Assessing the effectiveness of UAV data for accurate coastal dune habitat mapping

ABSTRACT Coastal dunes are considered some of the most threatened and vulnerable habitats in the European Union. Mapping the spatial distribution of these habitats is an essential task for their conservation. Advances in Unoccupied Aerial Vehicles (UAVs) facilitate the flexible acquisition of high-resolution imagery for identifying detailed spatial distributions of habitats within dune systems. This study aimed to assess the effectiveness of UAV remote sensing for mapping these habitat types. Specifically, we determined the impact of temporally acquired UAV-derived spectral and topographic information on classification accuracy. The work combined the multi-temporal UAV imagery with field observation data and used the Random Forest machine learning algorithm to classify dune habitats. Results showed that using multi-temporal UAV imagery increased classification accuracy compared to using uni-temporal UAV imagery (92.37% vs. 84.09%, respectively). Also, including topographic information consistently improved accuracy, regardless of the number of image sets used (the highest accuracy increased from 84.81% to 92.57% for a uni-temporal model). Temporal analyses showed that the data acquired in the middle period of the growing season were better than those acquired in the early or late periods. The methodology presented here demonstrates the potential of using UAV data for detailed mapping and monitoring of habitat types.


Introduction
The European coastline is frequently bordered by coastal dunes.They form a spatially narrow yet complex mosaic of dynamic habitats and are characterized by high floral diversity (Doody, 2013).They also provide important ecosystem services including coastal defense, maintenance of wildlife, water catchment, carbon sequestration and recreation (Barbier et al., 2011).
Coastal dunes are considered a highly dynamic environment.Formed by sand-sized sediment that is transported inland by onshore winds and trapped by vegetation, they exhibit a zonation pattern extending inland parallel to the beach (Bird, 2008(Bird, , 2010;;Hughes et al., 2014).Along beaches, the high tide line is marked by the deposition of a line of detritus, including seaweeds, and in this area vascular plants can establish.Around these drift lines, wind-blown sand can accumulate further, initiating the formation of the embryonic dunes which are unstable low mounds of sand, generally <1 m in height.The exposed sand of these mounds becomes vegetated by salt-tolerant species, such as sand couch (Elytrigia juncea) and lyme-grass (Leymus arenarius).As these dunes accumulate sand, their surface becomes less salty.This environment is where marram grass (Ammophila arenaria) begins to colonize.Marram grass has an extensive root system that stabilizes the sand and allows the building of larger dunes.These marram dunes, or white dunes, are partially stabilized hills that are located farther inland than embryonic dunes and can reach 15-20 m in height (Fossitt, 2000).The next phase in succession is that of fixed dunes, otherwise known as grey dunes, which, compared with marram dunes, provide a more stabilized and sheltered environment.In this zone, the vegetation cover is usually characterized by herb-rich grassland and heath communities.Within low-lying, seasonally flooded areas in a dune system, dune slacks can also develop.These nutrient-enriched areas can support diverse floristic and plant communities.Dune scrub/woodland may be found in more generally stable parts of a system (Delaney et al., 2013).
In the European Union (EU), dune habitats are included in Annex I of the "Council Directive 92/43/ EEC on the conservation of natural habitats and wild fauna and flora", also known as the Habitats Directive (HD).The HD is an EU legal framework that aims to protect a wide range of rare, endangered, and vulnerable habitats (those listed in Annex I) by requiring each Member State to monitor these habitats within their territories and report this to the European Commission on a six-yearly basis (Commission of the European Communities, 1992).
Within Ireland, eight Annex I coastal dune habitats are identified, and two other habitats are also often found in association with dune systems (Table 1).These dynamic habitats are vulnerable to natural hazards and human pressures, which can lead to habitat degradation and biodiversity loss (Delaney et al., 2013).In 2019, these ten Annex I coastal habitats were each assessed as having an Unfavourable-Inadequate or Unfavourable-Bad status (National Parks and Wildlife Service, 2019); these assessments were largely unchanged from the previous two HD reporting cycles in 2008and 2013(National Parks and Wildlife Service, 2008, 2013b).Two negative pressures associated with human activities in Irish dune systems are the use of vehicles and trampling, both of which can create networks of paths and tracks, resulting in damage to vegetation and increased dune instability (Delaney et al., 2013).Assessing and monitoring these coastal dunes is critical to identifying key conservation issues, measuring threat status, and developing and evaluating management and restoration programs (European Commission, 2022).
