The preservation of biodiversity has become a major challenge for sustainable development at national, European (Natura 2000 and Habitats Directive) and international levels (Convention on Biological Diversity, 2011-2020; SDG’s). To address the current conservation needs, there is an increasing demand to operationalise methods to assess the distribution of natural resources and habitats while integrating information on habitat condition; inform conservation planning and support the assessment and management of ecosystem services. Increased access to satellite imagery and new developments in data analyses can support progress towards biodiversity conservation targets by stepping up monitoring processes at various spatial and temporal scales. Free and open data policy is having a dramatic impact on our ability to understand how biodiversity is being affected by anthropogenic pressures, leading to increased opportunities to predict the consequences of changes in drivers at different scales and plan for more efficient mitigation measures. Satellite remote sensing is however no panacea, and little can be achieved without a robust understanding of the socio-ecological system considered. Similarly, access to relevant field-based information is key for satellite imagery to be properly calibrated, analysed and validated. Within this context, RS-enabled EBVs provide enhanced possibilities to monitor natural landscapes under increasing anthropic pressures. In this work we aim to demonstrate how the coupling of satellite remote sensing data with ground observations and adequate modelling can provide innovative applications and methodological developments to improve tangible operational solutions towards a better understanding and management outcomes for natural systems. In this work will share different cases and examples derived from monitoring complex systems in Europe (Mediterranean, Boreal, Alpine) to support innovative descriptors derived from various types of sensors. RS-EBV’s will be presented and discussed in terms of how this measures could become effective proxy’s providing key information like structure, function and composition at frequent revisit times and high to very high spatial resolution.

Area C - Board 272

Biodiversity conservation, increase food demand and climate change have been recognized as the greatest environmental challenges. Hence due to limited land resources and land availability natural ecosystem conversion to arable land continue, at certain degree. Embodied in agricultural settings open grasslands ecosystems host great deal of species and have been recognized as the most threatened ones. For example, grasslands species specialist are currently facing high rate of population decline as consequences of land conversion and infrastructural development. Furthermore, past and current environment conditions interact with species demographic traits - population size, density, structure or individual animal mass, length, body condition and shape population demographic rates. To capture this interactions we aimed to model demographic traits of an endangered small mammal species in agricultural area as a function of landscape structure, at the population level and individual level. We examined the RapidEye and Landsat 8 - derived variables and habitat carrying capacity on species habitat area and surroundings from population and individual animal level. To develop our habitat structure for each consecutive season, from 2013 to 2015, crop maps were created using a supervised random forest classification algorithm and data from the time series of satellite images. Average seasonal NDVI and AWEI from the time series of satellite images were calculated. Based on habitat use, management and type we developed habitat quality index on 20 localities within study area. To mark demographic traits on 20 colonies in central Banat area, region of Vojvodina - northern Serbia, we capture-recapture animals each consecutive year within 0.25 ha plots, measure body size and length and determine age and sex. To estimate population size we mapped habitat area with GPS Garmin eTrex Venture and extrapolate number of capture individuals to total area size on monitored localities. We developed population-level generalized linear models (GLMs) and individual-level conditional mixed-effects model (GLMMs) in R package Ime4 to model demographic traits. For GLMMs we included age and year as additional random effects to control for the variable crossing or nested effects. Top traits model included land system: composition (crop cover, AWEI), carrying capacity (habitat quality*habitat size) and productivity (NDVI). We found that land composition and carrying capacity determine endangered small mammal species population size. On the other hand system productivity determine individual’ body condition when controlling for variation in sampling year and age classes. This suggest that in agricultural landscape human induced conditions (e.g. crop rotation or water management) are the main drivers of population level differences in size and indirectly population structure. Environmental abiotic factors (e.g. dry season precipitation or temperature) affecting system productivity and drive individual level differences in body condition and indirectly influence population level dynamic. The often reported small mammal population decline should not be generalized independently of ongoing changes in weather conditions and decadal climate prediction. Conservation planning and management action should mediate ecosystem stability and include spatial adaption strategies in agricultural landscapes to respond effectively to predicted climate variability.
Key words: crop rotation, NDVI, demographic traits, rodent, AWEI

Area C - Board 273

Agriculture, because of its large proportion of land use and socioeconomic significance, plays a decisive role for the state of the environment in Europe. A Common Agriculture Policy compatible with nature and taking account of ecological conservation was afforded great significance. European Union (EU) agriculture needs to attain higher levels of production of safe and quality food while preserving the natural resources that agricultural productivity depends upon. High Nature Value farmland (HNVf) represents low-intensity farming systems with a significant presence of semi-natural vegetation, involving more traditional and low-intensity systems. Since 2003, the concept of HNVf is used to describe three types of farmland that are rich in biodiversity: Type 1 - Farmland with a high proportion of semi-natural vegetation; Type 2 - Farmland dominated by a mosaic of cultivated land and small-scale features; Type 3 - Farmland supporting rare species or a high proportion of European or World populations. Estimates suggest that over 30% of farmland in the EU may be classified as HNVf, therefore its conservation is one of the main objectives of the EU-Rural Development Policy. This study presents a Remote Sensing-based methodological and operational framework (using Sentinel-2 and Landsat-8 free and open access satellite multispectral data) based on the Google Earth Engine (GEE) platform for ecological monitoring in EU High Nature Value farmlands, using four different and representative case-study areas in Portugal: (1) “Extensively grazed uplands”: ‘Serra da Estrela’ Natural Park (‘Beira Interior’ region); (2) “Alpine meadows and pasture”: Pico Island Natural Park (Pico Island, Archipelago of the Azores); (3) “Cork Oak Savannas – Montados/Dehesas”: ‘Serra de Monfurado’ [Site PTCON0031 of Natura 2000 Nework] (Alentejo region); and (4) “Multifunctional Agricultural Areas”: “Charneca do Infantado” [territory managed by “Companhia das Lezírias”] (Ribatejo region).

Keywords:
Google Earth Engine; Ecological Monitoring; Landsat-8; Sentinel-2; Common Agricultural Policy; High Nature Value farmland.

