Our experimental site, the ECOSENSE forest, is a mixed temperate forest ecosystem located in the municipality of Ettenheim, Germany at the foothills of the Black Forest (reference point mixed plot tower: 48.26852173° N, 7.87821731° E, WGS-84, height 521 ma.s.l.). The climate is classified as Cfb (temperate air temperature, no dry season, warm summer) in the current Köppen–Geiger climate map (Beck et al., 2023). Mean annual precipitation sum and air temperature at the closest official weather stations are 911 mm (station Ettenheim/Ettenheimmünster 2.71 km to the south, 214 ma.s.l.) and 11.0 °C (station Lahr 11.37 km to the northwest, 156 ma.s.l.), respectively (DWD, 2023, reference period 1991–2020). The managed (low-impact) forest ecosystem is dominated by European beech (Fagus sylvatica L.), interspersed with Norway spruce (Picea abies L.), English oak (Quercus robur L.), European silver fir (Abies alba MILL.), Douglas fir (Pseudotsuga menziesii MIRBEL) and Scots pine trees (Pinus sylvestris L.). Two representative soil profiles were analyzed according to IUSS Working Group WRB (2022): In the southern part of the area a Dystric Stagnic Cambisol (loamic), derived from carbonate-free quaternary loess over strongly weathered mesozoic shell-limestone, is found. Below 40 cm redoximorphic features occur from seasonal water logging. In the northern part a Dystric Skeletic Cambisol (siltic, humic) was identified, derived from carbonate-free quaternary loess over colored/platy sandstone. Below 40 cm a strong platy structure and a high stone content occurs. Both profiles are strongly acidified with a pH (KCl) between 3.5 and 4.5. The forest floor is thin, consisting mostly of OL and shows a patchy OF layer.

In the ECOSENSE forest, we established three intensive measurement plots, of which one is dominated by European beech (“beech plot”), one by Douglas Fir (“Douglas fir plot”) and one contains a mix of the two species (“mixed plot”). Within each of these plots, we installed three scaffold towers that allow access to the tree canopies for sensor installation, manual measurements, e.g., of leaf gas exchange and destructive sampling (Fig. 2). For this purpose, towers were equipped with canopy-access platforms in 24–26 m height. The tower located in the mixed plot extends above the treetops to a total height of 46 m and accommodates an eddy covariance system for continuous measurements of net ecosystem CO2 and H2O fluxes over the forest (Sulzer et al., 2025) as well as other sensors and instruments capturing environmental variables (see Sect. 2.6 for more details). The three towers were placed in an arrangement so that the Douglas fir plot and the beech plot are located within the long-term turbulent footprint (measurement area) of the eddy covariance system at the mixed plot. The scaffold towers were constructed in April 2024. To keep the impact on forest soil, i.e., amount of sealed surface area, as low as possible, we opted for small concrete foundations (in the size of tower base areas, depth 0.80 m) and micropiles (eight per tower, in two different distances from the tower) as anchors for tower guy wires providing static stability. We also operate a weather station at a nearby clearing (see Fig. 3). An additional measurement plot with Silver firs (supplied with power but not connected to the communication network) is located 245 m southeast and downhill from the beech plot.

Central to our field site infrastructure is a custom-built measurement container (2 m × 4.8 m), mounted on a trailer chassis. It functions as the backbone for power distribution and network connectivity in the field, which will be described in detail in the following sections (see Sect. 2.4 and 2.5). It also houses gas analyzers that are not rated for outdoor use. Since some instruments only function at moderate temperatures and additionally produce considerable heat during operation, the container is ventilated and equipped with a redundant air conditioning system: if one unit fails, the second takes over automatically to maintain cool and stable indoor conditions and avoid overheating of expensive and sensible devices. The interior of the container is insulated with closed-cell rubber foam⁺ and covered with wooden paneling. This allows for the easy installation of shelves, lighting and additional hardware. The container's sheet metal construction provides inherent protection against lightning strikes. In addition to instruments and IT hardware, the container stores tools, first aid supplies, a defibrillator and work benches for preparing samples and performing small tasks. Pressurized gas cylinders are stored in a separate, lockable metal shed next to the measurement container.

