The development of monitoring systems focusing on the erosion of biodiversity is particularly critical in tropical ecosystems, due to the many factors leading to their degradation, including agriculture, climate change, and exploitation of natural resources. The Essential Biodiversity Variable (EBV) framework defines several classes of variables relevant for the monitoring of biodiversity, including community composition. Potential candidates for this class include taxonomic diversity (alpha diversity) and spatial distribution of species communities (beta diversity).

Remote sensing consistently showed strong potential for alpha and beta diversity mapping. Most studies focusing on biodiversity mapping in complex tropical ecosystems take advantage of airborne imaging spectroscopy, with inherent imitations in terms of spatial cover. The recent launch of the Sentinel-2 satellites opens perspectives for enhanced monitoring at regional scales based on high spatial resolution multispectral images. However, monitoring tropical rainforest with optical remote sensing is challenging for many reasons, including the complexity and diversity of the vegetation, and the conditions of acquisition suffering from strong cloud coverage. Despite the 5–days revisit period of Sentinel-2, some regions may never get cloud-free conditions. Cloud-free composites may be produced, but the possibility to take full advantage of the spectral information and map biodiversity from these composites needs to be assessed.

Here, we focused on two forested regions located in Cameroon and French Guiana. First, we produced a cloud-free composite from individual Sentinel-2 tiles. Then, we transposed a method initially developed and validated with airborne imaging spectroscopy to Sentinel-2 images. This method based on the spectral variation hypothesis allows estimation of both alpha and beta diversity. Finally, we used diversity plot networks from both study areas in order to validate biodiversity indicators. We highlight the importance of proper atmospheric corrections, and the strong potential of Sentinel-2 for tropical biodiversity mapping. More systematic studies need to be performed in order to conclude on the true contribution of such biodiversity products to tropical ecology and operational biodiversity mapping systems.

Lianas are key structural elements of tropical forests having a large impact on the global carbon cycle by reducing tree growth and increasing tree mortality. Despite the reported increasing abundance of lianas across neotropics, very few studies have attempted to quantify the impact of lianas on tree and forest structure. Recent advances in high-resolution terrestrial laser scanning (TLS) systems have enabled us to quantify the forest structure, in an unprecedented detail. However, the uptake of TLS technology to study lianas has not kept up with the same pace as it has for trees. The slower technological adoption of TLS to study lianas is the lack of methods to study these complex growth forms.

We propose a whole processing pipeline to extract lianas from TLS data and quantify their plot-level biomass. We present a method to automatically extract woody liana components from plot-level TLS data of tropical forests. We achieve this by developing a model based on random forest classifier that uses eigen values of all the points at multiple spatial scales as features. The presented liana extraction method has a high precision of 93% and a high accuracy (F1 score = 0.75). We tested the method across multiple forest types (from temperate to highly dense tropics) along an increasing gradient of liana infestation (from plots with no lianas to plots with very high liana density).

To calculate plot-level biomass from the extracted liana points, we present a method based on Quantitative Structure Modelling (QSM). QSM was originally developed and successfully applied to derive the volume estimates for trees. In this paper, we post-processed the liana pointcloud from the previous liana extraction step (manually to correct for misclassifications) and optimized the parameters of QSM to derive the volume estimates for lianas.
The advantages of the proposed semi-automated plot-level liana biomass estimation method is that it enables the monitoring of liana abundance and biomass over time in forest plots. In addition, it paves the way for establishing new allometric equations for liana biomass by facilitating non-destructive biomass measurements of more lianas especially the large ones (> 10 cm). Our method will facilitate new research to study lianas, which are now recognized as important components to many aspects of forest ecology.

Forest ecosystem mapping in mountainous regions with optical satellite data is a challenging task which greatly relies on image spatial and spectral resolution. Due to differences in phenology caused by elevation and topography, or inter-annual climate variability, single date classification approaches often fail to reach high accuracies. Increased temporal resolution of the latest globally available satellite images, such as those provided by Sentinel-2 and Landsat-8, along with newly introduced time-series processing algorithms, create new possibilities for overcoming these obstacles. Moreover, object-based classification approaches have been proven to considerably reduce processing time in large time-series datasets analyses or even over extended geographical areas. The Dynamic Time Warping (DTW) algorithm, initially developed for speech recognition, has been successfully used in recent studies for land cover mapping through time-series, and is also effective with small ground sample sizes and time-series that contain cloudy images. The latest permutated algorithm of the DTW is the time-weighted DTW (TWDTW), which also takes seasonality into consideration.

