnable areas are masked out with information derived from the yearly CCI Land Cover product. Different spectral indices adapted to BA detection are computed (BAI, GEMI). A synthetic index is then created by statistical analysis from the different LTDR spectral bands and derived indices. Seasonal variation of this index are subsequently taken into account for training monthly Random Forest (RF) models using the ESA BA product (FireCCI51) as reference. RF is a machine-learning classifier based on a random combination of decision trees, which classifies each pixel by the number of times that pixel is assigned to a particular category. RF models are created combining LTDR data from several years (both with high and low fire occurrence). In this study, 600 trees are created for each monthly model. Since the burned pixels are much more unlikely than the unburned, a sensitivity analysis is performed to obtain the appropriate monthly cut-off thresholds to improve overall accuracy and balance omission and commission errors. Then, they are applied to the full time series, converting the thresholds into BA percentages. The aim is to show similar BA between FireCCI51and LTDR during the MODIS era and be able to extrapolate backwards. This LTDR BA product will provide climate and earth system researchers with a consistent BA temporal series available since 1982.

Area A - Board 131

Fire is considered an Essential Climate Variable (ECV) by the Global Climate Observing System (GCOS). The existing relation between climate, fire and vegetation makes the fire disturbance an important component of the Earth System. Taking this into account the European Space Agency (ESA) created in 2010 the Climate Change Initiative (CCI) which aims to provide climate users with long term datasets of 21 ECVs. Fire Disturbance was included within those ECVs and, thus, the Fire_cci project was the responsible of the generation of burned area (BA) algorithms and datasets. The last global burned area mapping algorithm developed in the Fire_cci project, FireCCI51, showed very good results, obtaining slightly better performance than the NASA’s reference BA product MCD64A1 (Dice coefficient of 0.564 and 0.536, respectively). One of the differences was that the former offered BA information at 250 m while the later was released at 500 m spatial resolution.

The FireCCI51 is a two-phase hybrid algorithm that takes advantage of both active fire and spectral information. The first one provides an accurate location of active fire information based on thermal anomalies and the second gives a more persistent signal due to the charcoal and ashes that can be found after a fire event. Based on these two inputs the algorithm first detects highly probable burned pixels (seed phase). Then, in a second phase, a contextual growing is applied to fully detect the burned area patches. As FireCCI51 was developed based on MODIS 250 m bands, only two spectral regions were available at that resolution: the Near Infrared (NIR) and Red. However, only the NIR band is used by the algorithm which ensures a high capacity for adaptation to other sensors.

The main objective of the current study is to successfully adapt this algorithm to the Sentinel 3A and B OLCI sensor to ensure continuity of the European BA products within the Copernicus Climate Change Service (C3S). The moderate spatial resolution of this sensor (300 m) and the almost daily global coverage provided by the two satellites give a suitable scenario to generate a global BA product. Besides, based on the experience gained with the development of a previous MERIS BA product within the Fire_cci project, an improvement of the results can be expected compared to the product based on MODIS data due to the improved separability offered by the OLCI NIR bands in burned area detection. As for the moment no SLSTR active fire product is available, the new algorithm will use both MODIS and VIIRS active fire information. Complementary, the Land Cover CCI product (v2.0.7 at 300 m), whose time series is being extended in the C3S, will be used to generate a fuel mask to eliminate the unburnable areas.

Area A - Board 132