Area C - Board 274

Sentinel mission has a huge potential to support the implementation of long-term monitoring of ecosystem and agro-ecosystem traits though very high resolution optical and radar observation archives and new acquisition. Sentinel mission is presently the free of charge provider with finest space and highest temporal resolutions. This makes it extremely interesting for land and agricultural related monitoring. The quality of the Sentinel mission is going to paying off all the investments and it has positively refreshed the earth observation market since ever. In the last 20 years, from around 1988-1993, the Common Agricultural Policy (CAP) of the European Union has allowed for establishing set-aside fields (up to 5-15% or the total arable land according to the EU set-aside campaigns) to farmers. Such a measure was frequently accompanied by the compulsory establishment of native, permanent flora in the set-aside fields, along with ad-hoc funding for the farmers. Thus, set-aside consisted in profound modification of the environmental setups in lands dominated by arable fields. Such modifications lead to an increase of biodiversity in these areas, both for plant and animal biodiversity, and likely positively impacted a wealth of other environmental traits, including soil maintenance at the basin level, soil quality, erosion, water delivery to river and lakes, presence of insects helpful for agricultural production, etc. Similarly, it reduces the load of labour in low-productivity fields, that were likely retired, thus reducing the agricultural impact per unit weight of product and affecting many social traits. Thus, set-aside provided a valid opportunity toward an integrated sustainable land management. However, there is presently no public report of the areas under set-aside to monitor its actual benefit from an environmental point of view. Due to the spatial setting of the set-aside patches, the 10-meter NDVI from Sentinel 2 can be implied to map the presence of hedgerows and the adjacent fallow extension in non-irrigated arable areas. Set-aside mapping can be carried out by supervised classification of multiple NDVI maps. Set-aside are very frequently established as long and very narrow strips, thus making them easily distinguishable from the rest of the native and agricultural vegetation. The temporal resolution of Sentinel 2 data can allow for quantification of soil spatial trends and agronomy related aspects to overcome environmental threats that affect the ecosystem. With this contribution, we would like to start a Sentinel based set-aside inventory to promote integration of land cover data and improve the knowledge with consequential reinforcement of the academic, industrial and public land management collaboration in the agro-environmental sector.

Area C - Board 275

Over the last decade, many efforts have been focusing on providing land cover maps at various scales and with different levels of precision. In some cases, it is however more relevant to identify “land use” rather than “land cover” classes. Grasslands represent a vast and diverse land cover class. In the context of biodiversity assessment, all grasslands cannot be given the same biological value. A distinction should be made between natural grasslands and agricultural grasslands with different levels of intensity. With the increasing availability of satellite imagery with high spatial and temporal resolution, it should be possible to differentiate grasslands in terms of land use intensity and thereby provide valuable information for habitat monitoring and preservation.
This study focuses on Belgium, where 10 different types of grassland of high biological value have been recorded in the frame of the European network of natural and semi-natural sites (NATURA 2000) in addition to several other land use of grasslands associated with a high level of anthropic pressure: gardens, intensive pasture, hayfields and leisure (golf, soccer….). The aim of this study is to discriminate different land uses of grassland based on time series from Sentinel-1 and Sentinel-2 images for the year 2018. Some biotopes differ from floristic differences that are not observable by satellite remote sensing while other are accurately detected with optical images or SAR images based on specific spectral signature or water content. For the less obvious cases, we then focus on the land use intensity gradient based on the frequency and the intensity of mowing/grazing derived from the analysis of Sentinel-1 time series. This objective will be achieved by combining the information from the backscattering coefficient of gamma0 GRD images with the 6-day coherence derived from successive overpass of the satellite. The impact of the speckle was reduced thanks to the use of parcel-based smoothing. Validation results are not yet available, but preliminary analyses are convincing.

Area C - Board 276

Plant primary productivity is an essential variable both for biodiversity (as driver of ecosystem carrying capacity) and climate (both as driver of evapotranspiration and directly involved in carbon fixation). In situ measurements of Gross Primary Productivity (GPP) require large and costly infrastructures to estimate CO2 fluxes from eddy covariance estimates of net ecosystem exchange and respiration. Satellite medium resolution data (e.g. MODIS) are typically used to obtain global estimates of green FAPAR (the main component of GPP) at 300-m resolution. This resolution is problematic especially for biodiversity management at local scales, where it is important to appreciate habitat heterogeneity, as mountain and Mediterranean habitats. Higher spatial resolution sensors as Landsat or Sentinel 2 have lower temporal resolution and produce sparse and irregular time series due to atmospheric perturbations making more problematic analysis of time series.
We set up a fast and robust approach based on a harmonic model linearly fitted, to estimate Ecosystem Functional Attributes sensu Alcaraz et al. 2006, that are 3 features (mean, coefficient of variation, peak’s day of the year) of the GPP proxy seasonal dynamics plus the yearly anomalies that are relevant to compare resilience across the area of study.
Our harmonic model takes inspiration from HANTS procedures (for a review see Zhou et al. 2015) but add a year intercept to capture overall year variation and use the linear model framework (p-value, percentage variance explained) given that we use atmospherically corrected data were errors are normally distributed.
We used atmospherically corrected Landsat time series at 30 m resolution. We tested the approach over a large cloudy mountain region – the Peneda-Gerês National Park (Portugal) – data with simulated cloud cover, and a highly fire disturbed small forest – Bosco Difesa Grande (Apulia, Italy).
Within Peneda-Gerês, over 12 vegetation indices were tested and MSAVI followed by SAVI and EVI, showed the best fit over the data. With MSAVI-based model, 98% of pixels showed a significant model fit (p = 0.01) and 90% of them an explained variance higher that 58%.
The fitted data produce Ecosystem Functional Attributes that follow better the ecological expectation (i.e. higher mean value in pixel with aspect to south and higher inter-annual variation in pixel with aspect to north west) than original time series. Further using the LCCS land cover obtained with the service EODESM within the activities of Ecopotential H2020 project, it was possible to check that the most stable, (i.e. lower coefficient of variation across yearly anomalies), were the tree classes in opposition to forbs and grassland, as expected from the study of Keersmaecker et al. (2014).
We built an artificially cloudy data using the 400 cloud free pixel histories found in the Peneda-Gerês time series and used 100 cloud patterns found elsewhere in the scene. We compared how much the estimates of Ecosystem Functional Attributes are influenced by cloud cover using model fitted or original time series. The original time series produce biased estimates respected to equivalent statistics on unclouded time series, while fitted data produce unbiased estimate of similar variance.
To evaluate how our approach handle repeated critical pressure that modify the stability of the time series we applied our analysis to Bosco Difesa Grande, where we showed that the approach reach to track with the year anomalies parameters the repeated cycle of fire events (recorded in Italian official registry) and recovery, while other methods, as BFAST (DeVries et al. 2015), that require data interpolation to deal with irregular time series and are more parameter rich fail to detect multiple changes within 2-3 years over a same pixel history, as already noted by Forkel et al. (2013).
The model was written with open source libraries in python and is able to analyze 6 millions of pixel histories in about 2 hours using 8 threads on Xeon E7.
References:
Alcaraz D, Paruelo J, Cabello J (2006) Identification of current ecosystem functional types in the Iberian Peninsula. Glob Ecol Biogeogr 15:200–212 . doi: 10.1111/j.1466-822X.2006.00215.x
J. Zhou, L. Jia, M. Menenti Reconstruction of global MODIS NDVI time series: Performance of Harmonic ANalysis of Time Series (HANTS). 2015. Remote Sensing of Environment
163: 217-228
Keersmaecker W De, Lhermitte S, Honnay O, et al. How to measure ecosystem stability? An evaluation of the reliability of stability metrics based on remote sensing time series across the major global ecosystems. Glob. Chang. Biol. 2014;20(7):2149–2161. Available at: http://doi.wiley.com/10.1111/gcb.12495 [Accessed March 7, 2018].
DeVries B, Verbesselt J, Kooistra L, Herold M (2015) Robust monitoring of small-scale forest disturbances in a tropical montane forest using Landsat time series. Remote Sens Environ 161:107–121 . doi: 10.1016/j.rse.2015.02.012
Forkel M, Carvalhais N, Verbesselt J, et al (2013) Trend Change detection in NDVI time series: Effects of inter-annual variability and methodology. Remote Sens 5:2113–2144 . doi: 10.3390/rs5052113

