Spatially-distributed estimates of the faction of absorbed photosynthetically active radiation (FAPAR) from satellite imagery are essential for assessing global carbon balances. For FAPAR products, a target accuracy of 10% has been considered acceptable by the Global Climate Observing System (GCOS). Satellite-derived FAPAR products have been found to often cross this threshold, especially in forests. Sentinel-2 (S2) enables to derive FAPAR at decametric resolution, presenting new possibilities to assess the productivity and dynamics of forest ecosystems. To validate the product, ground measurements from different ecosystems are required. However, only few sites are equipped to provide permanent and robust FAPAR estimates. Additionally, investigations on the uncertainties of in-situ FAPAR have stressed to consider uncertainties in the validation process.
We present a validation study of the S2 FAPAR product in different forest ecosystems for the year 2017, including a temperate conifer-dominated forest in Southern Germany, a boreal deciduous forest in Northern Alberta, Canada and a tropical dry forest in Costa Rica. Incoming and transmitted PAR were measured with PAR sensors above and inside the canopy from which 10 min two-flux FAPAR estimates were calculated. Sensors were arranged in Wireless Sensor Networks (WSNs) to ensure synchronized data acquisition. Cloud-free S2 images were atmospherically corrected ("Sen2Cor" algorithm); FAPAR was retrieved using the “Biophysical Processor” (S2 toolbox). For the temperate site, recent investigations have revealed that FAPAR observations were affected by significant overestimation by up to 0.06 under solar zenith angles (SZAs) above 60° and by up to 0.05 during the presence of colored leaves, similarly to results obtained with canopy RTMs. To assess the influence of SZA, we investigated different temporal aggregation schemes, i.e. daily integrated FAPAR and FAPAR acquired at solar noon.
Our results showed high discrepancies between FAPAR ground measurements and the S2 FAPAR product. The highest agreement was obtained at the boreal forest, with a bias accounting to 15% (R²=0.64). We consider these discrepancies to be mainly attributed to different underlying FAPAR definitions (measured total vs. green FAPAR of the S2 algorithm) and the generic nature of the algorithm. As for uncertainties of ground data, an overestimation of FAPAR due to leaf color change was observed at all sites. Further, we found that the choice of temporal aggregation scheme reduced bias due to SZA at the conifer-dominated site. However, when taking sources of uncertainties into account, only the boreal site fulfilled the accuracy requirements set by the GCOS.
In contrast to absolute values, we found good agreement on seasonal dynamics, with the onset of leaf development and senescence well-represented by S2. Further, the spatial variability encountered across the sites was reasonably represented. Spatial variability attributed to species composition (i.e., contributions of spruces and beeches) at the temperate site was clearly reflected in the S2 product, thereby highlighting the sensor´s capabilities for forest applications. Overall, our experiences demonstrate the need for continuous, multi-sensor FAPAR observations for understanding uncertainties of ground data. We conclude that FAPAR monitoring performed with WSNs can be used to efficiently retrieve in-situ FAPAR to improve satellite-derived FAPAR products.

Area A - Board 45

Changes in land surface phenological events are among the most sensitive biological responses to global change. During recent decades, the spatio-temporal dynamics of land surface phenology have been assessed at regional and global scales using optical satellite systems sensitive to changes in vegetation greenness/productivity as monitored from vegetation indices like the Normalized Difference Vegetation Index data (NDVI). Recently, satellite passive microwaves have been established as an independent alternative to assess vegetation functioning using the vegetation optical depth (VOD), which is sensitive to the vegetation water content of both the leafy and woody components and is insensitive to cloud cover. Here, we analyze the spatial pattern and long-term trends of phenological metrics (start of growing season (SOS) and end of growing season (EOS)) derived from VOD and compare these with outputs from NDVI over the period 1992-2012 for the seasonally dry tropical regions (25°N to 25°S). Whereas VOD shows a generally later SOS/EOS and a shorter LOS as compared to NDVI, the specific differences between VOD and NDVI phenological metrics vary across vegetation density levels within and amongst sub-continents (with hot-spot diverging patterns in northern Australia, Sudano-Sahelian Africa, Southern Africa, and South America). NDVI data quality is deemed uncertain for calculation of phenology metrics for about 40% of the areas studied. For areas of adequate NDVI data quality, diverging trends in phenological metrics from VOD and NDVI are widespread (30% of the study area when combining the three metrics investigated). We also examined potential drivers of observed phenological trends, including variability in rainfall, fire, woody cover, and land cover. Taken together, phenological metrics derived from satellite passive microwaves have proven (1) to be an instrument for monitoring changes in ecosystem functioning in areas otherwise obscured by clouds in optical satellite imagery and (2) to be able to provide a complementary biophysical description of vegetation phenology of high relevance for studies of altered ecosystem functioning and structure caused by global change.

Area A - Board 46

Climate change modifies the functioning of terrestrial ecosystems leading to increased primary production, growth and greening of plants. In contrast, tropospheric ozone (O₃) damages forests reducing primary production. However, how the ozone threat to vegetation is affected by the interaction between climate change and air pollution is not well established.

To date, the European standard used to estimate the negative effect of O₃ on vegetation is the AOT40; however, in the recent years, a new metric has been proposed, the Phytotoxic Ozone Dose (POD). AOT40 only depends on the O₃ concentration in the air while POD accounts for environmental conditions that influence the O₃ uptake through the stomata. The O₃ uptake is estimated from photosynthetically active radiation, air temperature, vapor pressure deficit, soil water content and the plant phenology.

Plant phenology plays a pivotal role in the climate system as it regulates the gas exchange between the biosphere and the atmosphere. To efficiently estimate the O₃ uptake by forest, a consistent function describing the start (SGS) and the end dates (EGS) of the growing season is needed. Actually, in many risk assessment studies, this function is based on a simple latitude and topography model and it does not consider the inter-annual variability in the climate system as well as the influence of the climate change on the lengthening of the growing season and thus on the duration of the period during which plants can uptake O₃.

Remote sensing is a valuable tool for plant phenology estimation. In this study, two satellite datasets were used to estimate the SGS and EGS across Europe in 2011 with spatial resolution of 0.1 x 0.1 degree: biweekly composited GIMMS LAI3g and NDVI3g.

Results indicate a common spatial pattern in the cumulated stomatal uptake of O₃, although SGS and EGS estimates remarkably differ. At regional scale, climate is the main driver of the uptake of gases from the atmosphere to the vegetation. The climate change increased both growing season length (~7 days/decade) and stomatal conductance and thus enhanced the stomatal O₃ uptake by forests (+ 5.9%) over the time period 2000-2014, leading to an overall increase of potential O₃ damage on plants. POD increases primarily because of trend in phenology (+5.1%), soil moisture (+2.8%), temperature (+2.2%) and air humidity (+1.1%), while trend in O₃ alone would lead to a decrease of POD of 4.03%.


Three main reasons explain the increased risk for vegetation: 1) warmer temperatures lead to a longer growing season and to increased stomatal conductance, hence higher O₃ uptake, but also 2) warmer temperatures promote the chemical reactions that lead to O₃ formation, and 3) plants emit more volatile organic compounds at higher temperatures, which, in turn, can increase O₃ formation.

Our results indicate that improved estimates of O₃ fluxes require a better representation of plant phenology in the models used for O₃ risk assessment. The ongoing climate change leads to an increased oxidative O₃-based stress to the European forests, despite the reduction in O₃ concentrations. Forest protection against the negative impacts of O₃ thus requires new strategies that integrate both climate and air quality policies.

