The simulation of winter albedo in boreal and northern environments has been a particular challenge for land surface modellers. Assessments of output from the CMIP3 and CMIP5 climate models have revealed a large spread in simulated surface albedo and snow-albedo feedback strength in the boreal forest. Recent studies suggest that inaccurate representation of vegetation distribution, improper simulation of leaf area index, and poor treatment of canopy-snow processes are the primary causes of albedo errors. Most land surface models (LSM) represent global vegetation from a set of plant functional types (PFT), the number of which differs in each LSM. While several land cover (LC) datasets are commonly used to derive PFTs for use in LSMs and climate models, new LC and vegetation datasets with higher spatial resolution have become available in recent years.

The Climate Change Initiative (CCI) LC product at 300m resolution was developed by the European Space Agency (ESA) specifically to address the needs of the climate modelling community (ESA, 2017). The finer spatial resolution of the CCI product makes it inherently superior for LC mapping in heterogeneous landscapes than coarser resolution products, such as the GLC2000 dataset (1km resolution). GLC2000 is used to obtain PFTs for the Canadian Land Surface Scheme (CLASS) when employed as the land surface component in the climate and Earth system models of Environment and Climate Change Canada. CLASS is a physically based LSM with complete thermal and hydrological budgets. It has four broad PFTs: needeleaf forests, broadleaf forests, crops, and grass.

This study assesses the applicability of the CCI product for use in CLASS. Uncertainties in PFT mapping can be linked to uncertainties in LC datasets and/or the cross-walking (CW) tables used to convert the LC classes to PFT fractions. To assess the CCI LC dataset, we compare it with the GLC2000 and some high resolution LC datasets over Canada, as well as the vegetation continuous field (VCF) tree cover fraction data. We modify the default CW table from the CCI LC user guide (ESA, 2017) to generate the four CLASS PFTs, which are compared with those derived from GLC2000. We then conduct offline simulations with the latest version of CLASS (V3.6.1) using PFTs from CCI and GLC2000 respectively, and compare the simulated winter albedo with satellite observations. The results suggest that in comparison with the high resolution maps over Canada, the CCI product shows improved land cover distribution over that from the GLC2000 dataset. However, the CCI product appears to overestimate tree cover fraction along the taiga-tundra transition zone in northern Canada, which consequently results in a negative bias in the simulated winter albedo in that region. Reference sample data will be used to investigate the causes of the overestimation in tree cover fraction by CCI in the taiga-tundra transition zone. Methods to improve the CCI LC and thus PFT mapping using the high resolution datasets and the VCF tree cover fraction data will be explored and tested.

ESA: Land Cover CCI Product User Guide Version 2.0, available at: http://maps.elie.ucl.ac.be/CCI/viewer/download/ESACCI-LC-Ph2-PUGv2_2.0.pdf, last access: 10 November 2017.

Area A - Board 103

A new capacity to monitor land change across the conterminous U.S. will be launched by the U.S. Geological Survey (USGS) in 2019 to meet increasing demands for timely and frequent integrated land cover and land change information. Land Change Monitoring Assessment and Projection (LCMAP) is enabling nationwide tracking and characterization of annual changes in land cover, use, and condition and will be producing assessments of current and historical change. The goals of LCMAP include developing a capability to detect and characterize historical land change from data in the 1985-present Landsat archive, creating a suite of validated annual land-cover and land-change maps and area statistics, enabling the detection of land change in near real time, and modeling past and future change not covered by the satellite record. Historical land-change and land-cover LCMAP products are calculated from all available cloud- and shadow-free pixels in the USGS Landsat Analysis Ready Data (ARD) archive using the Continuous Change Detection and Classification algorithm developed by Zhu and Woodcock (2014). We are creating ten annual historical land change and land cover map products for the initial Version 1 release. In addition, the development of a low-latency monitoring capacity where change can be detected based on one to six months of observations is a high priority goal. The CCDC approach in the current historical LCMAP work flow needs a series of at least six high quality (i.e., cloud- and shadow-free) observations to support robust detection of change with low lead times. Methods for low latency change detection where existing models are compared to the most recent ARD observations to determine where conditions are changing rapidly are providing improved sensitivity for change detection. The USGS is defining requirements for an automated system that can alert relevant land managers to important emerging changes in cover or condition in their jurisdictions.