To provide power to the ECOSENSE forest, we tap into the municipal power grid at a high-voltage transformer (20 kV) via a three-phase transformer located around 700 m from the measurement container. Within the container, a three-phase (400 V) power line is routed in and split into multiple fused circuits, effectively distributing power across separate 230 V lines. Individual circuits are additionally protected with fault-current circuit breakers (RCDs) and their respective power consumptions are monitored with power meters⁺. Those also allow for remote surveillance and shutdown of (parts of) the power supply. Additionally, we use DC-RCDs to protect from DC failures. The entire system is equipped with surge protection, which diverts lightning and other events to ground or short-circuits them. Critical infrastructure and sensitive measuring devices run on uninterruptible power supply units (UPS) within the container to reduce the risk of damage during (short) power failure or current spikes.

For safety reasons, we decided to only provide 230 V AC power within the measurement container, reduce voltage there via AC/DC-converters⁺ and distribute power across the ECOSENSE forest at protective extra-low voltage (< 50 V DC, DIN norm VDE 0100-410:2018-10). For this, fused power lines⁺ in cable conduits⁺ were laid to each of the measurement plots. There, voltages are further converted by DC/DC-converters⁺, usually to 24 or 12 V, as required by measurement devices and data loggers. The various power lines are galvanically isolated to avoid failure of the entire electricity network in case of, e.g., malfunctional devices or external factors like lightning strikes. We emphasize that it has proven useful that each device connected to the power infrastructure is individually and appropriately fused to avoid negative mutual influence in case of broken equipment or human errors in operations among researchers of different projects. An overview of power supply, distribution and access points across the ECOSENSE forest can be found in Fig. 3.

Due to limited mobile phone network coverage in the ECOSENSE forest, we decided to install a satellite-based internet access point⁺. The receiving device was installed on top of the 46 m above-canopy tower (mixed plot). From there, an Ethernet cable⁺ connects to the main network switch⁺ inside the measurement container. The whole network in the ECOSENSE forest is directly connected to the network of the University of Freiburg via a VPN (virtual private network) tunnel. This approach facilitates easier access to research data, management systems and underlying servers (see Sect. 3), some of which are only accessible within the university network. Additionally, static IP addresses within a defined address pool can be manually assigned to devices in the ECOSENSE forest. This makes servers, computers and data loggers remotely accessible from the university network or through the university's VPN client. These are centrally managed by the ECOSENSE technicians. The remaining IP addresses available (253 addresses in total) are assigned dynamically to devices connecting to the network on site. The gateway⁺, which establishes the connection via VPN tunnel, is protected by a Fortinet firewall managed by the RZ (Computing Center) of the University of Freiburg. To restrict and track access from the outside, the University of Freiburg hosts a MAC filter, meaning only computers/servers with pre-registered MAC addresses can connect to the ECOSENSE forest directly. Another possibility for remote access is through the university's VPN client, which requires a login with a personalized user ID. To further enhance network security, the Wi-Fi access points are centrally managed, allowing continuous monitoring of connected clients. The network is secured using WPA2 encryption, which significantly increases the difficulty of unauthorized access. The LoRaWAN (Long Range Wide Area Network, see below) network is secured with AES-128 encryption. This allows us to prevent attacks against the ECOSENSE forest network from both outside and inside.

From the main network switch, the network spans to various other switches⁺, located at the three plots at ground level, on the tower platforms and at the weather station (see Fig. 3). For the most part, we laid patch cables for connections. Only for the long distance (> 450 m) from the main tower to the clearing with the weather station, we used a coaxial cable that was buried in a tube underground, together with the main power cable. For all important network nodes, intelligent and controllable managed layer-two network switches⁺ were procured to facilitate remote troubleshooting and allow for remote restart and restriction of data transfer over specific ports. To minimize the risk of surge damage, e.g., in case of lightning strikes, we included surge protectors⁺ at important network junctions and in front of sensitive and expensive devices, in particular protecting the main network switch and the server⁺ inside the measurement container. All sensors within the network can regularly retrieve the network time from the field site server to account for a potential time drift.

In addition to connection via Ethernet ports, we installed various Wi-Fi access points⁺ covering the three main plots (compare Fig. 3). In case an Ethernet cable is damaged or the connection is otherwise faulty, the Wi-Fi network automatically takes over part of the network traffic, offering redundancy and thus increasing network stability.

For newly developed sensors, we provide LoRaWAN as a local data transmission protocol. The PoE (Power over Ethernet) powered LoRaWAN gateways are installed on top of the measurement container and on each tower platform to increase transmission reliability and redundancy. In Europe, LoRaWAN gateways typically operate at 433.05–434.79 MHz (ISM band EU433) or 863–870 MHz (ISM band EU863). Due to its energy efficiency during transmissions, LoRaWAN provides an advantage for sensors that are autonomous or miniaturized. The gateways send their data via the network in the ECOSENSE forest to “The Things Network”, from where the measurements are then pushed to the server inside the measurement container.

