Wetlands, defined as areas where the land is saturated or inundated with water and is occupied by plants adapted to water-saturated conditions (Brinson 1993), provide a wide range of ecological functions (Mitsch and Gosselink 2015). Wetlands also serve as habitat for a diverse array of plant and animal species, supporting biodiversity conservation (Davidson et al. 2014). Over the years, wetlands have faced significant degradation due to human activities, as well as climate change (Yang et al. 2023), threatening their capacity to deliver their crucial services. In the past, wetlands provided 40% of the ecosystem services value on Earth (Xu et al. 2020), but this extent has been decreasing continuously for several decades (Yang et al. 2023).

In Central Europe, during the 20^th century, the melioration scheme with the aim of turning the area into agricultural land changed the lowland landscape dramatically. Artificial channels were built to drain the area, and traditional landscape management activities as grazing and mowing of wet meadows were replaced by intensive agriculture. Similarly, in the studied wetland – Čiližská Radvaň, where water regime degradation (decrease of groundwater table, absence of seasonal floodings), together with land-use changes, resulted in a high level of invasion of the degraded wetland habitats. The invasion of non-native plant species is considered as one of the major threats to the natural habitats (Pyšek and Richardson 2010; Rumlerová et al. 2016), including wetlands. Riparian and wetland habitats are among the most invaded ones (Walter et al. 2005; Pyšek et al. 2010; Medvecká et al. 2014) due to several reasons including high propagule pressure caused by floods (Baattrup‐Pedersen et al. 2013). Species composition of riparian habitats is shaped by the combination of environmental factors such as position within alluvia (Slezák et al. 2022; de Simone et al. 2025) and alien species presence (de Simone et al. 2025). Habitats that are closest to the river are usually the most invaded (Hood and Naiman 2000), especially in anthropogenetically modified rivers (Maskell et al. 2006). However, within the same habitat type, areas with good water regime that experience frequent and dynamic floods (Rood et al. 2005) or higher groundwater table are more resistant to invasive species (Petrášová-Šibíková et al. 2017; Mikulová et al. 2020).

Scientific research has highlighted the importance of wetland restoration in combating biodiversity loss and maintaining ecosystem services (Davidson et al. 2014). Wetland restoration projects worldwide aim to recreate or enhance wetland ecosystems that were previously degraded or destroyed (Zedler and Callaway 1999). In the Pannonian region, especially the Danube inundation, several LIFE projects were implemented in the last decades, focusing on the restoration of the water regime and traditional management (e.g., LIFE Dunajské luhy, LIFE Dynamic Life Lines Danube, LIFE Resistance). The studied locality is currently in an ongoing restoration process of water regime improvement and regular mowing. The effect of this revitalization has been monitored since 2023 with the aim of identifying the changes in extent of wetland habitats continuously.

Traditional field-based monitoring is often time-consuming and limited in spatial and temporal extent. The use of remote-sensing techniques in the case of wetland monitoring provides a very promising technique that allows inundation mapping of large areas (Lefebvre et al. 2019; Jussila et al. 2024). In smaller wetlands, drones have emerged as a suitable solution, offering high-resolution, flexible, and cost-effective monitoring capabilities for sensitive marine and wetlands (Ventura et al. 2018; Bhatnagar et al. 2021; Dronova et al. 2021). Drones enable even species-level vegetation mapping using advanced image analysis and machine learning, supporting management of invasive species and habitat conservation (Ruwaimana et al. 2018). However, mapping of habitats is often more complicated due to the fact that the same species can occur in a variety of habitats. To support Natura 2000 habitat classification, the Natural Numerical Networks (NatNet) method was developed by Mikula et al. (2023b). It is a graph-diffusion-type supervised deep learning classification algorithm based on forward-backward diffusion model and its numerical discretization. Earlier, the NatNet has been successfully applied in the classification of Natura 2000 forest habitats using Sentinel-2 satellite data (Mikula et al. 2023b). However, wetland habitats are typically smaller in scale and require data with higher spatial resolution. In the present study, RGB drone imagery was used for the first time to apply the NatNet method in the wetland context. Although there exist multispectral and even hyperspectral drones, we used only RGB drones and photogrammetry to generate orthophoto images and the digital surface model (DSM), mainly due to the reason of general usage of developed methodology by stakeholders. The method is implemented within NaturaSat, a software developed to support geobotanical research through the application of modern mathematical approaches.

