Global forest ecosystems absorb 20% of anthropogenic CO2 emissions and are expected to provide up to 25% of future emission reductions in the Paris Agreement. In the past decades, forests have been experiencing unprecedented climate conditions that impair tree health and increase their vulnerability to disturbances and may limit their mitigation potential. Insect outbreaks are thought to affect a much larger extent than wildfires, particularly in Northern Hemisphere forests, and to contribute to comparable losses of forest carbon stocks. In a warming world, insect outbreaks are expected to increase, potentially leading to increased tree mortality and long-term ecosystem degradation. However, to what scale warming and climate-change induced environmental changes affect insect outbreak occurrence and severity at large scales, and how they affect ecosystems is largely unknown.

This knowledge gap exists mainly because accurate and consistent large-scale information about insect outbreak occurrence and the subsequent forest response is still lacking. While in North America the US American and Canadian forest services conducted frequent aerial detection surveys over the past decades, in other countries the information is scattered and incomplete [1], hampering our understanding of present and future trends in ID regimes and their impacts on the world’s forests.
Satellite imagery is a cost-effective means of monitoring disturbances at the global scale. Insect outbreak detection requires monitoring of visible and near infra-red surface reflectance at moderate to high-resolution (to detect small patches) and frequent overpass (to accurately detect the timing and duration of outbreaks). Until recently, available satellite imagery could only provide a trade-off between these three aspects: early LANDSAT records (since 1972) with high spatial resolution (30m) and spectral coverage, but few images per growing-season [2]; and MODIS (since 2001) with daily to 8-day resolution, but moderate spatial resolution (250m to 1km, [3]). SENTINEL-2 opens the possibility for developing a new long-term and large-scale dataset of insect outbreak impacts on forests, advancing the knowledge on recent trends, patterns and drivers.

Here we want to evaluate the potential of SENTINEL-2 imagery to more accurately detect and track the evolution of insect outbreaks over forested lands in the Northern Hemisphere. We present preliminary results for a test region in Northwestern USA, which is covered by aerial detection survey data [4]. This region is highly prone to both bark beetle and defoliator attacks and annual insect-affected areas derived from aerial detection surveys are publicly available. Here we analyze the characteristic spatio-temporal patterns of surface reflectance measured from SENTINEL-2 and vegetation indices in insect-affected areas, with particular focus on the differences between defoliator- and beetle-affected areas.


[1] Kautz et al. Glob. Ecol. And Biog. 2017
[2] Towsend et al. Ecol. App., 2004
[3] De Beurs, K. & Townsend, P., Rem. Sens. Env., 2008.
[4] https://foresthealth.fs.usda.gov/portal

Area A - Board 42

The Amazon rainforest plays a key-role in the Earth’s ecosystem. In the last decades, illegal deforestation has increased at an alarming rate, threatening its ecosystems and strongly influencing climate changes. Spaceborne remote sensing represents a unique instrument for providing consistent, timely, and high-resolution products at a global scale, with no need for field activities. Moreover, rainforests are almost completely hidden from optical sensors by persistent cloud coverage during the wet season, which can last up to six months per year. Synthetic aperture radar (SAR), with its capability to acquire data independently from weather or day-light conditions, is therefore the most suitable technique to achieve such a high demanding task. The proposed monitoring concept combines data coming from two different SAR missions: Sentinel-1 and TanDEM-X. Sentinel-1 is part of the Copernicus Program of the European Space Agency (ESA) and consists of a constellation of two SAR satellites, operating at C band and with six days revisit time capabilities. They nominally acquire land areas in interferometric wide-swath (IW) mode, characterized by a swath width of about 250 km on ground, at a resolution of 5 m × 20 m, and in dual polarization (nominally VV and VH polarization channels). On the other hand, the German SAR mission TanDEM-X comprises two twin satellites, TerraSAR-X and TanDEM-X, operating at X band, with revisit time capabilities of 11 days. Both satellites fly in a close orbit configuration, acting as single-pass SAR interferometer and systematically scan the Earth’s land masses in single polarization Stripmap mode. Such acquisitions are characterized by swath width of about 30 km on ground and a resolution of 1.2 m × 3 m. The combination of Sentinel-1 and TanDEM-X data is the basis for the proposed new approach, aimed at exploiting the strengths of both systems: on one side, the large-coverage and short revisit time of Sentinel-1, and on the other one, the high-resolution and interferometric stability for forest detection of TanDEM-X. The proposed acquisition concept is based on a semi-automatic two-steps approach. Firstly, the so-called hot-spots (on ground regions characterized by a significant amount of on-going deforestation) are identified by means of Sentinel-1 data, which can provide a global coverage in IW mode once per month of the entire arc of deforestation, comprising all Amazon regions mostly affected by deforestation. The idea is therefore to generate lower-resolution forest/non-forest maps, and consequently forest change maps, by exploiting Sentinel-1 interferometric time-series and, in particular, by combining both the backscatter and the interferometric coherence in a machine learning-based classification method. Secondly, if the total amount of occurring changes surpasses a defined threshold, the detected hot-spot can be analyzed an constantly monitored by means of high-resolution TanDEM-X bistatic time series, which allow for a much more reliable detection of new clear cuts and deforestation activities by exploiting the bistatic coherence degradation due to volume scattering. In this work, we will present in detail the acquisition concept and the developed algorithms for detecting forested areas and new clear cuts, together with a large-scale proof of concept of the entire monitoring system.

Area A - Board 49

Forested areas strongly influence our ecosystem and play an important role in the carbon cycle, biodiversity richness, and climate regulation. Despite the efforts made by the United Nations with the Reducing Emissions from Deforestation and Forest Degradation (REDD) program to reduce deforestation and forest degradation activities, illegal deforestation, especially in the Amazon rainforest, is still increasing at alarming rates. The continuous monitoring of such areas is of particular importance to preserve them and to better understand their influence on the dynamic processes of our planet. Spaceborne synthetic aperture radar (SAR) represents a unique data source for the forest monitoring at a global scale and independently on weather conditions and solar light. The aim of this work is to exploit multi temporal SAR data stacks, provided by the SAR sensors TanDEM-X and Sentinel-1, to detect changes on the forested areas. In the case of the bistatic TanDEM-X system, where no temporal decorrelation takes place, the volume correlation factor, which quantifies the amount of decorrelation due to scattering from a volume, is a good indicator to distinguish between forest and non-forest areas. Such classification can be improved by applying an image segmentation approach. In the case of Sentinel-1, the combination of the radar backscattering and of the evolution in time of the repeat-pass interferometric coherence can be used to characterize and detect deforested areas. In both multi-temporal SAR data stacks, clear-cuts can be identified and isolated by means of morphological operators and pixel density indicators. A region growing algorithm, as well as different texture analysis operators, will help to evaluate the evolution of the detected clear cuts in time. An adhoc consistency analysis between subsequent images plays an important role in reducing false alarms within the multi-temporal change detection algorithm. By implementing a data fusion approach between image stacks provided by both sensors, the deforestation rate for a large time span can be estimated with an increased accuracy. Finally, by exploiting the sensitivity to the tree height of the bistatic TanDEM-X system and the short revisit-time of the Sentinel-1 satellites constellation, the identification of selective logging will also be addressed.