Area C - Board 277

A new era in the retrieval of plant traits has started by emerging the new satellites such as the Copernicus Sentinel families. Among these traits, leaf area index (LAI) is a key indicator of vegetation growth and an essential variable in biodiversity studies. Numerous literature has shown that radiative transfer approach has been a successful method to retrieve LAI from remote sensing data. However, suitability and adaptability of this approach largely depend on the type of remote sensing data, and the ecosystem studied. In this regard, the retrieval of leaf area index in a saltmarsh ecosystem is examined in this study using Sentinel-2 and RapidEye data through inversion of PROSAILH radiative transfer model. Field measurements of LAI and a number of other plant traits were obtained during two succeeding field campaigns in July 2015 and 2016 in the saltmarsh of Schiermonnikoog, the Netherlands. Sentinel-2 (2016) and RapidEye (2015) data were acquired concurrent to the time of field campaigns. The broadly employed PROSAILH model was inverted using a look-up table (LUT) which contained 500, 000 records. Different scenarios of band combinations, as well as different solutions, were considered to obtain the LAI estimates. The R2 and RMSE between measured and estimated LAI were used then to evaluate the retrieval accuracy. The removal of dead materials from the measured LAI improved the estimation accuracies. Our results showed that generally the LAI retrieved using the Sentinel-2 data had higher accuracy compared to RapidEye data. In particular, the SWIR bands of Sentinel were modeled best using the PROSAILH. Leaf area index was best retrieved using the NIR and SWIR bands of Sentinel-2 (R2=0.56, RMSE=1.7). Our results highlight the importance of proper parametrization of radiative transfer models and capacity of Sentinel-2 data, with impending high-quality global observation aptitude, for retrieval of plant traits at a global scale.

Area C - Board 278

Plant functional diversity (FD) is an important component of biodiversity, which characterises the variability of functional traits within a community. It can influence ecosystem processes and stability, making it highly relevant to understand its spatial and temporal variation. Usually, FD is assessed through highly labor-intensive field measurements, while assessing FD from space may provide an unprecedented way to monitor global FD change in a consistent, time and resource efficient way. Assessing functional trait diversity from space can improve our understanding about the large-scale relationship between functional diversity and environmental factors and provide a complementary approach to field experiments for assessing the biodiversity-ecosystem functioning links across space and time. The potential of operational satellites for inferring FD, however, remains to be demonstrated. Here we aim to develop the statistical method for upscaling functional diversity to a regional scale using ESA’s (European Space Agency) high-resolution Sentinel-1 and Sentinel-2 measurements (hereafter S1 and S2). The S2 spaceborne spectroradiometer provides global coverage of up to 10 m spatial resolution over 13 spectral bands, suitable for resolving fine details of variations in vegetation dynamics, and the C-band SAR (synthetic aperture radar) instrument onboard S1 satellites provides information about landscape topography and canopy structure that is complementary to the S2 optical instrument. We plan to compute community weighted means and diversity of functional traits based on a combination of in-situ measured and compiled trait and species abundance data over 209 forest plots across Europe. For these plots, we will extract S1 radar backscatter and S2 spectral reflectance measurements. We will explore the link between functional traits/diversity and S1/S2 signals with the aim of developing a robust statistical algorithm for upscaling functional diversity to a regional scale. The expected technical challenges include 1) to what extent we could resolve variations in canopy structure and physiological traits using spaceborne measurements with limited spatial and spectral resolutions; 2) how to statistically identify the correlations among different predictors and their non-linear relationships with response variables. The expected outcomes include functional diversity derived from field plot data across a wide range of climate conditions representing major European forest types and improved understanding of the link between spaceborne optical and radar signals with functional traits/diversity over different forest structures.

Area C - Board 279

Multiple researchers and institutions, including the intergovernmental panel for biodiversity and ecosystem services (IPBES), have acknowledged the global threat posed by biological invasions worldwide, and have highlighted the urgent need of up-to-date, useful and representative approaches to monitor invasive alien species occurrences and impact on the ecosystems. The increasing availability of remote sensing products in combination with free and open source software provide precisely a solution to tackle this challenge.

On one hand, remote sensing is gaining importance as an effective tool to monitor biodiversity in space and time. And on the other hand, free and open source software brings the transparent, ready-to-use and affordable setting required to ensure the involvement of researchers, managers and policy makers in the development of biological invasions monitoring and management schemes.

