The U.S. Midwest state of Minnesota has over 10,000 glacial lakes distributed across nearly 7° of latitude. Over the past several decades, aquatic communities in these temperate lakes have come under increasing environmental pressure as seasonal ice conditions have become more variable and unpredictable in response to warming winter and springtime temperatures. Systematic and quality controlled in situ lake ice observations of freeze- and break-up timing, duration, and thickness remain sparse at both local and regional scales across Minnesota. In the absence of in situ observations, satellite remote sensing offers an alternative measurement source to retrieve lake ice physical states and conditions. Because cloud cover is often persistent during winter and spring seasons at more northerly latitudes, microwave radar imaging can supplement optical remote sensing of snow and ice cover on lakes because of its all-weather performance, and its capability to provide spatially explicit information content on both surface and volume scattering. Since late 2016, the European Space Agency (ESA) / Copernicus Sentinel-1B synthetic aperture radar (SAR) mission has provided systematic, no-cost access to Interferometric Wide Swath (IW) C-band radar imagery over Minnesota’s vast lake system at a minimum revisit of frequency of 12 days. In this communication, we describe our technical approach to using SAR imagery from Sentinel-1B to characterize and interpret VV+VH dual polarisation backscatter over Minnesota lake ice during winter and spring seasons of 2017-2018 to inform under ice limnology applications.

During the winter 2017-2018 season, we used Google Earth Engine (GEE) to access Sentinel-1B imagery and visualize the development of ice cover on Minnesota’s lakes. This exploratory tool enabled our team to identify unique lake ice conditions from ground range detected (GRD) backscatter returns at a regional scale. We used Sentinel-1B imagery along with remotely sensed late summer water quality data derived from Sentinel-2 imagery to assist in designing two field campaigns during mid and late winter timeframes to ground truth ice conditions observed in Sentinel-1B imagery. These field campaigns were part of a larger project to increase our understanding of under ice limnological conditions. For the limnological sampling and experiments, we targeted two mesotrophic, two hypereutrophic and two dystrophic lakes identified from prior knowledge and the water quality maps created from Sentinel-2 imagery. For ice thickness and snow depth measurements, we targeted lakes from southern to northern Minnesota and used the GEE SAR maps to identify differences in ice conditions for field identification. During the late winter field campaign, we targeted the Sturgeon Lakes chain in northern Minnesota for more intense in situ field measurements including snow depth on ice, snowpack stratigraphy, ice surface and column type, and ice thickness in near coincidence (within 24 hours) with Sentinel-1B acquisitions. We also deployed the University of Michigan’s ground-based Wideband Autocorrelation Radiometer (WiBAR) X-band instrument to test a new approach for retrieving lake ice thickness using a passive microwave time-delay technique. The Sturgeon Lakes chain was targeted because of its large gradient in water clarity conditions (due to large differences in colored dissolved organic matter (CDOM) in close proximity. In addition to ground truthing Sentinel-1B VV+VH dual polarisation backscatter returns during the late winter field campaign, our team had a science objective to collect a suite of under ice measurements designed to characterize aquatic communities and in lake primary productivity in lakes with different optical water types.

Our initial results indicate that Sentinel-1B’s GRD VV surface backscatter provides useful information to differentiate between liquid water and lake ice formation. We observed an increasing backscatter amplitude signal from November to early February that corresponded to gradual ice thickening after initial freeze-up in mid-December. Comparisons between mid (early February) and late winter (early March) in situ field measurements confirmed that ice thickness increased in southern and central Minnesota but remained stable for the Sturgeon Lakes chain in northern Minnesota. Clear ice with little to no bubble structure (bubbles within the ice can modulate SAR backscatter amplitudes) was observed in the ice column at all measured lakes during both mid and late winter field campaigns. Thin layers of grey or snow ice near the ice-snow interface was observed in late winter for multiple lakes indicating the occurrence winter melt and refreezing cycles. Snow depth increased more so across northern Minnesota than southern and central Minnesota from mid to late winter. For the Sturgeon Lakes chain, we did not observe a marked increase in VV backscatter amplitude after early February. Our analysis also confirmed that the GRD VH backscatter provided much less information content as a single measurement on winter and spring lake ice conditions as backscatter amplitudes were of relatively low quality and close to Sentinel-1B’s noise floor. The passive microwave time-delay technique for recovering lake ice thickness from the WiBAR’s V and H polarisations is more straight forwards than complex radiative transfer modeling approaches, and performed exceptionally well, achieving a retrieval accuracy of less than 1 cm of the observed ice thickness for multiple lakes.

Concurrent biological measurements showed that lake snow and ice characteristics interacted with water column properties (such as colored dissolved organic matter and nutrient concentrations) to determine plankton productivity in the study lakes. Nutrient-rich lakes showed remarkably high rates of primary production even when covered by snow and supported active and dense zooplankton communities. On the other hand, lakes with high CDOM had very low primary production rates, regardless of snow cover conditions. These results indicate that although ice and snow cover conditions have large effects on lake biology, different lake types respond differently to seasonal ice cover and may show idiosyncratic responses to future changes in ice cover duration and snowfall patterns. Moving forward, our team plans to further investigate the full suite of in situ lake ice and under ice measurements collected during winter 2017-2018 field campaigns to evaluate the usefulness of SAR polarimetry techniques to exploit Sentinel-1B’s VV+VH dual polarisation phase information that preserves volume scattering in single look complex (SLC) backscatter returns.

Area A - Board 88

In this work the synergistic use of Sentinel-1 and 2 combined with the System for Automated Geoscientific Analyses (SAGA) Wetness Index in the content of land use/cover (LULC) mapping with emphasis in wetlands is evaluated. A further objective has been to a new Object-based Image Analysis (OBIA) approach for mapping wetland areas using Sentinel-1 and 2 data, where the latter is also tested against two popular machine learning algorithms (Support Vector Machines - SVMs and Random Forests - RFs). The highly vulnerable iSimangaliso Wetland Park was used as the study site. Results showed that two-part image segmentation could efficiently create object features across the study area. For both classification algorithms, an increase in overall accuracy was observed when the full synergistic combination of available datasets. A statistically significant difference in classification accuracy at all levels between SVMs and RFs was also reported, with the latter being up to 2.4% higher. SAGA wetness index showed promising ability to distinguish wetland environments, and in combination with Sentinel-1 and 2 synergies can successfully produce a land use and land cover classification in a location where both wetland and non-wetland classes exist.

