Main

Biodiversity is integral to maintaining healthy ecosystems and thus to supporting human health, food security, climate stability and agricultural productivity1,2,3. To effectively guide environmental policies and conservation efforts, we need tools that can rapidly and accurately inform about the current and future state of biodiversity4,5. Yet, contemporary biodiversity predictions remain inaccurate, particularly at fine spatiotemporal resolutions6, despite the increasing availability of extensive, long-term biodiversity data7,8,9,10, rapid advancements of technology to collect large biodiversity data11,12, and continuously improving modelling tools13,14. Reasons why reliable biodiversity prediction has remained so challenging include the inherently complex dynamics of ecological systems15, the diverse and often inconsistent sources of large datasets16,17, and the lack of modelling tools capable of rapidly converting the continuous data streams into information transferable to policy and management recommendations18.

As a partial solution to the challenge of achieving fine-resolution biodiversity data, much hope has been invested in the unparalleled potential of citizen science to provide data on a massive scale. Although the potential of citizen science has been repeatedly demonstrated19,20,21,22, data generated by citizen scientists are fraught with sources of biases and noise, potentially compromising the reliability of the resulting inference23. Most critical observer-based biases in citizen science relate to heterogeneity in participation, detectability, sampling and preference24. As it is difficult to reliably account for the variability in citizens in their skills of identifying species, as well as to quantify the spatiotemporal sampling effort, it remains hard to disentangle biological signals from these observation biases, especially if sampling effort is not carefully documented25,26.

Digital twinning refers to the concept of creating a digital counterpart of a real-world system. In ecology, digital twinning could mean building a dynamically updated digital model of a species’ distribution or an ecosystem’s state, based on continuously incoming observational data. While originally developed in engineering to simulate and optimize physical systems27, digital twinning is gaining interest in biodiversity research, where it can help integrate data, models and expert knowledge in near real time28,29,30. This approach holds promise for improving ecological forecasting and supporting timely environmental decision-making31. However, the development of digital twins (DTs) for biodiversity remains a complex and emerging research frontier, hindered by the complexity of natural ecosystems, the need to combine heterogeneous data sources and the technical challenges associated with generating and processing real-time biodiversity data streams.

This Article aims to demonstrate the applicability of DT approaches in biodiversity research for achieving accurate, real-time predictions of species distributions. We illustrate this through a case study in audio-based bird monitoring, showcasing how reliable real-time biodiversity predictions can be achieved through a DT approach that combines the strengths of citizen science, machine learning and high-performance computation. We build on recent approaches in data integration32,33 and integrated species distribution modelling34 to combine the continuous flow of new citizen science data with previous long-term data on bird spatial distributions, timing of migration and pattens of singing activity.

Our approach features a continuous model updating process, ensuring that the digital version and its predictions remain responsive to real-time changes in bird activity and environmental conditions. A core feature that distinguishes digital twinning from data integration, is that digital twinning goes further by maintaining a dynamically updated model that mirrors the real-word system as it evolves over time, here distributions, migrations and singing activity of birds, as well as citizens recording them. By relying solely on machine-learning-based bird classifications rather than citizen-based classifications, we remove an important part of observer heterogeneity and increase inclusivity by enabling ordinary citizens without bird identification skills to take part in data collection by making bird recordings. The technological innovations developed in this study not only reduce the time required to generate accurate biodiversity information for policy and management but also increase inclusivity by broadening the stakeholder community and the roles of the stakeholders. This approach empowers and engages citizens to provide pivotal contributions to both scientific research and environmental monitoring.

A tool for digital citizen science

We created a smartphone app called Muuttolintujen Kevät, henceforth the MK app, with the Finnish name meaning ‘The spring of migratory birds’ (Fig. 1). The app was launched on 30 March 2023, through a publicity campaign run in collaboration with the Finnish broadcasting company Yle. The MK app includes three recoding types: (1) direct recordings, (2) interval recordings and (3) point count recordings. The MK app was specifically designed to overcome two critical limitations of citizen science in biodiversity research.