Habitat mapping is a fundamental task for generating the data required for monitoring habitats under the HD.Most habitat mapping is conducted by ecologists in the field.While these surveys can provide detailed information, manual methods alone can limit the size and temporal scope of the survey, primarily due to cost (Mack et al., 1997).Remote sensing offers an advantage over traditional field-based mapping techniques as it may capture data more frequently and across larger areas (Jensen, 2014;Vanden Borre et al., 2011).However, monitoring coastal dune systems with freely available medium-resolution (>10 m) satellite imagery can be difficult, for example when mapping very narrow areas of patchy mobile dunes or identifying small-sized disturbed areas within the habitats caused by human tracks and coastal erosion (Ettritch et al., 2018;Marzialetti et al., 2019).For this reason, the classification output from this imagery can result in inaccurate spatial and botanical information about dune status; higher resolution data is needed to capture this detailed information.
Recent advancements in remote sensing platforms and sensors, such as the development of Unoccupied Aerial Vehicles (UAVs), have enabled high-resolution mapping of habitats and vegetation (Gonçalves et al., 2016;Kattenborn, Eichel, et al., 2019;Michez et al., 2016;Müllerová et al., 2021;van Iersel et al., 2018).UAVs can acquire centimeter-level resolution imagery.They are also characterized by high efficiency and flexibility in image acquisition of small (<20 ha) areas because they generally have low operational costs; this enables a high frequency of image acquisition that is suitable for environmental monitoring (Manfreda et al., 2018;Pajares, 2015;Yan et al., 2009).
In recent years, several UAV-based studies have implemented machine learning algorithms for image classification and habitat mapping due to the excellent classification results and fast processing time (Han et al., 2019;Lu & He, 2017;van Iersel et al., 2018).Random Forest (RF) is one of the most widely used algorithms (Belgiu & Drăgut, 2016;Pal, 2005).RF (Breiman, 2001) is a nonparametric, supervised machine learning algorithm that uses a collection of decision trees to solve  Commission, 2013;Fossitt, 2000).a classification task.Each tree in an RF model generates a prediction and the final prediction is determined by majority voting (Fawagreh et al., 2014).It allows large volumes of highly correlated variables to be ingested and processed without data redundancy affecting the model performance.The efficiency of RF in processing large datasets makes it appropriate for multi-temporal analysis of UAV images.The RF algorithm was selected as the preferred classification method for the present study due to the reported high accuracy in vegetation mapping using various remote sensing data (Amani et al., 2017;Barrett et al., 2016;Raab et al., 2018;Villoslada et al., 2020).
In habitat mapping, image acquisition timing is an important factor since plant communities within a habitat may be more distinct from the surrounding vegetation at a particular time of the growing season (Marcinkowska-Ochtyra et al., 2019).The use of multi-temporal images in habitat mapping can increase classification accuracy compared to maps produced from an image collected at a single time (Marcinkowska-Ochtyra et al., 2019;Schmidt et al., 2014;van Iersel et al., 2018).Several studies have used multi-temporal UAV images in mapping habitats and vegetation (Belcore et al., 2021;Michez et al., 2016;van Iersel et al., 2018); however, very few studies have been carried out on coastal dune systems (Taddia et al., 2019).
This study aimed to assess the effectiveness of UAV remote sensing for mapping complex coastal dune habitat types.More specifically, we aimed to (1) determine how classification accuracy varies when using multi-temporal images acquired throughout the growing season, (2) identify if and how the inclusion of UAV-derived topography improves the classification accuracy, and (3) identify the optimum period of the growing season for UAV data acquisition for the purposes of differentiating coastal dune habitats.These objectives were achieved by integrating highresolution UAV imagery with data from field surveys.The different habitat types were classified using the RF algorithm.