Area A - Board 89

Monitoring of conservation status of the habitats listed in Annex I of the EU Habitats Directive (92/43/EEC) has to be reported to the Commission every six years according to Article 17 of the directive. This is mandatory at the EU NUTS level 2. In order to supply accurate data to report the conservation status of the forest habitats in the Balearic Islands (NE Spain) within the period 2013-2018, we have combined a directed field survey, several multi-source data processing and modelling techniques. The objective is twofold: 1) to determine the distribution/range of the target habitats and 2) to assess the changes in conservation status based on a trend analysis of vegetation indices.

The Balearic archipelago is located in the western Mediterranean Sea, with four principal islands (Majorca, Minorca, Ibiza and Formentera) and other minor islands and islets. The climate is mainly Mediterranean although meso- and supra-Mediterranean climates occur in the mountains of Majorca. The predominant forest vegetation consists of the evergreen forests, woodlands and sclerophyllous shrublands dominated by Quercus ilex, Pinus halepensis, Olea europaea, Pistacia lentiscus and Juniperus turbinata.

A comprehended field campaign aimed at georeferencing the habitats and at assessing the conservation status was carried out over the study area during September and October of 2018. As a support for the assessment of changes in conservation status, we applied a trend-change detection analysis of NDVI Landsat-based temporal series. Image composites were generated using all available images from the USGS archive, in particular, images from Landsat-7 ETM+ and Landsat-8 OLI with cloud cover less than 80% and corrected to surface reflectance (using LEDAPS algorithm). Four scenes (path/rows) of the Landsat WRS-2 comprised the study area and the date range for eligible images was from May to September within the 2013-2018 period. An average of 175 images per scene were used to generate annual gap-free Landsat surface reflectance and NDVI composites using the maximum NDVI approach. The R-package “greenbrown” (Forkel et al. 2013; DOI: 10.3390/rs5052113) was then applied on the NDVI time series for trend change detection. Trend results were validated using the field data collected together with other field datasets available (data from fire statistics, forest management statistics and annual forest health surveys provided by the Government of Illes Balears).

Area occupancy of the habitats was modelled through Maxent at the spatial resolution of 20 m considering climatic variables, vegetation indexes and forest structure (Álvarez-Martínez et al. 2017; DOI: 10.1111/2041-210X.12925). Data from the habitats presence was obtained from high-resolution vegetation community mapping, most of them at the 1:15,000 scale. The Spanish National Forest Inventory was also used to obtain the presence of the main forest types. Climatic variables were sourced from the Climatic Digital Atlas of the Balearic Islands (200 m resolution). A set of 22 vegetation indexes were obtained from Sentinel-2 imagery (20 m resolution). We used most suitable observations (less % of cloud cover) for each pixel location from all available Sentinel-2 images acquired within July 15 2018 ± 30 days. Then, we generated monthly cloud-free gap-free L2A Sentinel-2 composites for the six tiles covering the study area using maximum NDVI approach. Forest structure was assessed through processing LiDAR data form the National Plan of Aerial Ortophotography open database (PNOA in Spanish) at 0.5 points/m2 resolution. After reducing collinearity drawbacks, we run Maxent for each habitat. Maxent applies a machine-learning technique called maximum entropy modelling and the fitted model expresses a probability distribution where each grid cell has a predicted suitability of conditions for the habitat (Phillips et al. 2006; DOI: 10.1890/07-2153.1). The predicted suitability and the omission and commission errors were combined to obtain the final habitat occupancy area.

Results support that combining multi-source data and modelling techniques are a cost-effective tool to report the conservation status of the habitats at the regional scale (EU NUTS 2 level). These techniques also provide powerful open-source easy-to-update methods to accomplish the mandatory report of conservation status based on Article 17 of the EU Habitats Directive, particularly for widespread habitats.

Area A - Board 90

The demand to characterize the terrestrial carbon pool and track its dynamics resulted in multiple attempts to model and retrieve forest aboveground biomass with spaceborne remote sensing datasets. Synthetic Aperture Radar (SAR) observations are possible regardless of solar illumination and the signal measured relates to some extent to structural properties of a forest, thus allowing in theory for repeated mapping and monitoring of the forest biomass. At C-band, the signal backscattered presents weak sensitivity to biomass, primarily in mature forests, and is affected by the environmental conditions at the time of image acquisition. Nonetheless, we have demonstrated in previous studies that the accuracy of the retrieved biomass increases and spatial patterns of biomass distribution are well represented when combining estimates of forest biomass from multiple acquisitions (Santoro et al., 2011, 2013 and 2015). Key for reducing errors in biomass estimates from C-band SAR backscatter observations is an algorithm that maximizes the information content on biomass, which is temporally stable, while reducing the component due to environmental conditions, which is time-dependent.
Among all spaceborne SAR datasets acquired so far, the C-band SAR data pool by Envisat ASAR and Sentinel-1 are a unique asset to support the estimation of biomass and its dynamics globally. Since 2002, wall-to-wall repeated data coverage has been achieved at high and moderate resolution and the data are freely available, thus fostering the understanding of the backscattered signal at C-band in forests and the development of retrieval algorithms. In this presentation, we illustrate achievements, current development and challenges on global biomass estimation with C.band SAR backscatter data based on 10 years of Envisat ASAR data and one year of Sentinel-1 data. In total, almost 300,000 ASAR scenes from Image Mode, Wide Swath and Global Monitoring mode and 250,000 Sentinel-1 scenes from the Interferometric Wide Swath mode have been pre-processed and assembled into a global database of C-band SAR observations. These are then ingested by the BIOMASAR algorithm to provide a temporal sequence of biomass estimates across the 2000 decade and for 2017. The BIOMASAR algorithm inverts a model of the forest backscatter relating this observable to the biomass of a forest with simple exponential function based on radiative transfer theory. Model training is typically undertaken with in situ observation. In a global perspective, this approach is not viable as the uncertainty of the model parameters would be large due to lack of reference data. BIOMASAR implements a self-calibration approach, in which the model parameters are estimated by identifying areas of specific canopy density. In addition, we have separate models to express the attenuation of the signal into forest canopies at a given wavelength modeled per ecological zone. While the retrieval is based on SAR backscatter data, optical imagery and LiDAR-derived metrics in the form of canopy density and height metrics are necessary to support the forest backscatter model training. Data contemporary to SAR observations are important to avoid calibration errors due to potential forest cover changes.
In this presentation, we shall present the advantages of combining multiple observations in a single retrieval scheme and indicate where gaps are still necessary (e.g., still limited LiDAR coverage, coarse representation of tree species). The assessment of the first global map of biomass from ASAR data acquired in 2010 indicates large errors and uncertainties at pixel scale; these decrease substantially when aggregating estimates at the level of provinces and countries (Santoro et al., 2015). We are currently investigating the temporal consistency of the ASAR-based estimates of biomass between 2002 and 2012 and will present results at the conference. Initial attempts to estimate biomass from Sentinel-1 data in boreal and tropical forest confirm the findings based on ASAR data. At the conference, we plan to present results from multiple sites and link with estimates from ASAR data in order to assess the possibility to quantify biomass pool dynamics over a decade. Results will be compared with existing statistics from national forest inventories and related data products (Liu et al., 2015).