Prospectively, network interconnection between plots should be changed to fiber-optic cables, as the amount of data transported is larger than expected, in particular due to sensors with a high temporal resolution or image-based sensors, that were in part not included in the initial project plan. This would also expand network capacity for future technical developments and collaborations that will likely further increase data traffic within the ECOSENSE forest.

We established a comprehensive monitoring system in the ECOSENSE forest across multiple scales, i.e., from leaf to stand level. More than 670 commercial and 430 self-built sensors monitor over 90 distinct parameters, fluxes, or processes generating upwards of 4500 time series that capture soil, tree and atmospheric processes with high spatial and temporal resolution. The volume of incoming data varies considerably throughout the year: on average, we receive 1 GB time series data from in situ sensors per month, while high-throughput instruments such as leaf angle cameras, phenocams, terrestrial LiDAR, isotope and volatile organic compound (VOC) analyzers, and (relaxed) eddy covariance systems generate substantially larger raw data volumes. The field site server handles over 66 TB of traffic per month, of which only 240 GB is transferred to the university, as most raw output is preprocessed on-site and only a small fraction is retained long-term. Campaign data, particularly the monthly LiDAR flights, can produce many gigabytes per flight, but only a much smaller processed version is retained. This processing is carried out outside our server infrastructure. With this large amount of data collected at different scales, we aim to improve upscaling approaches and modeling of ecosystem processes and stresses (see Sect. 4.1).

The following inventory (Table 1) provides an overview of the main measured processes and fluxes, based on ecosystem compartments in the ECOSENSE forest. Fluxes and processes shown in bold are measured by novel sensors, i.e., those developed within the ECOSENSE project.

A central feature of the ECOSENSE project is the high spatial resolution of forest measurements, enabled by the deployment of a wide range of sensing technologies. Both commercially available instruments and newly developed sensors were implemented to capture environmental processes across multiple spatial scales. The monitoring system was designed to cover macroscopic (> 10 cm), mesoscopic (< 10 and > 1 cm) and microscopic (< 1 cm) dimensions. Macroscopic sensors were primarily installed on the ground and within tree stems, mesoscopic devices extended measurements into branches and twigs and miniaturized lightweight sensors were attached directly to individual leaves.

In the following, we briefly describe a selection of notable sensors, measurements and systems, from above-canopy to ground-based. We refer to Table S2 for a complete list of all deployed devices, including manufacturers, model numbers, quantities and other useful information, such as associations with larger systems. Devices listed there and mentioned in the text are marked with ^#.

Monthly airborne LiDAR scans conducted via drone overflights generate geospatial point clouds used to monitor the forest's structure (Gassilloud et al., 2025). To further characterize vegetation structure, permanent terrestrial laser scanners^# on the ground provide a daily estimate of the plant area index (Calders et al., 2023). Monthly digital hemispheric photographs^# are used to estimate the LAI (Zhang et al., 2005). Both indices represent the amount of plant surface area – either total or leaf-specific – and play a key role in influencing ecosystem processes while providing insight into a changing vegetation structure. Complementing these structural measurements, canopy and below-canopy phenology is tracked with two phenocams^# mounted on the mixed plot tower (Brown et al., 2016).