Area A - Board 50

The Forestry Commission (FC) spends 20-40% of its budget monitoring, administering, and pursuing the establishment of young tree stock by aerial photography and on-site inspections. Capgemini currently works on an improved monitoring system combining the used of Space data processing and machine learning expertise. Indeed, the innovative solution presented here proposes to engineer features from SAR data as inputs and machine learning models from Data Science field. The inputs used are Sentinel 1 data products, which provide a high frequently measurements, and in-situ data form The National Forest Estate (NFE) and the National Forest Inventory (NFI) providing shapefiles. The applied methodology is a SAR-machine Learning framework relying on pre-processing with Sentinel-1 toolbox, reprojection and multi temporal filtering approach, exploratory analysis, feature engineering and statistical machine learning to obtain actionable insights accessible through a web application offering visual analytics.
The solution considered here is to use Sentinel 1 Synthetic Aperture Radar (SAR) data to identify young trees. As opposed to aerial photography, the Sentinel 1 constellation takes SAR measurements frequently: approximately every five days. Analysing this data will enable a much more timely response to violations of agreements and a close-knit monitoring of assumed woodland. Furthermore, SAR backscatter information can be used to distinguish clear-fell or unused land from regions with newly planted trees by tracking the temporal change in height of the land.
The solution proposed here is to engineer features from SAR data along with auxiliary as inputs for machine learning models. For the purpose of this research, three years of SAR data (2015-2017) have been downloaded. Then, the related data set has been pre-processed with Sentinel 1 toolbox from SNAP, a free software provided by European Spatial Agency. In order to filter the speckle, the preferred chosen approach is to remove it with a Multi temporal filtering.
To obtain a categorisation of the various woodlands in regions of interest, two publicly available shapefiles from the FC are used:
- National Forest Estate (NFE): The NFE contains the FC sub-compartment database and gives information for recording, monitoring, analysis and reporting across the entire FC estate.
- National Forest Inventory (NFI): The NFI includes interpreted forest types for all woodlands over 0.5 ha and interpreted open area information for areas over 0.5 ha surrounded by woodlands.
Every SAR image is split to cover individual polygons, and for every such combination of SAR image and polygon, a record in a data frame storing all combined data is created. Due to the complexity of the problem, it has been selected a simple region of interest for the first trial: Kielder (UK).
The aim of our experiment is to distinguish young trees (precisely conifer) from grassland on a polygon. To create a dataset, we used NFI data in order to keep only polygons containing these types of vegetation. A Python script has been developed retrieving a list with features and applies them to the dataset. The related output is a ‘feature data frame’ which contains all the computed features and a unique identifier for every SAR image. The features created together with the ground truth are used to train the model on the training data and this latter is applied to the test data.
The temporal step between two SAR data is tested and results of machine learning algorithms such as SVM, Gradient Boosting and K Nearest Neighbour are compared. According to our data, the machine learning algorithm Gradient Boosting has the best performance of all models tested. The model achieves a classification score of 87% for Kielder region. These scores are very promising and may already be sufficiently high for many applications.
The results then demonstrate that the SAR-Machine Learning framework is able to produce a machine learning model that can identify trees with a high accuracy. It must be applied on other regions to test the global capabilities of the machine learning model. However, it may not be feasible to build models for mountainous terrains, potentially due to the distortion of DEM. There is also an opportunity for improving results by adding meteorological data coming from UK MetOffice as input for the SAR-Machine Learning framework.
In the long-term vision, the tree identification model may be implemented as a tool to help business analysts with automatic recommended decisions derived from model predictions, as well as giving business intelligence and insight for data mining studies.

Area A - Board 51

The capacity of SAR data for deforestation monitoring is well known. However, this is usually restricted to long wavelengths such as L- and P-band radar as they can penetrate the forest canopy and provide information about the forest structure and thus its biomass content. While Sentinel-1 as a C-band system is not suitable for quantifying biomass, our implementation of a polarimetric change detection algorithm shows that Sentinel-1 data is indeed capable of picking up deforestation. It would thus be a mistake to disregard the potential of Sentinel-1 data for this purpose, especially as the open data policy and frequent global coverage will certainly prove valuable for forest monitoring.

We analyzed a three-dimensional datacube of dual-polarized Sentinel-1 SLC data (in latitude, longitude, and time) over a study area in Rondônia, Brazil for the year 2017. In order to include spatial context, our approach complements a pixelwise change detection algorithm based on statistical tests of the equality of consecutive complex covariance matrices (Conradsen et al. 2016). The algorithm relies on ensemble averages, which are usually a by-product of multilooking or traditional noise reduction techniques, such as boxcar or Gaussian filters. However, these filters reduce noise at a loss of spatial resolution. Instead, more advanced filters take into account pixel similarity as well as spatial proximity. They not only improve the image quality in general but also give better change detection results. One such filter is non-local means (Buades et al. 2011), which we extended to operate on three-dimensional datacubes.

We show that by using edge-preserving spatio-temporal filters in combination with pixelwise change detection we are able to detect even very narrow change patterns such as road construction, which would otherwise be lost in noise or disappear in the smoothing process. We validated our change detection results against the GlobalForestWatch (GFW) dataset (Hansen et al. 2013). Adding a spatio-temporal non-local means filter slightly reduced the omission rate with respect to the reference from 40% to 35% at an equal commission rate of 2.5%. These are not necessarily errors as the GFW dataset does not represent the ground truth. Visual inspection revealed an improvement especially at forest edges and therefore allows detection of also small-scale deforestation. The attached image shows a slice through the datacube as a composition of the original (left) and the filtered data (right).

Area A - Board 52

Wind storms are recognised as a highly damaging disturbance factor for forest ecosystems [4]. After a storm event, rapid and accurate information on the extent and severity of wind damage is crucial not only for emergency management purposes but also for assessing the ecological and economic impacts derived from these natural disasters. Unfortunately, current methods used for studying wind damage in forests do not always fulfil this requirement. This is because they either rely on passive sensors (e.g., Optical) [3], which are not available with cloud cover. Or they rely on airborne LiDAR which is particularly infrequent. In response to this, our work aims to generate a long-term and continuous method based on remote sensing data fusion techniques for monitoring forested areas affected by windthrow events. By combining advanced multi-frequency SAR data (X, C & L band), aerial photogrammetry techniques and ground-based forest information, this research focuses on the development of a novel monitoring system which allows to early detect the areas affected by wind-damage and to estimate the forest Above Ground Biomass (AGB) loss.

In European forests wind storms were responsible for 53% of the damage caused by natural disturbances during the last half of the 20th century [3]; causing a cost to the European Union countries of approximately 15 million euros per year [1]. The United Kingdom, and more specifically the west coast of Scotland, is one of the most affected regions by this issue due to its exposure to recurrent Atlantic wind and rain storm, being this recognised as a major threat to the management of British forests [1]. We are working along with the Scottish Forestry Commission to focus on regions highly exposed to wind.
In this project, besides using information related to known windblown events, we are synchronising our analysis with logging operations to generate dataset of well-reported logging. We are collaborating with the Queen Elizabeth Forest Park management team which allowed us to locate a study area (parcel 34039 in figure 1). The parcel is a 28.8ha mixed-conifer plantation [Sitka spruce (68.86%) and Norway spruce (19.42%)] which is planned to be completely logged within a time-frame of 5 months [January to May 2019]. Located near the village of Aberfoyle (56°09'40.3"N 4°27'19.5"W), this parcel sits in a relatively flat terrain ensuring the absence of any geometrical image distortions which conditions the use of SAR data.