Area A - Board 91

Global burned area maps have been generated systematically with daily, coarse resolution satellite data (Giglio et al. 2013; Roy et al. 2008). Since 2013, three new moderate resolution satellites with global land monitoring capabilities were launched, namely the USGS Landsat-8 carrying the Operational Land Imager (OLI) sensor and the COPERNICUS Sentinel-2A and Sentinel-2B carrying the Multi-Spectral Instrument (MSI) sensor. The production of moderate resolution (10 - 30 m) global burned area products would meet the needs of several user communities: improved carbon emission estimations due to heterogeneous landscapes and for local scale air quality and fire management applications (Mouillot et al. 2014; van der Werf et al. 2017).
Fire causes a land-cover change spectrally visible for a period ranging from a few weeks in savannas to over a year in forested ecosystems (Chu and Guo 2014; Roy et al. 2010) and because clouds, smoke, and other optically thick aerosols limit the number of available observations (Roy et al. 2008; Smith and Wooster 2005). The limited number of acquisitions obtained by moderate resolution satellites (10-16 days), compared to coarse resolution satellites (1-2 days), translates into potential burned area omission errors, because of the combined effects of the fire impermanent changes on the spectral signal and the reduced number of valid acquisitions due to cloud obscuration (Boschetti et al. 2015). The data fusion of the recently launched Sentinel-2A/B with the Landsat satellites has been proposed to overcome these limits (Boschetti et al. 2015; Roy 2015).
The present study estimates the probability of moderate resolution sensors and virtual satellite constellations (including Landsat-7/8, Sentinel-2A/B) to provide sufficient observations for burned area mapping globally, and by ecosystem.
We estimated the duration of the persistence of the signal associated with burned areas by combining the MODIS Global Burned Area and the Nadir BRDF-Adjusted Reflectance Product by characterizing the post-fire trends in reflectance to determine the length of the period in which the burn class is spectrally distinct from the unburned and, therefore, detectable. Results showed that, globally, the median burned area persistence time was estimated in 29 days and 86.6% of the global area, as detected by MODIS, can be detected accurately only for up to 48 days (Melchiorre and Boschetti 2018).
The full 17 year record of MODIS-Terra daily cloud detections were subsequently used to estimate the probability of cloud cover globally for each day of the year. The cloud probability was used at each location to estimate the minimum revisit time needed to obtain at least one cloud-free observation within the duration of the persistence time of burned areas. As complementary results, the expected omission error due to insufficient observations was estimated for each of the satellite combination considered making use of the geometry of acquisition for each of the sensors included in the virtual constellation.
The results of this work can be used as indications on the development of global algorithms for burned area detection at moderate resolution in terms of restriction on the temporal revisit time needed for global, systematic mapping of burned areas.

Area A - Board 92

With Sentinel-2B launched on March 2017, two twin Sentinel-2 satellites are providing global coverage every five days, therefore allowing the simultaneous study of the progression of active fire fronts with high spatial (up to 10m) and temporal resolution. For the large long-duration forest fire, the detection of active fires can support operational tactical fire management with spatially-explicit fire perimeter information. The objective of this research is to evaluate the effectiveness of Sentinel-2 MSI SWIR bands for active fire detection. The radiance of fire-sensitive bands of Sentinel-2 is compared comprehensively. A difference SWIR2-based index (∆AFI) is proposed to effectively monitor the active fire frontiers during the wildfire progression with obvious discrimination of topography effect. Updates based on the time-series burned area and visual inspection of the data corroborated indicated Sentinel-2 MSI spectral indices are adequate to discriminate active fire frontiers. Complemented by other missions of similar scope (e.g. Landsat-8/OLI), the integration of all available instruments could provide satellite-based spatially refined active fire detection data at a near-real-time wildfire monitoring.

For many years, the biomass burning related science has relied on coarser spatial resolution (≥ 1 km) satellite data to map and monitor fire-affected areas from sensors like Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS), MEdium Resolution Imaging Spectrometer (MERIS) onboard ENVISAT between 300m and 1 km [1, 6]. Their daily revisit cycle is suitable to observe the active fire and burn scars, however, the relatively coarse spatial resolution makes them unavailable to capture the small burned area extent and small active fire flame. Progression of fire fronts varied between individual fires, but in most cases remained within the scale of a few 30-m pixels [2].

As study areas, we used the 2017 Thomas Fire in the western United States, that burned approximately 114,078 ha and lasted around one month. A time-series Sentinel-2 MSI image (processing level 1C) acquired from 28th November to 28th December 2017 were filtered after the filter processing including cloud coverage filer and study area bounds filter. Seven Sentinel-2A and 2B MSI images were obtained as the data materials. Meantime, the biggest wildfire in California, Mendocino Complex Fire burned 185,800 ha from July 28th to November 7th, 2018, will also be studied and the multi-way active fire developments will be under investigation.