Fig. 1: The citizen science smartphone application MK.

a, The app has a continuously updating collective observation board where users can relate their detections to those of the other users (detections exemplified in the map for 1 to 4 April 2025. b, The machine-learning-based classifications are calibrated to a probability scale and highlighted with green colour if probability exceeds 0.90. c, In collaboration with national parks and municipalities, we implemented 580 permanent point count locations where citizens can make systematic 5-min recordings. d, To increase societal impact, user commitment and educational use in Finnish schools, we implemented a bird game through which citizens can learn bird vocalizations. e, The aggregated duration of recordings and number of detections per day peak during spring but remain continuous over the entire year. During peak days, the app has accumulated >1,000 h of recordings (with a median length of 33 s) which involve >100,000 detections. f, Among the 263 species that can be detected by the app, 110 have been observed with 90% confidence >5,000 times.

Source data

First, to mitigate the differences in species identification skills among citizens, all classifications are performed by a machine learning model, and thus bird identification by citizens is not required. Importantly, not only the classifications but also all raw audio data are submitted and stored in the MK server. This allows the reclassification of the audio with continuously improving machine learning models, as well as the manual validation of species detections if necessary. For this study, we fine-tuned a baseline BirdNet model35 for 263 Finnish bird species (all breeding species, non-breeding migrants and most common vagrants) using high-quality annotations generated by bird experts36. The model was calibrated specifically for the MK app data, and a 90% confidence score, which we used as threshold for the analyses presented in this Article, can be interpreted as 90% probability of correct classification.

Second, to mitigate spatial observation bias and preferential sampling, the MK app enables not only direct recordings, but also interval recordings and systematic point counts. In the interval recording mode, the app records 1 min every 10 min, continuing up to 12 h. This enables citizens to record, for example, overnight in their yard, including the very early morning hours when birds are most vocal. While the interval recordings do not remove the spatial bias of where the recordings are conducted, they largely remove the temporal preferential bias of when they are conducted. Even if the initiation of an interval recording would be triggered by bird vocalization activity, after the first 9-min break, the recorded minutes represent bird vocalization activity in a much less biased way than direct recordings. The permanent point count network was established in collaboration with Finnish national parks and municipalities and includes 580 preselected locations in which the citizens can conduct a systematic recording (Fig. 1c). The permanent point count locations mitigate spatial observation bias, as the citizens make recordings at preselected locations. They also partially mitigate the temporal bias, because the recording interval is 5 min long, and thus especially its latter part is less dependent on whether bird vocalization activity triggered the initiation of the recording. We have furthermore encouraged users to initiate point count recordings whenever they walk through the route, disregarding whether birds are vocalizing or not. To engage the users and support their education on bird sound identification, a gamified bird vocalization training feature was added to the MK app in spring 2025 (Fig. 1d).

The MK app rapidly gained popularity among Finnish citizens, with 315,609 individuals (5% of the national population) submitting at least one recording by 29 September 2025. By this date, the app has yielded 16.3 million recordings which contain 15.0 million bird detections with at least 90% classification probability. Most recordings and detections are made through direct recording, but a substantial proportion is also obtained through the interval and point count recordings (Fig. 1e). The detections involve 261 species, out of which 110 have been detected at least 5,000 times (Fig. 1f). In addition, the MK app is used actively for nature education in Finnish schools. For example, a single bird observation event organized on 7 April 2025 was attended by 3,900 school children representing 73 schools. Furthermore, by 13 October 2025, the bird game has attracted 34,248 users, who have together scored a total of 4.2 million identification attempts, each involving the selection of the correct vocalizing species from four candidate species.

A real-time biodiversity DT

We developed a DT that predicts spatiotemporal distributions of bird occurrences and their vocal activity across Finland, with a spatial resolution of one-hectare, and a temporal updating frequency of one day. The DT operates by updating a prior model each night using the latest data accumulated through the MK app (Fig. 2).

Fig. 2: Overview of the DT modelling strategy.

We parameterized a prior model by combining long-term bird observations with spatial and temporal predictors. a, Continuous recordings provide prior information about when birds vocalize, conditional on their presence. b, Long-term citizen science observations provide prior information about the timing of migration. c, Systematic transect line counts, as combined with data on land cover, forest structure and climatic predictors, provide prior information about the spatial distributions of birds. d–f, The continuously accumulating MK app data are used to update the detection model (d), the migration model (e) and the spatial distribution model (f) and, hence, knowledge of bird spatiotemporal distributions and singing activity. g, Probabilistic predictions by the three model components yield the probability that a given bird species is detected in a given MK app recording, as for this to happen (1) the bird should have returned from migration (or be resident), (2) the location should be part of the birds spatial distribution and (3) the bird should vocalize in a manner that leads to detection in the MK app.