Data fusion techniques stand out as crucial point due to numerous and heterogeneous datasets used in this project. The datasets can be classified into two major groups (1) Remote sensing data formed by multifrequency SAR acquisitions, airborne LiDAR and Drone-Photogrammetry images; and (2) Ground data composed by the historical forest inventory, the logging data and the tree parameters field surveys.
European Space Agency-Sentinel1 (C-band) images are regularly collected as Level-1 Interferometric Wide Swath (IW) - Single Look Complex (SLC) mode in dual-polarization (VV, VH), 10 m pixel resolution and at a centre frequency of 5.405 GHz. By using both Sentinel 1 (A, B) satellites, images of the study area can be obtained every 6 days. This permits the generation of a dense time-series that will be used for monitoring the progressive changes derived from the logging activities. ALOS PALSAR (L-band) data was provided by JAXA for two consecutive years [7th March 2016 & 26th April 2017]. These archive images, acquired at high spatial resolution (2.8 x 3.1 m) in fine beam - strip map mode and Quad-polarization (HH, HV, VH, VV), were used for performing a pilot assessment over the study area. COSMO-SkyMed (X-band) images will be collected on a monthly basis by the Agenzia Spaziale Italiana since November 2018. A total of 10 images will be acquired in two different modes (5 Ping-pong mode & 5 Spotlight mode) at 1-meter and 15-meter spatial resolution respectively. Lastly, Structure-from-Motion (SfM) photogrammetry techniques will be utilised for collecting three-dimensional information of the top of the forest canopy. Imagery will be collected using a Post-Processing Kinematic (PPK) enabled fixed wing Unmanned Aerial Vehicle (UAV) with an off-the-shelf 24mp RGB camera. A high resolution Canopy Height Model (CHM) will be derived from the photogrammetric point cloud created with the SfM software Pix4D.
Forest logging works will be done by a Tigercat Harvester machine equipped with the LogMate500© control system which will provide on-site information about the tree location and some tree parameters (e.g. height, diameter). Lastly, this data will be complemented by a fieldwork campaign carried out prior to the logging activities. Fieldwork campaign focuses on collecting tree parameters information for evaluating the reliability of both the harvester machine and forest inventory data.

In conclusion the outputs are: 1) designing a change detection able to perform a near-real time forest change detection analysis using SAR based change detector algorithms and 2) using the sporadic forest height and type maps together with the detection maps for generating accurate estimations of the timber volume and AGB loss derived from catastrophic wind damage.

Preliminary analyses carried out with ALOS and ESA-Sentinel 1 data showed promising results in the context of change detection. The estimation of AGB loss is still in progress however we hope we will have some results also on AGB loss by the time of the conference.



[1] Forest Research, 2018. Tree stability and wind risk to forests - Forest Research. Forest Research. Available at: https://www.forestresearch.gov.uk/research/tree-stability-and-wind-risk-to-forests/.

[2] Schelhaas, M., Nabuurs, G. and Schuck, A., 2003. Natural disturbances in the European forests in the 19th and 20th centuries. Global Change Biology, 9(11), pp.1620-1633.

[3] Schwarz, M. et al., 2003. Detection of Windthrow in Mountainous Regions with Different Remote Sensing Data and Classification Methods. Scandinavian Journal of Forest Research, 18(6), pp.525-536.
[4] Waldron, K., Ruel, J. and Gauthier, S., 2013. Forest structural attributes after windthrow and consequences of salvage logging. Forest Ecology and Management, 289, pp.28-37.

Area A - Board 53

MDA’s Forest Alert Service (FAS) is an operational service which detects and reports on subtle forest disturbances using MDA’s spaceborne RADARSAT-2 radar imaging sensor. Since 2016, this service has been helping Asia Pulp and Paper (APP) and Sinar Mas Agri Resources Technology (SMART), both of the Sinar Mas Group in Indonesia, to fulfill their forest conservation commitments. MDA’s FAS monitors 3.8 million hectares of concession forest every 24 days for APP and SMART and provides “alert” polygons to indicate areas where forest disturbances may have occurred. MDA’s forest alert polygons are validated as 85% to 95% accurate within conservation zones. In many cases, the polygons are used in support of community mediation activities to help prevent future conservation forest disturbances.

MDA’s forest alert polygons can be as small as 0.5 hectares. As this system is tuned to detect subtle changes, then forest alert polygons are defined as merely indicators that there has likely been some change, consistent with structural tree disturbances, somewhere within the alert polygon. The degree of subtlety is indicated by a computed strength indicator. Depending on the strength of the detection, the actual disturbance may range from loss of just a few small trees within the alert polygon, to a small clear cut encompassing the entire alert polygon.

Forest disturbances may occur from planned events (e.g. legal tree removal), unplanned events (e.g. illegal tree removal), events of unresolved legality (e.g. community activities) or natural causes (e.g. wind damage), among others. APP and SMART have field verification strategies to investigate detected forest alerts inside conservation zones and to mitigate validated alerts. In the case of APP, MDA’s alert polygons are inserted directly into APP’s geospatial database so that personnel in the field can begin assessment as soon as the polygons are delivered. Forest alert polygons are typically delivered within two working days after satellite data acquisition.

The primary source of false positives are flooding events, which can affect the radar signature on the ground in a way that appears similar to a forest change. Flooding events can be a particular challenge due to the typical frequency and extent of flooding that occurs in the areas of interest. It is, in fact, the persistent occurrence of clouds in these tropical zones that led APP and SMART to turn to MDA for its radar sensor, which images the ground through clouds, as opposed to optical monitoring. However, the persistent cloud cover leads to frequent rain and flood events, thus proving a challenge to improve the accuracy of radar-based monitoring. Methods employed to help reduce false alarms include examining repeating occurrences of low-strength alerts in particular areas. While this helps to reduce false alarms in areas that are prone to frequent flooding, work continues to identify and reduce false alerts that occur in areas where flooding patterns have not typically been prevalent.

With the success of the conservation monitoring program, the results are fed back into APP and SMART’s forest management programs. Monitoring of concession areas includes operational compartments; thus, MDA is able to adapt its monitoring program to help benefit forest operations management as well as conservation practices.