Good penetrability makes SWIR channel see through the fire smoke to detect thermal anomalies in benefit of high wavelength. Generic application products based on SWIR to detect the actively burning fire perimeters using ASTER and Landsat-7 Enhanced Thematic Mapper Plus were demonstrated [3, 8]. One approach built on the SWIR channel but complemented by visible and near-infrared channel data was proposed in [7] using Landsat-8/OLI data. Regarding the Sentinel-2, similarly, the fire-sensitive SWIR channel 12 with 20 m spatial resolution will be expected to exploit the emissive component in the 2.2 μm spectral wavelength. Single SWIR radiance of each image during the fire progression will be scaled between 0 and 1 first. The water is masked using the Hansen Global Forest Change Data [4]. To tackle the potential false alarm associated with other facilities like large reflective rooftops and heat exhausts and neutralize the effect of terrain, the difference of AFISWIR2 values between pre- and the ongoing-fire image is calculated. The histogram is calculated to determine the threshold to cluster the hotspot pixels as Figure showed. The index performance can be measured through time-series analysis, which updates the burned area based on the dNBR index of the SWIR1&2. Active area in previous imagery should be classified as the burned area in the next one. Unsupervised classification K-means is utilized to extract the burned scar.

Finally, to compare the performance of with Landsat-8/OLI data further, we also take the digital number (DN) folding into consideration. For Landsat 8/OLI’s, in some extreme conditions when large and/or very high-temperature fires occupy the pixel footprint, the increased radiance input to the detector can result in the excess of the maximum resolvable radiation on fire-sensitive SWIR channel 7, which causes pixel saturation [5, 7]. But, the same wavelength SWIR channel in Sentinel-2 does not show obvious DN folding, which implies a higher tolerance of analog-to-digital converter for Sentinel-2 than Landsat 8/OLI. The preliminary result shows that Sentinel-2 MSI SWIR bands have good penetrability for fire smoke, and is suitable to detect the active fire points using ∆AFI without obvious DN folding.

References
[1] Itziar Alonso-Canas and Emilio Chuvieco. Global burned area mapping from envisat-meris and modis active fire data. Remote Sensing of En- vironment, 163:140–152, 2015.
[2] Ivan A Csiszar and Wilfrid Schroeder. Short-term observations of the temporal development of active fires from consecutive same-day etm+ and aster imagery in the amazon: Implications for active fire product validation. IEEE Journal of Selected Topics in Applied Earth Obser- vations and Remote Sensing, 1(4):248–253, 2008.
[3] Louis Giglio, Ivan Csiszar, A ́goston Rest ́as, Jeffrey T Morisette, Wilfrid Schroeder, Douglas Morton, and Christopher O Justice. Active fire de- tection and characterization with the advanced spaceborne thermal emis- sion and reflection radiometer (aster). Remote Sensing of Environment, 112(6):3055–3063, 2008.
[4] M. C. Hansen, P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina, D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini, C. O. Justice, and J. R. G. Town- shend. High-resolution global maps of 21st-century forest cover change. Science, 342(6160):850–853, 2013.
[5] Ron Morfitt, Julia Barsi, Raviv Levy, Brian Markham, Esad Micijevic, Lawrence Ong, Pat Scaramuzza, and Kelly Vanderwerff. Landsat-8 op- erational land imager (oli) radiometric performance on-orbit. Remote Sensing, 7(2):2208–2237, 2015.
[6] Florent Mouillot, Martin G Schultz, Chao Yue, Patricia Cadule, Kevin Tansey, Philippe Ciais, and Emilio Chuvieco. Ten years of global burned area products from spaceborne remote sensing—a review: Analysis of user needs and recommendations for future developments. International Journal of Applied Earth Observation and Geoinformation, 26:64–79, 2014.
[7] Wilfrid Schroeder, Patricia Oliva, Louis Giglio, Brad Quayle, Eckehard Lorenz, and Fabiano Morelli. Active fire detection using landsat-8/oli data. Remote sensing of environment, 185:210–220, 2016.
[8] Wilfrid Schroeder, Elaine Prins, Louis Giglio, Ivan Csiszar, Christopher Schmidt, Jeffrey Morisette, and Douglas Morton. Validation of goes and modis active fire detection products using aster and etm+ data. Remote Sensing of Environment, 112(5):2711–2726, 2008.

Area A - Board 93

The climate change context, along with the increasing scarcity and deteriorating quality of the water resource leads to monitor different components of the water cycle. A particular attention has to be paid to some areas, particularly, agricultural lands which represent about 70% of the water consumption, and coastal areas subject to strong interactions with land. Thermal infrared (TIR) data from space are well adapted to these purposes, but the spatial variability of the surface requires that the complexity of both physical and biological processes involved must be assessed at a smaller scale which corresponds to the scale at which decisions concerning water management or implementation of policies devoted to the mitigation of climate change effect are effective. In addition, surface fluxes show short-time scale variability, which requires frequent observations to be done. The need of space borne systems combining both high spatial resolution and high revisit frequency in thermal infrared (TIR), which do not exist today, is now largely recognized by the scientific community and end-users, especially as SENTINEL and RESOURCESAT missions now provide high quality complementary data in the optical domain. To fill this gap a project, TRISHNA (Thermal infraRed Imaging Satellite for High-resolution Natural resource Assessment), is currently in the feasibility assessment phase, conducted by the French Space Agency (CNES) and the Indian Space Research Organization (ISRO).

Two scientific objectives drive the mission specifications, (i) monitoring of ecosystem stress of the continental biosphere and of water use with applications to agriculture and hydrology, and (ii) monitoring of coastal and continental waters. Four complementary goals enlarge the community aggregated around the project: (iii) urban microclimates monitoring (urban heat islands, mitigation of heat waves effects…), (iv) applications to Solid Earth/geology (detection of thermal anomalies, volcanology, peat fires…), (v) cryosphere monitoring (glaciers, polar regions…), and (vi) applications to atmosphere (water content, clouds…). These objectives and the expected applications will first be briefly reviewed at the symposium.