Study site description
The study area is a 37-ha coastal dune site located on the Magharees Peninsula (52°16ʹ21ʺN, 10°2ʹ15ʺW) County Kerry, Ireland (Figure 1).The Magharees Peninsula is a sand spit, about 5 km long, with an extensive area of fixed grey dunes containing several dune slacks (Delaney et al., 2013;National Parks and Wildlife Service, 2013a).This dune site is designated as part of the large Tralee Bay and Magharees Peninsula, West to Cloghane Special Area of Conservation (SAC code 002070, with an area of approximately 11,627 ha) (Figure 1), part of the Natura 2000 network (National Parks and Wildlife Service, 2018).
Six HD Annex I habitats are present within the study area (Table 1 and Figure 2) -annual vegetation of drift lines (1210), embryonic shifting dunes (2110), white dunes (2120), grey dunes (2130), dunes with Salix repens (2170), and dune slacks (2190).One of these, grey dunes, is a priority habitat that requires particular protection as its distribution is in danger of disappearance within the EU territory (National Parks and Wildlife Service, 2019).For botanical information on these six habitats at the study site, see Supplementary Material 1.
Other common habitats present within the study area, which are not listed in Annex I of the HD, were categorized based on the Fossitt classification scheme (Table 1 and Figure 2).The Fossitt scheme is a standard guide for identifying and describing habitats in Ireland (Fossitt, 2000).It is important to determine the locations and sizes of these habitats as some represent disturbance to the dune system (e.g.exposed sand, recolonizing bare ground, and the presence of nonnative species).

General workflow
A replicable workflow using multi-temporal highresolution UAV imagery and field data acquired during the vegetation growing season was used to characterize different habitat types.Several raster layers were generated from the UAV-derived photogrammetric products, including a raster orthomosaic and a Digital Surface Model (DSM).These layers, together with the field data as reference samples, were used to train the RF-based classification of each target habitat type.The trained model was then applied to the image mosaic to produce a spatially continuous map of habitats within the study area.These processes are illustrated in a flow diagram (Figure 3).Each step in the diagram is discussed in detail in the following subsections.
Step 1. Field Surveys Image acquisition.Field campaigns were conducted during the growing season in 2020: 26 May (M, early), 28 July (J, mid), and 15 October (O, late), under suitable flying conditions (i.e.clear sky with low wind speed).High-resolution imagery (6 cm) was acquired using a DJI Phantom 4 Multispectral UAV sensor (https://www.dji.com/p4-multispectral/specs) with five spectral bands -blue (450 nm), green (560 nm), red (650 nm), red edge (RE) (730 nm) and nearinfrared (NIR) (840 nm).Five Ground Control Points (GCPs) were evenly distributed throughout the study area to spatially reference the UAV-acquired images, following recommendations by Assmann et al. (2019) and Mesas-Carrascosa et al. (2015).The coordinates at the center of each GCP were measured and recorded using an Emlid Reach RS+ Global Navigation Satellite System (GNSS) device (https://emlid.com/reachrs/)with the observation time set to two minutes.All three flights were operated with a side and forward overlap of 75% and 80% respectively, and a flight altitude of 60 m.Invariant panels were also placed within the study area and imaged during the flights.These were subsequently used for radiometric calibration.These panels were used to normalize for variations in atmosphere and solar illumination across the three image acquisition dates.Reference data collection.Extensive ecological field surveys were carried out by experienced ecologists to collect ground truth data (i.e.reference points representing the actual location of a habitat type) to train the model and validate the results.The ecologists used the GNSS device to record the coordinates of the data points.Each point was located at the center of an area of relatively homogeneous habitat cover.The ecologists recorded the radius of this area around each point, typically 1-10 m, based on their expert opinion of its spatial extent (Figure 4a).Each point also included a record of habitat codes based on the Annex I and Fossitt classification systems and was photographed in the north-view direction.
First, the team collected reference points in the week following the first UAV survey to prevent disturbance to the site.Further site visits that were coordinated with the second and third UAV surveys facilitated the addition of more points and assessment of any significant vegetation changes from the previous survey.After verifying that there was no change in the habitat types that were recorded within the growing season (i.e. the habitat identification on the ground in the first visit remained correct in subsequent visits), the point data collected from all field campaigns were combined into a single set (n ¼ 418) and used to classify all acquired UAV images.The point data were converted into polygons using the radii recorded in the field (Figure 4b).Both the remote sensing analyst and the ecologists visually inspected the collected field data (i.e.polygons) on the screen to ensure that there was no overlap and to verify that they were aligned with the correct habitats as seen on the UAV image (Figure 4c).Due to time constraints related to field-based data collection, 198 additional reference polygons were created in a GIS environment to supplement those classes with insufficient field-collected samples.This process comprised manual and subjective analysis of the UAV-acquired images by the ecologists who had conducted the field survey and were thus familiar with the site and habitats.