Area A - Board 54

Large-scale fire activity in various parts of the Earth can result in fundamental changes of ecosystems. Destruction of vast amounts of vegetation through fire not only changes land cover types but also land surface properties. In addition to changes in natural environments, the loss of ecosystems like forests located over peatlands affects carbon storage balance, tied to atmospheric composition and thus climate. Indonesia contains more than 50% of all known tropical peat reservoirs on the planet, storing about 55-58 Gt of carbon in belowground peatland mass. Due to degradation and deforestation, Indonesia produces one of the highest carbon emission rates worldwide. One common way to clear forest is through “slash and burn” techniques, where forested land is extensively cut and burned. However, the uncontrolled spread of fire often leads to a large-scale loss of peat swamp forest. In addition, economically valuable trees are often logged for profit and canals are dug as means of transport to and from the logging sites. Canals lead to a particularly detrimental form of degradation, as it causes drying of peatlands and encourages the spread of fires, thus increasing the potential for forest and peatland loss.
The thermal bands of the new and freely available Sentinel-3 SLSTR enable the detection of active fires on both a daily basis and a global scale. This study assessed the correlation between detected Sentinel-3 active fires, MODIS hotspots, estimated Sentinel-1 burned areas, and Sentinel-2 active fires and burned areas over peatlands located in West Kalimantan, Indonesia. The combination of different ESA Sentinel products allows a better overview of fire activities over peatlands. Data from the different sensors were analyzed from August 7th – August 19th 2018, covering two weeks during the Indonesian dry season when fires were present. Processing of an almost cloudless Sentinel-3 image from August 13th resulted in 38 thermal anomalies within the peatlands. MODIS hotspots from the same day, acquired with a time delay of less than two hours, provided 37 hotspot detections. However, count does not consider the spatial location of the points. Hence, a join query including a search distance of 1.5 km (equal to 1.5 thermal band pixel from Sentinel-3) in ArcMap was used to identify fire detection pairs between Sentinel-3 and MODIS data. This distance was considered adequate, given spatial deviations resulting from the time shift between the image acquisitions, differing projections and spatial resolutions. Altogether, 29 of the 38 Sentinel-3 hotspots (76.3%) show a distance less than 1.5 pixels to the MODIS hotspots and were assumed as a correct fire detection by Sentinel-3.
We show that Sentinel-1 is able to detect burned areas with an accuracy of 87% over the entire two-week analysis period when compared with the 421 MODIS hotspots that occurred during the same time period and exceeded a confidence interval of 80.0%. Validation against Sentinel-3 data resulted in an accuracy of 84.2%. In some areas of the Sentinel-1 burned area classification a change is detected but no fire hotspots are identified. This can result either from fire events not detected by MODIS and Sentinel-3, or by changes not resulting from fire events.
Peatland fires in Indonesia have a high contribution to carbon release and should be monitored carefully. Since fires are highly dynamic, it is important to have high temporal coverage to detect as many fires as possible. The study on hand showed that Sentinel-3 is able to detect fire activities in tropical peatlands and is thus a remarkable addition to MODIS hotspots. Besides, the use of Sentinel-1 and -2 data allows the detection of burned areas and provides additional information about peatland fires.

Area A - Board 55

Forest fires in Iberian Peninsula have become a serious environmental problem for years (CO2 emission and climate change). Only in Galicia and Portugal, in October 2017, national information sources reported of more than 550,000 forest hectares calcinated by fires larger than 30 Ha. This fact suggests the use of quick and reliable techniques and methodologies, such as remote sensing, for correct discrimination of different areas within the forest fire. Sentinel-2 MSI sensor record data in a spectral domain which is very useful to wildfire applications like fire severity discrimination using spectral indexes. There is a group of spectral indexes designed, exclusively, to discriminate surfaces affected by fire. This is the case of NBR (Normalized Burn Ratio) and dNBR (Differenced Normalized Burn Ratio) algorithms, obtained from the near-infrared (NIR) and short-wave infrared (SWIR) reflectance. This algorithm can be reclassified into five different large classes which are NO FIRE, LOW, MODERATE, HIGH and VERY HIGH. The aim of this work is to check forest fire severity with satellite data and using fire severity information provided by EFFIS (European Forest Fire Information System) as ground truth. EFFIS also uses the dNBR algorithm to calculate fire severity but does not use Sentinel-2 information. It uses the MODIS (Moderate Resolution Imaging Spectroradiometer) sensor aboard AQUA and TERRA satellites and whose spatial resolution is, at least, 250 m (Sentinel-2's dNBR algorithm has a spatial resolution of 20 m). Results show a good synergy between both information, both in burnt surface and in severity classes’ discrimination. On the one hand, burnt areas estimation using Sentinel-2 is, approximately, a 15% lower than burnt areas estimated by EFFIS. On the other hand, NO FIRE and VERY HIGH fire severity classes obtained by EFFIS are, approximately, a 10% lower than NO FIRE and VERY HIGH classes estimated by Sentinel-2. However, LOW, MODERATE and VERY HIGH fire severity classes obtained by EFFIS are, approximately, a 10% higher than LOW, MODERATE and VERY HIGH classes estimated by Sentinel-2.

Area A - Board 56

Vegetation and peat fires contribute substantially to global emissions of greenhouse gases (GHG). According to latest estimates, net fire emissions (i.e. the part of fire GHG not balanced by regrowth) amount to 6% of global fossil fuel GHG emission. Improving the management of fires in frequently burning ecosystems can help reduce GHG emissions and thus contribute to mitigation of climate change. In order to implement, monitor and document success in fire management, timely and accurate data on fire extent and impact, as well as weather and burning conditions is needed.
Monitoring and analysis of fires over large and often remote areas is only feasible with the help of Earth Observation (EO) satellites. Over the last decades, availability of free EO data has increased enormously, as has the availability of computing power, network speed and web based geospatial visualization and analysis technologies. Thermal sensors on geostationary or polar orbiting sensors (Meteosat, MODIS, FireBird, S-NPP VIIRS) make it possible to estimate the intensity of burning up to every 15 minutes, while sensors in visible to short wave infrared wavelength on the Sentinel and Landsat satellite series enable the production of burned area maps with up ~20 m spatial resolution every week. Since collected raw data is being frequently updated and often requires extensive processing for the extraction of information, there is a demand for a sophisticated solution to simultaneously handle data pre-processing, analysis, and dissemination of results to be able to provide quick, reliable and user-friendly statistical information to assist decision-makers in different levels of management. Key features of the firemaps.net platform presented here are: Near real time monitoring of fire activity and carbon fluxes, weekly updated burned areas, daily analysis and forecast of relevant weather parameters, long time series of fire emissions to calculate baselines, fire risk and vulnerability maps and tools to monitor success of fire management planning and implementation. The coupling of remote sensing data with weather information and fire spread models also enables forecasting and detailed hindsight analysis of the behavior of wildfires. Here the rate of spread and fire intensity are key parameters to understand effects of fires on affected areas and to support management planning. To develop an new information product to analyze fire intensity, we assessed fire spread and fire radiative energy release rate (fire radiative power) over savanna fires using infrared sensors with different spatial, spectral and temporal resolutions, focusing on sensors providing a spatial resolution higher than 500 m. The sensors used offer either high spatial resolution (Sentinel 2, Landsat 8) for fire detection with low temporal resolution, moderate spatial resolution and fire radiative power retrievals with moderate temporal resolution over selected sites (FireBird satellites), or moderate spatial resolution, high temporal resolution, but no fire radiative power retrievals (S-NPP VIIRS I-band active fire product), or moderate to low spatial resolution and fire radiative power retrievals at high temporal resolution (MODIS).
We extracted fire fronts from Landsat and Sentinel 2 (using the Shortwave Infrared bands) and used the available fire products for S-NPP VIIRS, MODIS and Meteosat. From these results we derive metrics on fire behavior in our study areas. We relate our results to outputs of fire behavior models and to results to published values.
The presentation also covers processing workflows and challenges, as well as an overview of the implemented technologies to develop a modern geospatial information platform for monitoring wildfires.