The main mission specifications will then be presented. Additionally to previous work conducted to consolidate the revisit and resolution specifications, emphasis will be put on recent studies made for inventorying factors possibly perturbing surface temperature measurements. Significant advances obtained for characterizing and modelling TIR directional anisotropy and thermal hot spot effects will first be described and their impact on the choice of orbit discussed. Similarly, an original study of the impact of atmospheric turbulence on the accuracy of LST -largely ignored by the community till date- will be presented; it reinforces the need of high revisit, and allows to estimate the errors on LST measurements. An overpass time in the early afternoon, around 13:00 LST has been carefully justified to cope with the different objectives of the mission and to optimize the accuracy on retrieved fluxes. A baseline spectral configuration of 4 TIR channels within the range 8 - 12 µm is under study, which makes possible implementing both split-window and temperature-emissivity separation algorithms. The main specifications of VNIR instrument embarked aboard the same platform are presented, and the need of 6 bands, 4 in the VNIR (blue, green, red, near infrared at 485, 555, 670, 860 nm) and 2 in the SWIR (1.38 and 1.61 µm) justified. To cope with the requirement of global coverage at Equator with a single satellite, a revisit of 3 days is selected. The nadir resolution of 50 m is binned at 1 km over open ocean. The instrumental studies undertaken will be briefly presented.

The programmatic context of existing missions will be analyzed, only TRISHNA and the ESA LSTM mission providing high spatio-temporal capacities. The forthcoming phases of the CNES-ISRO TRISHNA project will finally be briefly mentioned, for a launch date foreseen at the 2024-2025 horizon.

Index Terms - Thermal infrared, satellite, TRISHNA, water stress, water management, hydrology, evapotranspiration, coastal water

Area A - Board 94

The experience gathered building and experimenting a “Witness” system that collects images about ongoing natural disasters in the first phases of emergency events from social media, and Twitter in particular, is presented. Through a set of filtering tools, which include also crowdsourcing tasks, the relevant images are provided to emergency operators and displayed on a crisis map.

Developments in the information and communication technology have an enormous potential to support emergency operators. Humanitarian organizations can gather data from a wide range of sources, make informed decisions and exchange real-time information with others.
During the Haiti 2010 Earthquake, in the immediate response and later recovery efforts, a new group was offering support to those working in the field: a group of ‘digital responders’. These online groups scoured social media searching for messages to help build situational awareness, mapped roads and areas from satellite imageries supporting the logistical operations and offered translations from Creole to English and back to support communication. The 2010 Haiti earthquake proved the value and potential of technology in disaster response, and shortly after, innovation and tech development for disaster response took a flight.

While these technologies have been used, and indeed seem to have a lot of potential to support disaster response, there is a difference between the potential of the developments and the actual impact. A key element in order to leverage the potential is a thorough understanding of what is actual being used in the field and why. It is of course appealing to show what is possible, and this generates a lot of attention, but in order to make a real impact, support the responders and the affected community a good understanding of what is being used ‘in the field’ and why is needed.

Recent cases in which a Witness platform, with its real time images extraction and filtering functionalities, could have made a difference or could have had added value are many, e.g.
the Greece fires in 2018, where initial focus on a ‘minor’ forest fire West of Athens while the blaze in the East was the deadly one, in the Florence and Michael storms in 2018, to provide a faster focus on storm surge with flooding than wind damage, in the Philippines, where a very deadly landslide in the south was initially not covered because of focus on direct hurricane damage and landslides in the North, and in Indonesia, where the focus was on tsunami impacted areas while more inland landslides made huge damage and victims.

Witness envisions a wide range of user groups related to emergency management, a user base much broader than that of a typical satellite analysis component. Potentially providing (near-)realtime technologies and using social media and crowdsourced data as source, the platform and its data are suitable to be made available to new user groups beyond the authorized Copernicus Emergency Management Services users and mapping operators, especially to disaster specific stakeholders that could benefit immediately from useful, validated disaster information (condition of roads, buildings, bridges ...). However, social media during crises reflect the disasters themselves: high pressure, fast changing situations where information overload, "noise" and information uncertainty are the rule. These highly complex circumstances pose challenges for digital applications.

The Witness system is composed by a set of concatenated tools that retrieve tweets with images or videos, geolocalize them, and filter them to select relevant tweets.
Platforms such as Twitter provide APIs that enable to retrieve posts that contain specific search terms. Since the quality of the extracted content is highly dependent on the quality of the search terms used, recent research in this field focused on the process of identifying and evaluating keywords in order to improve search results. One major problem is that this work mainly focus on the extraction of posts written in English.
Being capable to analyze social media content in various languages is a very challenging task.
To generate a multilingual crisis lexicon for additional languages, in a first step a Basic English keyword set has been generated for four specific emergency scenarios: flood, fire, earthquake, and storm, and another one containing very general crisis terms. A pure translation of these search terms using public translation services produces several low quality search terms, which lead to a noisy data collection when used to retrieve social media posts. To further filter the search terms according to their quality, these terms have been evaluated based on several evaluation metrics calculated on a large training corpus, consisting of 10,000 Wikipedia articles and several crowd labeled twitter datasets crawled during past crisis events. The result is a Search Term Lexicon used in Witness for five scenarios and 107 different languages.

Social media is not only used to share textual information but very often media like photographs or videos are disseminated among peers. Especially during emergency events people often take photos or capture footage to document the events and Witness is therefore focusing on retrieving images related to emergency events.
In Witness, we analyze the tweets, including contents, paces, and tags, to associate a geolocation as precise as possible to the extracted images. Geoparsing is based on Open Street Map, Stanford Named Entity recognition, and a disambiguation algorithm developed for Witness (CIME).
However, due to the vast amount of data that is made available through social media efficient further processing and analysis is required to make real use of media for emergency management. State-of-the-art computer vision and deep learning approaches allow identifying, sorting and classifying crisis relevant images. In an initial step, image duplicates are identified through a hashing approach. In a second step all remaining images are classified. The classification is based on a dedicated model on top of Google’s Tensorflow library. It is able to identify the relevance of photos in relation to floods and earthquakes. Moreover, the model is able to distinguish aerial from non-aerial photos. Results from several test cases show promising results, despite the complexity and variety of the processed images.