Image bands transformation
Principal Component 1 (PC1) First principal component that accounts for the greatest variance in the dataset (Jolliffe, 2002) Topographic variables

Elevation
Represents the height of the terrain.

Slope
Represents the rate of change of elevation.
features or variables that provide more information to help discriminate between classes.Sixteen raster layers were generated from the photogrammetric products for each field campaign.These layers were grouped into two datasets comprising spectral and topographic variables (Table 2).The spectral variables were derived from the orthomosaic and consist of five radiometrically calibrated surface reflectance bands, the first Principal Component Analysis (PCA) band and eight vegetation indices.In contrast, topographic variables such as elevation and slope were derived through morphometric analysis of the DSM.Reflectance bands were used as they can provide spectral information in the visible, RE and NIR regions of the image.The eight vegetation indices derived for this study have been found to have good performance in assessing vegetation cover, estimating chlorophyll content in densely and sparsely vegetated areas, estimating vegetation water content and providing evidence of correlation with biomass (Gitelson et al., 1996;Huete et al., 2002;Rouse et al., 1973;Xue & Su, 2017).The PCA technique was utilized as it reduces data dimensionality by creating new uncorrelated variables (i.e.Principal Components) that contain the most important information from the data as represented by eigenvalues (Jolliffe, 2002).Here, the first PC band, having an eigenvalue of more than 75%, was extracted from the reflectance bands (see Supplementary Table S2).Finally, information on elevation and slope generally supports the dune habitat classification as they can be determinants for vegetation composition and distribution.
The reference data polygons (various sizes) were used to extract pixels from each variable in Table 2.This study used a 17-pixel non-overlapping sliding kernel (equivalent to 1 � 1 m on the ground) to divide each reference polygon into several equal-sized subpolygons.The mean of the pixel values within each sub-polygon was calculated for every raster layer.The resulting mean pixel values and the corresponding polygon labels constituted the final reference dataset.This approach of a sliding averaging kernel reduces issues around reflectance-based adjacency effects and erroneous values due to geometric misalignment when using individual pixel values.

Step 3. Modeling and Validation
Several classification models were generated using the supervised RF algorithm to investigate the effect of vegetation phenology and topographic information on classification accuracy.The input for each model varied depending on combinations of variables (spectral variables only -RF S , spectral and topographic variables -RF ST ) and the number of field campaigns.In total, there were fourteen data combinations subjected to the RF algorithm (Table 3).In this study, the RF was implemented in Python 3.7 using the scikitlearn library (Pedregosa et al., 2011).The parameters used for each model were M var ¼ ffi ffi ffi ffi ffi M p for the number of variables at each split, where M is the total number of variables, and N tree ¼ 500 trees (Gislason et al., 2006).The relative importance of variables was analyzed based on Mean Decrease in Impurity (Breiman et al., 1984).
The RF technique was applied to each dataset to classify four of the coastal dune Annex I habitatswhite dunes (2120), grey dunes (2130), dunes with Salix repens (2170) and dune slacks (2190)-and five Fossitt habitats.The other two Annex I habitats recorded-annual vegetation of drift lines (1210) and embryonic shifting dunes (2110)-were not included in the classification as they occurred in small, isolated patches and only 13 reference points were collected.Including them would have resulted in a severely imbalanced training set and, therefore, a biased model (Khoshgoftaar et al., 2007).
To evaluate the model performance, the accuracy was estimated using k-fold cross-validation.This method estimates model performance by randomly splitting up the data into k number of divisions or S -spectral variables ST -spectral and topographic variables folds, wherein each fold is held out as the test set, and the other k À 1 folds are used as the training sets (Berrar, 2019).Five folds were used here, i.e. the model training and validation process was repeated five times.The average and standard error values were calculated from each 5-fold model.The final model to predict the class labels of the entire study area used all the training data.
The model performance for each class was also evaluated using confusion matrices and estimating derived evaluation metrics (Equations 1-3) known as precision, recall, and F1-score (Muller & Guido, 2017).Precision is the fraction of correct predictions for each class and recall is the fraction of true labels that were correctly classified.F1-score is a single value that reports the overall accuracy, considering both recall and precision.It is the harmonic mean (weighted) of recall and precision values.These three metrics were calculated as follows: (1) (2)

