AI will change EA practice – but are we ready for it? A call for discussion based on developments in collecting and processing biodiversity data

ABSTRACT

The opportunities and potential of advanced digitalisation involving the application of Artificial Intelligence (AI) in Environmental Assessment (EA) are often mentioned across international studies. However, it is essential for us in EA research and practice to comprehensively grasp the implications of this transformation and proactively prepare for the imminent changes. In this context and drawing on insights from biological sciences, this letter examines the established use, prospects and risks of these technological advances in the field of species, habitat and biodiversity related data and its analysis. We aim to initiate a thought-provoking dialogue across diverse groups of EA actors regarding the practical implications of AI for EA, highlighting new roles and evolving skills needed to guarantee quality and legal compliance. Central to this discussion is the origination of data, alongside the distribution of responsibilities across actors/stakeholders involved in EA with regard to data collection, sharing and interpretation. Key considerations regard the quality and integrity of AI-supported and systematically collected data and the prevention of potential manipulation. We emphasise the need to re-evaluate education and training programs, adapt practices, and enhance decision-making processes as initial steps toward establishing a focused research agenda.

KEYWORDS:

• Artificial intelligence
• environmental assessment
• biodiversity
• machine learning
• data collection
• species

Introduction

Digitalisation involves the introduction and integration of digital technologies and systems into various sectors of society, including environmental planning and its instruments in environmental assessment (EA). A growing number of sciences and disciplines related to EA, such as urban and conservation planning and management, particularly wildlife ecology and biodiversity monitoring, are already widely influenced by the application of advanced digital technologies, including artificial intelligence (Tabak et al. 2019; Schneider et al. 2021; Jetz et al. 2022; Yap et al. 2022; Xing et al. 2022; Tuia et al. 2022; Salman and Hasar 2023; Wild et al. 2023). We use the term artificial intelligence (AI) as an overarching concept of machines that mimic complex human cognitive functions based on training, learning and related development of skills as outlined further in the following sub-sections. Nested within the concept of AI is the subfield deep learning (DL). Deep learning is itself a component of machine learning that uses artificial neural networks with multiple hidden layers. Deeper networks with more hidden layers can learn more complex representations, allowing neural networks to approximate functions and capture patterns from input data which might rapidly change EA practice and challenge EA practitioners at the same time. In biological science, the application of AI combined with affordable hardware and advancing technologies is claimed to improve predictions of ecosystem dynamics by providing access to affordable long-term, high-resolution and large-scale data while offering great potential for efficient monitoring of global biodiversity (Lahoz-Monfort and Magrath 2021; Perry et al. 2022). Authors such as Tuia et al. (2022), Besson et al. (2022), and others discuss the multiple possibilities of machine learning and deep learning for monitoring wildlife and adapting conservation efforts accordingly.

Loss of biodiversity is considered to be the most significant global environmental threat, along with climate change, so the development of best practices for accounting for biodiversity in EA is particularly relevant (Figueiredo Gallardo et al. 2022). We discuss below how these new ways of collecting data, processing large amounts of data, and analysing data much more quickly are likely to contribute to overcome or at least minimize existing challenges in the field of data availability for both the assessment of environmental impacts on species and biodiversity as well as for monitoring of them in EA – for example, enabling learning loops from monitoring to scoping and providing baseline data for EA (González Del Campo and Gazzola 2020; Fonseca 2022).

Combining the discussion of novel technological advances in biological science, as one of the fields of rapid progress, with the way we deal with the assessment of various environmental issues in EA, we look at the opportunities and concerns associated with the application of novel advanced digital approaches. Looking at the scientific literature on EA, we find that several papers mention the potential of exploring AI (Banhalmi-Zakar et al. 2018; Bice and Fischer 2020; Bond and Dusík 2020; Fonseca 2022). However, these studies remain at an abstract level. Some papers and reports provide overviews of first pilot projects and applications of AI in EA (Fothergill and Murphy 2021; Ravn Bøss et al. 2021; Curmally et al. 2022) or the use of AI in EA research (Scott 2018; Ulibarri et al. 2019; Scott et al. 2020; Hileman et al. 2021). Current discussion among EA practitioners, e.g. in the IAIA working group on Principles for the Use of Artificial Intelligence, reflects applied concerns about the use of AI. It seems to be a strong need to explore the practical implications of AI in a specific EA context, as empirical research analysing EA practice shows the low awareness of technological advances and the controversial perception of these applications in some parts of the EA community (Uhlhorn et al. 2023).