Area A - Board 47

The joint retrieval of vegetation structure (LAI, fCOVER) and biochemistry (foliage chlorophyll, Cab, and water concentration) from directional reflectance spectra is an ill-posed problem. Retrieval algorithms based on either hyperspectral or multispectral bi-directional reflectance spectra typically show local biases related to prior assumptions of measurements or radiative transfer models used for calibration. It has been recently hypothesized that one specific structure quantity - the directional area scattering factor (DASF) - can be retrieved in a well posed manner using red-edge spectra. Here we extend this hypothesis to jointly retrieve the DASF and Cab using red-edge spectra.

The new approach is verified using both radiative transfer simulations from Monte-Carlo models and measurements over crops and forested landscapes. Initial results suggest that the joint retrieval provides results with an uncertainty below 10% for the DASF and 20% or (10ugcm-2) for Cab for moderate to dense canopies (fCOVER>0.5) when using <10nm spectral sampling. The use of these retrievals for constraining estimates of biophysical parameters will be discussed.

Area A - Board 48

Natural damages significantly influence forest ecosystem and often have strong effect on forest economical potential and management. Typically, severe structural damages are caused by winds or ice storms, but also heavy snow load can break trees over vast areas. Accurate delineation and early detection of forested areas affected by snow damages are critical for estimating economical losses, as well as for guiding further forest management activities.

To date, despite their notable versatility and robustness to imaging conditions, imaging radars were rarely used for snow damage mapping in boreal zone. In this paper, we evaluate the potential of satellite C-band SAR multi-temporal imagery for snow damage mapping in Finnish boreal forest.

The SAR imagery were represented by time series of Sentinel-1 scenes acquired in interferometric wide-swath (IW) mode. The study site was located in the south-eastern part of Northern Finland. Methods used in the study included: improved k-nearest neighbour method (ik-NN), logistic regression analysis and support vector machine classification. Snow damage data from 2018 were used in training and validation.

Our results showed an overall accuracy of 91% in delineating snow-damaged forest areas when using SVM and somewhat lower with ik-NN and logistic regression. SVM gave also the highest user accuracy (90%) while ik-NN gave the highest producer accuracy (80%). These results indicate that it is possible to develop an operational system for the snow damage mapping.

Additionally, stem volume estimation was performed using multitemporal Sentinel-1 data acquired before the damage events. This was done using i-kNN method. The relative RMSE, coefficient of variation, of stand level stem volume estimate (m³/ha) was 35.6% of the mean volume of 99.1 m³/ha, and that of the most common species Scots pine 37.5%. These results are competitive even compared with airborne laser scanner based. Further, we also tested the potential of estimating the total volume affected by the damage. The first results encourage to continue development both for damage area estimation and mapping as well as volume estimation. In such manner, it becomes potentially possible to estimate the total volume of damage affected forest and make assessment of economical impact of the damage.

Figure 1. An example of snow-load damage map (red) displayed on a volume map (rainbow coloring, white is non-forest land). Area size is 8 km by 12 km.

Area A - Board 49

The natural forest area in the Central Vietnam has spatially declined over time. Land cover and land use change in comparison with forest quality in terms of structural parameters confirmed the decline of natural forest area and its quality correlated significantly to land use change which causes serious effects to ecosystem provisioning services. The hypothesis was that if the land cover and land use change impact to the ecosystem services and if open-source satellite images could help to measure the changes over time series in the central Vietnam. We, therefore, conducted the random sample plot study along the newly constructed road through the natural forests in the central Vietnam. A 90 sample plot stretching with the size of 900m2 were established as ground truth inventory to measure forest structural parameters (height, diameter at breast height(DBH) of all tree species with the DBH ≥ 10cm). The coordination and topographical characters of sample plots such as aspects, elevation were recorded with GPS. We used the time-series satellite imageries (e.g. LANDSAT, Sentinel 2) to classify the vegetation cover indices and its structural parameter changes (basal area, number of trees, species and volume) along the newly constructed road where the forest inventory was done. The values of vegetation indices, derived from LANDSAT and Sentinal 2, such as the Normalised Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Advanced Vegetation Index (AVI) at random sample plots were processed in toolbox of ArcGIS 10.4 and ENVI 5.3. The mixed effect linear models in R-studio 2015 was used for modeling the changes of vegetation cover and forest structural parameters over time-series. The values of vegetation indices and forest structural parameters closer to the constructed road in comparison with other out-skirt random plots are significant different at p<0.05. This findings could help to retrieve the landscape and ecosystem provisioning services and to evaluate the impact of land use change and land conversion on ecosystem services over time series.

Keywords: Vegetation indices, forests, ecosystem services, mixed linear models, land use change

Area A - Board 50

The summer of 2018 has been extraordinarily hot and dry in many parts of Europe. In most countries, the months between May and August broke several heat records and even surpassed the conditions during the drought of 2003. This caused hundreds of wildfires, led to freshwater sparsity and severely damaged crops. In Germany, the time between April and October was the driest ever recorded, precipitation was 40% below the longtime mean while temperatures rose several degrees above the longtime mean.
Particularly damaged by drought were forested areas. In many parts of Germany, forests were discolored due to a lack of water. This was especially visible at forest borders and on hill slopes. In the undergrowth, young trees, which lack the extensive root network of older trees, were endangered by drought damage. It is estimated, that in German forest plantations 85% of saplings died, causing an economic damage of 5.4 billion euros. The exact degree and extent of damage to the forests will only reveal itself in 2019, when it is clear, which trees were unable to resprout after losing their leaves early in summer 2018. Another factor will be the amount of precipitation during the winter months and if this will suffice to compensate for the water deficit accumulated during the summer and autumn months.
However, not all forests were susceptible to the same degree. Environmental factors determine, how exposed trees are to water shortage. We hypothesize, that forests growing on lime soils suffer less from water deprivation due to higher soil water retention compared to sandy soils. Additionally, forests situated in uplands might be less susceptible to drought due to lower temperatures and a higher propensity for thunderstorms. Sentinel-2 data is used to asses the spatial distribution of drought damage in forested areas in northern Bavarian, Germany. In addition, it is tested how environmental factors such as soil type, slope and elevation influence suffered drought damage.
Furthermore, the remotely sensed data is combined with field data from an extensive field campaign from the Summer 2018. The field data holds information on tree species and amount of damage sustained during the drought. All plots will be revisited in 2019 to investigate tree survival.
Additional to analyzing the environmental factors driving the severity of drought effects, this study aims to detect differences and correlations between remotely sensed drought damage (i.e. damage to the upper most vegetation layer) and drought damage measured in the field (i.e. drought damage to the lower vegetation levels).