The main aims of the present study are: (1) to evaluate the use of the Natural Numerical Network for the automatic identification of wetland habitats from RGB drone imagery, and (2) to assess the success of revitalization by analysing changes in habitat extent between 2023 and 2024 in a recently restored wetland.

The first objective of this study was to train the Natural Numerical Network for the classification and identification of wetland habitats in the Čiližská Radvaň area, in the Sloval Republic. Three distinct wetland habitats were selected for the study, each represented by 30 representative squares. These training squares ensure comprehensive coverage of the spectral variability in each habitat cluster, thereby enhancing the model’s ability to distinguish between different habitat types. The training phase of the NatNet resulted in a high classification accuracy, achieving a success rate of 97.34%. This level of accuracy means that the algorithm was able to correctly classify most of the data points, with only four misclassified points. Notably, misclassifications occurred between the Typha angustifolia and Typha-Phragmites habitats, as well as between the Typha-Phragmites and Phragmites australis habitats. These habitats share dominant species, and the differences are mainly in the proportion of species covered. These misclassifications likely reflect overlapping characteristics of vegetation in ecotonal areas where spectral similarities disable the algorithm to find sharp distinctions.

Despite the existence of these few misclassified points, the overall success rate of 97.34% is considered sufficient for NatNet to be applied to habitat classification. The success rate was obtained with the set of parameters where K₁ = 3800, K₂ = 4300, and δ = 0.002 (see also the Suppl. material 1 on NatNet mathematical details). The number of K_i-s in the diffusion coefficient (2) was reduced to two by applying a dimension reduction approach using the Principal Component Analysis (PCA), where only the first two principal components were utilized. PCA was used to catch the maximum variance in the data, thereby enabling a transformation of the coordinate system considering the direction of the maximal variance. The use of classification results and the construction of relevance maps provide valuable information on the spatial distribution of wetland habitats, allowing more precise monitoring of habitat area development in the future.

In the Čiližská Radvaň wetland, the major vegetation type is Phragmition communis (Lk11 – reed beds) with two associations, Phragmitetum vulgaris and Typhetum angustifoliae and transitional parts where Phragmites australis and Typha angustifolia occur together. Typha, as well as Phragmites species, are highly characteristic for wetland environments due to their ability to thrive in waterlogged soils and shallow waters. The robust root system of Typha helps in sediment trapping and nutrient cycling, making it a vital component of wetland ecosystems. Fig. 6 shows RGB drone imagery of Čiližská Radvaň wetland and relevancy maps for NatNet trained on November 2023 data (Fig. 6A, C, E, G) and on October 2024 data (Figs 6B, D, F, H), in rows always for one of the analysed habitats. In both columns, the results for Phragmites australis, Typha-Phragmites, and Typha angustifolia (with and without considering canopy height) are shown.

The pure Phragmites habitat in the Čiližská Radvaň wetland is a relatively rare habitat, with only a few locations identified. It occurs mainly along the water channel, on places with the highest water levels in the wetland. It occurs in shallow water and along the edges of bodies of water, contributing to the ecological health of the wetland. Despite this limited presence, the relevancy map for the Phragmites habitat effectively captures these areas, even identifying some regions that were not initially segmented (Fig. 6A, B). This highlights the sensitivity of the Natural Numerical Network in detecting this habitat, which is very important due to the presence of species in harder accessible terrain. The relevancy map also shows a notable transition between the Typha angustifolia habitat and the pure Phragmites habitat. This transition is visualised as black regions in the Phragmites polygons and white colour of the corresponding part on the Typha angustifolia relevancy map (Fig. 6E, F, G, H).

In our study, the Typha-Phragmites dominated habitat is a specific one due to its unique composition, representing a mixture of Typha angustifolia and Phragmites australis, two species commonly associated with wetland ecosystems. This mixed habitat is situated along the old channel, and also in broader surroundings, where the ecological conditions support both species (Fig. 6C, D). These two species frequently coexist in locations with oscillating water table, creating dense vegetation stands. However, the Natural Numerical Network can detect slight optical differences in the Typha-Phragmites mixture, enabling it to represent this habitat on the relevancy map accurately. In this representation, the interior of the Typha-Phragmites polygon (blue polygon) is white, indicating a high relevancy coefficient and a confident classification of this mixed habitat type.