Temporal analysis of coastal dune habitats
Two tests were prepared to determine which time of year is optimal, in terms of classification accuracy, for image acquisition for coastal dune habitat mapping (Figure 5).The first test (Test 1) was applied to the M_RF ST , J_RF ST and O_RF ST training models produced above.For each of these models, data from the image dates that were not used in training were combined to generate the test dataset.So, the M_RF ST model was evaluated with the combined July and October data.The generated test dataset was randomly resampled 1000 times to create 1000 different test datasets.In other words, the M_RF ST model was evaluated using 1000 test datasets, which generated 1000 accuracy values.Next, the J_RF ST model was evaluated using the combination of the May and October data.
Finally, the O_RF ST model was evaluated using the combined May and July data.Therefore, each of these models has 1000 accuracy values.For the second test (Test 2), each model was reproduced but this time with only 80% of the data (sub-polygons) .The remaining 20% of the data from each of the three campaigns were pooled and used to generate the test dataset.In this way, each model was assessed using the same test dataset.Similar to Test 1, the test dataset was randomly resampled 1000 times to create 1000 different test datasets.These 1000 test datasets were used to evaluate each Test 2 model.The mean and standard error were computed for each of the models from Test 1 and Test 2.

Acquisitions and variables used in the classification
The results obtained from 5-fold validation are presented in Table 4.The mean accuracy values of all models were above 80%, with MJO_RF ST showing the best performance (94.26%) and O_RF S the least good (84.09%).The computed standard error (SE) showed that mean accuracy based on k ¼ 5 had low variability across the data.The models that used only spectral variables (RF S ) achieved higher accuracy when combining data from two or more field campaigns.This pattern was also seen in the results from the models using both spectral and topographic variables (RF ST ) with just one exception: JO_RF ST, which had a slightly lower accuracy (92.54%) than J_RF ST (92.57%).
All the models that used both the spectral and topographic variables (RF ST ) had a mean accuracy that exceeded 90%, with the highest accuracy acquired,   which was for MJO_RF ST , being 94.26% and the lowest, which was for M_RF ST , being 91.35%.When comparing the RF ST models with the corresponding RF S models, the addition of topographic data always improved mean accuracy, which was also supported by the variable importance ranking (see Supplementary Table S3).The accuracy increase ranged from 2-8%, depending on the number of image sets considered in the model.The increase in accuracy for single image models was greater (M, +6.77%; J, +7.76%; O, +7.59%), than for the multiple image models (MJ, +5.01%; MO, +3.08%; JO, +2.78%; MJO, +1.89%).In other words, the more image sets considered in the model, the smaller the benefit of including the topographic variables.Moreover, analysis of the variable importance of uni-temporal RF ST models indicates that, aside from the topographic variables, NDWI and GNDVI were the two most important variables common to these models.In contrast, NIR and RE bands were among the variables of low importance (see Supplementary Table S3).
The model performance for each habitat differed between datasets, as presented by the confusion matrices, recall, precision, and F1-score values (Figure 6 and Table 5).Overall, all models were able to predict most of the four Annex I dune habitats correctly (F1-score >0.75).Misclassifications among Annex I habitats primarily occurred between dunes with Salix repens (2170) and dune slacks (2190) (Figure 6).White dunes (2120) were also misclassified as ED1 and ED3, probably due to large areas of exposed sand in all three habitats.This result for white dunes (2120) was consistent with its slightly lower precision values (Table 5).
When topographic variables are included in the models, the F1-scores increase for most of the habitat types, with some classes reaching F1-scores >0.90 (Table 5).In general, both marine littoral sediments, LS1 and LS2, showed the highest F1-score values at 0.96 and above (Table 5).In contrast, degraded versions of Annex I habitats-bare sand (ED1) and recolonizing bare ground (ED3)-had the lowest F1-scores for all the models, specifically due to low recall values (Table 5).
During the growing season, different months were identified as optimal for discriminating between the different habitats.For example, using the RF ST models, both white dunes (2120) and dunes with Salix repens (2170) can be best differentiated in July as the F1-scores of 0.91 and 0.90, respectively, are slightly higher than the F1-scores for May and October (Table 5).In contrast, October was the best time to identify grey dunes (2130).Finally, dune slacks (2190) achieved the same F1-score (0.92) across all acquisition dates (Table 5).