Area A - Board 104

The availability of Global Land Cover (GLC) maps has significantly increased over the last few years. This is a result of both the modern satellite missions capability of providing frequent pass, high-resolution, and global coverage Earth Observations (EO), as well as the emerging consideration of the land cover as a key variable into a number of pressing scientific and societal concerns, including natural resources management and climate change.
Examples of the above are a number of high-resolution products produced by different national and international EO programs. These include multi-classes GLC datasets with a spatial resolution equal or lower than 30 m e.g. the GlobeLand30 (GL30, 10 classes) and the Fine Resolution Observation and Monitoring of Global Land Cover (FROM-GLC, 8 classes). Higher resolution (20m or better) thematic datasets are also available, but with smaller number of classes e.g. the Global Human Settlement Layer (GHS, 2 classes) and the Global Urban Footprint (GUF, 2 classes).
On one hand, these GLC maps supply users with consistent land cover information, thus empowering reproducible land cover-based applications at a global scale. On the other hand, for regional applications the accuracy of GLC maps does not always meet the users' requirements, thus making the use of regional land cover maps, where available, often preferred. In this context, the accuracy assessment of GLC maps by means of comprehensive validation procedures still represents a pivotal task in order to promote the use of GLC map for local applications.

The evaluation of the classification accuracy of land cover maps is traditionally performed through comparisons with a reference dataset by means of a confusion matrix, from which a number of global accuracy indices can be derived. However, global accuracy indices do not provide insight into the spatial variability of disagreement between the target and the reference map, hence classification errors might not be distributed evenly across a region of interest. This effect might mask misspecifications in the validation procedures due to a not proper consideration of the errors spatial variability e.g. during the sampling design.
The aim of this work is to empower traditional accuracy assessment procedures with exploratory components for analysing the spatial variability of classification errors by means of spatial association statistics and mapping.
The study focuses on the validation of the GL30 that is the most frequently updated (2000, 2010, and 2015 announced) high-resolution GLC multiple-class map currently available. The validation is carried out at a regional scale for the Lombardy Region (Northern Italy). The DUSAF land cover vector maps at a scale 1:10000 is employed as the reference for the classification accuracy. Indeed, it represents the highest resolution land cover map available for the study area.

A preliminary accuracy assessment using traditional procedures based on the confusion matrix is performed by considering a number of indices suggested in the literature e.g. the Overall Accuracy, the Kappa, etc. The global accuracy of the GL30, according to Overall Accuracy index for the study area resulted in around 73%. The spatial characteristics of classification errors are initially investigated by using the join-counts statistics, which is a method working for categorical (binary) classes analysis. Basically, this statistic counts frequencies of having neighbouring pixels of the same type (i.e. pixels affected by a classification error) and compares them to expected in case of spatial randomness for the errors distribution. The results indicate a globally significant spatial association for the classification errors in the study area. However, the join-count informs nor on location neither on the type of detect spatial association patterns of errors.
Further spatial association analysis is performed by generating vector maps containing disagreements between the GL30 and the DUSAF. Classification errors are here expressed in terms of GL30 class area prevalence at each land cover class DUSAF polygon, computed by means of overlay statistics. The obtained disagreement variables (numerical) are used to compute Local Indicators of Spatial Association (LISA) in order to investigate both positions and magnitudes of spatial patterns for the classification errors. LISA consists of local statistics that allow detecting significant non-random patterns in the spatial distribution of a variable. The spatial association features generating the patterns can be described and mapped by looking at each LISA values and significance to provide insights into local disagreement issues within the study area. Spatial association tests are running considering both intraclass and interclass classification errors. Results consist of vector maps with associated LISA values and significances for each considered disagreement variables. These maps can be adopted to re-design part of the traditional accuracy assessment procedures or eventually correct them for local effects due to spatial association patterns.
The analysis is carried out by the exclusive use of free and open source GIS software such as QGIS and GRASS GIS, coupled with custom Python scripts based on powerful spatial analysis libraries such as the PySAL.