In the following sections, we, as an interdisciplinary team, discuss the implications of AI for EA, drawing a link between advances in biological sciences and their relevance for EA. We focus our exploration on a specific thematic area, emphasising species and biodiversity in the context of digitalisation as they are environmental issues which deserve more attention given the complex developments (climate change, energy transition, etc.) our society and natural environments are facing. Due to article length constraints, we have omitted discussions of other AI-supported developments that could potentially influence EA, such as AI-driven text generation, and leave this for future exploration. Our letter is anchored by two central guiding questions: How might the integration of AI for data collection and analysis reshape our practices within EA? What developments and related debates should we as EA practitioners be aware of?

We follow the structure along the process looking at data collection/capture, data processing, data analysis and interpretation, linking these novel opportunities to EA procedural steps, benefits and challenges.

Automated biodiversity data collection, processing, and analysis – innovations relevant for the consideration of species and biodiversity in EA

Data collection – particularly in the field of species and habitats – remains a persistent constraint affecting both the quality and effectiveness of EA, consequently impacting the subsequent follow-up process aimed at learning and adapting measures, especially if they inadequately mitigate harm to environmental resources (González Del Campo 2012; Cilliers et al. 2022). In contrast to fields like air quality monitoring, which have a well-established tradition of automated data collection for EA applications (Hejlová and Voženílek 2013; Kumar et al. 2015), biodiversity data, for the most part, rely on in-person observations by experienced experts during EA processes.

Looking at biological science, we see that several automated technologies have significantly improved the collection of biodiversity data, providing richer and more detailed information on species distribution, abundance, and ecological processes. Advances in technologies such as acoustic sensors, unmanned aerial vehicles (UAVs) and camera traps now allow for rapid, non-invasive, and high-resolution collection of sound and image data. Three developments lead to this large-scale adoption of sensor networks in data collection relevant for biodiversity monitoring (Speaker et al. 2022) but also relevant to other contexts for the identification of information on other environmental issues:

1. Consumer market driven development of UAVs and camera traps led to significantly reduced costs and highly intuitive operation (Glover-Kapfer et al. 2019).

2. The open sourcing of sensor hardware and software reduced upfront costs and allowed for a fast community-driven development of monitoring tools (Hill et al. 2019).

3. The large amount of raw data produced by these sensor networks can be processed and stored using consumer IT hardware and available software that automate machine learning workflows (design, training, testing, deployment and operation) (Feng et al. 2019).

Consequently, according to authors such as Farley et al. (2018), ecology has entered the realm of big data due to developments in sensor technology and decreasing costs.

Two major developments will influence nature conservation and environmental planning in the near future:

1. Multi-modal machine learning workflows combine different sensor outputs like picture and sound to allow for even more robust automatic species recognition (van Klink et al. 2022).

2. Edge computing¹ moves the analysis onto the sensor itself. The smart camera trap automatically takes a picture and processes the species recognition on the device without any internet connection (Kays and Wikelski 2023). Only pictures of target species are saved. This reduces the amount of raw data generated and complies with the EU General Data Protection Regulation as no human pictures are saved.

Opportunities arise not only from minimising disturbance to wildlife behaviour during observation periods (by reducing human presence in the field) but also from the continuous availability of raw data representing real-life situations (Perry et al. 2022). However, additional on-site field observations may still be necessary, especially under specific circumstances such as assessing abundance (population size at a particular location, percentages in comparison with the total population). Currently, AI performs well in species detection but is still weak in accurate abundance estimation (Rhinehart et al. 2020; Bohnett et al. 2023). There are ongoing developments, towards individual song recognition and spatial sound analysis, aiming to enhance abundance estimates by providing a more precise assessment (Ghani et al. 2023). However, these capacities are not yet adequate for the population estimates required for EA. Technological advancement may soon bridge this gap. This example emphasises the importance of continuous interdisciplinary collaboration on AI application evolution to keep EA practitioners informed about technological developments and opportunities.