Recently, several scientists from multiple European institutions funded by the ERA-Net BiodivERsA have created a consortium (DIARS) to work on the detection, monitoring and impact assessment of plant invasions through the combined use of remote sensing and free open source software. Using hyperspectral and LiDAR data, they have developed easy-to-implement and open source methodologies that allow monitoring plant invasions from inconspicuous invasive alien mosses across sand dunes to invasive alien trees in temperate forests. This set of open source methods is now available in the DIARS toolbox, a web-based set of guidelines accompanied with real case datasets perfect for capacity building on plant invasions monitoring and impact assessment of communities worldwide.

This toolbox features clear guidelines to process and analyze ground and remote sensing data to: (1) map the actual distribution of invasive alien plants; (2) model the potential distribution of invasive alien plants; and (3) assess the impact of invasive alien plants on ecosystem functioning. The toolbox uses Free and Open Source software (GRASS GIS and R) and provides a dataset specially designed to allow computation even at small processing power as well as the “iSDM” R package to help inform the sampling of invasive alien species.

The toolbox will help unified biological invasions monitoring schemes while hosting an online community of researchers, managers and policy makers that will improve our capacity of early detection and management of invasive alien species worldwide. The DIARS toolbox can be found at http://diarsproject.github.io/DIARS/HomeDIARS.html

Area C - Board 280

Forest disturbances (i.e. deforestation and degradation) are a serious problem significantly contributing to greenhouse emissions and biodiversity loss [1]. Quantifying its impact on the carbon budget is challenging because of the diversity of definitions, varying scale of the changes, and many drivers that are applying pressure on the forests. As a result, measuring and mapping of forest degradation is still a technical challenge [1-3]. Radar (SAR) data is very promising for detecting and monitoring forest degradation considering its sensitivity to above ground biomass and forest structure [4]. There is also evidence that the integration of sensors in combination with field data can provide more precise results when monitoring changes in forest [5].

Most definitions from the literature agree that forest degradation is related to the detriment of forest structure. We investigated whether it was possible to assess significant differences in forest structure and link them to different levels of disturbance. And if these differences could be analysed over the time. There are two main objectives of this research. First, we investigated the potential of integrating multi -frequency Synthetic Aperture Radar (SAR) data: ALOS PalSAR-2, Sentinel-1B and TanDEM-X, in combination with field data to identify and classify variances in structure associated to forest disturbance. We assessed the capabilities of ALOS-2 and Sentinel-1 to identify multi-temporal changes in forest structure associated to forest degradation. The study area is located in Bajo Calima a secondary forest in the Pacific coast of Colombia. A wet tropical forest that has been under pressure from gold mining and selective logging at different intensities.

Our methods included: (i) the collection of forest structure information (e.g. tree height, diameter and species density), which was after classified by statistical differences using K-means classification. Second, a support vector machine (SVM) algorithm to retrieve forest structure information using the multi-frequency SAR data: ALOS PalSAR 2 (L-Band), Sentinel-A and B (C-band), and TanDEM–X (X-band). From the SAR data, we considered variables as backscatter, Polarimetric decompositions and parameters, and interferometric coherence. Third, multi-temporal analysis of Sentinel-1 and Alos-2. Results showed a good approximation to categorise these forests by degrees of disturbance and also indicated that multi-temporal SAR observations in the bands L and C are promising to identify variations in forest structure due to forest degradation.

REFERENCES

[1] P. Smith, M. Bustamante, H. Ahammad, H. Clark, H. Dong, E. Elsiddig, et al., "Climate Change 2014: Mitigation of Climate Change. Contribution of Work-ing Group III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change ed O Edenhofer et al," Forestry and Other Land Use, 2014.
[2] FAO. (2012b, 11 November 2012). Forestry and climate change: Roles of forests in climate change. Available: http://www.fao.org/forestry/ climatechange/53459/en/
[3] I. D. Thompson, M. R. Guariguata, K. Okabe, C. Bahamondez, R. Nasi, V. Heymell, et al., "An Operational Framework for Defining and Monitoring Forest Degradation," Ecology and Society, vol. 18, p. 20, 2013.
[4] A. L. Mitchell, A. Rosenqvist, and B. Mora, "Current remote sensing approaches to monitoring forest degradation in support of countries measurement, reporting and verification (MRV) systems for REDD+," Carbon Balance and Management, vol. 12, p. 9, April 17 2017.
[5] V. De Sy, M. Herold, F. Achard, G. P. Asner, A. Held, J. Kellndorfer, et al., "Synergies of multiple remote sensing data sources for REDD+ monitoring," Current Opinion in Environmental Sustainability, 2012.
[6] K. Krishna and M. N. Murty, "Genetic K-means algorithm," IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 29, pp. 433-439, 1999.
[7] O. Antropov, Y. Rauste, T. Häme, and J. Praks, "Polarimetric ALOS PALSAR time series in mapping biomass of Boreal Forests," Remote Sensing, vol. 9, p. 999, 2017.
[8] J. García-Gutiérrez, D. Mateos-García, M. Garcia, and J. C. Riquelme-Santos, "An evolutionary-weighted majority voting and support vector machines applied to contextual classification of LiDAR and imagery data fusion," Neurocomputing, vol. 163, pp. 17-24, 2015.

