Greece Case Study

Structured framework for post-construction ecosystem monitoring of wind farms

The Greek case study focuses on the potential conflict between the requirements and ambitious goals set by the EU and Greek policymakers about the development of wind energy on the one hand and the protection of biodiversity on the other. The Case Study aims at improving the monitoring capacity in the broader wind farm areas, targeting mainly birds that are considered the most vulnerable to wind energy development biodiversity element. To achieve this, it assesses the potential of novel machine learning algorithms that provide automatic species identification in audio recordings collected in the field. At the same time, the implementation of CS2 develops a framework for the assessment of changes occurring in ecosystem services related to wind farm development.

Main Objectives

The exploration and assessment of the potential applicability of automated avifauna species identification systems using sound recordings to improve relevant monitoring projects conducted in wind farm areas

The development of a framework for the analysis and assessment of ecosystem services in wind farm areas

Specific Objectives

The utilisation of acoustic avifauna species identification technology in bird monitoring programmes during the operation of wind farms

To investigate the potential of utilising avifauna monitoring data obtained through sound identification technology to comprehend the impact of wind farm operation on the behaviour of wind farm sensitive species populations during different time periods

The evaluation of ecosystem services as they transform and evolve following the development of wind farms, considering the main biotic and abiotic environmental components and other existing land use and anthropogenic activities

To identify methods of using the case study findings in institutions with a vested interest in wind farm development and are directly or indirectly involved in the permitting and implementation of such projects or on the other hand, in the protection and conservation of bird species and their habitats.

Pilot implementation of acoustic monitoring of birds

The case study focuses on the pilot implementation of acoustic monitoring of birds around installed wind farm areas to test the feasibility of implementing the species audio identification algorithms in the context of wind farm monitoring programs. This would provide a valuable supplement to the conventional ornithological monitoring approach based on human observations, which is the current practice in most ongoing windfarm bird monitoring activities.

For the practical implementation of the pilot study, two areas of interest in Greece were identified. The first is in the South Evia region, and the second is in the Thrace region. Both implementation areas include habitats for important avifauna species that are particularly vulnerable to wind farms, such as birds of prey and vultures. Additionally, these areas are either in close proximity to, or part of, protected areas for birds (Natura 2000 Special Protection Areas). At the same time, they are also priority areas for the development of wind farms due to their high wind potential. As a result, they have both intense wind farm development, but also significant prospects for further future expansion of wind energy infrastructures.

In each study area, specific wind farms were selected as the main study areas. Furthermore, respective control areas, sharing similar characteristics, but at a greater distance from installed wind turbines, have been identified.

Special audio recording devices were deployed in appropriate locations around the selected wind farm areas and the control areas. The devices were set to record the morning and afternoon audio landscape. The recorders operated for 10 hours each day for 4 months. This time frame allowed the coverage of the bird’s spring migratory and early summer breeding periods (mid-March to the end of June 2024). Recordings were collected in batches of about 20 days each. In addition, every 20 days, observation transects were executed by expert ornithologists to ground-truth the acoustic data and allow comparisons with traditional observation methods.

Field data collection also included the installation of a meteorological station in each of the two areas of interest, providing constant measurements of weather parameters.

Naturalis Biodiversity Center (Naturalis), based in the Netherlands, is the technical partner in the GUARDEN project that has undertaken the processing of all audio data that are collected during the implementation of the case study. Naturalis has developed a specialized deep-learning bird audio identification algorithm to identify bird vocalizations and produce predictions of the respective bird species along an audio recording.

The analysis of the species identification predictions derived from the abovementioned procedure attempts to provide answers to the following questions:

Can acoustic monitoring of bird populations be a reliable supplementary or standalone method for recording bird populations in wind farm areas?
What parameters affect the results of the method and in what way?
Can acoustic monitoring of bird populations provide information on the potential impact of operating wind farms on bird species populations?

The outcomes of the analysis are expected to increase the capacity of stakeholders related to the development of wind farms and environmental protection to appreciate the significance of data collected with acoustic avian monitoring methods make more biodiversity-aware decisions.