Area A - Board 52

The Horse River Fire (HRF) started on May 1, 2016, approximately 10 km southwest of Fort McMurray, Alberta, Canada. The extremely dry conditions in northern Alberta, coupled with low relative humidity and high winds contributed to the rapid expansion and unprecedented growth of the fire.
RADARSAT-2 MF1 and FQ18 data was acquired before, during, and after the HRF. Analysis of RADARSAT-2 MF1 (HH) imagery of the HRF indicated a significant increase in the backscatter for regions that been impacted by the fire. The backscatter increase was related to the change of the tree canopy that was in turn was related to the burn severity. Following the commonly-used classes, the burn severity was segmented into four categories: High, Medium, Low, No Change. Backscatter variability was observed for the different burn severity regions, but the most significant was the High burn category.
RADARSAT-2 FQ18 data was used to investigate the backscatter change for different polarizations and to identify the dominant scattering mechanisms for different burn severity categories. A burn severity map was obtained from the Alberta Wildfire Management. The burn severity map was clipped to the spatial extent of the FQ18 image and a subset of the FQ18 was selected for analysis. The choice of the subset was based on a representative area that encompassed all of the burn severity categories and was characterized by flat to gently rolling terrain. The Single Look Complex data were converted to a 4x4 coherency matrix format and then symmetrized to a 3x3 matrix. The coherency matrix was filtered using a 7x7 Lee polarimetric filter, followed by the Freeman-Durden decomposition.
Prior to the HRF impinging on the spatial coverage of the FQ18 data, the increase in HH, HV, and VV was consistent with the seasonal increase that was observed in the MF1 imagery. Although the fire had burned the study area in early May, there was no obvious difference in the backscatter in the June 4 image. However, the June 28 image displayed a significant increase in the backscatter for all polarizations. In general the largest increase in backscatter was associated with burn severity as follows: High > Moderate > Low > Unchanged.
The Freeman-Durden decomposition was used to understand the change in the scattering mechanisms that were associated with different burn severity. Figure 1 shows a comparison of the double bounce (DB), volume (Vol), and surface scattering (Sfc) for high burn severity (H-H), and a control region that was ~ 20 km from the edge of the fire so not burned. In terms of scattering strength, volume > surface > double bounce. The dominance of volume scattering was consistent with expected scattering behaviour for a forest. The strength of the three scattering mechanism increased dramatically between June 4 and June 28, especially for High burn severity. Although volume > surface > double, the largest increase in the strength of the scattering mechanisms was double bounce (~ 8 dB), followed by surface (~ 4 dB), and finally volume (~ 2.5 dB). The dominance of double bounce scattering for H-H was attributed to a lack of foliage which resulted in scattering from tree trunk and the forest floor.
From a harvesting perspective, the classification of High burn severity areas in the aftermath of a forest fire is important. The higher the burn severity, the greater the tree mortality, but tree mortality is related to insect infestation. If trees are harvested before insect infestation, the timber is useable, but the value of timber drops dramatically once insect infestation has occurred. Therefore, if high burn severity areas are mapped as soon as possible after a fire, the areas that should be harvested first can be identified.

Area A - Board 53

Information related to the impact of wildfire disturbances on ecosystems is of paramount interest to account for environmental loss, to plan strategies for facilitate ecosystem restoration and to monitor the dynamics of vegetation restoration. Besides being natural hazards, wildfires could determine the reduction of vegetated surface and consequently reduce soil protection provided by the tree root system and therefore triggering landslides activation. Proper evaluation of post-fire vegetation restoration should not exclusively rely on fire severity and post-fire conditions, but also consider the plant conditions before the wildfire occurred. For this purpose, phenological metrics can represent a good candidate to monitor and quantify vegetation recovery after wildfire disturbances.
Satellite observations have been demonstrated to be a suitable tool for wildfire disturbed areas monitoring, allowing both the identification of burned areas and the monitoring of vegetation recovery. The launch of new satellites represents nowadays an opportunity to further improve the quality of wildfire-disturbed areas monitoring and many research efforts are moving toward understanding the added value brought by these new technologies. In particular, the MSI sensors onboard the Sentinel-2A and 2B satellites provide for the first time free data access with unprecedented trade-off in spatio-temporal resolutions (10-60 m pixel size and 5-days revisit time).
In this framework, the present contribution aims to exploit the benefit of Sentinel-2 time series data for the monitoring of post-fire restoration areas, with particular attention on vegetation phenology.
To achieve this goal, we focus our attention on the case study of the area surrounding Napoli (Italy), where a significant number of wildfire events took place during summer 2017 after a prolonged drought period. A database of Sentinel-2 acquisitions was collected for the reference tile T33TVF during the period 2016-2018 covering the pre-fire, fire and post-fire phenological seasons. Sentinel-2 images were processed in order to derive smoothed time series of Leaf Area Index (LAI) along the reference time span. LAI biophysical variable values were computed using the biophysical processor available in SNAP, while a Whittaker smoothing operator was applied to the time series to avoid residual noise due to residual cloud contamination errors.
A 2017 semantic map over the study site, indicating the burned areas extent and the temporal occurrence of fire events, was derived by applying the methodology proposed by Filipponi (2018). Starting from this product, the main fire events were isolated and identified as single region of interest (ROI). Representative LAI average time series were then calculated for each ROI and each analyzed year. A set of phenological features relying on agronomical expert knowledge were extracted from LAI time series during the pre- and post- fire seasons. Later, for each ROI, LAI average time series and extracted phenological features were used as descriptors and clustered, in order to characterize the different post-fire LAI response behaviors. Clusters were finally compared against local topographic and ancillary information, to qualitatively rank vegetation restoration dynamics and to identify the environmental drivers mainly influencing the fire restoration process.
The use of temporal signatures and phenological features derived from Sentinel-2 time series demonstrated to be an effective tool for tracking vegetation cover dynamics after wildfire disturbance. Results lead to isolate different vegetation recover trends and identify the main environmental drivers involved in the restoration process, highlighting the key role played by vegetation restoration monitoring and stimulating the development of downstreaming operational services, indispensable for the provision of fundamental information to decision makers and needed for the implementation of recovery plans and strategies over fire affected areas.

Area A - Board 54

Recently, UAV Laser Scanning (UAV-LS) has become available that produces ultra-high density point clouds (>1000 points/m2) on local plots. Such point densities have only been achieved with Terrestrial Laser Scanning (TLS) in the past. Therefore, UAV-LS is an interesting data source for fast forest structure characterisation, both in horizontal and vertical dimensions. Additionally, the fast mapping speed over larger areas compared to traditional field methods makes re-assessment and monitoring to investigate temporal patterns more viable. This is relevant for example to monitor forest recovery and regrowth.

We tested UAV-LS for forest structure assessment in two separate experiments and compared it to TLS data. The first experiment was aiming at modelling explicit 3D volume and structure of single trees in a 180x100 m forest plot covering 5 stands of different composition in forested study area located in the Netherlands. The study area contained 2 deciduous and 3 coniferous species, and 5 canopy structure types. TLS data was acquired within 2 days on 58 scan positions, while UAV-LS data consisted of a flight with 9 flight lines captured within approximately 2 hours. After co-registration of TLS and ULS point clouds of 134 single trees were extracted. Quantitative Structural Models (QSM) were built from these point clouds. The models were optimized based on fitted cylinders extracted from the point clouds. Results show a good agreement of UAV-LS and TLS for branches with diameter >40 cm.

The second experiment consisted of a multi-temporal campaign with UAV-LS and TLS that covered the forest leaf flush period and associated structural changes. In total 3 flights were performed and TLS was acquired in 12 locations in the study area at each flight event. We derive Plant Area Volume Density (PAVD) profiles from both data sources. PAVDs have so far only been derived from TLS or Airborne Laser Scanning (ALS). They represent an explicit characterisation of the vertical distribution of plant material in the canopy that is extremely difficult or impossible to derive with other field methods. We compare the products based on the two data sources in order to find out if UAV-LS can complement TLS canopy assessment.

Area A - Board 56

Vegetation fraction cover (fCOVER) is an essential variable for monitoring forest canopy processes as evapotranspiration (ET). ET plays a major role in the hydrological cycle and in the land surface energy balance. It may be used for quantifying water stress. At local scale ET can be estimated from detailed ground observations (eddy covariance towers), but these measurements are only representative of very limited homogeneous areas. When regional information is required, e.g. for monitoring the extension of drought or water resources, flux tower measurements cannot be used directly but ET may be estimated using potential evapotranspiration formulas and soil water balance equations driven by remote sensed fCOVER .

We implemented a simple soil water balance model, SimpKcET, with the objective of mapping daily evapotranspiration and water stress over large territories and long periods of time. Evapotranspiration and water balance were calculated from potential evapotranspiration and precipitation provided by meteorological reanalysis such as SAFRAN (Meteo France) or ERA5 (ECMWF) and precipitation products such as MSWEP (Princeton University). Transpiration and evaporation were considered. Transpiration depends on a canopy coefficient which account for vegetation fraction cover (fCOVER) and soil moisture. Evaporation depends on the fraction of soil surface and antecedent precipitation. Water stress is simply defined from the ratio of actual ET to potential ET.