Area C - Board 281

Forests globally are playing a fundamental role in global biogeochemical cycling, ecosystem services, harboring biodiversity and for societal well-being. However, forests are vanishing in many parts of the world and often at the expense of remaining biodiversity-rich frontiers. Intact Forest Landscapes (IFL) are the largest terrestrial ecosystems, which remain for decades without the signs of human activity – they are important global biodiversity hotspots. The IFL boundaries were mapped for the year 2000 using strict criteria that exclude all remotely detectable signs of human alteration and limit the minimal IFL patch size to 500 km². The global IFL map was updated for the years 2013 and 2016, revealing a dramatic decline of the wilderness area. Overall, 9.3% of the IFL area was lost due to conversion and fragmentation from the year 2000 to 2016, and the annual IFL reduction area increased by 17% during the last three years. However, it was unclear how such IFL losses affected IFL patches of highest conservation priority globally. It was also unclear, what were the major causes of IFL losses from 2000 to 2013 and recently from 2013 to 2016. While the incentives about conservation of IFLs are still in discussion and measures couldn’t be assumed simultaneously for all valuable forest areas, we identified those IFLs that require urgent actions. All remaining IFLs are important in terms of functions and services they provided, but their conservation values and vulnerability to anthropogenic degradation distributed disproportionally. In order to highlight the IFL areas with the highest conservation priority, we constructed a panel of composite indicators of conservation values (biodiversity, carbon storage, hydrology and local climate regulation, and importance to the indigenous cultures), resilience, and vulnerability for each IFL patch globally. The metrics to construct conservation value indicators were derived from the most up-to-date global and regional datasets. The IFL resilience was estimated as a function of the patch area and configuration, while vulnerability was analyzed using observed patch degradation from the year 2000 to 2016. The IFL patches were ranked by their integral conservation value, resilience, and vulnerability for each biogeographic region separately. The resulting IFL preservation priority map helps the environmental community to focus on the forest wildlands with the highest conservation value. However, it’s even more important to get information on the IFL fragmentation and conversion drivers which should be addressed by a conservation campaign. We implemented a sample analysis to analyze the IFL fragmentation drivers within each selected high priority wilderness patch with the most up-to-date free available satellite imagery (Sentinel-2A, Landsat-8). The proximate causes of IFL change (such as industrial logging, mineral resources extraction, agriculture and tree plantations establishment) were estimated from a random sample of 50 plots 1km² each within areas of the IFL patch that were fragmented or degraded during the last 16 years. In areas where public data exists, IFL fragmentation was attributed to a specific industry or company. We also highlighted resource extraction operations that are certified as being environmentally responsible while causing significant loss of wilderness and related ecosystem conservation values. Primarily from 2000 to 2016 agriculture was the dominant proximate cause for IFL loss in tropical America, industrial logging in Africa, South-East Asia, southern boreal/temperate North America and Eurasia. Human-induced forest fires were the cause of IFL decline in North America and Eurasia, while mineral resources extraction in Australia and northern boreal Eurasia. Our analysis will help local, national, and international environmental organizations and concerned citizens to address identified causes of forest wilderness degradation. We believe that continuation of IFL monitoring is critically important to protect the last and vanishing forest wilderness on our planet. Spatial data and information on methodology and approaches of IFL mapping can be accessed http://www.intactforests.org/.

Area C - Board 282

Precise and up-to-date information on tree species distribution in managed and natural forests is of critical importance to sustainable forest management, biodiversity conservation, and climate change mitigation. At the individual tree level, local-scale studies have obtained reasonable classification accuracy for tree species using remotely-sensed data with high spatial and/or spectral resolution. Due to the empirical approaches adopted in many studies; however, predictive models developed at the local scale are often not generalizable to larger geographic extents. In addition, sensors with high spatial or spectral resolution (e.g., commercial spaceborne sensors, airborne LiDAR and hyperspectral instruments) are cost-prohibitive for large-area mapping endeavors. On the other hand, freely available medium spatial resolution multispectral data, such as those from Landsat or Sentinel-2, allow for systematic and dynamic Earth observation over large geographic extents for a number of applications, ranging from land cover mapping to forest structure characterization. Due to the confounding influences of background signal and forest structure, large-area species classification with medium resolution imagery remains an open problem that requires additional research. In this research, we present a classification framework based on ensemble-tree based techniques for mapping the dominant tree species at the provincial level for British Columbia, Canada. In addition to Landsat reflectance composites, the predictor variables include historical forest change information, forest structural properties, and geographic/topographic variables. By using class probabilities rather than hard labels, a probability threshold is established to control the behavior of the classifier. When spectral, topographic, and forest change information cannot offer a confident prediction, the classification system can be switched to an interpolation mode, where the most dominant species in the regional window will be used to populate the prediction unit. Via this implementation, we have a quantitative means to enforce the association between species distribution and climatic gradients in our model, leading to an improvement upon a pure classification approach.

Area C - Board 283

The diversity and composition of forests is an indicator for biodiversity and an important parameter for sustainable forest management. Currently there are no detailed up-to-date and area-wide mapping products based on Copernicus Earth Observation satellites publicly available. Within the 4D-FORMAT project Sentinel-1 SAR data and field measurements will be integrated for the development of exhaustive forest soil-moisture products. In order to enable a species-specific characterization of soil-moisture conditions, area-wide maps of tree species are derived from Sentinel-2 optical imagery and structural data. We present a workflow to derive comprehensive maps of tree species (Deciduous-, Larch-, Pine-, Spruce/Fir- and Dwarf-Pine stands) at medium to high resolution for selected parts of the Austrian Alps. It is a novel approach of mapping tree species over a large extent in mountainous environments, which has been stated as a major challenge in related work.
Both pixel- and object-based classification methods are explored to investigate their suitability for mapping tree species. The mountainous terrain of the Alps poses several challenges in the classification process, specifically the timing of phenological events, cast shadows by mountains and frequent cloud cover during satellite overpass. Atmospherically and topographically corrected multitemporal Sentinel-2A/B Level 2A imagery is integrated with digital elevation models and structural data (e.g. tree height and crown coverage) derived from airborne laser scanning. We employ a Random-Forest classifier and explore how the use of additional classification outputs such as classification margin and confidence score can be used for iterative reclassification and enhancement of the results. To our knowledge these measures have only been used to quantify map reliability so far. To address issues related to the regional characteristics imposed by the mountain relief and topography, we further investigate if a relationship between low accuracy and shadowed areas or structural features such as crown coverage exists.
Since we aim at developing exhaustive products, we employ several strategies to validate our products. The validation is based on manually digitized reference data by experts. We differentiate five tree species and forest types (Deciduous-, Larch-, Pine-, Spruce/Fir- and Dwarf-Pine stands). First results yield high user- and producer accuracies for four of five classes under investigation (>90%) based on validation with internal, independent data. Confusion with Spruce/Fir-stands is high for all other classes and leads to lower class specific accuracies (<90%). A spatial cross validation resulted in worse performances in certain regions. The common denominator of these regions is their position at the margin of our study site and a poor model performance in detecting Spruce/Fir-stands. Therefore we conclude that area-wide products need to be validated by several means in order to avoid positively biased accuracy metrics by spatial autocorrelation. We show that a simple cross validation on the reference data might lead to an overestimation of accuracy metrics. Our findings underline the importance of choosing and thoroughly documenting an appropriate validation strategy.
Ongoing work is dedicated to the differentiation of deciduous tree species and explicitly incorporating phenological information such as date of green-up and duration of the growing season. Furthermore we plan to apply our workflow to different regions in the Alps in order to assess the transferability of the methodology in regions were training data might be more sparse.