Previous research until now, has mainly focused on applying DTW classification for crop type or rural regions, while only few studies have used the algorithm in forest areas, mainly focusing on the analysis of phenology curves and forest disturbances. To our knowledge, no study has yet examined the applicability of the DTW to forest species classification.
Within this context, in this study, a TWDTW classification is performed to assess the algorithm’s capability to map forest species and overcome elevation and topography-induced phonological variability.

An object-based TWDTW algorithm is implemented to a set of Sentinel-2 yearly time-series indices, in order to classify forest species in the mountainous region of the Northern Pindos National Park in Greece, including 4 Natura2000 areas. Spatially continuous and homogeneous forest patches were delineated as individual segments, through the segmentation of a high resolution RapidEye image. The indices used for the time series analysis were the Normalized Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI) and the Normalized Difference Water Index (NDWI). The accuracy results in terms of classification accuracy (more than 80%) indicate that the proposed methodology can effectively classify the different forest species at different elevations and aspects.

Open landscapes that are not cultivated for intensive agriculture production are rated as essential hotspots for biodiversity in Europe. One particular landscape type, heathland dominated by Calluna vulgaris, is among the most threatened anthropo-zoogenic landscape formations in Europe. Heath is characterized by small-scale spatial patterns of habitat diversity. These patterns result from intense disturbance regimes induced by human land use, such as fire or pasturing. Effects of such land use are short-term successional processes of regrowth in open landscapes. The evolved spatial heterogeneity is a prerequisite of a high value of biodiversity particularly with regard to invertebrates. This value of biodiversity of heather as well as of other open habitats such as pioneer grasslands, sandy-xeric-grasslands or fresh meadows are protected under the European Natura 2000 network. In particular, on former military training areas, open habitats are still established in large areas and need to be managed in order to maintain a favorable conservation status and conceive the potential of change in habitats under different successional phases. Habitat management (burning, mowing and grazing) is therein implemented for compensating an observed decline in natural process dynamics in frequency and extent.
In order to evaluate the diversity of landscape structures, the spatial extent of habitats and their management conditions, open landscapes can be monitored using Copernicus Sentinel-2 imagery. The large spatial extent and the dense temporal resolution of Sentinel-2 optical imagery allows for a detailed repetitive mapping of habitats and their conservation status that can further be translated into definitions of biodiversity indicators and essential biodiversity variables. However, the repetitive spatial representation of open habitats is complex since species turnover, environmental gradients and habitat quality is varying continuously over the time-space domain. As a consequence a habitat unit is always sampled by references that are spatially segregated and spectrally determined for a distinct terrestrial unit. The spectral-spatial reference samples are the basis for every machine learning algorithm that need to be re-collected and refined for a specific area and time. Thereby the sampling of references spectra itself is a crucial and time-consuming step for an accurate prediction of habitats in remote sensing imagery.
We introduce a novel combined spatial-spectral sampling and classification procedure (Habitat Sampler) that collects reference samples for habitat units and conservation parameters on the basis of two assumptions: a) habitats are determined by representative spectral continua and b) habitats are spatially aggregated (distance decay of similarity). As a consequence, the reference samples for habitat units can be extracted automatically from multispectral image time series using a cascade of machine learning models that describe the spatial structure of habitat composition with the highest iteratively optimized accuracy. An iterative learning forms sampling units incorporating spatial coherent and spectral similar pixels that are evaluated by machine learning classifier. Reference samples are not needed a priori to predict the spatial distribution of habitats. The procedure automatically generates spatial clusters of distinct habitat units as inputs for a high number of different classification models. This model cascade is subsequently assigned to a single species or habitat spectrum a posteriori. Such identification spectra can be derived from existing spectral libraries or the image itself in order to determine an individual characteristic spectral profile per habitat unit. The resulting cascade of models is tested against each spectra and a final model ensemble is selected that predicts the probability of each habit unit per image pixel.
The habitat sampler was used to map 15 open land habitat type probabilities and their management status on 4 former military training areas. We used a Sentinel-2 multispectral time series in an automated process chain framework that integrates image pre-processing and habitat mapping in one workflow. Using the same identification spectra, the habitat sampler outperforms random Forest and Support Vector Machine classification by 15% on average in overall accuracy assessment with external validation samples. In heathland areas, heath age structure i.e. old and young stands as well as management measures i.e. times after mowing and burning could be delineated in imagery. The procedure will further be developed to be implemented in the variable workflow for European biodiversity monitoring initiatives.