The top tower platform also hosts a high-resolution, non-imaging spectrometer system^# for continuous and spectrally resolved measurements of vegetation reflectance and sun-induced chlorophyll-a fluorescence above the canopy. In addition, the same platform supports an above-canopy eddy covariance system^# that continuously measures wind fluctuations, air temperature fluctuations and density changes of CO2 and H2O and calculates ecosystem sensible heat, CO2 and H2O fluxes on-site. Approximately 10 m below, a relaxed eddy accumulation (REA)^# system (Kunz et al., 2025; Sarkar et al., 2020) is installed to measure VOC fluxes in campaign mode, complemented by soil VOC fluxes (Kreuzwieser et al., 2025). At canopy height (approximately 26 m) a network of 26 leaf-angle cameras^# (Kattenborn et al., 2022, 2024; Kremer et al., 2026) monitors leaf movement to investigate radiative transfer, relationships with water supply and related physiological plant stress. Leaf gas exchange, leaf δ13C discrimination and biogenic VOC (BVOC) emissions from 24 leaf enclosures located in the sun and shade canopies of European beech and Douglas fir trees are continuously monitored using a gas sampling system comprising a zero-air supply and automated switching, measurement and flushing units (Werner et al., 2021). Six cavicams^# (Brodribb et al., 2016) measure branch shrinking and swelling, enabling continuous monitoring of xylem water potential. Along the tower profile, a custom-built gas sampling system draws air from 12 inlets at different heights for sequential on-site measurements of CO2 and H2O concentrations. With this data we calculate CO2 storage below and inside the canopy used to correct above-canopy eddy covariance measurements (Foken et al., 2012). Inside the measurement container, the air sampled along the height gradient can also be directed to other analyzers usually connected to the leaf gas exchange system mentioned above, e.g., to a carbon isotope and gas concentration analyzer^# and a proton-transfer-reaction time-of-flight mass spectrometer^# to determine δ13C values and VOC concentrations, respectively. At ground level, eddy covariance measurements are conducted beneath the canopy (Douglas fir and beech plots) (Paul-Limoges et al., 2017). An additional REA system is deployed in the Douglas fir plot. This distributed network enables atmospheric quantification of CO2 and H2O fluxes across different forest strata. In addition to the measurement plots, three intensive soil measurement plots were established near the Douglas fir and mixed towers. At these plots, soil CO2 efflux is estimated using the gradient method (Maier and Schack-Kirchner, 2014), which applies moisture-specific gas diffusivity to determine fluxes at increasing radial distances from three selected tree stems. To continuously monitor soil CO2 dynamics, one of the intensive plots was equipped with low-cost soil CO2 sensors enclosed by specially manufactured gas-permeable membranes, while the other two plots were equipped with established soil CO2 sensors.

Alongside these measurements, an extensive sensor network was deployed across the site to monitor ecosystem functioning at multiple spatial and temporal scales. The most represented sensors at the site include 507 soil moisture probes^# distributed across multiple soil depths to quantify infiltration and water availability, 58 sap flux sensors^# to assess tree-level water use and 20 soil CO2 sensors^# to monitor carbon cycling and related processes such as soil CO2-efflux and water retention. Precipitation is recorded using 20 tipping-bucket gauges^# and 13 air temperature and humidity sensors^# provide reference data and microclimatic/vertical differences.

In addition to the commercial instruments, a suite of newly developed sensors was deployed to expand measurement capabilities and overcome persistent challenges in environmental monitoring (details in the following paragraph). Rainfall redistribution in the forest is measured by 5 throughfall trough systems^# below the canopy and 12 stemflow collar systems^# around selected trees. Spread in a stratified sampling pattern across the ECOSENSE forest, 60 self-developed infiltration samplers measure infiltration of both canopy throughfall and stemflow from the forest floor into the soil (Dedden and Weiler, 2026). A total of 33 self-built dendrometers^# were installed to record stem radial variation and growth dynamics (Dumberger et al., 2025). Furthermore, newly developed leaf-scale sensors^# (Wallrabe et al., 2025) were implemented. These consist of three complementary modules that can function as part of a system but can also be deployed independently: (i) a minimally invasive gas sampling cuvette (ECOvette^#) for CO2, δ13C, H2O and VOC fluxes, connected via tubing to a centralized gas analyzer (Frey et al., 2025b), (ii) a sensor module measuring leaf and ambient temperature (Frey et al., 2025a), relative humidity and PAR (Klüppel et al., 2026) and (iii) a chlorophyll fluorescence sensor (Baghbani et al., 2025, 2026). They are operated by newly developed microcontroller platforms that enable unified control (Shinde et al., 2025) and data transmission via the LoRaWAN network⁺ (Bäumker et al., 2019). There are currently 22 chlorophyll fluorescence sensors^#, 27 ECOvettes^# and another 22 self-built needle-cuvettes^# installed on Douglas fir trees distributed across the forest. All of these are equipped with PAR and temperature sensors^#. Cuvettes additionally measure air humidity inside and outside the leaf enclosure.