Area A - Board 57

Forestry has been suffering increasing economic damage from windthrows since the 1990s. This is due to more frequent storms with high wind speeds. As similar extreme weather conditions are to be expected increasingly against the background of climate change, the forest administration needs timely information on the quantity and location of wind damages in order to react promptly. In the first step, this includes the removal of impairments of the infrastructure as well as the quick balancing of losses. On this basis, timber processing can be planned to prevent a further loss of timber value as well as the consequential ecological damage caused by pest infestation.
The autumn storms Xavier and Herwart, which hit the north-east of Germany on October 5th and on October 29th in 2017, caused considerable damage to the forests in this region. In the ground softened by prolonged rainfall, large-crowned trees were unable to withstand the storm, especially as most deciduous trees were still densely foliated.
In this study we analyze the potential of recent satellite missions with high revisit rates (Sentinel-1 and -2 and PlanetScope) to implement a rapid assessment of the impact of the two storm events.
For the study area (approx. 8000 km²) nearly cloud free images before the storm, after the first event, and after the second event were processed. Moreover, 800 field survey reference areas are available. Additionally to the input spectral bands a number of vegetation indices and difference indices between the dates were used as predictor variables.
For a first bi-temporal classification of the windthrow areas we used the Random Forest (RF) classifier in combination with the reference data-set (70 % cal. / 30 % val.) for windthrow and non-windthrow areas. The results can be used to detect true and false positive and negatives as well as to evaluate to most contributing layers within the different sensors.
For Sentinel-2 (10/20 m at 13 spectral bands) the last cloud free before-storm-image was acquired on September 7th and the first image after the storm events on November 6th. This comparably long period hampered the detection of changes because all deciduous trees were defoliated within this time span. The advantages of the high radiometric standard of Sentinel-2 could not fully be exploited. For PlanetScope (3.7 m at RGB and NIR) the closest cloud free before- and after-storm-images were acquired much closer to the actual event on September 18th, October 8th, and October 30th. While the shorter time span before the storm event had a clear positive impact on the results, the time lag of only one day after the storm appeared to be too short In contrast to clearcuts that drastically change the canopy structure the amount biomass, the change after the windthrow event is not immediately as obvious. It rather develops over the following days with changes in the vitality and water content of the trees thrown. Also, for the Sentinel-1 data with a revisit time of less than a week, the mere falling of a tree had less influence on the SAR-signal than the following biophysical changes.
Mainly because of the higher spatial resolution and the higher temporal resolution, the PlanetScope classifications resulted in much higher detection of true positives. Additionally, the higher spatial resolution of PlanetScope made it possible to detect smaller windthrow areas. Still, in order to take full advantage of the higher spatial and the temporal resolution, a radiometric normalization algorithm will be applied to the PlanetScope data. A special focus for optical and SAR data is the use of temporal normalization of time series before and after the storm event.
However, it has to be highlighted that Sentinel images are free data with a high radiometric standard, while PlanetScope data for operational applications would be a matter of financial resources.

Area A - Board 58

Natural forest disturbances caused by wind and snowstorms are reduced to infrequent extreme events in the context of Mediterranean forests. However, when they occur, they can seriously compromise forest ecosystem services. The most immediate impact of storm damage is a massive conversion of living forest biomass to dead biomass, element that can interact with other forest disturbances, thereby aggravating the negative impact on forest functions. Indeed, previous studies have demonstrated that the increased amount of dead biomass caused by storms increases wildfire risk and intensity, particularly in dry climates, and may also increase the risk of suffering insect pest outbreaks. Early detection and quantification of this type of impacts on forest ecosystems is essential for planning and executing short-term management actions to restore the affected area (e.g. logging and fuel reduction), and to avoid the occurrence of even bigger damages.

In this study we develop a cost-effective methodology for conducting a regional assessment of forest storm damage using Sentinel-2 imagery, working on a large forest area recently affected by a winter storm occurred in January 2017 in the Mediterranean basin. The study area is the forest area dominated by Aleppo pine (Pinus halepensis, Mill.) in the region of Valencia (eastern Iberian Peninsula).

A field campaign, aimed at collecting information about the amount of storm damage (in terms of dead biomass) in plots georeferenced with submetric accuracy, was carried out over the study area from December 2017 to January 2018 (n = 80 plots). In parallel, all the available Sentinel-2 images of the seven tiles of the Sentinel-2 grid that comprised the study area, acquired within the period from May to August, of years 2016 and 2017, and with cloud cover less than 20%, were downloaded and corrected to surface reflectance using the Sen2Cor L2AProcess algorithm. Then, we generated yearly cloud-free gap-free L2A Sentinel-2 composites, with all the spectral bands resampled to 20-meter spatial resolution, for the seven tiles covering the study area using the maximum NDVI approach. A set of 22 vegetation indices were obtained from the pre-disturbance (2016) and post-disturbance (2017) composites. The absolute and relative difference of Sentinel-2 bands and vegetation indices between pre- and post-disturbance composites were fitted to the field data to find out their suitability for storm damage severity discrimination. Finally, the most suitable indices were selected to calibrate a Random forest regression model predicting damage severity with R-square of 0.796 and RMSE of 10.347.

Results support that Sentinel-2 data may contribute to identify and quantify accurately the impacts of disturbances causing partial damage to forest stands, such as storm severity. They also encourage to use this methodology to assess vegetation changes caused by other forest disturbances, such as large-scale disease outbreaks (e.g. pine processionary caterpillar), having important applications in the field of early damage evaluation during or after catastrophic events.

Area A - Board 59

In May 2016, an area of 432,000 ha of Prey Lang, which is one of the last major lowland rainforests of continental Southeast Asia, situated in central Cambodia, was turned into a Wildlife Sanctuary (Prime Minister Sub-decree). According to this Sub-decree, the government agreed to protect the wildlife habitats and ecosystems of the Prey Lang Sanctuary, which is home to the indigenous Kuy people as well as various threatened animal and plant species. Despite this official protection status, illegal logging, mining activities and industrial agriculture still pose threats to the sanctuary. In order to shield the sanctuary from further destruction, local community members form the Prey Lang Community Network (PLCN) to patrol the forest, recording any illegal activities using smartphone technology. The patrolling happens arbitrarily several times every month and PLCN members use clues such as the sound of chainsaws or the appearance of logging trucks to direct their inspections patrols. Direct actions are taken such as peaceful discussions with illegal loggers as well as the confiscation and demolition of chainsaws and logged wood. Findings are reported regularly to policy makers or the general public and shared over social media.
In this study, we want to investigate to what extent the freely available Google Earth Engine-based Delta-rNBR (Δ-rNBR) script (Langner et al. 2018; Fig. 1) – a near real-time forest disturbance monitoring tool – can be useful for the PLCN. The Delta-rNBR tool can work with Landsat as well as Sentinel-2 imagery. For this study, a combination of the Landsat 7 and 8 sensors is used, analyzing canopy cover disturbances occurring within the calendar year 2018. In a first step, it is analyzed how well the Delta-rNBR (Fig. 2 – showing the 2015 condition) correlates to the field data of single tree removals (Fig. 3) or conversions of larger areas as collected by the PLCN. In that context, indications from the Delta-rNBR tool are compared to results from the GLAD (Global Land Analysis & Discovery) alerts footprint system of Global Forest Watch as well as to ground data stemming from the PLCN. Furthermore, this study investigates to what extent the Delta-rNBR tool allows optimizing actions of the PLCN members on the ground, helping them to direct their patrols to areas showing recent signs of canopy cover disturbances. In this respect, we plan to analyze the time gap between the most recent disturbance indications as detected by the Delta-rNBR tool and the timing of the actual logging activities on the ground as derived from interrogations during the patrols – as far as this information is available.