The model predictions are a product of three probabilistic components. First, the migration model yields the probability ({p}_{{\rm{M}}}) by which the species is present from the point of view of their migratory behaviour (with ({p}_{{\rm{M}}}=1) for non-migratory species), given the latitude, year and the day of the year. Second, the spatial distribution model yields the probability ({p}_{{\rm{S}}}) by which a given location is part of the species distribution during the non-migratory period. Third, the detection model yields the probability ({p}_{{\rm{D}}}) by which, conditional on a species being present, it vocalizes in a way that the MK app detects it with at least 90% classification probability. The detection model is parameterized in terms of the day of the year, time of the day, and the length and type of recording. The product of these three probabilities ((p={p}_{{\rm{M}}}{p}_{{\rm{S}}}{p}_{{\rm{D}}})) yields the probability the species is observed in a given MK app recording.

We inferred the prior migration model using long-term citizen science data on species observations (Fig. 2b). We quantified prior knowledge on bird species’ spatial distributions by fitting the joint species distribution model Hierarchical Modelling of Species Communities (HMSC)37 to long-term data on transect-line surveys, using as predictors 1-ha-resolution raster maps of land-cover variables, forest structure variables and climatic variables (Fig. 2c). The prior detection model was inferred using 4-year-long continuous passive audio monitoring (PAM) data from seven Finnish research stations38. We used logistic regression to model the probability by which a passive audio recorder would detect a vocalization of a given species as a function of the day of the year and time of the day, conditional on the species being present at the location (Fig. 2a).

We used the MK app data to update the migration functions and spatial distributions at daily intervals. This is a computationally intensive task, which we simplified by modelling species independently and updating the prior in stages. The first stage is to translate the detection model from PAM to the MK app data; this is achieved using a probit model that accounts for the length and type (direct, interval or point count) of the recording. The second stage updates the parameters of the migration model using MK app data directly, with the overall shape of the migration probability curve over time shrunk to the prior using a functional penalty to promote stability and improve forecasting. Finally, the spatial distribution is updated using a local-likelihood method, in which each cell is updated based solely on nearby data. This approach allows cells to be updated in parallel and is critical for scalability. The updating of the spatial distribution component is conducted directly at the level of the model predictions through spatial smoothing, not at the level of prior model parameters that map, for instance, the environmental affinities of the species. Full details are available in the Methods.

Example predictions

The DT continuously updates its long-term knowledge on bird spatiotemporal distributions through the newly accumulating citizen science data. As illustrated in Fig. 3 for two example species, the posterior predictions of species distributions can deviate substantially from the prior predictions both at large and small spatial scales. This indicates that the DT undergoes substantial learning. For the common gull (Larus canus), the DT increases the contrast between high-prevalence areas (lakes and coastal areas) and low-prevalence areas, thus changing the predictions consistently over large spatial scales (Fig. 3c). For the sedge warbler (Acrocephalus schoenobaenus), the posterior predictions deviate substantially from the prior predictions at higher spatial resolution, as illustrated for the capital area in Fig. 3d–f. Sedge warblers breed in reedbeds, which are not well represented in the transect line data and which are not distinguished in the habitat classification used to make prior predictions. This leads to poor predictive performance of the prior model, leaving room for substantial improvement by the DT in areas with abundant MK app data such as near the capital. The DT also learns to predict bird temporal dynamics. By tracking the daily arrival of migrants, the DT can accurately infer the timing of spring migration and the associated spatial dispersal (Fig. 4a). This results in highly dynamic spatiotemporal distributions, such as for the garden warbler (Sylvia borin), where the distribution changes from almost universal absence to widespread presence within 2 weeks (Fig. 4b,c).

Fig. 3: Example illustrations comparing posterior and prior predictions of spatial distributions.

a–f, National-level distributions for the common gull (Larus canus) (a–c) and smaller-scale distributions for the sedge warbler (Acrocephalus schoenobaenus) around the capital area (d–f). The prior model is based on long-term bird data only, whereas the posterior model also utilizes observations acquired by digital citizen science through the MK app. The spatial predictions are shown for the prior mean (a and d), posterior mean (b and e) and the difference between posterior and prior mean (c and f).