Spatial data are most of the time stored and represented as either continuous field or categorical data. While it is not a new concept, spatial regions that provide quantitative information within irregular polygons are still rarely used. This data structure has however a great potential to describe natural habitats, especially in terms of interoperability. In this study, we will show the advantages and inconvenients related to the use of spatial regions for ecosystem models. As a case study, we show how spatial regions are used to map ecotopes in Belgium and how they relate to the Copernicus EAGLE concept.
Ecotopes are the smallest ecologically functional units. They can be mapped by intersecting a large number of thematic layers (soil type, land cover, topographic types). However, this creates a lot of polygons with an exponentially increasing number of categorical values combination. The analysis of those polygons is therefore difficult and many artefact are created because of the different spatial resolution and boundary description of input database. On the other hand, the spatial region concept provides a framework to map irregular polygons with quantitative variables that can be used directly into computer models. Those polygons are derived from geographic object-based image analysis (grouping adjacent pixels of similar properties to create a partition of irregular polygons). The segmentation algorithm uses the spectral information from aerial images as well as height and hillshade information derived from a 1 m resolution LIDAR point cloud. Each polygon is then characterized using a set of continuous fields: quantitative variables (climate, elevation, slope, artificial light, contextual variables…) are either interpolated or averaged at the level of the ecotopes depending on the resolution of the input layers; categorical variables (soil types, …) are provided as proportions inside the ecotopes. The land cover components are derived by supervised machine learning classification of a combination of high resolution orthoimages, LIDAR data, Sentinel-1 data and Sentinel-2 data. Based on several fitness to purpose analysis, we show the interest of the region-based model as well the parameters that can influence the quality of habitat and/or biotope models.

Many migratory species are facing increasing environmental and anthropogenic pressures. Important migration pathways and resting sites are gradually lost due to habitat degradation and transformation leading to reduced survival and poor reproductive success. Observations of dramatic declines in population size across taxa in conjunction with the unique challenges faced by migratory species, such as non-stationary habitat requirements, have prompted the parties of the UN Convention on the Conservation of Migratory Species (CMS) to pledge to identify, monitor, restore and conserve habitats needed for successful migration. Remote sensing approaches are among the most promising tools to pursue such identification and monitoring of spatially and temporally dynamic habitats. Yet, matching the spatial and temporal scales of available remote sensing imagery and animal habitat use remains challenging.

We examined habitat use of one large avian herbivore: the greater white-fronted goose (Anser a. albifrons) which migrates annually from central Europe to its breeding grounds in the Russian arctic and back. During spring migration the animals need to maximize their energy uptake by carefully timing their migration with vegetation growth at suitable stopover foraging sites. In recent years notable changes in their migration routes and timing have been observed, the reasons for which remain unclear. We tracked well over 150 individuals by means of high-resolution GPS transmitters and accelerometers in order to identify suitable stopover sites and characterize their behavior and space-use within.

We aimed to address two main questions: a) Which landscape elements are important habitat factors inside the stopover sites (local scale)? and b) Can suitable stopover sites be predicted based on the knowledge acquired at the local scale (continental scale)?
Habitat factors were derived based on 20 m resolution optical (Sentinel 2 and Landsat 8) and SAR (Sentinel 1) time series. Annual summary statistics of Sentinel 1 time-series were used as proxy variables of land-use intensity and for gap-filling monthly composites of Sentinel 2 and Landsat 8. Vegetation indices from the optical time-series were used as proxies for the biophysical habitat state during the time of stopover visits.

Habitat preferences and resource use of the geese was modeled by means of behaviorally informed step selection functions both on the local scale (a) and on the larger continental scale (b). Results indicate that white-fronted geese generally select for young vegetation on intensively managed agricultural fields, but there are large differences between stopovers and individuals. Building on these results, we identified the composition of habitat factors in typical stopover sites which served as the basis for continental scale modeling of stopover site selection. Further, we aim to reveal how the birds select for suitable stopover sites in relation to available alternatives. Building on these results, spatially and temporally targeted conservation schemes come into reach through the combination of remote sensing based mapping of temporal trajectories of habitat change with the now available models on stopover site selection.