Figure 1: Graphical user interface of the Delta-rNBR tool, which allows the user to analyze any geographic region or period, given there is coverage by the Landsat or Sentinel-2 sensors. The mapped area (greyish) shows the (semi-)evergreen forest cover. Figure 2: Subset of the southern part of the Prey Lang Wildlife Sanctuary; grey color: masked non-evergreen forest, red color: indications of forest canopy disturbances. Figure 3: Situation on the ground of a location that was detected using the Delta-rNBR tool.

Reference: Langner, A.; Miettinen, J.; Kukkonen, M.; Vancutsem, C.; Simonetti, D.; Vieilledent, G.; Verhegghen, A.; Gallego, J.; Stibig, H.-J. Towards Operational Monitoring of Forest Canopy Disturbance in Evergreen Rain Forests: A Test Case in Continental Southeast Asia. Remote Sens. 2018, 10, 544.

Area A - Board 60

Ecosystem functioning is mainly characterized by fluxes and budgets of energy, water and carbon, where vegetation and soil properties regulate the exchange of energy and matter at the soil-vegetation-atmosphere interface. Natural hazards that result of natural events but also from human activity prevent sustainability of ecosystems.
Almost all hazards can be characterized by three phases: pre-crisis (before the event), crisis (during the event) and post-crisis (after the event). Risk management can thus be structured into three phases as well, the first one dealing with risk assessment, the second related to short term forecasting and disaster monitoring.
A pilot study for monitoring forest cover state and capacity for sustainable functioning before/during/ after natural hazards using capabilities of MSG geostationary satellite for land surface monitoring is performed. Two types of forest disturbances are examined, first of forest fires and second of disease that leads to forest desiccation. The spatial and temporal evolution of functional and structural characteristics of forest canopy before and after the hazard is monitored by MSG LSA SAF biogeophysical products: Land Surface Temperature (LST), Evapotranspiration (ET) and Fraction of Vegetation Cover (FVC) are used as indicators of natural hazards effects. LST as a key surface variable links the processes of energy and water exchange between the forest and the atmosphere, and characterizes energetic aspects of forest functioning. ET as a water balance component indicates the disturbances of stand-scale water balance and FVC reflects the disorder in forest structure after forest fire and disease.
Two real situations of natural hazards are analyzed to illustrate the capacity of MSG LSA SAF products to monitor in near real-time forest state and anomalies. The effects of a large forest fire lasting from 24-28 August 2017 in Kresna Gorge, SW Bulgaria is explored testing pre-fire, fire and post-fire canopy processes and structure. Classified as an environmental catastrophe it strongly has influenced canopy state/structure as seen by MSG.
The second example for efficient MSG application is a real problem in silvicultural practice in Bulgarian with conifer forests the lower forest belt after the injured by a snow storm in 2015. Broken trees become affected by a disease and subsequent desiccation. A selected region with progressive mass wilting of conifers is tested by space/time evolution of LST, ET, FVC deviations from their means (10 years averaged) before the storm (e.g. 2012, 2013, 2014 years) and after the storm and disease (e.g. 2016, 2017).

Area A - Board 61

In Finland, information on forest loggings are collected in the form of the compulsory forest use notifications. The actor that performs the logging is responsible by law to prepare the notification of the intended logging for the Finnish Forest Centre. The notifications that include the borders of the cutting and the cutting type. The information is used to update the forest resource database of the Centre. The information provided by the notifications is not reliable, since not all the loggings that have been declared are implemented and there are also loggings, for which notifications has not been done.

Satellite based monitoring can be used to provide information on true cuttings. A special challenge is poor availability of reference data. Our solution was to use the notifications as reference. We developed an approach that could use also unreliable reference, i.e. the cutting notifications as training data for the change detection process.

Copernicus Sentinel-2 images from years 2015, 2016 and 2017 in five study areas in Ähtäri, Virrat, Liperi, Parikkala and Ruokolahti were used as satellite data. AutoChange method that uses hierarchical unsupervised classification was applied as the principal method. The pixels of Sentinel-2 data were classified as unchanged, clear cut or thinning cutting using forest use notification data and the feature classes that were produced by AutoChange. The approach was based on the assumption that the probability of forest logging is higher in an area where forest use notification has been done than in an area where it has not been done.

The approach was tested in the five study sites. Reference data for accuracy assessment of the results was collected using visual analysis of very high resolution Pléiades and WorldView-2 satellite images, that were received through Category-1 project from the European Space Agency. The clear cuts were detected with an overall accuracy close to 90 % using the developed approach. For detecting thinning cuttings, the accuracies were lower and varied between the sites. Based on the results, approach for using forest use notification data was better applicable for classification of changes with high magnitude, i.e. feature classes representing clear cuts, whose proportions inside and outside the notification areas was clearly different. The uncertainty of the reference data affected more detection of thinning cuttings, that were also more difficult to be separated from uncut forest using the spectral features.

The study was funded by the Ministry of Agriculture and Forestry of Finland.

Area A - Board 62

Three-quarters of Finland’s land surface area (22.8 million hectares) is filled with forests. In Finland, forests serve as a resource for the nature conservation as well as for the forestry industry. In addition, forests compose a great part of the country’s biomass and carbon sinks. At present, remote sensing technology is the main instrument in acquiring up-to-date and near real time information on the forests. Moreover, open and free Earth observation (EO) data volume increases constantly. In future also hyperspectral satellite missions (e.g., Sentinel-10) will start providing remote sensing data to support the needs of forestry and other natural resource management practices. However, to this day it is unclear what will be the additional value of hyperspectral remote sensing data compared to multispectral data in forest variable and biomass estimation. Therefore, we investigated the influence of spectral and spatial resolution of remote sensing data on forest variable estimation in the boreal forests of Finland.

We used the remote sensing data by Sentinel-2 (10 bands, resolution 20 m) and hyperspectral AISA imager (128 bands, 400–1000 nm, resolution 0.7 m). As reference data, we used forest resource data provided by the Finnish Forest Centre and additional independent in situ measurements. We applied machine learning based regression models to relate the forest variables of interest with the remotely sensed data. Based on recent studies, we selected Gaussian Process Regression (GPR) and Support Vector Regression (SVR), which have proven to work well with hyperspectral and multispectral remote sensing data. Regression estimations were performed for seven different forest variables: mean height, stem count, stem biomass, leaf area index (LAI), basal area, leaf biomass and main tree species. Estimation accuracies were mainly examined with absolute and relative root-mean-square errors (RMSE).