Next, to automated data collection, the effective impact of AI becomes apparent during the analysis of vast datasets which as a consequence can allow using data collected in longer time series as baseline data on the state of the environment in EA instead of short-term in-field collected data only. Recent studies in biological science demonstrate that AI offers promising solutions in tackling biodiversity challenges through automated species identification and classification (Besson et al. 2022; Wägele et al. 2022).

Most existing systems to date are based on deep learning, a subset of machine learning utilizing multilayer artificial neural networks. These are supervised deep learning algorithms requiring high-quality labelled data to learn species identification from raw data (Borowiec et al. 2022). Human experts must go through large datasets and correctly identify the species on this training data. The accuracy of the final model output mainly depends on the quality and balance of this training data. Consequently, it is crucial to thoroughly document and openly publish training data. In the case of biodiversity data, this entails obtaining permission from data owners (species experts, companies, agencies) to use and publish their data accordingly (Urbano and Cagnacci 2021). Similarly important, some species data necessitate confidentiality and obscuring exact locations to prevent potential disturbance or harm (Bubnicki et al. 2023). From initial application contexts and studies, we observe that these aspects are critical discussion points, warranting future studies to explore diverse perspectives of the groups of actors involved in EA, including authorities, consultants, and NGOs, to explore concerns and simultaneously consider opportunities. Within this context, questions regarding human capacities and roles of various actors arise, which we will discuss in the subsequent section, together with opportunities and challenges for EA practice.

Implications of these novel approaches of smarter data collection, processing, and open data repositories for theEA community and practice

As addressed by introductory studies into the field, such as Fonseca (2022) and González Del Campo and Gazzola (2020), the innovative approaches outlined above have the potential to overcome methodological and procedural limitations, along with resource constraints (financial and expert availability), which have often limited the availability and quality of baseline data (Gachechiladze-Bozhesku and Fischer 2012; Dias et al. 2022). Procedural steps most influenced by novel approaches in smarter data collection and processing encompass scoping, the zero variant, the consideration of alternatives, monitoring, and consequently, the availability of data for future environmental assessment. This availability could also support learning and indirectly influence tiering across planning levels. Similarly, coordinated application in other contexts such as the monitoring obligations following the FFH Directive for instance seems promising. In the following, we delve into the opportunities, challenges, and responsibilities across various actors how to deal with these emerging topics.

Firstly, innovation related to these novel approaches can influence the assessment of the state of the environment and likely have significant impacts based on the availability of long-term data. Data collection for both baseline data as well as monitoring purposes concerning species and habitat conditions has been significantly constrained in practical application due to its heavy reliance on partly highly skilled experts performing fieldwork. Future workforces may lack the necessary taxonomic expertise, as highlighted by experts like Stroud et al. (2022), making it challenging to fulfil this role. Consequently, technological innovations have the potential to reduce the necessity for human-centred data collection. At the same time, they have a substantial impact on the actors involved in EA and their skills needed to interfere in future planning processes. Experience in AI-driven data compilation underscores the indispensable role of experts in fully leveraging AI’s potential and achieving the accuracy and validity needed for commissioning purposes.

Secondly, this links the discussion to the responsibilities associated with the introduction, surveillance, and exploitation of these novel data collection and interpretation options, ensuring comparability and maintaining high-quality standards while profiting from their benefits. Managing and maintaining sensor networks and aggregating produced raw data remains a labour-intensive task and requires specific skills as outlined above. Only the supervision or handling of these processes by environmental authorities and adequate training of EA consultancies might secure the suitability according to legal standards and provide insights into the quality assurance, which again could help to increase acceptability of EA actors of these technological advancements.