Fig. 4: Example illustrations of spatiotemporal predictions for the garden warbler (Sylvia borin).

a, Comparison of migratory timing between the prior model and the posterior model for 2024. b,c, Posterior predictions for the species distribution in the beginning (b; 15 May) and in the end (c; 1 June) of the realized migratory period in 2024. The black dots in b and c show the MK app detections for the exemplified days.

Evaluation of predictive capacity

We evaluated the predictive capacity of the DT with two different test datasets: MK app recordings for the next day, and manual point counts for the next day. For both test datasets, the DT approach substantially improved predictive capacity, as compared with the prior model (Fig. 5). We performed both evaluations for those 89 species for which the MK app data contained at least 5,000 detections in 2024. For both evaluations, we updated the DT model using the MK app data up to the previous day and then used the test data to evaluate the predictions of both the prior and the DT models.

Fig. 5: Comparison of DT (posterior) and prior models in terms of predictive power.

a–c, Evaluation of the capacity to predict future MK app data. We updated the DT dynamically until the present day (a) and contrasted the next day’s prior and posterior predictions to actual MK data (b and c). The results are shown for those 89 species that were observed at least 5,000 times by the end of 2024. d–f, Evaluation of the capacity to predict future point count data by experts. We used the DT predictions from the end of April 2025 to select point count locations that showed the greatest differences between the prior and DT predictions (a), and then compared the next day’s prior and posterior predictions with the actual point count data (e and f). The results are shown for those 73 species that were observed at least 10 times in the expert point counts. The dot size in e and f is proportional to (p(1-p)), where (p) is the species’ prevalence in the point count data and, hence, larger dots show cases where the AUC can be calculated more reliably. The DT makes substantially better predictions for most species (mean difference in AUC between posterior and prior predictions 0.06 for both types of prediction), and especially for species for which the prior model is poor (b and d) and that are migratory (c and f). The P values and slopes in b, c, d and f originate from a linear model that includes as explanatory variables both prior predictive performance and the proportion of time that the species spends as a resident.

Source data

For the evaluation against future MK app data, we used the year 2024 as the evaluation period. Based on the location, time, type and duration of each MK app recording, we predicted detection probabilities for each species by both the prior and the DT models, with the DT incorporating data up to the previous day (Fig. 5a). The DT substantially improved the next-day MK app predictions for bird detections, with the mean area under the curve (AUC) across 89 species increasing from 0.71 to 0.77. The improvement was most pronounced for migratory species and for species with initially poor prior model predictions (Fig. 5b,c).

To further evaluate the difference between the posterior and prior predictions against fully independent data, we performed manual point counts by bird experts in preselected locations from 7 May to 7 June 2025. The bird experts were seasoned volunteer birdwatchers, whose capacity to identify birds from their vocalization has been demonstrated, for example, by providing high-quality survey data to the national line transect or point counting schemes. The manual point count locations were selected algorithmically to represent different combinations of prior and DT predictive probabilities, prioritizing sites where the prior and posterior predictions were most contradictory. These locations were visited by bird experts, who performed a total of 1,185 5-min point counts without knowing the prior and DT predictions by which the locations were selected. This test confirmed that the DT leads to improved predictions: the mean AUC across those 73 species that were available for this comparison increased from 0.62 to 0.67 for the expert point count data. This was again the case especially for species for which the prior model was poor (Fig. 5e) and that are migratory (Fig. 5f).

We further compared the DT predictions with those based on the eBird39 global citizen science project. For each survey week, we extracted species occurrence probabilities from the eBird Status and Trends Weekly Abundance Maps released in summer 2025, which represent data accumulated through 202340. For those 53 species for which eBird-based predictions were available, the mean AUC was 0.62 for eBird-based predictions and 0.67 for DT predictions.