As the volume of social media data extracted automatically over a crisis situation can easily reach tens of thousands posts, where many of them can be false positives, e.g. posts with images in which the event is mentioned but not providing any useful information, crowdsourcing is also employed to improve the overall data quality of the result, as crowdsourcing has demonstrated to play a key role filtering non-relevant content.
In addition, crowdsourcing techniques have been applied to improve the location precision of social media data. Geolocating images and video is very challenging, especially for volunteers with not experienced on geolocation. To ease the process Witness provides candidate locations that have been automatically extracted from the social media content. In this way the volunteers start with a good estimation of the position and the goal is to improve the accuracy, thus reducing the complexity of the volunteer task and reducing the time to complete it. Preliminary results show that a community of volunteers with non-experience in mapping are able to validate and estimate the position of social media content with a precision up to few meters.

Finally, all posts are displayed on a webGIS platform, providing an interface to select areas, to filter posts by the precision of the location and relevance, and to retrieve the actual posts directly from Twitter (see figure).

The Witness platform also provides new methods to help the disaster related actors define “hot spots” and “areas of interest“ to analyze the activity in social media in an unsupervised machine learning approach. The platform automatically collects social media activity before and during the disaster. Interested only in social media data that carries geocoordinates, this particular component limits itself to the content that can accurately be geospatially pinpointed. The content of the social media activity is analyzed using the self-learning algorithm Latent Dirichlet Allocation (LDA) to identify all disaster-relevant social media activity. The algorithm results in a general understanding of the kind of content that is shared in different disasters while requiring hardly any information about the disaster itself. The most relevant results are geospatially analysed, producing a layer of local hot spots (clusters with a distinctive amount of disaster related data) and cold spots (distinctive absence of disaster related data).

Witness has been experimented in some of the recent flooding events, such as the Florence and Michael storms that hit US in 2018. The results of these live case studies have provided material for further consideration and future research work. In particular, while the case studies confirm that the percentage of useful images is very small, they also confirm the potential of providing real-time (within minutes) situational awareness on the ongoing events. On the other hand there is need for further analysis on how to combine the different tools in order to get relevant and well located images taking into consideration the workload of the Witness toolchain and of the involved operators.

Acknowledgments
This work has been partially funded by the European Commission H2020 project E2mC - Evolution of Emergency Copernicus services, under project No. 730082.

Area A - Board 95

Time series of ground deformation are used to describe motion produced by various natural and anthropocentric processes, such as earthquakes, volcanic eruptions, landslides, subsidence due to resource exploitation, and uplift due to fluid injection. The original Multidimensional Small BAseline Subset (MSBAS) technique simultaneously processes multiple ascending and descending Differential Interferometric Synthetic Aperture Radar (DInSAR) data sets and produces approximate two-dimensional (2D), horizontal east-west and vertical up-down, deformation time series with combined temporal resolution over overlapped area. A modified version of the software presented here and called MSBAS3DT produces three-dimensional (3D), horizontal north-south, east-west and vertical, components of motion occurring along the surface tangent plane. A constrain that limits the motion to the surface tangent plane direction reduces degrees of freedom from three to two, allowing to precisely resolve three components of motion from two, ascending and descending DInSAR data sets. The set of linear equations that comprises MSBAS3DT is rank deficient when acquisition times of ascending and descending data sets differ and is solved in the least-square sense by applying the singular value decomposition (SVD) and the zero-, first-, or second-order Tikhonov regularization. The MSBAS3DT source code is written in C++ and is linked to the linear algebra package (LAPACK) library that provides SVD support. For demonstration of capabilities, MSBAS3DT is used to compute 3D deformation time series of a large deep-seated landslide in Bukavu, DR Congo, by simultaneously processing 343 ascending and descending Cosmo-SkyMed and Sentinel-1 SAR images acquired from March 2015 to October 2017. Validation of the deformation measurements is provided by in situ measurements and by comparison to 2D time series produced with the original MSBAS. In addition to mapping landslide, this software could potentially map ice flow and other processes that produce motion along the surface tangent plane. The MSBAS3DT software can be downloaded from http://insar.ca/.

Area A - Board 96

Image correlation remote sensing monitoring techniques are becoming key tools for providing effective qualitative and quantitative information suitable for natural hazard assessments, specifically for landslide investigation and monitoring. In recent years, these techniques have been successfully integrated and shown to be complementary and competitive with more standard remote sensing techniques, such as satellite or terrestrial Synthetic Aperture Radar interferometry. The objective of this article is to apply the proposed in-depth calibration and validation analysis, referred to as the Digital Image Correlation technique, to measure landslide displacement.
The activities have been carried out for the Montescaglioso landslide case study (Italy), a complex landslide occurred the 3rd December 2013 over an area about 500,000 m2 wide. It was characterized by a maximum depth of the failure surface of 40 m and an estimated volume of about 8 millions of cubic meters.
The availability of a multi-dataset for the Montescaglioso landslide, characterized by different types of imagery, such as LANDSAT 8 OLI (Operational Land Imager) and TIRS (Thermal Infrared Sensor), high-resolution airborne optical orthophotos, Digital Terrain Models and COSMO-SkyMed Synthetic Aperture Radar, allows for the retrieval of the actual landslide displacement field at values ranging from a few meters (2–3 m in the north-eastern sector of the landslide) to 20–21 m (local peaks on the central body of the landslide).
Furthermore, comprehensive sensitivity analyses and statistics-based processing approaches are used to identify the role of the background noise that affects the whole dataset. This noise has a directly proportional relationship to the different geometric and temporal resolutions of the processed imagery. Moreover, the accuracy of the environmental-instrumental background noise evaluation allowed the actual displacement measurements to be correctly calibrated and validated, thereby leading to a better definition of the threshold values of the maximum Digital Image Correlation sub-pixel accuracy and reliability (ranging from 1/10 to 8/10 pixel) for each processed dataset.