The forest variable estimations with both algorithms and the reference data by the Finnish Forest Centre were successful. Compared with previous stand-wise estimations of forest variables in Finland, the machine learning approach used in this study was more accurate. The estimation accuracies of the algorithms were similar. However, the faster SVR algorithm was found to be more practical. Best estimation accuracies were obtained for the same variables that were given most accurately in the reference data. Respectively, the weakest estimations were for the same variables that had the worst accuracy in the reference data. Results of this study showed that stand-wise forest variable estimations have approximately the same accuracy with multi- and hyperspectral remote sensing imagery, and that the hyperspectral data improves only the estimation accuracies of main tree species and mean height. In stand-wise forest variable estimations, the importance of spatial resolution was minor.

Area A - Board 63

Poplar is one of the fast-growing and wood producing trees which are increasingly considered as an important resource. In France, the leading European country in terms of plantations, poplar cultivation is a local key industry. However, since the economic slump, wood prices are at their lowest. Planted surfaces are in continuous regression.
The future of poplar is mainly based on replanted areas after harvesting. Thus, spatially explicit information of newly-planted and disappeared surfaces provides essential baseline data for industrial and socio-economical dynamics.
Accurate and regularly up-to-date maps of poplar plantations are not yet available at the national scale. The update rate of the French National Forest Maps is unsuitable for this species because of its short rotation cycle (from 15 to 20 years).
Since the availability of high spatial and temporal resolution Sentinel time series, new opportunities for monitoring poplar plantations over large areas have come up.
In the present study, we performed a Random Forest supervised classification on Sentinel-2 time series to identify poplars over a first study site : Tarn Et Garonne department in southwest France (~3730 km²). We also added multi-temporal SAR features (σ0 VV, σ0 VH) extracted from the Sentinel-1 images to examine the complementarity of optical and radar information. In order to deal with the limited number of labelled and confident samples, a semi supervised Active Learning (AL) technique is then applied repeatedly. It consists on selecting new training samples guided by the algorithm needs to enrich the classification model and adapt it to a larger scale (iterative classification of other departments in France). The AL approach is based on entropy measure as a criteria of sample informativeness.
In this work, Sentinel time series dated from 2016 and reference samples were grasped from polygon-eroded national forest database (IGN BD Forêt®) and in situ observations with a stratified random sampling.
The preliminary results underline that 2 % of additional actively-learned samples from a neighbouring department (Gers) are good enough to enrich the previously developed model. The achieved classification performance (OA ~ 78 %) is equivalent to that of a random model built with only department-specific samples (OA ~ 80 %). Moreover, the new model is adapted to both study areas.
In this case, Radar information has not improved significantly the classification accuracy of poplar plantations but it will be considered for stand age distinction due to its sensitiveness to the vegetation structure.
Further work will be conducted to extend the developed model, with the AL approach, to a more contrasting study sites in terms of poplar variability, climatic conditions and silvicultural management practices.

Area A - Board 64

Detailed knowledge about tree species composition is of great importance and the potential of multi-spectral earth observation (EO) data for tree species mapping is well known. Until recently, most research focused on spectral and spatial resolutions. However, the temporal resolution can decisively influence the ability to distinguish between individual species. In 2015, the European Space Agency (ESA) launched its first Sentinel-2 satellite (S2A). Together with its twin satellite launched in 2017 (S2B), EO data are now provided with unprecedented spectral, spatial and temporal resolution. Here, we investigate the potential benefits of using high temporal resolution data for classification of five coniferous and seven broadleaved tree species in a highly mixed and diverse Central European forest. To run the Random Forest (RF) classification, cloud-free Sentinel-2 (S2) acquisitions between August 2015 and October 2017 within the overlap of adjacent S2 swaths of orbits 122 and 79 were analyzed in a 2-step approach. The available scenes were first used to partition the study area into six broad land-cover classes (Coniferous Forest, Broadleaved Forest, Grassland, Agriculture, Urban Areas and Waterbodies). Subsequently, additional RF classification models were created separately for the coniferous and the broadleaved class, taking into account all possible permutations of the S2 scenes. Each model includes a stepwise recursive feature elimination. The model quality was assessed in terms of out-of-bag overall accuracy (OOB-OA). Compared to the best mono-temporal result, the multi-temporal analysis approach improves the OOB-OA for the coniferous and the broadleaved tree species by more than ten percentage points. Remarkably, a combination of 6-7 scenes achieves a model quality equally high as the model based on all data; images from April until August proved most important. The most important spectral variables to distinguish between tree species are located in the red (conifers) and SWIR bands (deciduous), respectively. Furthermore, a semi-supervised learning approach is employed in an area of overlapping S2 swaths in order to improve classification accuracies for regions covered by a single orbit only. Opportunities and limitations of exploiting the higher temporal resolution in orbit overlaps are explored. Overall, the study highlights the potential of multi-temporal S2 data for species-level classifications in middle European forests. In a next step, the modeling approach will be extended to include data from different sensors and of higher spatial resolution.

Area A - Board 65

In 2016-2017 an extensive bark beetle outbreaks were observed in Polissya region of Ukraine, a climatic zone of higher humidity and lower temperatures with mostly monoculture pine-tree management forests. In particular, in the Drevliansky natural reserve more than 5000 hot spots were detected covering the 2.5 % of the forest area (470 ha). Trees, attacked by bark beetles colony, are dying at a very fast pace, in up to 8 weeks. Therefore, it is extremely important to develop methods for early detection of tree damage.
The foliage color change from green to gray is reported on the early stages of the disturbance by bark beetles, which might be detected on the remote sensing data under certain conditions. The aim of the research was to develop the algorithm of Sentinel images processing to track the beginning of tree fading effect, caused by the bark beetles. The objectives of the research include: 1) to analyze the spectral signature of canopy fading and tree mortality caused by back beetles in Polissya region; 2) to examine the conditions that may favor the spread of back beetle damages.
In this regard, the Drevliansky natural reserve is an interesting object to study the bark beetle spread factors and dynamics, because the immediate sanitary loggings where not applied there due to the legislation restrictions. With an available set of damaged forest areas detected in 2016-2018 and the aerial imagery in addition to Sentinel-2 time series, the retrospective analysis was conducted to range the “age” of damaged areas and to track the early features of pest attacks. Support Vector Machine, Random Forest, Principal Component Analysis and Maximum Likelihood Classification methods were applied for Sentinel images. The accuracy of 86% was received for Maximum Likelihood Classification, applied for Sentinel-2 images, obtained in July-August.
The results of the stand condition analysis confirmed that, flow accumulation, age group, stand composition (percentage of Pinus sylvestrus) and tree height summed with elevation effected the bark beetle distribution in the study area which is important for risk assessment and decision making.

Area A - Board 66

The current studies showed that warming temperature, repeatable dry and warm summers and mild winters are driving bark beetle outbreaks. The expansion of pine and spruce beetle outbreak is visible across North America, Europe and Siberia. In general, forest becomes less resilience to droughts and more vulnerable to insect attacks. The increase of the bark beetle’s population is one of major forest disturbance factors in the Tatra Mountains. The number of trees affected by bark beetles is increasing gradually since 2003 (Grodziki W. et al., 2010). In 2018, due to the warm spring, the bark beetles become active already in April.