Temporal analysis of dune habitats identification
The RF-based models trained with data from different months to the test data (Test 1) each achieved mean accuracies of above 70% (Table 6).However, the accuracy of the model trained on the July data was higher than that of the models trained on the May or October data.The results of Test 2 followed the same trend as those of Test 1.The J_RF ST model had the highest performance, followed by O_RF ST and M_RF ST (Table 7).These results suggest that the best month for image acquisition was found to be the middle of the growing season for coastal dune vegetation.The mean accuracies obtained from all models evaluated using Test 1 data ranged from 0.74 to 0.81, which was slightly lower than the accuracy range when Test 2 was used (from 0.78 to 0.86) (Tables 6 and 7).

Distribution and extent of dune habitat types
The computed area of the different habitat types, as classified by RF models based on three single image acquisition times, and the spatial classification output for July are shown in Figure 7.The study site was dominated by grey dunes (2130), which covered almost 60% of the total area.This was followed by dunes with Salix repens (2170) and dune slacks (2190), which had about 12% − 15% each of the area.White dunes (2120) covered around 5% − 6%.All Fossitt habitats (i.e.dune scrub and woodland, exposed sand, recolonizing bare ground, shingle and gravel shores, and sandy shores) corresponded to the smallest area proportions within the study area and collectively accounted for less than 10% of the total area.The proportions of habitats did not differ markedly between models.Dune slacks (2190) were surrounded mainly by dunes with Salix repens (2170).White dunes (2120) were generally located parallel to the shore, alongside LS1 and LS2 habitats.Scrub composed of the invasive species sea buckthorn (Hippophae rhamnoides) was classified as habitat CD4 on the east side of the study site.The presence of human disturbance (beach access paths) can also be observed in the coastal dune environment.