The general workflow for empowering traditional accuracy assessment with spatial association analysis is outlined and discussed. The implicit consideration of errors spatial variability within any accuracy assessment procedures is suggested as best-practice in order to properly and fully exploit GLC maps into high accuracy-demanding applications.

Area A - Board 105

In the context of the Climate Change Initiative (CCI) supported by the European Space Agency (ESA), the Land Cover (LC) team aimed at demonstrating the potential of the Copernicus Sentinel-2 (S2) mission for regional and continental LC mapping at a decametric spatial resolution to address the requirements of a high spatial resolution LC map expressed by the climate science community.

In 2017, large-scale and timely S2 processing capabilities were demonstrated by the research team by releasing a first 20-m prototype LC map of 2016 over the whole Africa with a legend of 10 classes, only 9 months after the last S2 acquisition. 72 TB of S2-A observations from December 2015 to December 2016 were downloaded, pre-processed and classified using the CCI LC Moderate Resolution classification chain. This pioneering experiment faced several challenges including the big data management issues, the pre-processing chain development for improved and cloud-screened S2 surface reflectances and the precise definition of a scalable land cover typology.

Building on this experience and on a fruitful collaboration with the Mexican Space Agency and Mexican scientific public entities active in statistics, LC and biodiversity mapping, a regional prototype map of Mexico and Central America was produced at a spatial resolution of 10m, using the full archive of S2-A and B between 2016 and 2018.

The decametric spatial resolution required significant methodological adjustments to the pre-processing such as the topography and adjacency effects that had to be taken into account in the atmospheric correction. The increase in S2 data acquisition provided by S2-B was fully exploited as time series to capture the landscape elements diversity in this region. The 10-m spatial resolution also imposed revisiting the stratification, training data collection and classification strategy to map the LC consistently through the strong seasonal and ecological gradients of this region. Finally, a regional workshop with local experts was organized in Mexico, with the aim of sharing experience on LC mapping with S2, collecting feedback on the classification and on building a validation database of more than 1500 objects, visually interpreted independently on very high-resolution imagery.

Classification and assessment results of the S2 prototype LC map of Mesoamerica will be presented and discussed, highlighting the potential of S2 for global land cover mapping at 10m, the lessons learnt and the remaining challenges ahead.

Area A - Board 106

We are living in a golden era for land cover mapping. The opening of the Landsat archive has lead to an enormous amount of new studies and possibilities enabling to study regional and global land cover and its changes at a spatial resolution of 30m since the eighties, herewith complementing the various existing medium resolution land cover maps. More recently, the advent of the COPERNICUS programme and the Sentinels has even raised the bar, as it is now possible to map land cover globally at 10m, using 5-daily optical time series complemented with C-band SAR data.

New possibilities however also lead to new difficulties. Land cover mapping efforts combining Sentinel and Landsat data require processing chains to cope with i) sensor differences such as differences in geometry, spatial resolution, spectral response, … ii) incomplete Landsat data archives for the eighties and nineties, iii) lower revisit times compared to medium resolution sensors and iv) drastically increased data volumes.

On top of this an additional challenge exists in ensuring consistency in various aspects. Firstly, multi-year processing chains should ensure temporal consistency, meaning that changes detected between different years should represent real changes and no classifier artefacts. Secondly, high resolution land cover maps should be consistent with their lower resolution equivalents, hence consistency should also play across scales both in the designated class as well as the used legend system.

In this presentation, we will cover these various challenges while mapping 30 years of land cover changes in Senegal using Landsat and Sentinel time series thus complementing the 100m land cover maps of the COPERNICUS global land service. We will focus especially on the mechanisms to ensure consistency across scale and time by comparing and validating various methods and resulting maps. Finally, we will zoom in as well on the particularities of the land cover in the Sahel region and the benefits of using high resolution time series for land cover mapping in this area.