Area A - Board 97

Sentinel-1 is a constellation of two spaceborne C-band Synthetic Aperture Radar (SAR) sensors developed and operated by the European Space Agency under the European Commission’s Copernicus programme. Thanks to the global coverage ensured by S-1 [1], it is possible to assess globally the presence of radio frequency interference (RFI), which is a cause of annoyance for the users. Known sources of RFI include: the Canadian RadarSAT-2 system, ground (civil and military) radars and weather radars operating at the same wavelength. Furthermore, the level of interference could worsen in the near future, as the 5350-5470 MHz band will possibly be shared with wireless LAN systems. Hence, there is a strong need for a mapping and classification system of RFI in the C-band.

In order to perform thermal noise estimation, Sentinel-1 acquires blocks of signal-free echoes in the initial and final section of each data take. Since the beginning of the mission, the analysis of the noise echoes has shown that Sentinel-1 instruments are sensitive enough to capture the different Earth surface EM emissivity between sea and land. Furthermore, it was shown [2] that these signal-free echoes can be used to extract relevant statistics for performing RFI detection and mapping. Recently, ESA increased the number of noise measurement available, also including the first signal-free echoes of each TopSAR burst [3]. Since April 2018, signal-free measurements in the C band are available with a rate of about 1 Hz in correspondence with TopSAR acquisitions. The availability of these measures allows a world-wide mapping of noise power and RFI sources as shown in the attached figure (Sept. 2018 data).

In this paper we describe an automatic screening system generating global Earth surface EM emissivity and RFI maps with a periodicity of 12 days (S-1 repetition cycle). For tuning the system, the measured Earth surface EM emissivity power is calibrated against real measures from scatterometers operating at the same range of frequencies such as MetOp Advanced Scatterometer (ASCAT). As there are various sources of radio interference, the statistical properties of the measured signal are exploited to classify the occurrences according to their nature. Analysing the strength and geographical distribution of the RFI, the most affected regions are identified and probability maps are extracted. Finally, the results of our work are compared with the RFI maps produced by the SMOS mission (L-band).

[1] https://sentinel.esa.int/web/sentinel/missions/sentinel-1/observation-scenario

[2] Monti-Guarnieri, A., Giudici, D., & Recchia, A. (2017). Identification of C-Band Radio Frequency Interferences from Sentinel-1 Data. Remote Sensing, 9(11), 1183.

[3] De Zan, F., & Guarnieri, A. M. (2006). TOPSAR: Terrain observation by progressive scans. IEEE Transactions on Geoscience and Remote Sensing, 44(9), 2352-2360.

Area A - Board 98

Large efforts are being placed into optimizing water use for crops to promote sustainable use of finite resources. As a result, field scale (<30m) metrics of water stress, such as soil moisture (SM) and evapotranspiration (ET), are desired to support decision making and enhance decision support tools. The objective of this research is to develop robust field scale metrics of soil moisture and ET using fused Landsat-8, Sentinel-2, Sentinel-1, and PALSAR-2. As an application test we examined management practices intended to conserve water resources and inform irrigation scenarios at a production hot spot in the USA. Semi-empirical modeling and polarimetric decomposition techniques were scaled to the region. Results show that early and midterm growth stages have relatively high accuracy (RMSE +/- 5%). Anecdotally, efforts to downscale more coarse drivers of water stress to field scale, such as MODIS, to support daily interpolation were largely influenced by default cloud mask (C/FMASK) performance. Quad pol SAR data, such as is planned with NISAR and available from PALSAR-2, were superior for SM compared to dual pol techniques or coherence time series. Ultimately, the growth in moderate resolution data streams will propel the next generation of decision support tools. Continued delivery and continuity, open access policies, and short latency are critical for successful implementation.

Area A - Board 99

Remote sensing estimates of Leaf area index are important determinants of many crop related metrics like water use efficiency, crop yields, and crop water productivity. These estimates are usually derived from vegetation indices (VI) using top of atmosphere or surface reflectance red and NIR bands. Literature has shown that these estimates are scale-dependent. Harmonization of Sentinel-2 and Landsat 8 products is important for providing a baseline for operational monitoring of agricultural system dynamics and for consistent estimates of vegetation indices and variables dependent on them. Recently, NASA has generated a harmonized sentinel and Landsat product set for selected global test sites. Among these products are the S10 (Sentinel-2 MSI surface reflectance at full resolution), the S30 (Sentinel-2 MSI harmonized surface reflectance resampled at 30m over the Sentinel-2 tiling system), and the L30 (Landsat-8 OLI harmonized surface reflectance resampled at 30m over the Sentinel-2 tiling system). In this research, we compare remotely sensed estimates of LAI from these products, generated using different algorithms, with field measurements of LAI and Photo-synthetically Active Radiation (PAR) over the Bekaa valley in Lebanon. The field measurements were collected over a period of five months (May-September 2017) at the time of several satellite overpasses (Landsat and Sentinel-2) for a large range of agricultural crops are different locations in the Bekaa. A total of 520 fields (5600 ha) were surveyed with at least three LAI measurements per field. We evaluate the correlation of the measurements with results of LAI calculated using different equations from the literature (using SAVI, NDVI). The research provides insights into the many factors affecting the discrepancy between the equations and the measured values. The research also compares the performance of developing a generic LAI-VI relation for agricultural crops (which are more useful when only the general land cover is known) vs. deriving a crop-specific LAI-VI (which are useful when crop-specific Land cover information is available). The performance of the different LAI-VI relations in both situations is compared for the three harmonized products (S10, S30 and L30). The field measurements will be also compared to the 5-day Landsat-8 OLI - Sentinel-2 MSI harmonized surface reflectance resampled at 30m (M30) once available.