In this study we compared the results obtained using vegetation fraction cover derived from different sources of remote sensing data.
• We derived fCOVER from the Enhanced Vegetation Index (EVI) available since 2000 as part of the MODIS vegetation index products (250 m – 1000 m spatial resolutions). We developed relations between EVI and fCOVER over a synthetic dataset simulated by using the SAIL radiative transfer model.
• We also used the vegetation fraction cover products that are provided by the COPERNICUS Global Land Service. Products are available at 1 km resolution from data of the VEGETATION sensor starting in 1999 and at 300 m resolution from the PROBA-V sensor starting in 2015.

Modeled daily evapotranspiration was evaluated over flux tower data on the holm oak (Quercus ilex) forest of Puechabon near Montpellier and the mixed holm oak (Q. ilex)/ Aleppo pine (Pinus halepensis) in Fontblanche near Marseille. Both sites are situated in South of France and are part of the ICOS network. We also evaluated ET calculations at the yearly time scale over the watershed of the Fontaine de Vaucluse karst system near Avignon. This system (1000 km2) has a unique spring as outlet and annual evapotranspiration was estimated by comparing annual flow and precipitation over the watershed. Around 80 % of its surface is covered by forests. SimpKcET was also implemented over South East France at 1 km resolution and compared to other ET products (MOD16 and SSEBop).

Evapotranspiration was estimated with a root mean square error between 0.4 mm d-¹ and 0.6 mm d-¹ depending on model inputs. Influence of fCOVER in the results was related to different factors: differences in time dynamics of fraction cover (COPERNICUS products exhibit longer periods with maximum cover), differences in stability of the estimation (MODIS vegetation indices product presented large temporal noises, while COPERNICUS fCOVER products were provided after temporal smoothing), estimation of fraction cover from MODIS EVI required to account for the dominant type of vegetation as the relation between EVI and fCOVER was changing depending on SLA (specific leaf area). However, we also showed that the variability in fCOVER estimations was generating less uncertainties in ET estimations than rain estimations from various spatialized rain products and soil water properties (soil available water capacity) from various soil maps.

Area A - Board 57

The main aim of the SAT4EST project is to design system and to build non-state forest focused service dedicated to the local government administration in Poland. Forest in Poland covers around 29,5% of the country. About 80% of the total forest area is publicly owned and the remaining 20% belongs to private owners. Forest is managed according to the forest management plan, taking into account the sustainable forest economy. The privately-owned forest is managed by the local administration (district governor) that is responsible for continuous monitoring of forest areas. Up to now, there is a lack of operational tool, which supports local administration in monitoring of forest cover and forest changes at the regional level. A time series of Sentinel-2 images is used to derive a set of EO-based products, which meet the user requirements and obligations. These products are: a) forest and woodlands cover derive on the yearly bases, b) forest change mapping on the regular bases, c) forest type (coniferous and broadleaf), canopy coverage and forest condition (derived based on the combination of spectral indices). There are also products available on demand like assessment of the forest damage due to windstorms, fires, insect damage areas.
The Sentinel-based products are integrated in the SAT4EST system. The system consists of the remote sensed data component, ancillary data component, data processing and map server components. The users are able to compare various EO-based products with existing ancillary data, to upload own datasets, calculate statistics and generate reports. The additional functionality of the service is that the users are able to familiarized with the satellite images by viewing present and archive satellite images (i.e. Sentinel-2 and Landsat). The service is at the pre-operational stage – it has been implemented and tested in three selected pilot districts in Poland. More details on the SAT4EST system and service is available on www.sat4est.pl.
The SAT4EST project is funded by the European Space Agency (ESA) through the Polish Incentive Scheme Programme.

Area A - Board 58

The pine wood nematode (PWN), Bursaphelenchus xylophilus, is a significant threat to European forests. Native to North America, it was first detected in Europe in 1999. Since the first detection was made in Portugal, eradication and containment efforts were put in place to little effect. New detections in the Madeira Islands and Spain attest to the difficulties arising from multiple factors, including insufficient early detection methodologies.

To address these challenges, this paper present a method for detection of PWN infected forest in Sentinel-2 images. The forest disturbance caused by this pest often happens on subpixel scale and affects noncontiguous trees, hence it results in a very weak signal in remote sensing data. Its detection therefore requires a judicious choice of mathematical methodology combined with prior knowledge. We propose an algorithm which consists of partially supervised preclassification of pine forest segments, followed by multispectral unmixing to extract the signatures of healthy forest and background soil and understorey, before we can test for weak disturbance signal by a detector which is guided by knowledge of the disturbed forest signature.

The development and verification of the method relies on medium-resolution data from Sentinel-2, for which the method is intended in an operational setting, and high-resolution data from WorldView and RapidEye. Preliminary results show that the problem is challenging, but feasible, and that exploitation of multitemporal data that can track the evolution of the disturbance may strengthen the approach.

The PWN is an intrusive species which infects pine trees and brings pine wilt and death to trees which are significantly stressed. It is for this reason considered to be one of the top five most damaging biotic agents in European forests and all European Union member states are required to do annual surveys to ensure that the PWN has not entered their forests. A direct consequence of the pine wilt disease is dehydration [Kuroda, 2008], which again leads to changes in physical properties like color, texture and internal structure of the needles. These are features that can theoretically be observed in multispectral imagery, and satellite remote sensing therefore provides a fast and cost effective alternative to manual ground surveys and aerial photography missions for PWN detection.

A comprehensive pilot study by conducted by the Joint Research Centre documented that PWN infected trees can be discriminated using various features from hyperspectral and high-resolution multispectral images [Beck et al., 2015]. The authors of the current paper tried to develop a demonstrator PWN detection service based on the normalised difference vegetation index (NDVI) computed in Sentinel-2 images in the ESA-funded Silvisense project (ESA contract no. 4000117722/16/INB), but experienced high false-alarm rates and problems with detecting small objects and weak signals in mixed pixels. A follow-up study was conducted in the master’s project of [Hansen, 2018], who investigated refined procedures to extract weak signals from mixed pixels at medium resolution. Among the tested methods were algorithms for multispectral unmixing.

Spectral unmixing is a well developed technique in hyperspectral imaging [Boucas-Dias et al., 2012], where algorithms from blind source separation are among the methods used to determine signatures and fractions of the pure materials, also known as endmembers, that constitute the remote sensing measurements in mixed pixels. The number of spectral channels recorded by hyperspectral sensors is much higher than the number of mixed sources, and the endmembers can therefore be reliably determined. Spectral unmixing of multispectral data, with their more modest number of channels, represents a much bigger challenges in the general setting, which is probably the reason why this problem has been largely overlooked in the literature.

Among the few approaches to multispectral unmixing, we find the algorithm in [Karoui et al., 2012], which relies on the assumption that the data under investigation contains at least some pure pixels, that can be identified in a clustering procedure. It was found in [Hansen, 2018] that this condition was not met by the coarse Sentinel-2 resolution, and the method therefore failed. The method proposed in this paper is inspired by the multispectral unmixing approach in [Benachir et al., 2014], which relaxed the assumption to presence of pixels with a mixture of two materials. We attempt to simplify the task even further, by letting the unmixing algorithm focus only at pre-classified forest pixels, whose signatures are assumed to be a mixture of pure pine (healthy and diseased) and a background with low variation in a local neighbourhood.