4D-FORMAT (http://4dformat.mountainresearch.at/) is funded by the 13th Austrian Space Applications Program of the Austrian Research Promotion Agency.

Area C - Board 284

Forests provide several ecosystem services such as carbon storage, climate regulation, biodiversity and wood production. The use of wood fuel as a source of renewable energy constitutes an important part of the objectives for 2020-2030 in European energy and environmental policies. Many countries in Europe have an operational forest inventory based on a statistical sampling approach and field measurements. This method provides accurate estimations of forest resources but is not designed for spatially continuous maps and estimations at stand level.
The current trend to meet growing need for spatialized forest inventories is to use spaceborne remote sensing data: The Sentinel-1&2 images every 5 or 6 days (C-band SAR and optic), the SPOT6/7 annual mosaics (very high resolution, optic), together with the ALOS annual mosaics (L-band SAR) constitute a great asset to monitor forests over large territories. For forest monitoring, most studies consist in evaluating this data separately. The first objective of our study is to assess the synergy between these different data sources and the improvement that occurs when combining their measurements for estimating major forest variables generally used in forest monitoring. The second objective is to evaluate the effect of forest management and species composition on the quality of the estimations and deduce what stratification is needed to build accurate models. The perspective is to develop a generic semi-automatic method to be applied over large areas.
The first test site is located in the south of France within the Landes de Gascogne forest. This forest is the largest European pine forest and is characterized by a flat topography and even-aged stands. Ground data was collected in 2016 (GLORI project, INRA) over 84 stands of maritime pine (Pinus Pinaster Ait.) from 3 to 72 years old. Several dendrometric parameters were measured: tree number density, diameter at breast height (DBH), basal area (BA), and height. Aboveground biomass (AGB) was estimated using allometric equation at tree level. The remote sensing data consists of multi-temporal Sentinel-1 data, one mosaic ALOS PalSAR data, 2 dates of Sentinel-2 and 1 date of SPOT in 2016. Haralick texture index from optical and radar data were computed. We extracted 29 features from these images. Several inversion algorithms were tested: best estimates were obtained using Support Vector Regression. Random Forest was also tested but needs larger dataset to be really efficient and Multi-linear Regression provided the worst estimations. The regression can work with all the input features, however, we built an automated selection method to reduce their number and improve the results. The best relative RMSEs obtained on AGB, BA, DBH, density and height were respectively 26%, 25%, 19%, 23% and 12%. The results indicate that multi-temporal Sentinel-1 is the best data source to estimate the forest variables on even-aged stands. Furthermore, combining optical or radar textural indexes with Sentinel-1 and ALOS-PALSAR backscatter data improves the accuracy by 5-10% compared to SAR data only.
The next two test sites are located in the national forest of Orléans, at 110 km south of Paris, and the national forest of Saint-Gobain, 110 km north-east of Paris. These forests are characterized by a generally flat topography and mixed coniferous and broadleaved stands. The dominant species is oak (Quercus petraea, Quercus robur, Quercus pubescens), and the main secondary species are hornbeam, Scots pine, beech, linden and tilia. Ground data was collected (ONF's project) in 2016 for Saint-Gobain over 71 plots and in 2017 for Orléans over 111 plots. All collected samples are oak-dominated forests, where oak basal area represents 30 to 100% of the total basal area. DBH, tree density and BA were measured. We used the same 29 remote sensing features as for the first study on maritime pine, but the data was normalized separately for the two sites before we applied the processing chain that uses feature selection and SVR. Initially we only selected the samples which have oak basal area higher than 75% (93 samples). DBH was estimated with 21% relative RMSE, BA with 26% relative RMSE, and Density with 34% relative RMSE. The estimations of these three forest variables have similar errors than for maritime pine if we compare the same ranges of values for density. We also tested the method using all 182 samples to see the effect of mixed forests and the quality of the estimations is degraded. Therefore, we tested to stratify the samples according to the different species mixtures, results are much better. Finally, the analysis of remote sensing features used in the SVR shows that ALOS L-band backscatter is less used in these mixed forests than for maritime pine stands. In return, SPOT panchromatic textures (very high spatial resolution) and Sentinel-2 spectral index are more important for the model. We have to determine if this is due to the heterogeneity (compared to plantations) or the different species present. Furthermore, the first tests suggested that the temporal information from Sentinel-1 and Sentinel-2 (both signal intensity and spatial texture index) could improve the results for these mixed broadleaved forests.
To conclude, we built a method combining open access SAR and optical data for the estimation of forest variables. The results are very good for coniferous plantations, and the preliminary results obtained on mixed semi-natural forests are very promising. Further work is ongoing to determine what level of species stratification is needed to improve estimates on these more complex forests. Also, we study the site-specific effect on models between these 2 oak test sites on one hand, and two maritime pine sites on another hand.