The relevancy map for the Typha-dominated habitat (Fig. 6E, F) in this wetland illustrates a high degree of relevancy in the Typha angustifolia polygons (green polygons), which is visualised by almost all white coloration. This white colour indicates high relevancy of the presence of Typha species, confirming that the spectral signature closely aligns with this habitat type. Small areas in the polygons show zero relevancy in the relevancy map of the Typha habitat, reflecting the transitional zone between Typha angustifolia and Typha-Phragmites habitats. The relevancy map in Fig. 6E also shows regions of white colour, which extend beyond the boundaries of the Typha angustifolia polygons included in the training set. However, following a comprehensive inspection of the area with high relevancy on the map, we concluded that a larger presence of Typha angustifolia is only partially correct and that the NatNet model trained on spectral RGB data produced for Typha angustifolia false positive results in certain locations.

The spectral characteristics of the Typha habitat exhibit similarities to those of the invasive species in the surrounding area, which leads to the tendency of the spectral model to overestimate Typha relevancy, highlighting the need to refine the approach. Nevertheless, a notable distinction is observed in the height of the canopy, which can thus be used as a differentiating factor. Thus, we incorporated additional structural information such as vegetation height. It was calculated using the difference of DSM and DTM models (Suppl. material 1: fig. S5) and helped us to reduce these overestimations and better delineate Typha stands from neighbouring habitats.

An estimated threshold value was determined for the minimum height of the Typha canopy, and this information was combined with the relevancy map in Fig. 6E, F. The resulting map, presented in Fig. 6G, H, indicates a significant and correct reduction in the higher values of relevancy. The integration of such data enhances the ecological accuracy of the model, thereby improving the effectiveness of the Natural Numerical Network in detecting and monitoring these typical wetland species.

The second aim of our work was to compare the extent of wetland habitats after approximately one year since the initial monitoring. To accomplish a comparative analysis for the assessment of the ecological changes and evaluate the impact of restoration and management efforts on the wetland over one year, we leveraged high-resolution drone imagery collected in October 2024, along with detailed habitat relevancy maps generated from this imagery (Fig. 6B, D, F, H). These maps were calculated using the NatNet trained on representative squares created in November 2023, but using the statistical characteristics inside the squares calculated from October 2024 data. New training of the network is natural because drone images captured on different dates often suffer from inconsistent illumination, shadows, or atmospheric effects (like haze), which is important in classification tasks when using RGB imagery. The newly trained NatNet success rate was equal to 100%. Such a high success rate was obtained with a set of parameters K₁ = 2800, K₂ = 2500, and δ = 0.002. That optimal NatNet was used to obtain the relevancy maps for the habitats in October 2024.

To validate these relevancy maps, a systematic approach was applied using randomly generated points in the high-relevancy (white colour) as well as in low-relevancy (black colour) areas of the map for each habitat. Field surveys were then conducted to assess the accuracy of these classifications. During the field work on November 2^nd 2024, 58 validation points randomly generated in high and low relevancy areas were checked, and 7 of them were found to be inaccurate. Table 1 summarises the outcomes of the validation procedure applied to relevancy maps, including confusion matrices, and accuracy, precision, recall, and F1 scores for each habitat. For Typha, high-relevancy areas yielded 14 correctly classified versus 3 incorrectly classified points, while low-relevancy areas achieved full classification success (10/10). Field research indicates that misclassification occurs between the Typha species and invasive species due to their similar spectral characteristics when invasive species, as well as Typha, are not flowering. Within the Typha–Phragmites habitat, both high (9/9) and low-relevancy (6/6) points were classified without error. By contrast, Phragmites high-relevancy areas exhibited a higher misclassification rate (7 correct vs. 4 incorrect), whereas low-relevancy areas again achieved full accuracy (5/5). The incorrectly classified points were between the Phragmites canopy and the invasive species. Overall, when considering all habitat types, classification accuracy was 0.88 and F1 score 0.90. In Table 2, we present the same metrics – Accuracy, Precision, Recall, and F1 scores – for the classification of all habitats.