The method proposed here consists of three stages. In the first stage Identify a local segment of pine pixels, which hopefully contains a two-material mixture: healthy pine and a soil/understorey background. Thresholding of the NDVI is used for simple prescreening of pine pixels, which later be elaborated into a more reliable pine classification. A spectral angle mapper (SAM) algorithm [Kruse et al., 1993] is then used to delimit a segment of nearby forest pixels, whose similar spectral signature should certify that they have a similar background. The SAM segmentation exploits prior knowledge of the spectral signature of pure healthy pine.

In the next stage we perform multispectral unmixing targeted at healthy pine and background. It is here assumed that there might be forest disturbance present in the segment, but in a minority of the pixels. This stage extracts the spectral signatures of healthy pine and backgroundfor the local segment without the use of prior information. In the final stage we detect any disturbance signal by using the signatures of healthy pine and background extracted in stage 2, together with prior knowledge of the signature of wilted pine. The signatures used as prior information have been measured in field studies.

In the results we demonstrate that the proposed method is able to detect PWN down to single tree level in stands of infected pine forest when the assumptions are met. In the verification we use field data from Condeixa-a-Nova and Coruche, Portugal, together with high-resolution images from Worldview-3 and RapidEye and medium-resolution images from Sentinel-2. We also discuss the role of temporal information in the detection of PWN in view of the temporal characterisation of the pest [Zarco-Tejada et al., 2018].

[Beck et al., 2015] Beck, P.S.A., P.J. Zarco-Tejada, P. Strobl and J. San Miguel-Ayans (2015): «The feasibility of detecting trees affected by the pine wood nematode using remote sensing», JRC Technical Report no. EUR 27290 EN, European Commision, Joint Research Centre, Institute for Environment and Sustainability.

[Hansen, 2018] Hansen, S. (2018): «Forest health monitoring by detection of biotic agents in satellite images», Master’s thesis in Energy, Climate and Environment, Department of Physics and Technology, UiT The Arctic University of Norway, Tromsø, Norway.

[Boucas-Dias et al., 2012] Boucas-Dias, J.M., A. Plaza, N. Dobigeon, M. Parente, Q. Du, P. Gader and J. Chanussot (2012): «Hyperspectral unmixing overview: Geometrical, statistical, and sparse regression-based approaches», IEEE. J. Sel. Topics in Appl. Earth Observ. Remote Sens., vol. 5, no. 2, pp. 354-379.

[Karoui et al., 2012] Karoui, M.S., Y. Deville, S. Hosseini and A. Ouamri (2012): «Blind spatial unmimxing of multispectral images: New methods combining sparse component analysis, clustering and non-negativity constraints», Pattern Recognition, vol. 45, no. 12, pp. 4263-4278.

[Benachir et al., 2014] Benachir, D., Y. Deville and S. Hosseini (2014): «Blind spatial unmixing of multispectral images: An approach based on two-source sparsity and geometrical properties», 2014 IEEE Int. Conf. Acoustic, Speech and Sig. Process. (ICASSP), pp. 3171-3175.

[Kruse et al., 1993] Kruse, F.A., J.W. Boardman, A.B. Lefkoff, K.B. Heidebrecht, A.T. Shapiro, P.J. Barloon, A.F.H. Goetz (1993): «The Spectral Image Processing System (SIPS)—interactive visualization and analysis of imaging spectrometer Data», Remote Sens. Environ, vol. 44, pp. 145–163.

[Zarco-Tejada et al., 2018] Zarco-Tejada, P.J., A. Hornero, R. Hernándex-Clemente and P.S.A. Beck (2018): «Understanding the temporal dimension of the red-edge spectral region for forest decline detection using high-resolution hyperspectral and Sentinel-2A imagery», ISPRS J. Photogramm. Remote Sens., vol. 137, pp. 134-148.

Area A - Board 59

Lianas are woody climbing plants that use trees as structural support for ascending to the canopy and constitute up to 40% of the woody stems and about 25% of the woody species. They allocate more resources to canopy development, stem and root elongation than to their structure, resulting in a high leaf to stem ratio compared to trees. As a result, lianas can significantly attenuate light and can contribute up to 40% of the canopy leaf cover. Since light is a major limiting resource, competition for light is one of the main reasons for reduced tree growth and regeneration. Attempts to disentangle aboveground and belowground competition between lianas and trees have shown that aboveground competition affected the allocation of biomass in trees and resulted in poorly developed crowns. Recent study has shown that liana load on trees alter the allometry of trees by decreasing tree slenderness resulting in shorter and thicker stems.
Since lianas are disturbance-adapted plants, liana abundance is likely to increase with increased forest disturbance. In forests with high liana density, this change in tree allometry owing to liana load could result in uncertain biomass estimates. In addition, change in biomass allocation patterns of trees have important implications for ecosystem functioning. With liana abundance increasing in the tropics, quantifying liana impact on forest structure and biomass is crucial.
Terrestrial laser scanning (TLS) can measure various forest structural parameters with high spatial accuracy non-destructively. We collected TLS data between February 2016 and April 2016 from a tropical secondary forest in Gigante Peninsula, Panama. We selected trees with no liana load and with heavy liana load (50 – 100% lianas in the tree crown) of the following species: Apeiba membranacea, Alseis blackiana, Cordia bicolor. We derived various crown structural parameters and total biomass for all the trees from TLS data.
We compared the different structural parameters for these trees with and without lianas to quantify the liana impact on the tree structure. Our results indicate significant differences in the crown structure of trees with and without lianas. Trees without lianas have larger crown projection area, crown height and branch volume than the trees with lianas. We also show that lianas could change the crown allometry of trees resulting in the reduction of carbon storage potential of trees by as much as 50%.

Area A - Board 60

SENSITIVITY OF PASSIVE MICROWAVES TO CARBON STOCKS IN TROPICAL FORESTS

David Chaparro1,*, Grégory Duveiller2, Maria Piles3, Alessandro Cescatti2, Mercè Vall-llossera1, Adriano Camps1, Dara Entekhabi4
*Corresponding author: david.chaparro@tsc.upc.edu

1 Universitat Politècnica de Catalunya, CommSensLab & IEEC/UPC, Jordi Girona 1-3, E-08034, Barcelona, Spain.
2 European Commission, Joint Research Centre, Directorate for Sustainable Resources, Ispra, Italy.
3 Image Processing Lab (IPL), Universitat de València, 46980, Valencia, Spain.
4 Department of Civil and Environmental Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, United States.


Monitoring vegetation carbon in tropical regions is essential, as the contribution of tropical forests to the Earth’s carbon balance and to climate change mitigation is still uncertain [1, 2]. Frequently, optical-infrared vegetation indices, as well as LiDAR data, have been used to this purpose (e.g., [1, 3, 4]). Nevertheless, optical techniques sense only the top of the vegetation canopy, and are masked by clouds. Alternatively –and complementarily- to VIS/NIR indices and Lidar data, microwave-based satellite missions are insensitive to atmospheric conditions and provide information of the attenuation that different vegetation layers exert to the soil microwave emission. This attenuation is represented by the vegetation optical depth (VOD) parameter, and related to the vegetation water content and to the vegetation biomass. VOD has been applied for biomass estimation using frequencies ranging from 4 to 30 GHz (C- to K-bands; [5, 6]).
As lower microwave frequencies have greater penetration depths, it is suggested that L-band VOD (1.4 GHz) should improve the capacity to retrieve biomass information. In that sense, recent research has focused on the applicability of L-band VOD to estimate carbon stocks [6] and has reported that L-band VOD enhances carbon estimates in specific regions and periods [7]. Still, exhaustive research to compare and assess the capacity of VOD at different frequencies to capture carbon variability is still incipient.