Area C - Board 285

Biodiversity and habitat loss are often caused by forest conversion to agricultural land use types. Assessing the impact of forest conversion and deforestation on biodiversity and habitat quality hence becomes crucial. The recent Earth Observation data revolution and the resulting availability of frequent, high-resolution satellite data from the Copernicus Programme and its inherent capability to map large areas on a regular basis opens great opportunities to improve biodiversity and habitat monitoring.
Here, we suggest an approach to combine remotely sensed deforestation events derived from Sentinel-2 with the InVEST habitat quality model to assess the spatial dynamics of habitat quality and degradation in North-East Peninsular Malaysia. In addition to the deforestation events, the OpenStreetMap road vector and the GLC2000 land cover product have been used.
The InVEST ecosystem services model for habitat quality applies spatially explicit decay functions, which enable a wider biodiversity impact modelling based on deforestation events whilst considering edge effects of cleared land.
The results indicate that the impact of deforestation on habitat quality in the study area is severe and reducing the potential biodiversity, assuming that based on regional conditions the species richness in natural forest landscapes is higher than in agricultural dominated landscapes. A comparison between natural forests and land cover mosaics with a majority of agricultural land cover types was carried out since agricultural land also provides habitat to many species but is less suitable for high species richness than natural forests.
Overall, the combination of EO-derived deforestation detections with the InVEST habitat quality model is a suitable tool to analyse the impact of forest conversion on habitat quality and potential biodiversity. The aggregation of the habitat quality index to a Landscape Biodiversity Score allows for quick comparison between different levels of administrative boundaries and support decision makers in assessing the regional discrepancies in habitat quality and potential biodiversity. This could lead to an improved use/application of scarce resources in environmental management and the reduction of further biodiversity loss or even an increase in biodiversity. However, further research is recommended into the use of additional data such as biomass maps to enhance the representation of the spatial variability of biodiversity.

Area C - Board 286

Monitoring vegetation phenology helps to detect the changes in ecosystem functions, providing baseline data to track vegetation dynamics related to events such as drought, fire, spring frost, land use changes, climate oscillations, etc. The recent establishment of the USA National Phenology Network (USA-NPN; https://www.usanpn.org/), the Pan European Phenology Project (PEP725; http://www.pep725.eu/), the GLOBE phenology project (https://www.globe.gov/web/phenology-and-climate), suggest the need for a greater understanding of biological responses to a changing environment at different geographical scales (Lim et al., 2018). Remote sensing observations of seasonal greenness dynamics take advantage of the potentialities of high temporal resolution satellites (like Sentinel-2) and represent a valuable tool for studying vegetation phenology at scales consistent with ecosystem-level processes and regional climate information.
Remotely-sensed phenology is the study of the timing of vegetation seasonal pattern of growth, senescence and dormancy, in a spatially aggregated form (e.g. pixel size of metres to km). Observed remotely-sensed phenology patterns are the response of heterogeneous land surface conditions, integrating multiple species, age classes and canopy layers within the ecosystem .
Satellite-based green indices, such as the Normalized Difference Vegetation Index, NDVI and the Enhanced Vegetation Index, EVI, represent effective proxies of vegetation photosynthetic performance by exploiting the interaction of visible light with leaf pigments, and of near-infrared (NIR) energy with internal leaf and canopy structures. NDVI is the most common green index used to study vegetation; however, it tends to saturate over dense canopies, like forested areas, losing sensitivity. EVI was hence proposed as a modified NDVI, having a larger dynamic range and atmospheric and soil background correction. As a consequence, EVI is more responsive than NDVI to detect forest seasonal variations, especially for dense and large canopy background such as broadleaved forests.
The time series analysis of vegetation indices allows for quantifying intra-annual changes in vegetation activity timing and intensity, relating these changes to environmental processes and dynamics, measuring the start, the end and the length of the growing season, along with assessing the plant communities rhythms of growth, greening and senescing.
The phenological timing of plant communities is regulated by the seasonal period strategic for growth and reproduction according to synchronizing (e.g. temperature, rain, frost, drought risks, topography, latitude) and asynchronizing (e.g. resource competition, seed dispersal, pollination) factors. The synchronizing factors tend to homogenize the phenological behaviour of different species, such that plant communities can be identified by a characteristic phenological pattern. Macroecological investigations have shown that similar phenological responses characterize species belonging to similar eco-regions due to their plastic response to some environmental conditions such as temperature, water availability or photoperiod. Remote sensing provides ideal data to base regional vegetation phenology classifications on, since they consistently measure vegetation processes and functions in time and space.
Given the huge amount of remotely sensed data, effective computing strategies are necessary to exploit the phenological information provided by long-term time series and to reduce data redundancy and processing complexity. Several studies used phenological clustering to classify pixels with an identifiable seasonal behaviour to identify NDVI-based pheno-regions with similar vegetation phenology and climate, to provide an early warning system for differentiating between normal and abnormal phenology, etc.
Within this framework, we applied a multivariate clustering approach to a Sentinel-2 EVI time-series (2016–2018) for investigating the phenological variability of forests in Italy and examining the different mechanisms of phenology response to biophysical drivers. The objectives of this study are: (i) classifying the Italian forests into phenological clusters (i.e. pheno-clusters), (ii) identifying the main phenological characteristics and the forest compositions of each pheno-cluster and (iii) exploring the role of climate and physiographic variables in the phenological timing of each cluster.

Area C - Board 287

Earth observations from Synthetic Aperture Radar (SAR) can provide unique information related to forest structure and condition. Furthermore, SAR has many potential applications in forest monitoring systems, particularly where clouds have impeded optical observations. Currently, there is a reliable, freely-available, provision of SAR datasets, such as Sentinel-1, and there are plans to have more observations in the near-future (NISAR, BIOMASS). Given SAR’s enhanced earth observation characteristics, there is broad interest in using SAR datasets for decision support systems, such as deforestation early warning systems. However, the application of SAR is still underutilized in operational settings. What is preventing users from using SAR data in their decision support systems?


Here, we discuss the results of a collaboration between SERVIR, a joint program between NASA and USAID, and SilvaCarbon, the United States’s contribution to the Global Forest Observation Initiative (GFOI), to build capacity in using SAR for forest monitoring. The SERVIR-Global community joined efforts to develop a hands-on guide to start utilizing SAR in forest applications. We worked with world-renowned SAR experts to provide targeted trainings and develop the SAR Handbook: Comprehensive Methodologies for Forest Monitoring and Biomass Estimation. This handbook consists of approachable theoretical background and applied content that includes step-by-step trainings for using SAR to monitor deforestation and degradation, estimate forest biomass and forest stand height, monitor mangroves, and design a suitable sampling strategy. Additional complementary materials were also developed, such as quick-reference one-pagers, which break down the most important and commonly used SAR principles into easy to understand graphics, and short animated videos, which use narrative, illustration, and humor to explain complex SAR concepts. We present these groundbreaking results, that will lower the barrier to start using SAR technology for forest applications. The materials generated from this collaboration can be found on the servirglobal.net website and all software and data used in the trainings are freely available. We will also share lessons learned throughout the process and discuss challenges that have yet to be overcome.