This challenge leads us to, thirdly, the framing conditions and standardised formats, which guarantee transparency and maintain quality control. In contrast to other environmental aspects where data is often standardised through national and international laws (e.g. water and air quality) and gathered by official authorities at national and sub-national levels, the collection of biodiversity and species data exhibits a diverse approach. Using well-documented metadata is essential for correctly interpreting the data and conducting meaningful comparisons with data sets from different origins. In the context of EA, this is especially important, when automated workflows must be accepted by authorities and the public. In this context, project developers and environmental consultants as the users of the technologies need to assure the comparability of data collected as well as guarantee the transparency on quality-related aspects. Raw data needs to be run through reproducible pipelines and must always be accompanied by standardised open metadata formats like the Darwin Core Standard (Wieczorek et al. 2012). To provide an example, the picture of a lynx is useless, when information on where and how the camera trap was placed, for how long, and with what settings, is missing.

Looking from data collection to data interpretation, critical voices refer to the traceability of the AI supported analysis and its suitability as a basis for legal commissioning processes (e.g. according to nature protection laws) and monitoring according to EU Directives such as the EIA Directive, the Habitats Directive, or the Water Framework Directive. Implementing and trusting AI workflows often suffers from the black box effect (Díaz-Rodríguez et al. 2023). Recent developments strive for a more explainable AI (xAI) (Machlev et al. 2022; Samek 2023). xAI uses frameworks that help humans understand and interpret predictions made by AI models. This also allows for identifying unbalanced training data that might lead to biased model performance towards species with frequent observations.

Just like the data standards for the raw data, we need data standards to document the used AI models, software version, and used settings as far as possible. AI models need constant development, and this may challenge its use in terms of regulation or guidance, if specific models must be approved by authorities. We thus may face a balancing act between standardising the AI models (as might be required in legal commissioning and assessment processes) and ensuring sufficient flexibility for the models to be further developed and adopted to novel conditions. Herewith, EA consultancy might need the capacities of specialised conservation scientists to cooperate with. At the same time, environmental authorities require skills to develop quality standards and control mechanisms themselves – linking point three to point one and two again.

The discussion on data collection challenges is further extended by, fourthly, addressing the need to negotiate responsible institutions for the costs associated with systematic data collection. Normally, data collection related to EIAs is managed by the project developer, and the project developer is also responsible for monitoring, if the project permit includes this condition. While novel options arise through continuous data collection, the questions of presumably shared responsibilities and (economic) resources as well as knowledge and capabilities reach a new turning point. The discussion on data collection, capacities, and resources related to it is also directly linked to questions of responsibilities and legal framing conditions for data accessibility and storage. So far, accessibility of data from past EA processes is challenging, as the latest research on digitalisation in EA practice shows (Lyhne et al. 2022; Geissler et al. 2022; Uhlhorn et al. 2023; Garigliotti et al. 2023). While the collection of environmental reports faces obstacles, including legal constraints and a shortage of digital platforms, the sharing of raw data is infrequently practiced. This scarcity is, in part, attributed to conservation risks mentioned earlier. Despite these challenges, the sharing of raw data holds significant potential, aligning with the original intent of monitoring for learning from the impacts of superior or previous planning decisions. Aggregating raw data and processed data in open and general repositories would allow consultants and companies in EA to share the cost related to the collection and data processing effort and lead to more trust in data quality and workflows. While full-scale open data sharing may pose challenges for conservation purposes, a viable alternative could involve implementing restricted access within a protected network, involving official EA authorities and accredited consultancies with specialized expertise in conservation and impact assessment. To sum up, we see several benefits deriving by novel data gained from longer time series, more locations and faster analysis timeframes. They are accompanied by the need to act proactively in order to allow a sound and safe exploitation. The questions now pertain to the abilities and willingness of the EA community to transform their workflows, along with the adequacy of framing conditions, including legal standards. These factors may either facilitate or hinder a responsible advanced digital data sharing and usage within the community.