The functioning of all above-described environmental sensors is regularly challenged by high variability in relative humidity, intensive rainfall and storms, fog (condensation), air temperatures below 0 °C in the winter months, high air temperatures and UV irradiation during the summer months, tree motion and the presence of animals and insects. Under these conditions, all sensors and electronics need to be robust and suitable for long-term and continuous measurements or regular campaigns. We therefore discuss some challenges encountered and the lessons learned from operating the sensor network in the ECOSENSE forest. To ensure water resistance and enhance long-term durability, sensor boards were coated with Plastic70. The electrical contacts remained uncoated for maintenance and reflashing of microcontrollers. Parylene C was used to prevent gas diffusion of newly developed leaf cuvettes. Within the ECOSENSE project, we additionally work on novel, nanostructured perfluoroacrylate-based coatings preventing biofouling and soiling on sensitive sensor surfaces. The exposure to wind and resulting tree motions are particularly challenging for measurement devices attached to branches and leaves, as it can induce vibrations and sudden displacements at high speeds and accelerations. To mitigate this, sufficient slack was left in the power cables and gas sampling tubes connected to affected sensors, allowing them to absorb wind-induced forces. In addition, predetermined breaking points were incorporated into both cables and tubing to reduce the risk of damage. A range of sensors attached to branches and leaves operate on batteries instead of wind-prone power cables, but this introduces other challenges: batteries are difficult to replace in the canopy and add weight that stresses leaves and twigs, counteracting our lightweight design approach. To enhance overall system resilience, we therefore work towards energy autonomy. Wind, however, not only affects sensors but also our towers. We found that they move with unexpected variability under windy conditions. Therefore, in spring 2025, we added tube insulation as tower padding to minimize the impact of towers on surrounding trees and in particular sample branches. Another source of equipment malfunction arises from various animals present in the forest, such as voles, birds, wild boars, snails and insects. Issues arose, e.g., from insects feeding on leaves equipped with cuvettes and voles gnawing on sensors and cables or nesting inside infiltration sampler systems. Combined measures of increased physical barriers and chemical repellents, such as metal grids in the soil, cable channels with steel wool and chili spray on cable insulation, proved effective. Overall, regular maintenance and repair remain tasks whose time expenditure should not be underestimated and should be considered in project planning.

The research data life cycle of the various, heterogeneous data streams originating from all instruments can be structured into a hierarchical data acquisition framework (Fig. 4). The framework resembles a funnel, starting broad with diverse sensors measuring numerous environmental variables across the ecosystem and gradually narrowing as data are transmitted, unified and processed. Ultimately, all streams converge into centralized servers and databases, ensuring structured, consistent and accessible datasets for scientific analysis and long-term archiving.

The physical layer (1) represents all environmental and ecosystem variables and fluxes measured. Sensors directly measure (2) variables or related proxies and transport the data (3) using different communication protocols (e.g., SDI-12, analog, RS232) to recording devices (4), mostly commercial or self-built data loggers. These devices transmit the data (5) via Ethernet, Wi-Fi, or LoRaWAN to various receivers (6), such as the server in the measurement container or “The Things Network”. The server offers a range of services (see also Table S1), e.g., FTP server (file transfer protocol), HTTP interface (hypertext transfer protocol), MQTT (Message Queueing Telemetry Transport) broker, virtual machines with specialized software, e.g., “Node-RED” for processing. All data streams, with the exception of campaign data, which are typically processed off-site, converge at the FTP server hosted on the field site server. The server functions as a central hub for initial data handling (7), i.e., parsing and preprocessing data with Python scripts into a unified format to ensure consistency across data sources and reduce data volume. Once formatted to fit the database structure, the processed data are uploaded directly from the Python script via a dedicated API call, and transmitted to the database within the university network over a VPN-secured, satellite-based internet connection (8). There, time series data are stored (9) in a specific SQL database (“Aquarius”⁺). Spatial or complex data are temporarily held locally for further processing and made available later on the shared network-attached storage (NAS). The university-based NAS infrastructure functions as a centralized repository, where project members can access (10) data internally – either through file-based access or the SQL database – and interactively explore it using tools such as Grafana (Grafana Labs, 2025, compare Fig. 5). While pure data access is currently restricted to project members, derived data products (see Sect. 4) may already be available. Completed datasets are published in open repositories to ensure open-source accessibility for the wider research community. Meanwhile, the acquired data are actively employed (11) for scientific analyses, including ecosystem model evaluation, model forcing and training of deep learning algorithms (see Sect. 4).

The end of the research data life cycle comprises the structured data publication and archiving, ensuring that data remains Findable, Accessible, Interoperable and Reusable (FAIR, Wilkinson et al., 2016) for future research. This is implemented and supported via the University of Freiburg's publication platform “FreiData” (https://freidata.uni-freiburg.de/, last access: 5 October 2026), which provides storage, registers digital objects and generates their DOIs. For large remote-sensing datasets, such as aerial imagery and LiDAR point clouds that often exceed a volume of several gigabytes, storage and archiving are provided by the domain-specific hubs “deadtrees.earth” (aerial imagery, https://deadtrees.earth/, last access: 5 October 2026) and “3Dtrees.earth” (LiDAR scans, https://3dtrees.earth/, last access: 5 October 2026), which are hosted at the University of Freiburg and supported by the National Infrastructure of the Earth System sciences (NFDI4Earth). These platforms also enable rapid access, interactive web visualization and AI-based data analytics (Möhring et al., 2025; Mosig et al., 2026). In parallel, several data streams recorded in ECOSENSE are ingested in specific databases, including the global PhenoCam Network (https://phenocam.nau.edu/, last access: 5 October 2026) and the European Fluxes Database (https://europe-fluxdata.eu/, last access: 5 October 2026), contributing to broader coordinated environmental monitoring frameworks.