Area C - Board 288

Fires alter ecosystem patterns and processes, impact climate (i.e. atmospheric chemistry and biogeochemical cycles), plus affect human health (Bowman et al., 2009). The worldwide phenomenon has a very dynamic role, causing both negative and/or beneficial transformations of biodiversity. These variations are shaped by the complex relationship between fire, climate, and humans. Only few of the remote sensing sensors cover long enough time spans to analyse these changes over time. Knowledge of long-term fire activity and patterns can, however, support our understanding of fire-prone ecosystem and assess their current and future adaptation to the projected increase in fire risk, as a result of climate change (IPCC, 2013). Here, the Advanced Very High Resolution Radiometer (AVHRR) heritage sensor comes into play. We used the unique potential of AVHRR for climatological studies to retrieve a 30 year active fire record for Europe.

Full-resolution imagery on-board morning and afternoon NOAA-9 through 19 (except NOAA-15), MetOp-A and B platforms were employed, which have been received, archived, and homogenized at the University of Bern, Switzerland (46.93 ̊N / 7.41 ̊E). We expanded on the original two-channel contextual approach of Giglio et al. (1999) to calculate our AVH18 active fire product. A rigorous comparison was performed using coinciding MODIS derived active fire maps (MOD14/MYD14) and calculated binary active fire maps from Landsat 5 TM data. To achieve a high-quality time series of active fire, we will discuss the needs for that as well as highlight the caveats (e.g. low saturation temperature of the MIR channel and missing on-board calibration of the reflective solar channels). The major limiting factor for long-term fire records are however, differences in diurnal sampling over time due to orbital drift of the satellites (i.e. fire observations observed at different local times) resulting in spurious trends (Giglio, 2007). We investigated the impacts of drifting on the active fire time series for Europe and corrected for the diurnal sampling bias by employing data from the European Forest Fire Information System (EFFIS; San-Miguel-Ayanz et al., 2012). Results of the long-term active fire are shown for different European ecosystems and are discussed in relation to the mentioned caveats of the AVHRR/2 and AVHRR/3 sensor characteristics, observing, and environmental conditions dating back to 1985.

Bowman, D. M. J. S., et al. (2009), Fire in the Earth System, Science, 324(5926), 481-484.
Giglio, L., et al. (1999). Evaluation of global fire detection algorithms using simulated AVHRR infrared data. International Journal of Remote Sensing, 20, 1947–1985.
Giglio, L. (2007). Characterization of the tropical diurnal fire cycle using VIRS and MODIS observations. Remote Sensing of Environment, 108, 407–421.
IPCC (2013) Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change, (ed. by T.F. Stocker), G.-K. Plattner), M.M.B. Tignor), S.K. Allen), J. Boschung), A. Nauels), Y. Xia), V. Bex), and P.M. Midgley) Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA.
San-Miguel-Ayanz, J., et al. (2012). Comprehensive monitoring of wild- fires in Europe: The European forest fire information system (EFFIS). In J. Tiefenbacher (Ed.), Approaches to managing disaster - Assessing hazards, emergencies and disaster impacts (pp. 87–108). Rijeka, HR.

Area C - Board 290

Wildfires is an important factor influencing the vegetation dynamics (spatial and temporal) because it has different responses to fire disturbance and different degrees of regeneration in post-fire situations regarding the different types of species which can influence the vegetation land cover. This research presents an evaluation methodology of post-fire dynamics of the vegetation between the years of 2016 and 2017 in the Centre region of Portugal the area which had the highest percentage of burned areas in 2017 in the months of June and October. The modelling approach was based on Biophysical algorithms (FAPAR, LAI, CAB, CW and FVC), vegetation radiometric indices (NDVI, SAVI and WDVI) and environmental data which have been processed on the software SNAP, available on Copernicus RUS, through Sentinel 2 images acquired for the months of January, February, March and April of 2016 and 2017 in order to obtain the vegetation dynamics post-fire. The outputs obtained from the vegetation indexes and biophysical outputs allowed to measure which types of species growth faster in different areas and their capacity to respond to post-fire events. In other hand, it allowed to identify the probability of invasive species response capabilities by identifying their radiometric value and mapping them. Thus, the findings of this study allowed to measure and identify that areas where Pinus halepensis and Pinus pinaster have a higher occupation it will take approximate 15 years of recovery time under optimum conditions (RTOC) and the Quercus about 20 years and since these species don´t have the capacity to burst from its woody structures or rhizomes it may take longer to recover thus becoming more dependent on the interval between fires. On the other hand, when wildfires are recurrent, these trees don’t have time to reach maturity compromising seed production and, consequently, the colonization of a place after its occurrence allowing invasive species to become predominant.
Besides these findings, this study allowed also to compare different radiometric responses regarding different vegetation indexes and other physical variables, such as slope, aspect and elevation in order to validated it through a ROC curve combined with field observations. These results are important to preventive and reactive services of wildfires and to measure the vegetation capabilities to post-fire in different areas with different land cover under different physical variables.

Area C - Board 291

Mast seeding is the variable and synchronous production of abundant seeds by plants. In forests, mast seeding has important ecological consequences, ranging from forest demographics, to nutrient cycling, trophic cascades, human allergies and disease vector populations. Predicting the occurrence of mast years can be useful to time forest management operations, activate allergy alerts, and improve vegetation and ecosystem models.
Here we show the first application of remotely sensing imagery to detect ongoing mast seeding and predict upcoming mast years in conifer and broadleaf mountain forests in Italy. Landsat and Sentinel-2 imagery was used in conjunction with seed collection data from traps deployed in the field, in order to explore correlation between seed abundance and spectral features. We used image data both from the year of seed production, and from one and two years before. Our hypothesis is that summer climate in the years preceding seed maturation affects resource availability and floral induction, which could be associated to changes in forest structure and physiology directly captured by vegetation indices (e.g. decrease of cover after summer drought), or indirectly associated to spectral and texture image features.
Parametric and non-parametric models were built against 15 years of data in seed production, together with other independent variables such as weather, forest structure, and site characteristics.
The existence of a significant relationship between seed production and image features will allow the development of predictive mast seeding models for the forest cover types analyzed herein.

Area A - Board 55