Area A - Board 107

During the past decade, there has been a remarkable development in the availability, quality, spatial resolution and processing capacity of global coverage satellite imagery. This development has led to improved opportunities in quantifying, monitoring and characterizing Earth’s resources through multi-sensor approaches. The Eastern Arc forests are one of the 35 global biodiversity hot spots supporting diverse pools of flora and fauna, while simultaneously functioning as a key source of ecosystem services for the local communities. Their forest cover is highly fragmented and despite their wide protected status, deforestation and degradation remain as main threats to these forests. For sustainable management of the ecoregion, spatially explicit estimates of the coverage and condition of the remaining natural forest patches are needed. There have been multiple estimates on the largest remaining forest patches, but the distribution, quantity and quality of the smallest remnants has been largely unknown. It is estimated that fragments smaller than 10 ha contain a substantial amount, over 30,000 ha, of the remaining natural forests. Knowledge of their spatial distribution is crucial for understanding, monitoring and protecting the ecological functions of the Eastern Arc forests.

In this study, we estimate how combining optical and Synthetic Aperture Radar (SAR) satellite imagery will improve the detection of natural forests in the Eastern Arc Mountains of Tanzania and Kenya. We have used Sentinel-1 SAR, Sentinel-2 MSI, Landsat 5 TM, Landsat 8 OLI and ALOS PALSAR 2 imagery and derive spectral and textural variables for forest remnant detection. Our aim is to build on the previous work and test how well the combination of different data sources and machine learning algorithms can detect the smallest remnants (1ha-10ha), while simultaneously estimating how each datasets improve the detection process. Preliminary results show that multi-source satellite imagery provides best estimates on the forest patches and the small forest patches down to 1ha are mapped with good precision. This research will provide spatially explicit information on the distribution of old-growth natural forest patches in the Eastern Arc area, crucial for assessing the quality, monitoring the dynamics, modelling the ecological functions and supporting the sustainable management of this valuable biodiversity hotspot. Furthermore, this research brings information on the potential of synergetic use of optical and SAR imagery in detecting detailed land cover characteristics. The final results and discussion will be presented in the session.

Area A - Board 108

A priori knowledge of forest and non-forest areas is essential for the correct detection of forest degradation and deforestation. This paper presents the method developed with a minimum pre-processing to assess the extent of land use/land cover in Mato Grosso state, Brazil. The proposed method is based on Linear Spectral Mixing Model (LSMM) applied to PROBA-V images to derive vegetation, soil and shade fraction images for regional analysis. Fraction images can be used for mapping land use/land cover due to the following characteristics: a) vegetation fraction images highlight the forest cover conditions and allow differentiating between forest and non-forest areas similarly to vegetation indices such as the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI); b) shade fraction images highlight areas with low reflectance values such as water, shadow and burned areas; and c) soil fraction images highlight areas with high reflectance values such as bare soil and clear-cuts. We used 100 m composites of 5 days (S5-TOC - 5-daily global composites, Top-Of-Canopy) for the study area for regional analysis (totalizing 72 images). For the analysis, we built the composites corresponding to the three endmembers with the highest fraction values during the year 2015, 2016 and 2017 for the Mato Grosso State using 100 m dataset, based on the assumption that: a) the greatest values of vegetation composition highlight the areas that were used for agricultural activities during the year; b)the greatest values of shadow/water highlight the areas occupied by water bodies during the year and c) the greatest values of soil enhance areas without vegetation cover.. In that manner we could map automatically the land cover easily identified such as forest, main crops, pasture and Savannas vegetation in the study area. The Mato Grosso state southern flank of the Brazilian Amazon region is experiencing high deforestation rates and it has been observed forest degradation by fires particularly during extreme dry years and in this context, Mato Grosso is extremely interesting for the purposes of this study. PROBA-V images with spatial resolution of 100 m acquired in the dry season, between June and October, were used to assess the land cover changes due to deforestation and forest degradation by fire. In a first step, a forest/non-forest mask was extracted from PROBA-V fraction images acquired in 2015 to avoid detecting and mapping previous disturbed areas in the following years (2016 and 2017). Then deforested areas and burned areas were identified and mapped from the soil fraction and shade fraction images, respectively for the 2016 and 2017 years. Finally, forest areas degraded by fires were mapped by combining a forest mask with the resulting burned area and recent deforestation maps. Our results show the land use land cover map, new deforested areas and forest degraded by fire during the year 2015, 2016, and 2017. The results are validated by comparing with results derived from one Landsat OLI image acquired in 2017 and supported by fieldwork information collected in 2017. Therefore the proposed approach can be potentially used operationally for detecting, mapping and monitoring land use/land cover changes using PROBA-V 100 m resolution datasets for, regional and local analysis.