The aim of this study is to detect a spatial and temporal pattern of forest degradation caused by beetle outbreak and to monitor forest changes over time using a time series of Sentinel-2 data. The study is carried out within the Tatra National Park in Poland, where forest covers around 152 square kilometres. The forest change detection is carried out using a combination of normalized difference vegetation index (NDVI) and normalized difference moisture index (NDMI). NDMI is calculated based on NIR (824 nm) and SWIR (1610 nm) bands. The forest changes are derived on the regular based, from one image to another, to detect new changes and to monitor forest affected areas. The analysis of Sentinel-2 images revealed drastic increase of the forest affected areas over the last three years. Figure 1 presents the forest damage areas which occurred within a year, between August 2016 and August 2017. The study demonstrats that most susceptible to periodic bark beetle outbreaks are mainly secondary spruce forest. To understand the current pattern of forest damage, a set of archive Landsat data acquired in 2000, 2005 and 2010 has been processed and analysed.
Fig. 1 Forest damaged by bark beetles during the period 2016-2017 based on Sentnel-2 data (yellow outlines).

Area A - Board 67

The increasing negative effects of climate change and invasive species on forests worldwide require the development innovative methods to monitor and measure quantitatively the change in forest health over vast regions. Forestry managers need new forest inventory information to respond to risks produces by stressors provoked by climate change. Recurrent droughts results in increased damages due to populational bursts of otherwise secondary herbivores. It is necessary that the estimation of structural variables will be complemented with the monitoring of physiological processes in order to describe the response to these stressors. Sustainable Forest Planning requires not only information about forest composition and its structure but also about vegetation condition in a spatial and temporal context. Remote sensing technologies have been shown to provide the means for wide coverage measurements with reasonable accuracy. In particular, new spectral indices obtained from high-resolution hyperspectral and thermal imagery have been demonstrated as candidate indicators for the early detection of physiological changes linked with diseases. Also a stress indicator suggested in several studies is the temperature of the canopy as an indicator of tree transpiration (stomata conductance rate).
In this study, a pilot 95 ha Pinus halepensis stand in Comunitat Valenciana (Spain) has been used to develop a model to detect stressed trees, with Tomicus destruens induced attacks and girdling of different stages.The potential detection of changes at early stages using different spectral vegetation indices derived from high-resolution hyperspectral imagery and temperature from with thermal imagery were studied. Concurrent with the image captures, attack and physiological stages of individual trees were surveyed in detail for the study area. Some of the indicators would also apply for the early detection of trees affected by the invasive Bursaphelenchus xylophilus,a very serious threat for European forestry. The results obtained in this study are discussed in the context its early detection. Preliminary results show differences in reflectance between 500-600 nm comparing healthy and declining trees after girdling (figure1).
The overall design of the study is to combine airborne physiological variables derived from hiperspectral and thermal sensors. The results also were compared with medium spatial resolution imagery from Sentinel 2. The forest status monitoring was also explored by comparing temporal rate of change of the Sentinel-2 derived spectral indices over areas with declining and healthy trees.

Area A - Board 68

Parallel with the lower price of Light Detection And Ranging (LiDAR) technology a newer research area has been emerged in the forestry. Normalized digital elevation model (Canopy Height Model, CHM) based single tree detection methods using local maxima can be misleading, because deciduous trees may have more peak points, and second canopy level may be hidden under the first level. This abstract describes a voxel-based approach which attempt to take all the vegetation points into consideration for identifying stems.

The airborne laser scanning point data were acquired in leaf-off winter period in order to get as many returns on stems as possible. The average point density is about 32 points/m2 using 4 overlapping flight strips with 8-8 points/m2. The study areas are 50 m x 50 m square-shaped forest scenes in Erdősmecske (South Hungary). Field measurements were performed after the clear cutting of the scenes. The position of the stumps was recorded by decimeter-level accuracy. The species were also recorded; Pedunculate oak (Quercus robur) is the dominant species but European beech (Fagus sylvatica), European hornbeam (Carpinus betulus) and Turkey oak (Quercus cerris) were also recognised. Besides field measurements, Fugroviewer software was used to identify tree stems in the 3D point cloud. This dataset was crosschecked with the field measurements, and it was suitable to utilise as an alternative reference data.

The geographical space of the scenes was split up to a set of volumetric elements (voxels). The points of the vegetation class were indexed by column, row and layer based on the location of X, Y and relative Z (above ground) coordinates in the 3D voxel grid system. The 3D voxels stored binary data (0, 1) according to the presence or absence of any vegetation point in the space. In order to identify the stems, algorithms searched for objects with large, nearly vertical features where the point density was relatively high. The upper part of the stems has branches, and the undergrowth can confuse the algorithm. Therefore, the lower 3-meters-high zone above the ground was filtered for the data processing where the herbaceous vegetation was sparse. Tree stems were identified as the grid points where (1) the point density has a local maxima, (2) the density of relevant voxels has a local maxima or (3) the relevant voxels occur in large number within a search radius.

More than 30.000 models were calculated combining many input variables such as (1) minimum and (2) maximum height of filtered zone, the (3) height and (4) width of voxels, the (5) search radius, (6) minimum tree height, (7) applied filtered (binary erosion and dilation on voxels in different shapes) and (8) thresholds which was able to handle the different point densities. A metric was applied to compare the efficiency of models based on the ratio of the True/False Positive/Negative matches. The analysis was also performed by the combinations of flight strips. Using the 32 points/m2 point density the 90% percent of the tree stems were found, and 99% of the modelled tree was valid.

Area A - Board 69

The EUMETSAT LSA SAF service is one of the few (if not the only) that generates evapotranspira-tion estimates in near real time at regional/continental scale by ingesting SRS derived information into a physical model of energy exchange between the surface and the atmosphere (Ghilain et al., 2011, 2017). ET is a metric of ecosystem functioning and it is related to water, carbon, and energy cycles. Thanks to trend analysis, continuous monitoring of evapotranspiration, presents a great po-tential for the early detection of droughts and for the preparation of mitigation strategies, it can be combined with other variables like precipitation and/or potential ET to estimate severity, extent and duration of a drought at regional scale. In hydrology and water management, a potential utility of the product is to help hydrologists and water managers to estimate the water balance at district and basin scale. Indeed, hydrologists have a quite accurate knowledge of the amounts of precipitation over the basin and in many cases also streamflow measurements at gauged stations. Nevertheless, for the estimation of water released as water vapour, very often the estimation is realised on the ba-sis of potential ET which may overestimate the real amount of ET and lead to an over/under estima-tion of the other terms of the basin water balance. Therefore, LSA SAF ET could bring value, as an appropriate areal estimate of the loss of water through evapotranspiration. In agricultural manage-ment, early drought detection can be necessary, and highly demanded. Although precipitation and vegetation indices are most used, the indicator that can be derived from ET acts as an early warning indicator, as the lack of real evapotranspiration compared to potential evapotranspiration quantita-tively give an estimate of the dryness of the soil, before any senescence of the vegetation.

In this contribution we present examples of the LSA-SAF evapotranspiration products generated with the latest version of the algorithm which has been improved thanks to the use of new variables derived from earth observation by MSG SEVIRI. We will highlight the potential of ET products in droughts detection and monitoring as well as in hydrology.

Area A - Board 70