Key Stakeholders

The Multistakeholder Partnership (MSP) that supported the implementation of the Greek case study included representatives of the following:

The stakeholders came together to be informed and discuss the UC implementation and contribute to the development of the UC through their comments and opinions.

Results and Conclusions from the Implementation of the Case Study

The Use case has practically examined the applicability of the acoustic monitoring program implemented in the two areas of interest through analysing the data derived by processing of audio recordings from special microphone devices deployed in the study areas. The analysis included the following:

Analysis of synchronous audio-based human–machine monitoring results

The comparison of the species list compiled by expert ornithologists during observation transects with the results obtained by processing the audio recorded at the same time by the EchoAves algorithm, showed a small rate of species recognition coincidence (35% on average for S. Evia region and 28% respectively for Thrace). In addition to the false-positive detections, the algorithm failed to identify a large percentage of the species recorded by the ornithologist (false negatives). These discrepancies could be partially attributed to the noise produced by the movement of the observer, obscuring the soundscape that the algorithm had to work on. To address this issue is essential that a more careful design for collecting ground truth data is applied in future applications.

Percentage of species detected by only the algorithm, only the expert ornithologist or both in each of the transects conducted in Evia (left) and Thrace (right) regions. The central table shows the parameters of each transect and the total number of species detected by the ornithologist and the algorithm.

Pie charts of the distribution of the species (top) and acoustic detection events (bottom) in Evia (left) and Thrace (right) regions according to the results of the validation exercise: “Only_No”: None of the acoustic samples validated as true, “Only_Yes”: All acoustic samples were validated as true, Both_Yes_No: A portion of the acoustic samples validated as true (“YES”) and the rest as false (“NO”).

Validation of results on a species level by human experts

The evaluation exercise consisted of randomly selecting approximately 30 audio detection samples for each species identified, across each of the confidence score 0.1 range bins that were reported by the algorithm. Expert ornithologists listened to the respective audio clip, labeling the detection event as “Yes” or “No” (True or False, respectively). This procedure was repeated for each area of interest, resulting in the validation of ~ 7,500 detection samples in total. The outcomes of the exercise were used to estimate the general performance of the algorithm for each species, to exclude species with no validated samples from the final dataset, and set a confidence score acceptance threshold for species with both positive and negative validation. The confidence score threshold (0 for “Only_No” species, 1 for “Only_Yes” species and 0 to 1 for “Both_Yes_No” species) was then used to filter the raw dataset to produce a validated base for further analysis.

The results of the validation exercise showed that falsely identified bird species accounted for 68% of the species identified in the S. Evia and 46% respectively in the Thrace region. However the analysis of the individual identifications showed that the algorithm provides correct detections at a rate of 91% – 97%. It is therefore evident that relatively few incorrect species detections significantly impact the reliability of the produced species list.

Taking in account the above the following conclusions are drawn:

The performance of available audio-based bird monitoring systems and their deep learning species identification algorithms is substantially hampered by a high number of false species identifications, a weakness that critically compromises the reliability of such systems for autonomous use in wind farm monitoring and underscores the critical requirement for human expert supervision and validation.
The confirmation of even a small, representative sample of the algorithm’s output by expert ornithologists can provide the basis for the significant improvement of the overall accuracy and trustworthiness of the final acoustic monitoring dataset, allowing for a more reliable understanding of bird species occurrence and facilitating more accurate further analyses concerning the significant environmental parameters affecting avian populations.

Modeling bird activity and species richness

The “validated dataset” derived from thresholding the raw results according to the confidence score thresholds derived from the validation exercise was consequently used to model the “Total Activity” (number of identification results per hour) and “Species Richness” (number of species detected per hour) against the variability of environmental parameters (mean hour temperature and wind speed), a time related variable (hour of the day) and wind turbine proximity metrics (distance to the nearest wind turbine, inverse distance weighted Index).

The statistical modelling approach employed was based on Semiparametric Generalized Additive Models (GAM), which allow for the combination of parametric (linear) and non-parametric (non-linear) relationships between bird detection frequencies and different environmental variables. Four models were constructed for the two areas of interest and each of the two response variables (Total Activity and Species Richness).