Having defined the layered architecture – from physical measurement through transport, processing, storage, access, quality control and FAIR publication – we now elaborate how this operates from four end-to-end streams from the forest to data publication. We depict four different flows as data are primarily acquired from those distinct sources: (1) commercial sensors which are connected to loggers communicating via Ethernet or Wi-Fi, (2) novel sensor systems, built on low-power microcontrollers transmitting data through wireless LoRaWAN nodes and a LoRaWAN gateway linking to “The Things Network”, (3) scientific instruments producing complex data, such as gradient gas sampling setups and eddy covariance systems that analyze and partially process data on-site using dedicated virtual machines, and (4) campaign-based data collection, where measurements are gathered during field campaigns and stored locally on internal memory devices. These four exemplary data streams, their transmission and application are depicted in Fig. 6.

It is worth noting that the current top-down workflow shown in Figs. 5 and 6 is planned to become a loop. One of the key future directions involves feeding outputs from modeling and deep learning processes back into the field operations. For example, insights gained from pattern recognition could trigger adjustments to a sensor's measurement interval or the number of sensors measuring. This adaptive feedback loop would not only enhance the scientific value of the data but would also improve the efficiency and intelligence of the monitoring setup.

As mentioned above, we aim to ensure the FAIR principles for all data collected. Essential components of FAIR data are the systematic documentation and preservation of metadata associated with all spatial and temporal datasets (Wilkinson et al., 2016). To foster standardized and consistent research data management across all sub-project, the data manager organized a series of workshops covering: (1) an introduction to research data management (RDM) structure, including data collection and transmission; (2) file data upload, including input data parsing, metadata, QA/QC, and time series storage in Aquarius; (3) technical data management, covering FAIR principles, documentation tools and formats, and both commercial and open-access software (e.g., GitLab); (4) geographic information systems, centered on a GIS tutorial introducing users to the ECOSENSE QGIS project; and (5) Aquarius, focusing on the web-based GUI for finding, filtering, visualizing, and downloading data. Acquisition frequencies are set by the respective researchers according to their specific research questions, and in some cases are further constrained by sensor characteristics (e.g., heating phase of sap flow sensors). Rather than homogenizing frequencies across all datasets, we instead rely on Aquarius' processing capabilities to aggregate raw data as needed. For instance, high-frequency precipitation data (1 min resolution) are aggregated into hourly and daily sums. Since measurement frequency requirements vary across modeling approaches and applications, users are expected to further process the data to fit their specific needs. Measurement units are standardized through a parameter and unit list distributed to the consortium at the onset of the project. While this approach worked well for programmable data loggers, other cases required calculation of time series within the database to convert sensor outputs into standardized units (e.g., analog signals from mV, or hPa to kPa).

A key component of the established framework is the assignment of unique sensor IDs that are linked with meta data. Each sensor and its measured time series are accompanied by key contextual information, including spatial coordinates, instrument identifier, data ownership, measured parameter, unit of measurement, sampling interval, observation period and time reference. Due to predetermined and unalterable database architecture, it is not possible to integrate all metadata into the active datasets using a standardized schema. Therefore, upon completion of each time series, metadata are appended in the form of XML or JSON schema files, in accordance with an appropriate metadata standard, i.e., “DataCite” (Brase, 2009). The specific metadata standard is selected in consultation with the Central Data Facility (CDF) of the University of Freiburg and NFDI4Earth, with the goal of ensuring long-term interoperability and broad usability. Our choice of a relational database, compared to alternatives such as HDF5 or ZARR, as the primary data storage system reflects the structure of our time series data, which consist predominantly of tabular sensor time series rather than large multidimensional gridded datasets. However, similar to standards such as the HDF5 group recommendations, we follow the general principle of self-describing, co-located metadata by storing core contextual metadata (as listed above) directly alongside the observations.