Area A - Board 100

Combination of data acquired by Landsat 8 and Sentinel-2 remote sensing satellites can provide high temporal resolution globally (3-5 days) (Li & Roy 2017), which is critical for various applications requiring dense data time series. Previously, such (or better) high temporal resolution at global scale was available mainly for remote sensing sensors, which acquire daily data over Earth’s surface, but at coarser spatial resolution (> 250 m) (Justice et al. 1998, 2013). The latter, for example, includes space-borne remote sensing sensors, such as MODIS, VIIRS, AVHRR, SPOT-VEGETATION. Taking into account an increased frequency of observations at moderate spatial resolution (<30 m), the assumption is that methods and models developed for generating products for coarse spatial resolution sensors can be ported to moderate spatial resolution sensors (Landsat 8/OLI, Sentinel-2/MSI). However, such a transition is not always straightforward due to larger data gaps caused by clouds presence and uneven coverage, sensor characteristics and increased spatial resolution (at least at the order of 10, when going from 250 m to 30 m).
In this work we focus on a crop yield assessment/forecasting application. The hypothesis is that satellite-based features, such as vegetation indices (VIs) or biophysical parameters derived at a single date or accumulated over some time period, can be correlated to crop yields (Johnson 2016). Since the reference data on crop yields are mainly available at regional scale, the corresponding empirical models are built by averaging satellite-based features over those regions and correlating these derived variables to crop yields (Becker-Reshef et al. 2010, Franch et al. 2015, Kogan et al. 2013). It is assumed that there is a homogeneity within the region in terms of crops grown and agricultural practices applied and, therefore, the averaging should be performed for satellite data acquired at the same (or approximately the same) stages of crop growth, meaning that the data are normalized. This is usually the case for coarse spatial resolution remote sensing sensors, whose high temporal resolution enable high likelihood of obtaining cloud-free data over the Earth’s surface (Whitcraft et al. 2015). This is also evidenced by multiple successful applications of applying coarse spatial resolution satellite data to crop yield assessment and forecasting (Becker-Reshef et al. 2010, Franch et al. 2015, 2017, Johnson 2016, Kogan et al. 2013).
However, this is not the case for moderate spatial resolution satellite data (<30 m). Irregular spatial coverage, when the area in question is covered by several “stripes” sensed at different times, and high revisit cycles lead to discrepancies in dates of cloud-free observations. Therefore, satellite data should be normalized for the further processing.
In the case of crop yield assessment, one of the important questions is as follows: How do discrepancies in dates of satellite observations influence the performance of empirical crop yield models at regional scale? The present study aims to address this question by building empirical models for winter wheat yield assessment with Landsat 8 and Sentinel-2 data in Ukraine. Through simulation modelling and analysis of real datasets, we show that satellite data normalization is critical in building robust crop yield models. Not performing satellite data normalization may lead to poorer performance of the empirical crop yield model, which would be attributed not to the lack of correlation with satellite-derived variables, but rather to observation irregularities.
References
Becker-Reshef, I., Vermote, E., Lindeman, M., & Justice, C. (2010). A generalized regression-based model for forecasting winter wheat yields in Kansas and Ukraine using MODIS data. Remote sensing of environment, 114(6), 1312-1323.
Franch, B., Vermote, E. F., Becker-Reshef, I., Claverie, M., Huang, J., Zhang, J., ... & Sobrino, J. A. (2015). Improving the timeliness of winter wheat production forecast in the United States of America, Ukraine and China using MODIS data and NCAR Growing Degree Day information. Remote Sensing of Environment, 161, 131-148.
Franch, B., Vermote, E. F., Roger, J. C., Murphy, E., Becker-Reshef, I., Justice, C., ... & Baret, F. (2017). A 30+ year AVHRR land surface reflectance climate data record and its application to wheat yield monitoring. Remote Sensing, 9(3), 296.
Johnson, D. M. (2016). A comprehensive assessment of the correlations between field crop yields and commonly used MODIS products. International Journal of Applied Earth Observation and Geoinformation, 52, 65-81.
Justice, C. O., Vermote, E., Townshend, J. R., Defries, R., Roy, D. P., Hall, D. K., ... & Lucht, W. (1998). The Moderate Resolution Imaging Spectroradiometer (MODIS): Land remote sensing for global change research. IEEE Transactions on Geoscience and Remote Sensing, 36(4), 1228-1249.
Justice, C. O., Román, M. O., Csiszar, I., Vermote, E. F., Wolfe, R. E., Hook, S. J., ... & Tschudi, M. (2013). Land and cryosphere products from Suomi NPP VIIRS: Overview and status. Journal of Geophysical Research: Atmospheres, 118(17), 9753-9765.
Kogan, F., Kussul, N., Adamenko, T., Skakun, S., Kravchenko, O., Kryvobok, O., ... & Lavrenyuk, A. (2013). Winter wheat yield forecasting in Ukraine based on Earth observation, meteorological data and biophysical models. International Journal of Applied Earth Observation and Geoinformation, 23, 192-203.
Li, J., & Roy, D. P. (2017). A global analysis of Sentinel-2A, Sentinel-2B and Landsat-8 data revisit intervals and implications for terrestrial monitoring. Remote Sensing, 9(9), 902.
Whitcraft, A. K., Vermote, E. F., Becker-Reshef, I., & Justice, C. O. (2015). Cloud cover throughout the agricultural growing season: Impacts on passive optical earth observations. Remote Sensing of Environment, 156, 438-447.

Area A - Board 101

Earth observation data does not always come in the spatial and temporal resolution required by the user. Usually, there is a trade-off between the ground sampling distance and the revisit time of optical satellites. The synergistic use of different satellite sensors can mitigate these effects. For this purpose a new set of methods have been developed, commonly referred to as spatio-temporal fusion models. The most popular of them is the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM), originally developed for the fusion of Landsat and MODIS images. Spatiotemporal fusion models such as STARFM allow the creation of synthetic products combing the higher spatial resolution of sensors such as those on board of Sentinel-2 and Landsat 8 with the more frequent revisit times provided by instruments like OLCI and MODIS. The purpose of this study is to explore the differences and the similarities between the fusion products of both sensor pairs and to quantify their accuracy in respect to a set of reference images. Thus, it will be possible to give insights into the quality of fusion products and also into the peculiarities of specific sensor combinations. The study area covers approximately 3000 km² west of Madrid (Spain). To derive the fusion product, one pair of fine and coarse resolution images is used (acquired before the date of prediction) together with one coarse resolution image at the date of prediction. The fusion is performed separately for Sentinel-2 & Sentinel-3 (OLCI) and Landsat 8 & MODIS. The output is compared to a set of original Sentinel-2/Landsat 8 images collected on the day of prediction. Such comparison has not been performed before on sensor level and can contribute to the future development of multiple sensors fusion products based on images from more than two sensors both for deriving higher resolution surface reflectance and densifying time series.

Area A - Board 102