The results of modeling consistently highlighted known ecological patterns, both in the S. Evia and Thrace regions such as:

The increase in the activity of avifauna species at average values of the temperature range and the decrease of bird activity and variability in extremely low or high temperatures (bell-shaped effect).
The monotonous reduction of bird activity and/or detectability with increasing wind intensity.
The higher activity of a high number of bird species during the morning hours is followed by a less intense evening vocalization period that is significantly reduced during night hours.

Plots of the parametric coefficient (for the linear term) and the smooth functions (for the non-linear terms) obtained by modeling the Total Activity (TA) and Species Richness (SR) variables against the gradient of environmental and turbine-proximity variables in Evia (EV) and Thrace (TR) regions.

The modelling results, however, did not reveal meaningful correlations between the type and quantity of acoustic recordings of avifauna species and the proximity to the wind turbines. It remains open for future investigation whether the observed absence of correlation is due to the technical design parameters (mainly the number of recorders) of the monitoring program, or indeed, such correlations do not exist.

Parallel achievements of the Greek Use Case

Multistakeholder partnership

The Use Case established a multistakeholder partnership (MSP) that brought together national and regional groups, organisations, and authorities relevant to the Use Case’s subject. Through dedicated presentations and Q&A sessions, the participants were provided with information on the organization and progress of the Use Case implementation and its results. Through interactive exercises and open dialogue sessions, the stakeholders exchanged knowledge and experience and expressed their views and opinions, successfully contributing to the development of the Use Case, shaping its parameters, and evaluating the achievement of its objectives.

Ecosystem services modelling and scenarios for future development

The Athens Harokopio University Team established an analysis framework for mapping the future trajectories of biodiversity and ecosystem services (ES) in each of the Areas of interest by leveraging participatory approaches and advanced modelling techniques, such as Bayesian Belief Networks (BBNs) and deep Species Distribution Models (deep SDMs). The final outputs include ES maps, as well as the trade-offs and synergies emerging across ES in the areas of interest, both at the current state of the social and ecological systems, but also at the future ES trajectories under different management and climate scenarios. The methods and results obtained by this analysis are presented in Deliverables 3.6 and 3.7 as well as through the interactive maps included in the Decision Support Application (DSA) of the GUARDEN project (see below).

Augmented Reality Application Development

The Augmented Reality Application for the Greek Use Case was created by leveraging the technical capabilities of the CirculAR application developed by ICCS/NTUA. The application aimed to showcase the Use Case implementation and provide an immersive and engaging way for the user to navigate the results obtained by the acoustic monitoring program. The demo application can be downloaded and used on IoS and Android mobile devices through App Store and PlayStore, respectively.

Decision Support Application (DSA)

The DSA comprises the main demonstration tool of the Use Case, presenting the results obtained by the monitoring program through interactive web maps, data tables, and graphs, allowing users to filter appropriately and deeply explore the data. The tool includes all results of the ecosystem services mapping for different scenarios and information derived from other tools of GUARDEN applied or supporting the Greek Case Study, such as GeoPl@ntNet mapping of habitats, species occurrences provided by citizen science projects Pl@ntNet and Minka, Landscape mapping in the two study areas etc.

Benefits from the GUARDEN project

The outcome of the case study implementation enhances the ability of windfarm owners, monitoring services providers, and regulatory authorities to understand the potential but also relevant limitations of novel algorithmic approaches on acoustic bird surveillance systems through the practical novel application of this approach in environmental monitoring scenarios of installed windfarms. The Use Case assessed the complementarity of this approach with traditional human-based bird monitoring techniques and assessed the effect of environmental parameters on its results.

At the same time, the analysis of ecosystem services distilled the knowledge and perspectives of a diverse group of stakeholders to provide insights into the trade-offs and synergies between different environmental variables and human activities that may occur in wind farm development areas.

The Use case implemented methods and techniques in the collection of biodiversity data, its processing, validation, and analysis. It also used novel applications for communication to promote collaboration between groups with different and often conflicting opinions and approaches.

In the wider context, the implementation of the Greek Use Case provided support to the decision-making capabilities of relevant stakeholders, contributing to the promotion of sustainable wind energy production.