In general, the scientific value and (re-)usability of environmental data depend on the data quality. Real-time monitoring enables early detection of anomalies or sensor malfunctions, minimizing data gaps and poor-quality measurements. After data collection, the implementation of quality control procedures is an important aspect within the ECOSENSE project. Rather than removing questionable measurements, nearly all observations are retained and assigned a quality flag on a per-data-point basis. Three categories are applied: good, fair and poor. Data flagged as good have passed certain checks and are deemed reliable for analysis. Data flagged as fair may be plausible but come with uncertainties that might require inspection in further analysis, while poor data should be treated with substantial caution and might not be suitable for subsequent analysis. The only values permanently removed from the dataset are those resulting from major technical failures, where it is certain that no valid measurement was obtained. The type of quality control applied depends on the measured parameter, but common approaches include variance analysis, threshold testing and cross-checking with related variables (e.g., substantial increases in soil moisture must be preceded by precipitation). While some of these processes are automated through scripts, much of the evaluation still relies on expert judgement via visual inspection. Ultimately, the goal is to publish all data with quality flags.

Analysis and interpretation of the wide range of data collected is still ongoing. A current list of publications associated with ECOSENSE can be found on our website (https://uni-freiburg.de/ecosense/publications, last access: 5 October 2026) and by searching for ECOSENSE in FreiDok, the institutional publication repository of the University of Freiburg (https://freidok.uni-freiburg.de, last access: 5 October 2026). After publication, the data used in the respective paper will be made available via FreiData (https://freidata.uni-freiburg.de/communities/ecosense_sfb1537, last access: 5 October 2026).

To illustrate the usage of the data collected at the ECOSENSE forest, we provide a short summary of current publications and manuscripts: Lee et al. (2025) revealed pronounced spatial heterogeneity in forest floor VOC emissions, with strongly enhanced toluene emissions under waterlogged, low-redox conditions, demonstrating the importance of localized emission hotspots for ecosystem VOC budgets. Lee et al. (2026) further showed that tree stems are a highly variable source of VOCs across six major European tree species and, by scaling the measurements with the ECOSENSE forest inventory, quantified the contribution of stem emissions at the ecosystem level. Taking advantage of a variety of tree physiological measurements including our novel in situ leaf cuvette system, Dumberger et al. (2026) investigated inter- and intra-specific differences in leaf gas exchange. They found pronounced differences in seasonal dynamics of stomatal regulation and VOC emissions between the two investigated species and sun and shade leaves, demonstrating the value of data at high temporal and spatial resolution. Making use of extensive soil respiration measurements, our detailed tree species inventory and data on soil physical properties, Brzozon et al. (2026) investigated drivers of soil respiration patterns. They found that temporal patterns were strongly affected by meteorological conditions, while tree species distribution explained some of the spatial heterogeneity. Stock et al. (2026) combined continuous measurements of active chlorophyll fluorescence with eddy flux and meteorological data to investigate the impact of cold spells on forest carbon dynamics. They found species-specific differences of photoinhibition in response to low air temperatures and highlight the value of chlorophyll fluorescence measurements for early stress detection in forest trees.

The collected data are further used to evaluate models and explain ecosystem responses to environmental changes using data and models. Therefore, it is essential to obtain representative ecological and biophysical knowledge from measurements at small scales, e.g., leaf-level, derive course-effect relationships, put these into functional and process-based relationships and apply rules about how triggers and processes are temporally and spatially related, considering bidirectional interactions and feedback loops (see e.g., Mahecha et al., 2010). Both functional relationships as well as scaling laws have been integrated into physiological-oriented, process-based models such as LandscapeDNDC (Grote et al., 2011a; Haas et al., 2013). This dynamic vegetation model has been initialized, applied and will be further developed with data obtained at the ECOSENSE forest. It has been demonstrated that the model covers full carbon-, nitrogen- and water balances as well as growth responses of various forests (Cameron et al., 2013; Grote et al., 2011b; Mahnken et al., 2022). The intriguing aspect of this model is that it not only considers microclimate in high vertical resolution, but is also capable of representing very small soil- (e.g., N2O, Cade et al., 2021) and plant-related fluxes (incl. BVOCs, Havermann et al., 2022). Such models can serve as an instrument to analyze small scale observations and to use this knowledge to estimate larger-scale impacts of particular environmental influences. Thereby, it is possible to evaluate previous knowledge, estimate unmeasured fluxes (gap-filling), test the plausibility of measurements at different scales, judge the importance of specific findings for the ecosystem response and extrapolate responses beyond observed impacts in scenario analyses. At the same time, measurements will reveal mismatches between real-world responses and the model's virtual reality, which is important information to adjust the model's functionality and parameterization. Confirmatory or deviating results could also guide a reduction and/or intensification/redistribution of sensors in order to increase their efficiency.