Discussion

The need for accurate and real-time biodiversity predictions has been much advocated14,15,41, but achieving this has remained challenging28. This Article demonstrates the feasibility of constructing a real-time DT of biodiversity and shows, with independent test data, that the DT improves predictions about the current and future states of biodiversity. By combining long-term data with a continuous stream of smartphone-based citizen science data, our DT addresses the challenge of generating reliable predictions at a high spatiotemporal resolution6. Such predictions are much needed under the UN Convention on Biological Diversity’s Global Biodiversity Framework to detect biodiversity changes and to promptly implement the necessary environmental management and policy actions. Although the DT developed here is aimed at quantifying changes in species distributions rather than directly identifying their potential drivers or recommending management or policy actions, it provides a foundation for making informed progress in these directions.

Citizen science can provide massive amounts of biodiversity data. For example, the platforms eBird39, iNaturalist42 and Pl@ntNet43 have recruited some 1.1 million, 8.9 million and 8.2 million users, respectively. These extensive citizen science datasets have not only provided an invaluable resource for biodiversity research but have also stimulated the development of numerous statistical methods to address data quality issues, such as sampling biases and detection errors23. For example, although eBird’s data collection procedures involve systematic quality control and quantification of user skills, using these data for prediction and inference requires statistical approaches that carefully account for confounding factors and changes in the observation process. The best practice recommendations for using eBird data involve choices related to filtering the data for complete checklists, performing spatial subsampling and using filters for observation effort44. The predictions based on eBird data that we utilized in our comparison are not updated automatically in real time, but periodically by Cornell Lab of Ornithology data scientists, who provide Status and Trends products based on data accumulated over several years.

A core feature of the MK app is that it was directly developed to overcome the outstanding challenges of citizen science23. First, to tackle the issue of variable and often unknown sampling effort, the MK app quantifies the location, time, type and duration of each recording and implements standardized interval recordings and permanent point count routes. Second, to remove observer heterogeneity in species identification, the MK app uses machine-learning-based classifications with well-calibrated estimates of uncertainty. These characteristics of the MK app data facilitated their straightforward integration into a predictive DT approach. Furthermore, by storing raw audio data, the DT enables reclassification of past observations with continuously improving classification models, ensuring that they remain useful and accurate over time. Despite the above-mentioned features, the MK app data have some of the biases that are characteristic to citizen science datasets. Most importantly, the direct recordings are triggered by bird vocalizations that are of interest to the users. As shown in our previous analysis, some users target only new species that they have not recorded before, whereas other users provide data that are comparable to PAM45. Accounting for such variation in user profiles provides an important challenge for future work. Another limitation is that the MK app is based on audio only, omitting visual observations of birds.

Global biodiversity databases have major spatial biases, which influence our understanding of biodiversity and hamper its protection46. Many areas remain understudied due to barriers related to wealth, language, geographical location and security47, making it difficult to implement large-scale biomonitoring programmes that would require transport of specialized experts and equipment at appropriate times. To fill such spatial and temporal gaps in biodiversity data, the potential of citizen science has been previously recognized48. Our approach is highly scalable both computationally and in terms of smartphone technology and can be easily extended across geographical regions, given the widespread global ownership of smartphones. Even for birds, one of the best-studied taxonomic groups, and in Finland, a country with exceptionally well-documented biodiversity, our DT approach demonstrated substantial improvements in predictive performance. Thus, in regions where biodiversity is less well studied, our technology offers strong potential to rapidly improve ecological knowledge and inform conservation efforts.

Our DT approach may not generalize straightforwardly to many existing citizen science data streams, as the seamless integration between the data collection and the real-time predictive modelling was enabled by the fact that the MK app was specifically designed to serve this purpose. While building an operational DT such as the one presented here may initially require more effort than most other citizen science platforms, its capabilities go well beyond what static systems can achieve, as it provides a dynamic approach for forecasting biodiversity. Compared with this effort, the improvement in predictive power that we reported here may appear moderate: the AUC improved from 0.62 in our prior model to 0.67 in the DT. However, we argue that this improvement is substantial, as the AUC value increased by 42% if compared with the baseline value of 0.50. Instead, the low AUC values are explained by the fact that the predictive task that we targeted is highly challenging. Namely, our test data concern variation in species detections over a small geographic area (where all the species generally occur) and over a short period (during which all the species were generally present), making it highly challenging to predict in which samples the species were present and in which they were absent.