This study aims to contribute on filling this gap. It compares the sensitivity of different VOD products (from L, C, and X-bands) and an optical vegetation index (EVI) to the above-ground carbon density (ACD). It quantifies the relative contribution of ACD and forest cover (FC) proportion to the VOD/EVI signals. The study is conducted in Peru, southern Colombia and Panama, for which ACD maps have been derived from airborne LiDAR [4, 8, 9]. VOD datasets have been retrieved from the Soil Moisture Active Passive (SMAP) mission (L-band), and for the Advanced Microwave Scanning Radiometer 2 (AMSR2; C- and X-bands).

Results confirm the enhanced sensitivity of L-band VOD to ACD when compared to higher frequency bands and EVI, and are consistent through most of the regions and geographic features within the study area (Figure 1). In addition, it is also found that L-band VOD shows greater sensitivity to ACD than VOD at C- and X-bands in very dense forests, although in general terms the sensitivity of all VOD bands decreases under high biomass conditions.

Results presented in this work support the use of microwave VOD for large-scale forest carbon mapping. Continuity of low frequency microwave observations (L, C, X, Ku, and Ka bands) at much improved radiometric and spatial characteristics is now being proposed by the Copernicus Microwave Imaging Radiometer (CIMR) high priority candidate mission [10].


REFERENCES
[1] Baccini, A., et al. 2017. “Tropical forests are a net carbon source based on abovegroung measurements of gain and loss.” Science, 358(6360):230-234.
[2] Le Quéré, C. L., Andrew, R. M., Canadell, J. G., Sitch, S., Korsbakken, J. I., Peters, G. P., ... & Keeling, R. F. (2016). Global carbon budget 2016. Earth System Science Data, 8(2), 605-649. doi: 10.5194/essd-8-605-2016.
[3] Sun, Z., Peng, S., Li, X., Guo, Z., & Piao, S. (2015). Changes in forest biomass over China during the 2000s and implications for management. Forest Ecology and Management, 357, 76-83. doi: 10.1016/j.foreco.2015.08.013.
[4] Asner, G. P., Clark, J. K., Mascaro, J., García, G. G., Chadwick, K. D., Encinales, D. N., ... & Balaji, A. (2012). High-resolution mapping of forest carbon stocks in the Colombian Amazon. Biogeosciences, 9(7), 2683. doi:10.5194/bg-9-2683-2012.
[5] Liu, Y.Y., et al. 2015. “Recent reversal in loss of global terrestrial biomass.” Nature Climate Change, 5:470-474.
[6] Brandt, M., Wigneron, J. P., Chave, J., Tagesson, T., Penuelas, J., Ciais, P., ... & Rodriguez-Fernandez, N. (2018). Satellite passive microwaves reveal recent climate-induced carbon losses in African drylands. Nature Ecology & Evolution, 2, 827-835. doi: 10.1038/s41559-018-0530-6.
[7] Vittucci, C., Ferrazzoli, P., Kerr, Y., Richaume, P., Guerriero, L., & Laurin, G. V. (2016b). SMOS forest optical depth intercomparisons over pan-tropical biomes. In Geoscience and Remote Sensing Symposium (IGARSS), 2016 IEEE International (pp. 5311-5314). IEEE. doi: 10.1109/IGARSS.2016.7730383.
[8] Asner, G. P., Mascaro, J., Anderson, C., Knapp, D. E., Martin, R. E., Kennedy-Bowdoin, T., ... & Potvin, C. (2013). High-fidelity national carbon mapping for resource management and REDD+. Carbon Balance and Management, 8(1), 7. doi: 10.1186/1750-0680-8-7.
[9] Asner, G. P., Knapp, D. E., Martin, R. E., Tupayachi, R., Anderson, C. B., Mascaro, J., ... & Llactayo, W. (2014). Targeted carbon conservation at national scales with high-resolution monitoring. Proceedings of the National Academy of Sciences, 111(47), E5016-E5022. doi: 10.1073/pnas.1419550111.
[10] MET Norway & CIMR Expert Team. (2018). CIMR. The Passive Microwave Satellite Mission for EU Copernicus. Retrieved from: https://cimr.eu.

Area A - Board 61

Whether based on terrestrial devices or more recently on drones, laser scanning data have shown their promising capacity to estimate the 3D geometry at tree and forest plot level, and can be considered currently as a true disruptive technology for the estimation of Above Ground Biomass (AGB). Beyond a significant support [well-recognized advantages] for AGB estimation, laser scanning data can also provide relevant inputs for the parametrization of forest scattering forward models, for which the 3D geometry of scatterers is needed.
Unlike for AGB estimation which is mainly driven by trunk volumes and wood densities, smaller elements like branches and leaves can also significantly impact scattering, depending mostly on the ratio between the smallest dimension of a given scatterer and the effective wave-number. While a test case defined with a single scatter would easily give the ratio beyond which small scatterers can be neglected regarding their impact on microwave scattering, the more realistic simulation of a forest stand including million of scatterers per hectare is much more complex, given in particular the antagonist effects of individual and bulk scattering, which are not necessarily monotonic functions of the dimensions in the resonant regime.
In order to address these questions, the present study consists first in a sensitivity analysis performed in the cases of two test sites of tropical forest, related to the AfriScat and TropiScat experiments (in Ghana and French Guiana respectively) at which laser scanning and multi-frequency radar data have been acquired. Given the complexity of the 3D reconstruction from the laser scans (especially for the higher and smaller scatters often masked), auxiliary information based on allometric equations have been also used in order to fully parametrize the MIPERS4D model which outputs can be also compared with time series at P, L and C-band frequencies. Depending on processing methods for the 3D reconstruction, on experimental design (terrestrial, drone or combination), and on integration levels with standard allometries, several scenarios based on the associated mock-ups have been considered. While previous results focused on the link between changes of dielectric permittivity (highly variable with water content under the effects of evapo-transpiration) and temporal decorrelation, we propose here to use the coherences at the three polarizations and several tomographic layers in order to build a metric to assess whether the uncertainties due to geometric dimensions is significant or not. Beyond the progress related to a better understanding of microwave interactions with natural medium, this study contributes to better define the requirements for the future terrestrial or drone based laser acquisitions.

Area A - Board 62

Following the TropiScat and AfriScat experiments that took place in French Guiana and Ghana respectively between 2011 and 2017 in the framework of the preparation phase of BIOMASS mission, a new campaign TropiScat-2 has been set up since March 2018. This campaign, located in the experimental site of Paracou in French Guiana, has three major interrelated objectives: First, to extend the time series provided from previous TropiScat and AfriScat missions, mainly aiming at characterizing P-band temporal decorrelation, a key element for BIOMASS mission acquisition scenarios. Second, to investigate the possible synergies between BIOMASS and other sensors, such as SENTINEL1 that provides C band data acquired with a short revisit time. Finally, to deepen our understanding of the underlying physics, based on Electro-Magnetic (EM) models requiring a detailed characterization of the environment, both in regard to the 3D geometry of the scatterers and their water content (dielectric constants).
This new campaign started by a feasibility study that allowed us to select different types of C-band antennas which were tested in March 2018 according to several configurations by adjusting the acquisition time and studying the coherence and intensity variability over many days. Preliminary results demonstrated an important effect of environmental conditions on the evolution of intensity and coherence. In particular, differences between day and night coherences were highlighted and explained by the convective motion of forest scatterers. Indeed, daytime coherence values obtained from repeat-acquisitions were already low after few seconds onwards, while we were able to obtain high coherence values (>0.6) up to several hours during night.
In addition to its multi-frequency features, this campaign can also rely on auxiliary measurements not only from the flux-tower (fluxes and meteo data) but also dedicated to the forested area of interest, considering laser scanning measurements from terrestrial and drones systems but also moisture probes for soil and trees. These measurements are particularly interesting for a better parametrization of the aforementioned EM simulations (using herein the MIPERS-4D model) and comprehensive cross-comparisons with the C, L and P bands time series, the latter range of frequency being particularly relevant to analyze the various vertical layers extracted from tomography. Although limited to a specific test case of tropical dense forest, we show that the cross-analysis between multi-frequency and environmental measurements enable to support physical relations that can also serve the design of retrieval algorithms or acquisition scenarios for future spaceborne missions.