A complementary approach to the described process modeling is a data-driven analysis of interrelations among measurements. The breakthrough in neural network-based machine learning, typically referred to as deep learning or artificial intelligence (AI), also facilitates using a different form of inputs. When traditional statistical and machine-learning approaches use tabular data, deep learning can mix those with time series, images and even sound and video information (Pichler and Hartig, 2023). However, such flexibility comes at a cost: firstly, the amount of data required to fit deep learning models is substantially larger than for the less flexible traditional regression-like approaches; secondly, the resulting multi-dimensional representation of data in such a network typically remains extremely opaque, despite attempts to deduce the main relationships implicitly encoded (through “explainable AI”: Adadi and Berrada, 2018; Briegel et al., 2020; or symbolic regression: Udrescu and Tegmark, 2020). What we get is a representation of the interrelationship of observed data. Since the neural network is ignorant of biology and physics, some of these relationships will be indirect or spurious and this is where the complementarity with the process-based models arises. We can efficiently explore, for specific environmental conditions, how deep-learning models predict the tree or stand response. These data-driven relationships and hypotheses emerging from them can subsequently be tested with physiology-oriented models. Discrepancies between the two approaches can then be investigated in detail. Is the process correctly depicting our biological understanding for this specific situation? Did our process-based understanding lack cardinal components? Or is the lack of auxiliary data the reason why the neural network makes deviating predictions?

To ground model-based analyses in field data, all sensor systems, infrastructure and trees in the three-hectare core zone of the ECOSENSE experiment are precisely georeferenced using GNSS (Global Navigation Satellite System). In the current setup, the main purpose of georeferencing is to ensure consistent relative co-registration of sensors, trees, infrastructure and repeated 3D observations within the ECOSENSE forest. This enables sensor observations to be linked to individual trees, canopy positions and plots, and provides the spatial basis for tree- and plot-level modeling as well as digital twin and virtual reality (VR) applications. Temporal synchronization is tailored to the relevant process dynamics, typically within seconds to minutes, enabling the precise coupling of rapidly interacting processes such as changes in irradiance and resulting responses in chlorophyll fluorescence or leaf temperature. Regular terrestrial and drone-based laser scans capture the architecture and seasonal dynamics of every tree, creating a detailed and living 3D structural model of the forest. This dynamic 3D representation together with the sensor-derived data streams form the foundation for the ECOSENSE 3D digital twin, developed in collaboration with the XR Future Forests Lab at the university campus. Using advanced engines like Unity (Unity Technologies, 2025) and Unreal (Epic Games, 2025), the 3D digital twin is brought to life in immersive VR, offering realistic and interactive experiences of the ECOSENSE site (Fig. 7). VR opens new dimensions for teaching and outreach. It allows hundreds of students to virtually “walk” through the forest without setting foot on site and can bring ECOSENSE to people around the world. Hidden sensors, buried in the soil or perched high in the canopy, become visible and invisible processes like CO2 fluxes, sap flux and soil-plant-atmosphere interactions are animated in space and time. As the temporal coverage of the ECOSENSE forest grows, the 3D digital twin will also allow users to “time travel” through the site's development and dynamics, enabling an interactive and immersive experience of ecological change across years.

During the GNSS-assisted inventory of the ECOSENSE site, approximately 1400 trees not only received precise coordinates and a unique identifier, but also a durable physical tag featuring a scannable QR code and an alphanumeric ID. These QR codes can be accessed with any smartphone or tablet and link directly to the ECOSENSE database (Fig. 8). This includes structural properties of that tree as obtained from LiDAR campaigns (e.g., crown volume, height, position) and the sensor-based data streams. Moreover, this system transforms the forest into a living, responsive archive: users in the field can not only instantly retrieve data but they can also input data – whether it is recording a mortality event, updating diameter measurements, or documenting sensor maintenance.

For students and researchers alike, the QR code acts as a gateway into the tree's digital history. Linked metrics like height, crown dimensions, or growth rates provide insight into individual development. Moreover, time series from associated sensors, such as sap flux or soil moisture, can be directly accessed and visualized on mobile devices, contextualizing tree responses to environmental conditions. In essence, these QR codes turn each tree into a smart node within the digital twin of ECOSENSE, revealing processes and ecological dynamics in real time or from the past.