Area A - Board 109

Accurate and reliable information on the spatial distribution and extent of agriculture is highly relevant for assessing production prospects, to support agricultural monitoring, for food security assessment and as essential input for decision-making in agricultural management and early warning systems, such as ASAP (Anomaly hot Spots of Agricultural Production) of the Joint Research Centre. In this context, agriculture is not only defined as the land that is suitable for crop, but also the land that is suitable for livestock production, either naturally grown or cultivated.

Remote sensing significantly contributes to monitoring the spatial extent and distribution of agriculture, either independently (i.e. GMRCA) or as part of a multi-class land cover classification (i.e. GlobCover). In this sense, Global land cover datasets provide information based on homogeneous input data and methods for large areas, but pay the price of lower spatial resolution (usually from 300 to 1 km). On the other hand, regional and national datasets are able to provide a higher spatial detail but are affected by lower level of availability and geographic comparability. The recent ESA’s Sentinel images together with the opening of the Landsat archive allow moving towards a new era of higher-resolution products such as the 30 m Globeland30 and the S2 prototype Land Cover-ESA.

Despite the availability of such wide number of products, previous research highlights considerable disagreement among them, especially with respect to agriculture for continental Africa. Hybridization or data fusion, defined as the process of integrating existing land cover datasets into a single one, represents a worthy alternative to reconciling the strengths of individual sources and overcoming some of its main limitations.

This research follows on previous products based on land cover hybridization and targeting the specific information needs of crop and rangeland monitoring for food security early warning by incorporating the latest high-resolution datasets (i.e. CGLC-LC100 V1.0 Copernicus) and by identifying the most adequate low-resolution land cover products in those countries where high-resolution products are not available.

The proposed methodology exploits a powerful multi-criteria decision-making (MCDM) method such as TOPSIS (Technique for Order Preference by Similarity to an Ideal Solution). To do this, each dataset is evaluated and weighted, at country-level, according to the following five criteria: timeliness criterion, spatial resolution criterion, FAOSTAT criterion or comparison with statistics, accuracy assessment criterion and expert knowledge criterion.

The final product is two hybrid agricultural static maps at 250m, one for cropland and another for grassland for the nominal year 2016. These masks represent for each country, the most recent and highest quality land cover/use data for agricultural monitoring.

Area A - Board 111

The interferometric SAR (InSAR) coherence represents the normalized cross-correlation coefficient between the interferometric pair of acquired images and is the key-quantity for assessing the quality of the interferogram. Moreover, it is a valuable input for a variety of other scientific applications, such as land usage classification, biomass estimation, soil moisture monitoring or forest detection. TanDEM-X is currently the only spaceborne SAR system able to provide, every 11 days, high-quality single-pass interferograms, which are not affected by temporal decorrelation. The bistatic information is fundamental for properly separating all decorrelation contributions which depend on the system parameters (such as SNR, quantization, and ambiguities) and on the acquisition geometry (volume and baseline decorrelation). Furthermore, both temporal and volume decorrelation phenomena affect the coherence measurement in repeat-pass systems, such as Sentinel-1, and cannot be separated, since different combinations of the two can be ascribed to different scattering mechanisms. The aim of this work is to analyze TanDEM-X repeat-pass interferometric time-series, aimed at understanding and modeling the temporal decorrelation at X band for different land cover classes. By exploiting bistatic systems such as TanDEM-X, the combined use of single- and repeat-pass data, derived from the same multi-temporal stack, allows for the isolation of volume and temporal decorrelation contributions. In this sense, the information given by TanDEM-X is unique: for instance it could support the modeling of InSAR coherence temporal dynamic for other sensors as well, by considering the link between different bandwidths. This topic will be investigated by comparing the temporal evolution of the coherence in overlapping time-series acquired by TanDEM-X (X band) and Sentinel-1 (C band) over the same ground area. Finally, a proper modeling of temporal decorrelation can also help to improve the accuracy of the currently developed techniques based on the analysis of the repeat-pass coherence, in particular when applied to land cover classification purposes.