Successfully protecting nature requires collaboration between governments, businesses and civil society, with a key question being how to engage a larger part of society in supporting nature49. Citizen science has great potential to engage people more actively in environmental monitoring50, and our DT addresses major challenges associated with large-scale spatial monitoring in citizen science51. Moreover, mobile-based approaches may attract younger participants, making them an important means of increasing public understanding of science52. Thus, these approaches can be effective in motivating participants to sample biodiversity in more meaningful ways, potentially reducing some of the biases inherent in how citizen science data are collected53. The MK app has substantially promoted citizen engagement and helped reconnect citizens with nature through extensive school collaboration, media coverage, the possibility of sharing results through social media, and educational features such as the bird game. In particular, the MK app has gained popularity among ordinary citizens who do not necessarily recognize any bird sounds themselves, as it enables them not only to learn which birds vocalize in their surroundings, but also to contribute valuable biodiversity data that are immediately used for research and monitoring. This inclusivity, together with the DT approach, greatly enhances the ability of citizen science to provide reliable, real-time information on global biodiversity, helping to bridge the current time lag between research and policy.

Methods

The MK smartphone app

The MK mobile application and its technology infrastructure were developed collaboratively by the University of Jyväskylä, CSC – IT Center for Science and the University of Helsinki. The MK app is built upon an open-source technology stack and developed using the Flutter mobile application framework. The MK app is freely available on Android and Apple mobile devices in Finland and Sweden. At the core of the application’s architecture lies a server-side, customized Camunda BPM hyperautomation platform, providing a robust solution for anonymous user participation in research and data collection processes. This architecture addresses challenges related to European Union data protection regulations (General Data Protection Regulation), ensuring secure application use and enabling the transfer of data for research purposes.

The operational workflow is initiated by a user recording bird sound—including direct, interval or point count recordings—along with metadata such as an anonymous participation key, location, recording length and timestamp. These data are transmitted via Internet connection to an application programming interface running within a secure computing environment provided by CSC. The audio files are stored in CSC’s object storage system Allas, while the metadata are saved in a MongoDB database. The workflow directs the audio files to several virtual machines running in the cPouta cloud service, which performs bird classifications. The results are returned to the user, who can voluntarily assess the correctness of the identifications and provide feedback to further develop the classification model. The backend stores observation data and results for scientific purposes and retains the original audio files, allowing reprocessing with future classification models and manual validation of observations.

The machine-learning-based model for bird classification

The bird species classifications are produced with a convolutional neural network that consists of a pretrained convolutional base of EfficientNet B0 architecture from BirdNET-Analyzer35 and a classification head that we fine-tuned with vocalizations of 263 Finnish bird species36. Although the list of the selected 263 species is not the full list of all 496 species ever recorded in Finland, it contains all breeding species, non-breeding migrants and most common vagrants, making it unlikely that a citizen records a species not included in the classification model. The training dataset combined targeted recordings from Xeno-canto54, soundscape recordings from eight sampling locations in Finland, targeted field recordings by Harry J. Lehto and selected mobile phone recordings produced by MK app users. Labels for training data were collected through Bird Sounds Global annotation portal (https://bsg.laji.fi).

The classification model analyses the recordings in 3-s segments. The audio signal is converted into spectrogram images with an overlap of 1 s between consecutive segments using short-time Fourier transform. For each segment, the model predicts detection probabilities for all species. The model predictions were calibrated with species-specific logistic regression models. We selected 80 vocalizations per species from the MK phone recordings uniformly across confidence bins ranging from 0.2 to 1.0. The binary labels (presence/absence of the species) were provided by a bird expert who listened to the recordings. The predictions of highly unlikely species are penalized on the basis of location and day of the year to remove obvious misclassifications (for example, migratory species detected during winter) from the data.

The citizen science campaign

The MK application was launched in collaboration with Finland’s national public broadcasting company Yle, which substantially amplified its visibility in national media. The first public mention occurred on 12 April 2023, during the Metsäradio (‘Forest radio‘ in Finnish) programme, which focuses on forestry, nature and outdoor lifestyle. Subsequent coverage included the Luontoilta (‘Nature evening’ in Finnish) radio broadcast on 4 May, and a featured theme on Yle’s special television programme Muuttolintujen Kevät (‘Spring of migratory birds’ in Finnish) on 10 May. The application was also highlighted in Yle’s main evening news broadcast, which reaches an average television audience of approximately 750,000 viewers—roughly 14% of the Finnish population. In addition to traditiona