Area A - Board 63

Remote sensing is often used to derive forest attributes. Especially radar data is sensitive to the structure of forests. With TanDEM-X data a high-resolution data-pool covering the entire Earth can be used for research based on X-band images. The bistatic single-pass acquisitions of TanDEM-X provide an invaluable basis for forest mapping and monitoring. The radar waves penetrate to a certain extent into the forest. This penetration potential causes more interaction with forest stands compared to optical imagery and is influenced by forest density and its structure. In the present study, the potential of single-polarisation bistatic TanDEM-X images in stripmap mode was evaluated for the prediction of forest stand parameters. These data formed the base for the generation of the global TanDEM-X DEM and hence, are available in comparable configuration for most parts of the land surface. Specifically, the aim was to predict forest attributes indicating structure and density, i.e. basal area, 90th percentile tree height and wood volume over deciduous and coniferous forests because they describe the key variables that are relevant in forest management. Linear regression models were tested on TanDEM-X coherence and backscatter data as independent variables. Available field plot data were used to train a leave-one-out cross validation model. In order to better understand the influence of forest phenology and sensor geometry (baseline, incidence angle) on the SAR signal, acquisitions of different seasons and from different orbits were considered.

The results show that forest attribute predictions from TanDEM-X bistatic data are a function of forest type, volume decorrelation and imaging geometry. ANOVA and Wilcoxon tests were performed and confirmed the significant impact of forest type and viewing geometry on the X-band SAR signal and hence, the model output. In general, leaf-off conditions resulted in lower RMSEs of the regression models, both, for coherence and backscatter data. Best prediction results were obtained by using TanDEM-X coherence information. 17 out of 20 available scenes showed significant correlations (p < 0.05). The tree height had a relative RMSE between 6 % and 22 %, basal area between 11 % and 25 % and volume between 22 % and 34 %. Significant results for basal area and volume could, however, only be obtained over deciduous forests. The coefficient of determination, R², for these two attributes ranged between 0.16 to 0.43 for basal area and 0.27 to 0.67 for volume. Tree height had higher R² values ranging from 0.28 to 0.73. In comparison to coherence, backscatter showed limited prediction abilities. Significant correlations for backscatter were only reached in five out of 20 scenes. In general, our study underlined the potential of single-polarisation bistatic X-band data from the TanDEM-X mission for the prediction of key forest variables. The accuracy of the achieved results are comparable to studies using dual- or full-polarised X-band or C-band data. However, the detected relations are only valid and transferable in a constant environment with regard to forest type and phenology as well as viewing geometry.

Area A - Board 64

Land cover maps constitute a major information source for numerous spatial analyses of ecological and socioeconomic relevance. Typically, classified maps depict the spatial arrangement of classes within a landscape without considering variability within each class, although structure and variation within the classes may well influence landscape’s potential and functions. Specific functions of a secondary forest, for example, depend very much its vertical and horizontal structure. As a consequence, to better understand the multiple functions of “forest classes”, their within class structure should also be analyzed. Nevertheless, comprehensively assessing “forest structure” requires either time intensive fieldwork or cost intensive LiDAR campaigns, or both. By the term of “forest structure”, we are referring here to the vertical distribution of vegetation layers. Important forest structure indicators such as height and complexity allow for conclusions about former disturbances and biodiversity as well as carbon sink capabilities. In this study, we investigate the potential of combining different sources of information, optical and SAR imagery with on-site measurements, to relate forest structure indicators to canopy surface roughness, as one strongly dependent but remotely observable structural indicator. The study is conducted in Harapan Rainforest in Jambi Province, Sumatra, Indonesia, which is one of the last areas dominated by a diverse lowland rainforest and habitat for many endemic species. Indicator variables for forest structure were recorded on 70 field plots on transects following defined UAV flight paths where multispectral images were taken. We hypothesize correlations between indices from three-dimensional photogrammetric point clouds based on UAV captured images, interferometric coherence from TanDEM-X, and field observed indicators of forest structure. The goal is to distinguish different structural forest types by means of characterizing their canopy surface roughness from the mentioned remote sensing data sources. Different studies have shown the potential of photogrammetric point clouds derived from UAV captured images to analyze canopy structure with high resolution. First results show that each source depicts spatial variability in forest structure independently. By combining such information, we expect to produce more detailed information on forest structure. Further, detailed forest structure maps in combination with Sentinel-1 disturbance mapping should provide a cost-efficient and large scale applicable approach to forest structure change monitoring.

Area A - Board 65

We have recently demonstrated [1] that a notable difference can be observed between VV and HH responses coming from dielectric dihedral-type structures, commonly present in urban areas. Actually, at the first order, we stipulated that any dihedral response can be modeled as two successive reflections: one occurring over the ground and the second one over the façade. For each specular reflection over an interface between two dielectric media, a Brewster (or a Pseudo Brewster) angle exists. This angle indicates the location of a null (resp. a minimum) of the Fresnel reflection coefficient at VV polarization when dealing with a lossless media (resp. a lossy media). The Fresnel reflection coefficient at HH polarization remains monotonic with respect to the incidence angle. When both of these specular reflections are combined, we observe a strong difference between VV and HH and from about 20° to 70° for the incidence angle, as illustrated in Fig.1. We named this phenomenon the Double Brewster angle Effect (DBE).
In this article, we propose to investigate now the forested areas in order to determinate if this DBE effect is also observable in this environment. Our rationale is that this effect should be seen provided that the contribution of trunks-ground response is strong enough.

To do so, we reviewed several research articles and scientific reports presenting radar campaigns over forests. We thus studied different type of forests (boreal, tropical,continental), at different frequencies, from P-band to C-band (See Figs 2 and 3).We found that a systematic difference is observed between the backscattering coefficients at HH and VV, so that HH ≥VV with a difference ranging from around 6-7 dB at P-band to 1-2dB at C-band. Indeed with increasing frequencies, the penetration through the canopy is lowered so that the trunk-ground contribution is less visible. In addition, the descriptive parameters of the ground and the trunk play an important role: a tilt, a low permittivity or a strong roughness may affect the intensity of the double-bounce signature. In this case, the signature of the forest tends to VV=HH (See Table 1).
DBE is then clearly visible on forested areas and it provides now a new light on the understanding of the signature of these environnements.

[1] The double Brewster angle effect, Laetitia Thirion-Lefevre and Régis Guinvarc’h, Comptes-Rendus Physique de l’Académie des Sciences, Volume 19, Issues 1-2, pp 43-53, 2018.