Conclusions and suggestions for further research

More generally, AI emerges as a transformative force with immense potential to reshape conventional approaches in environmental assessments. Like development in other disciplines (Howard 2019; Yüksel et al. 2023), AI will undoubtedly influence EA practice. Our exploration of the rapid development of digital technologies in the domain of conservation of species and biodiversity reveals a vast spectrum of possibilities for smarter and more efficient data collection, analysis and integration in to report generation. Alongside biodiversity, new digital technologies are increasingly being deployed in water, climate, and various other environmental issues, indicating a broad range of applications across the comprehensive environmental framework outlined in the regulations and frameworks that guide EA practice.

AI and digital technologies seem to constitute parts of the solution to current criticisms concerning data on which assessments are based and related deficiencies in data collection, documentation and transparency in EA practice. Thus, they potentially contribute to shorten permit processes and speed up green transition (as, e.g., called for by the European Commission 2022; Geissler and Jiricka-Pürrer 2023). Their application may ease efficient monitoring of global biodiversity and improve predictions of ecosystem dynamics (Lahoz-Monfort and Magrath 2021) and at the same time facilitate learning and tiering as originally intended in EA. However, digital technologies and AI may also induce new challenges, which means that we as an EA community need to proactively address and discuss critical implications along with the potentials. In order to balance between exploring advances and their advantages and at the same time maintaining EA quality and standards, we additionally need new funding for two big challenges: 1) We need publicly funded and run data repositories driven by interdisciplinary teams for high-quality training and reference data sets 2) We need additional funding to keep our ‘traditional’ monitoring schemes running in parallel with new digital monitoring tools in order to keep our monitoring time series valid (or establish them further) and transform them into the future. This might then allow also a transformation of EA while at the same time trying to assure quality and traceability of decisions taken.

Based on the previous sections, we summarise that AI technologies within few years will result in a wider set of baseline data at lower cost and higher speed than currently. This will likely A) improve our knowledge about species populations and their status, B) qualify significance determinations, and C) potentially speed up planning processes. But it will likely also lead to D) a transformation of the classical biologists towards digital dexterity as field studies become automated, E) ‘technification’ of discussions around accuracy and validity of collected data and thereby quality of EA, F) challenges to public acceptance and legitimacy of EA as the technification may be perceived as a black box to parts of the public, G) need for new training of competences among actors involved in EA to exert quality control and interpret data, and H) require updates of legislation and guidance related to data management.

Our key message is that the development of AI technologies is not only relevant for a few specialists but for many of us involved in EA. We will engage with AI technologies differently based on pre-knowledge, newly acquired expertise and professional responsibilities. Questions for future responsibilities and capacities of the diverse actors involved in EA come up. While consultancies may fear that some of their core resources are less needed, we see that engagement by EA actors in these new technological advancements is absolutely necessary. Trained experts are needed to advance the algorithms and provide quality control, but who is to be responsible for the training? To this aim, an understanding of the digital advancement is essential as well as skills of species identification itself; however, such competence profile is seldom, especially with profound knowledge of EA processes. We thus need, as a field of related actors, to raise attention to digital developments and the discussion on how we want digital technologies to support us, control the quality of their application as well as implications for effectiveness, rather than wait and see how we will be changed by the technological development. The current work on principles for the use of AI in IA among members of the IAIA (see session(s) at the IAIA 24 conference in Dublin) is a good example of the discussions we need to engage in, but the current extent of discussions is far from sufficient for us to make a proactive stance on digitalisation and AI.

We therefore suggest an outline of topics for a research agenda on digital technologies in IA. They go beyond the field of biodiversity and include interrelated topics that necessitate interdisciplinary and potentially transdisciplinary research:

1. Standardisation of application of AI in EA: Application of advanced digitalisation in EA on the one hand requires standardisation and uniformity of terminologies and content of EA reports. How much do we as a field accept to standardise across actors and contexts and how to standardise smoothly? When is standardisation needed, on the other hand, to ensure the quality of data collection and comparability of AI-based interpretation of data, and under which circumstances is the latter possible at all?