Area A - Board 112

The Lancang-Mekong river, known as the Lancang river in China, the Mekong reiver outside China. The Lancang-Mekong Basin is a trans-boundary river with an area of 795,000 km2, including territorial parts of six countries: namely China, Myanmar, Laos, Thailand, Cambodia, and Vietnam. With a total length of over 4350 km, the Lancang-Mekong River is the longest river in Southeast Asia. It originates from the glacier melting of Qinghai Tibet plateau at the elevation of 5200 m, and eventually flows into the South China Sea at Mekong Delta in Vietnam. More than 72 million people benefit from this river. Consequently, the Lancang-Mekong basin has an outstanding practical significance for the ecological and economy development of alongshore area. However, the current land use/cover in the Lancang-Mekong river basin is in a very critical situation. Large patches of primary and secondary forests have been destroyed in Laos, Myanmar and Cambodia. Crop rotation is replaced by single cropping of rubber, cashew, sugar cane, and eucalyptus etc. Social and economic transformation, urbanization and interregional cooperation brought by increasing human activities also play an important role in land use/cover change in the Lancang-Mekong river basin. The land use/cover change have influenced climate, precipitation, the energy balance, carbon budget, and hydrological cycle in the basin.
Remote sensing has long been recognized as an effective tool for broad-scale(such as global scale, regional scale and basin scale) land use /cover mapping. At present, eight land cover thematic datasets( such as USGS with 1km, UMD with 1km, BU with 1km, GLC2000 with 1km, Globcover 2005 with 300m, GlobCover 2009 with 300m, GlobCover 2010 with 250m, Globeland30 with 30m) at a global scale have been developed with resolution ranging from 30m to 1km. In recently years, remote sensing scientists are interested in land use /cover change, climate variation, and urbanization in the Lancang-Mekong river basin. Remote-sensing technology has the potential to monitor the environmental changes in basin scale. However, the Lancang-Mekong Basin is very large, and covers numerous climate zones and eco-regions, and needs seven MODIS tiles, or over 50 Landsat frames to cover the complete north–south-trending basin. Furthermore, the almost persistent cloud cover over the Lancang-Mekong Basin for large parts of the year, and optical remote sensing images are unavailabe in part of regions. The landscapes have complex spectral and textural characterization in the Lancang-Mekong river Basin. Because of these factors, there is reported about high resolution(≤10m) land use/cover products in the Lancang-Mekong river Basin in recent years. As new satellites and sensors become avaiable. the Sentinel-2A/B are optical satellites, which respectively launched in 2015 and 2017. The Sentinel-2 has multi-spectral data with 13 bands in the visible, near infrared and shortwave infared with respectively spatial resolution of 10m, 20m, and 60m. Multi-source remotely sensed images with high temporal, spatial and spectral resolution in the Lancang-Mekong river basin have obtianed by the ESA Sentinels satellites in recently years.
Lots of research has shown that the land use /cover information play a role in climate variation, ecology and environment destroy, and naural hazards. The land use/ cover is fundamental information for natural resource management, environmental change studies, urban planning to sustainable developemnt, and many other societal benefits in the Lancang-Mekong river basin. Bacause of region area large, complex topography, cloudy and rainy environment in the Lancang-Mekong river basin, so far, the high resolution land use/cover products have not presented. The relation research has became urgent. More specifically, the whole research may include the following some parts. (1) employing Sentinel Application Platform (SNAP) toolboxes to pre-process the data sets over the Lancang-Mekong river basin (2) Developing object-oriented random forest (OORF) Classifcation algorithm for mapping Land use/cover in basin scale . (3) Mapping high resolution land use/cover in the basin using the presented alogrithms in the basin.

Area A - Board 113