The study of the wetland habitats using relevancy maps resulting from NatNet classification is an effective tool for monitoring changes in habitat extent. We were able to observe a shift in vegetation type occurrence after one year of revitalisation activities, which could be hardly possible by using only field research and permanent plots. In Fig. 7A, C, we present three cases of expansion of the Typha-Phragmites transitional habitat indicated by the relevancy maps and documented by field survey and photographs of the place. The left column figures show details of the October 2024 orthophotos of Čiližská Radvaň wetland along the old channel. In the middle column, there are details of relevancy maps for the Typha–Phragmites habitat. The photos in the right column show the view from the red-marked location where high-relevancy pixels extend beyond the 2023 boundaries, indicating a measurable increase in Typha–Phragmites cover.

A quantitative comparison of the 2023 and 2024 relevancy maps indicates apparent shifts in vegetation types (Table 3). Pure Phragmites australis canopies were mapped over a larger area in 2024 than in 2023. As shown in Table 3, the area of high relevancy has nearly doubled. This expansion is likely linked to higher water levels in the canal due to the revitalization activities. Since Phragmites typically thrives in the wettest parts of wetlands, this development can be considered a natural response to changing hydrological conditions.

Table 3 also shows a modest increase in the mixed Typha–Phragmites habitat. The growth is limited, however, because water-level fluctuations near the canal tend to suppress this mixed community, allowing pure Phragmites to dominate in those areas. As illustrated in Fig. 6, Typha–Phragmites stands have expanded mainly in the central part of the study region. The fourth column in Table 3 shows that the area covered by Typha has decreased. This decline appears to result from former Typha-dominated areas gradually transitioning to mixed Typha–Phragmites habitat.

Both dominant species profit from higher water levels in the locality after revitalization, which is also visible from the expansion of the Typha–Phragmites habitat in the central part of the wetland, where this community is slowly replacing areas dominated by Calamagrostis epigeios and invasive species. Typha angustifolia (and association Typhetum angustifoliae) prefers more stable waterlogged conditions (Oťaheľová 2001) in comparison with Phragmites australis (association Phragmitetum vulgaris). Phragmites australis can also grow in dynamically changing conditions, which are currently occurring at the revitalized site. This is reflected in the significant increase in the total area occupied by Phragmites habitat and the rise of mixed Typha-Phragmites habitats at the expense of pure Typha habitat.

Our results provide spatially explicit evidence that revitalisation measures, particularly hydrological improvements and recurrent mowing, have facilitated an increase of Typha–Phragmites and Phragmites-dominated vegetation and that the drone–NatNet workflow reliably captures these changes. We confirmed that the relevancy map approach not only delineates existing habitat boundaries but also supports quantitative, year-to-year monitoring of wetland restoration outcomes.

Although the study focuses on a single area, namely the Čiližská Radvaň wetland in the Slovak Republic, and on one year of monitoring, this concentrated approach allowed for a detailed evaluation of habitat dynamics during the early phase of restoration. The area of the wetland is not large-scale, but for the purposes of revitalization monitoring we found the extent suitable, since the restoration process is challenging and is usually spatially limited to smaller areas. One year of monitoring can capture initial stages of vegetation change after revitalisation, and these trends could be different in the next years, depending on the continuous succession and other factors. Our results cannot predict the wetland development in the future, but can serve as a methodological basis for the next monitoring since we proved the possibility to identify single vegetation associations and shifts between them.

While the limited spatial and temporal extent naturally constrains broader generalisation, our study provides a solid and well-documented baseline for future multi-year and multi-site comparisons. Described methods could be transferred to other wetlands with similar habitats; however, local phenology must be considered when choosing the dates for drone data sampling.

All authors were involved in conceptualization and study design and methodology creation. K. Mikula led the project and M. Šibíková co-led the project. K.Mikula and A.A. Ožvat designed the NatNet classification algorithm for drone data, A.A. Ožvat and M. Kollár implemented NatNet in NaturaSat software and K. Mikula, A.A. Ožvat and M. Kollár performed all data analyses in NaturaSat software. M. Šibíková and J. Šibík were responsible for field sampling, GPS measurements, and field validation. J. Papčo and J. Sigmund were responsible for drone flights and drone data processing. All authors participated on the interpretation of results, text writing, and preparation of Figures.