Area A - Board 66

The potential of X-band interferometric synthetic aperture radar (InSAR) heights (e.g. from the TanDEM-X mission) for vegetation canopy height estimation has long been recognized, e.g. for the estimation of aboveground biomass. This is based on the assumption that short wavelengths penetrate only marginally into the forest canopy and thus are an approximation of the canopy surface. However, substantial penetration was frequently observed, which affects this approximation. In addition, penetration depth differs between acquisitions of different geometry and in relation to the structure of the forest stands. The penetration of the X-band wave results in a height bias of the InSAR height compared to a true surface height. Consequently, an assessment and a correction of the penetration depth is necessary when comparing InSAR height models with other height models (e.g. LiDAR) or in multi-temporal analysis of InSAR height models.
We applied a semi-empirical penetration depth model on different TanDEM-X data in order to assess their penetration resulting in the aforementioned height bias. The penetration depth model is based on the fact that the penetration is related to the phase normalized coherence. The height bias was retrieved from the coherence magnitude due to the unique relationship of coherence magnitude and phase under certain conditions. For instance, the penetrated volume must be infinite, where substantially higher volumes compared to the penetration depth can be considered infinite. This was true in the majority of the analyzed forests. Therefore, we extracted the volume coherence from different TanDEM-X data and modeled the height bias. The outcome of the model was used to compensate for the penetration in InSAR heights resulting in comparable quantities (i.e. canopy heights) with a LiDAR model considered as true surface. These results can be also used to understand and quantify InSAR height differences due to different penetration depths. We evaluated the estimation of the height bias by compensating it in highly accurate TanDEM-X DEMs and compared the original InSAR height and the height bias compensated InSAR height with a LiDAR digital surface model. In general, the compensation of the height bias based on penetration from the original InSAR height resulted in heights with a lower root mean squared error (RMSE) and mean error (ME) close to 0 compared to the original InSAR heights. The RMSE was 6 m to 13 m before the compensation and decreased to about 5 m to 7 m. The ME was -12 m to -5 m indicating a substantial height bias due to the penetration and improved to about ±1 m in all studied forests. This suggested that the height bias was accurately modeled, which can be of high relevance when InSAR heights are used for canopy height estimation or used in multi-temporal analysis in order to detect deforestation.

Area A - Board 67

Openings in forest canopy cover play a crucial role in the rejuvenation of forest structure and the maintenance of biodiversity. However, if human disturbances affect the habitat of certain species, gaps may have detrimental effects on the ecosystem. Airborne laser scanning (ALS) and digital aerial photogrammetry (DAP) are three-dimensional remote-sensing data sets that both have the potential to characterize canopy gaps. While ALS is the more well-known data source of the two, DAP is less expensive to acquire.
The objective of this study was to examine the accuracies of ALS and DAP in the boreal forest of northern Alberta. While previous authors have conducted similar studies, they have all taken place in dense-canopy tropical and temperate rainforests. The current area of interest (AoI) is a 1 km² expanse of boreal forest situated in northern Alberta, characterized by highly variable vegetation cover, ranging from low vegetation density wet lands to densely forested, drier uplands. Thus, traditional definitions of gap, and approaches to gap detection must be reconsidered and were tested for their applicability in this study. In addition, the study area contains linear features that are the consequences of large-scale oil exploration.
A fixed-height approach and variable-height approach were applied to three canopy height models (CHMs) extracted from the two data sets: CHM_ALS, CHM_DAP, and CHM_Hybrid, which is a combination of both DAP and ALS data. Validation was conducted based on field measurements, supplemented by visual image interpretation.
Overall accuracies for ALS-derived gaps were 90% and 93% for fixed- and variable-height approach respectively, compared to 63% and 82% for the DAP derived gaps, and 64% and 82% for the gaps derived from the hybrid data set. Large errors of omission were produced by both the DAP and Hybrid data sets (15% - 46%). We found that especially small openings (< 200 m²) were incorrectly classified by DAP and Hybrid when using the fixed-height approach, and showed large errors of omission (> 90%). Markedly better results were achieved in these smaller gaps when using the variable-height approach. Accuracy only varied by 3% when using the variable-height approach with ALS data, and was distinctly higher for all gap-size classes. Number and average size of the gaps detected varied clearly between the approaches, with ALS detecting more than twice the number of openings when using the fixed-height approach than the DAP/Hybrid data sets. Average gap size detected by fixed-height approach was less than half the gap size of each model’s corresponding variable-height approach results.
Lower overall accuracies and the omission of small gaps are attributed to the method of data acquisition by DAP, which characterizes the top of canopy and doesn’t penetrate to the forest floor reliably. Thus, this optical technology is more vulnerable to occlusions, shadows and tree sway, optical effects which negatively affect the image matching process and thus the quality of a detailed CHM.
This study demonstrates that ALS is an appropriate means for monitoring human disturbances in the boreal forest, and that DAP data is not yet a viable alternative to ALS data in the context of detecting and mapping gaps in this setting. However, given the improvements that were achieved in this study compared to previous studies, it is possible that with further software development, DAP will soon be a cheaper and more easily accessible means to monitor forest structure dynamics.

Area A - Board 69

In this study, forest changes were investigated in a time series of almost seven seasons of synthetic aperture radar (SAR) images. The forest changes were mainly analyzed with regard to forest height changes and above-ground biomass (AGB), where the latter is one of the key variables necessary to continuously map in order to improve the understanding of the global carbon balance. The potential of using time series of TanDEM-X data for mapping forest height changes, estimation of site index, and biomass growth and loss has been demonstrated in numerous papers [2], [3], [1], [4]–[6]. However, changes that include biomass loss followed by biomass growth during several years before a second acquisition used for comparison, has been a challenge. Also, the effects of varying configurations of the satellite system (e.g. inconsistent height-of-ambiguity, polarization and incidence angles) motivate additional research.
A time series, extending from 2011 to 2017, consisting of 14 TanDEM-X acquisitions in VV-polarization was used. The SAR images were acquired in StripMap mode with a resolution of a few meters. The single-pass setup of TanDEM-X allows bi-static acquisitions, which enabled interferometric processing of each acquisition without temporal decorrelation. The observed variables, mean backscatter, interferometric coherence and interferometric phase height, were investigated with respect to forest growth but also different kinds of forest degradation, registered as thinnings and clear-cuts. Forest change estimates obtained from the semi-empirical two-level model [1], a possible candidate for global upscaling, are also considered.
The acquisitions cover the hemi-boreal test site Remningstorp, located in southern Sweden. The field data consists of circular 40 meter radius plots where every tree has been calipered in 2010, 2014, and 2016. In addition to the diameter measurement of every tree, the species and position of each tree has also been recorded, and the height was measured on a subsample of trees.
The preliminary results indicate that despite a certain variability in each acquisition, the overall trend of forest growth can be estimated fairly well. Distinct changes due to clear-cuts are also easily distinguishable. Smaller changes happening in-between trees, or below the upper canopy, can sometimes be identified but in some cases, there was little change of the response in the SAR data. The reasons for this are still studied. Complementary uncertainty metrics to quantify the changes will be added when the analyses have finished. Although preliminary results are promising, more research is required in order to understand the underlying causes, if this SAR monitoring should be used operationally at a large scale.

References
[1] M. J. Soja, H. J. Persson, and L. M. H. Ulander, “Estimation of forest biomass from two-level model inversion of single-pass InSAR data,” IEEE Trans. Geosci. Remote Sens., vol. 53, no. 9, pp. 5083–5099, 2015.
[2] H. J. Persson and J. E. S. Fransson, “Estimating Site Index From Short-Term TanDEM-X Canopy Height Models,” IEEE J. Sel. Top. Appl. earth Obs. Remote Sens., vol. 9, no. 8, pp. 3598–3606, 2016.
[3] H. J. Persson and J. E. S. Fransson, “Analysis of tree height growth with TanDEM-X data,” in 35th EARSeL Symposium, 2015, no. 1, pp. 1–6.
[4] M. J. Soja, H. J. Persson, and L. M. H. Ulander, “Modeling and Detection of Deforestation and Forest Growth in Multitemporal TanDEM-X Data,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 11, no. 10, pp. 3548–3563, 2018.
[5] H. J. Persson, M. J. Soja, L. M. H. Ulander, and J. E. S. Fransson, “Detection of thinning and clear-cuts using TanDEM-X data,” in Proceedings of IEEE International Geoscience and Remote Sensing Symposium, 2015, pp. 1–4.
[6] M. J. Soja, H. J. Persson, and L. M. H. Ulander, “Detection of forest change and robust estimation of forest height from two-level model inversion of multi-temporal, single-pass InSAR data,” in IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2015, pp. 3886–3889.

Area A - Board 70