2. A new role for specialists in EA: As data collection is increasingly automated, there is a decreased demand for specialists to conduct on-site observations. Instead, specialists should focus on ensuring the appropriate application of technology for the intended purpose, conducting quality checks on data collection, and interpreting the data gathered. Can we anticipate new forms of collaboration, different dialogues, and a redefined role for specialists within the EA processes?

3. Changing responsibilities and roles: How will roles and institutional responsibilities in EA practice need to change? How will AI affect the distribution of tasks between public bodies and private developers in curating and managing data for the description of the current state, the consideration of alternatives and particularly the monitoring of unexpected impacts and the effectiveness of measures? How to implement the polluter pays principle accordingly if new knowledge and capacity building is needed to work with new digital approaches?

4. Understanding of digital technologies: How much expertise would be needed to ensure quality and validity of results? How to deal with a potential lack of understanding of technologies with increased degree and sophistication of digitalisation? Who in EA processes should understand and be able to explain technical aspects of digital technologies used?

5. Acceptance and legitimacy of decision-making: Increased technification of processes runs the risk of being perceived as ‘black boxes’, especially in AI applications. How should our field address the issue of transparency in this context? Furthermore, what implications do these ‘black boxes’ have for the acceptance of EA results by authorities and the public. How might they affect power dynamics? Under what circumstances do we risk jeopardizing the legitimacy of EA within the context of digital transformation?

6. Effectiveness of EA: How will digital technologies influence substantive, normative and transformative effectiveness leading also to incremental change and learning? Will we be able to use digital technologies as a lever to increase effectiveness, or do we risk losing focus on effectiveness in our own transformation?

7. Motivation and identity: As the role of specialists, writers and coordinators of EA processes will change, how does it affect our ideals of best practices, our identity, and our motivations? Would roles of software programmers be more prominent in EA practice, and would they need an enhanced understanding of values, conflicting objectives/interests and trade-offs concerning the environmental issues subject to EA in context to other policy and societal goals?

8. Training and competences: The rapid digital development means a need for rapidly changing skills. How do we ensure that we as a field have sufficiently updated skills? How should we change educational programs to ensure the upcoming need for competences?

9. Learning and coping with uncertainty: Will these novel developments provide a real option to introduce and continuously apply adaptive monitoring as recommended by several scholars for dealing with uncertainty?

10. Research and Evaluating IA: How will this change research and evaluation of IA and also of procedural steps under researched so far such as monitoring and quality-related questions? What are new research designs supported by AI? Which risks but also chances for EA research quality, transparency, replicability, and legitimacy does this entail?

We recognize the need for future interdisciplinary exchange to delve into the application of AI at various steps of the EA process, comprehensively exploring its opportunities, e.g., in scoping, assessing alternatives, evaluating significant environmental impacts and monitoring. At the same time, these studies must profoundly investigate risks and vulnerability to manipulation associated with AI application involving interdisciplinary exchange on quality and transparency.

With this letter, we extend an invitation to refine, critique, challenge, and/or act on the proposed topics. This letter adds to other ongoing initiatives that collectively have begun to explore the transformation awaiting us in the years to come. We anticipate and encourage broader dialogues fostering reflections among actors/stakeholders in EA on the implications of digital technologies for our field and practices. We aspire to see a surge in proactive decisions concerning how to effectively engage with and adapt to the realm of digital technologies.

Acknowledgments

The experiences and challenges identified in this letter are in part gained from a number of research projects, including the project “International trends in EIA and SEA research and practice 2.0” funded by the German Environmental Agency (UBA) (grant agreement number, FKZ 3721131010) and the DREAMS project funded by the Innovation Fund Denmark (grant agreement number, 0177- 00021B DREAMS).

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The work was supported by the Innovationsfonden [0177- 00021B DREAMS]; Umweltbundesamt [FKZ 3721131010].

Notes

1. “Edge computing: a distributed computing paradigm that brings computation to the ‘edge’ of a network by processing and analyzing data in real-time on the same device that collects the data, rather than sending all data to a centralized location for processing.” (Kays et al.& Wikelski 2023, 2).