Discussion
This study focused on the development of an automatic classification method using high-resolution    2020) achieved a similar level of accuracy for the classification of dune vegetation using RF.Our study, using RF, yielded more accurate results than Suo et al. (2019) achieved when mapping another Irish dune system using supervised Maximum Likelihood.This result confirms the high potential of using RF to map complex habitats, such as coastal dunes, using UAV (Feng et al., 2015;Sotille et al., 2022;van Iersel et al., 2018).
A major strength of RF is that it is an ensemble classifier, i.e. it can efficiently handle a large number of variables with a lower risk of overfitting than models using single classifiers (Miao et al., 2012).This means that several variables can be incorporated into the model, providing more information about the target classes and improving the classification.In this study, elevation and two spectral variables, NDWI and GNDVI, appear to be the most discriminating variables for dune habitats (see Supplementary Table S3).NDWI, an index related to plant water content, can distinguish dune slacks (2190), as they are generally characterized by saturated soil and moisture-loving plants (National Parks and Wildlife Service, 2019), whereas, GNDVI is highly sensitive to changes in chlorophyll content among different vegetation types (Gitelson et al., 1996).
The methodology was also designed to assess the impact of temporally acquired UAV-derived spectral and topographic data on classification accuracy.The inclusion of the UAV-derived topographic variables (elevation and slope) increased classification accuracy (Table 4), and this has been demonstrated in other vegetation and habitat-related studies (Fraser et al., 2016;Scholefield et al., 2019).In the present study, the inclusion of these topographic variables helped discriminate between dune habitats, especially dunes with Salix repens (2170) and dune slacks (2190), thus increasing the overall accuracy of the model.While habitats 2170 and 2190 contain similar plant communities and are even categorized as a single habitat type in the Irish Fossitt classification system (Fossitt, 2000), their distinct topographic characteristics enable their discrimination.Dune slacks (2190) are predominantly located at the lowest elevation in the site and on flat terrain, whereas dunes with Salix repens (2170) can be found within or adjacent to dune slacks but at slightly higher elevations (National Parks and Wildlife Service, 2019).Moreover, the misclassifications were reduced between spectrally similar white dunes (2120) and those degraded areas in a dune system (ED1 and ED3) after adding topographic data.This was shown by the improved precision values in those three classes, indicating that more predictions in the RF ST models were correct (compare RF S and RF ST , Table 5).White dunes (2120) are characterized by partially stabilized hills of sand with an incomplete vegetation cover.In contrast, ED1 and ED3 include areas on generally flat terrain that are regularly trampled or driven over.Hence, these areas have less vegetation cover (Fossitt, 2000).However, discriminating between white dunes (2120) and grey dunes (2130) remains a challenge even after the inclusion of topographic data (Figure 6).This is probably because, within the study site, both habitats are often characterized by hills of sand with the presence of Ammophila arenaria (see Supplementary Material 1).
Other vegetation mapping studies have used separate sensors to acquire optical and topographic data; for example, some have used LiDAR (Bazzichetto et al., 2016;Marcinkowska-Ochtyra et al., 2019;Ward et al., 2013;Yousefi Lalimi et al., 2017).However, combining optical and topographic data acquired by different sensors can be costly and challenging as it requires geometric alignment between the datasets.In contrast, UAV-based topographic data can be obtained free of charge during the photogrammetric processing with a high level (±3 cm) of vertical accuracy (Cohen & Héquette, 2020).
Previous studies have shown that using multitemporal satellite or UAV images can, in many cases, improve classification accuracy compared with using uni-temporal images (Aghababaei et al., 2021;D. J. A. Wood et al., 2022;van Iersel et al., 2018;Vilar et al., 2020).This is confirmed by our results (Table 4).The improved overall accuracy was due to the increased correct classifications of the four Annex I habitats.As these habitats are characterized by different vegetation, their classification can benefit from the acquisition of multi-temporal images which provide spectral response variations throughout the growing season.
Interestingly, the models using both spectral and topographic variables (the RF ST models) show only a slight increase of 3% in accuracy when more temporal data were included (91.35% vs. 94.26%).In the models that only use spectral variables (RF S models), the accuracy improved by up to 8% when data from two or more acquisition times were used (84.09% vs. 92.37%).Therefore, from a practical point of view, using the RF ST model, with UAV data acquired from a single time, is deemed sufficient for accurately classifying dune habitats as it can have a similar level of accuracy to data from multiple times.
The best acquisition time for each Annex I dune habitat varied as indicated by the F1-scores in Table 5.Hence, it can be assumed that there was no single optimum acquisition period for classifying these individual habitats.However, our analysis showed that the model trained on July data achieved the highest overall accuracy on two test datasets (Test 1 = 0.81, Test 2 = 0.86, see Tables 6 and 7).The difference between the two accuracy values could be due to Test 2 models being tested with a dataset that contained samples from the same month as the data used to train the model.This higher overall accuracy obtained in July may be explained by the fact that this period of the growing season is a time of maximum photosynthetic and floristic activity in the dune vegetation.Hence, spectral differences between the dune vegetation types are enhanced.For example, the dune-binding species Ammophila arenaria changes its color in July and August as the golden flowers appear.This result is consistent with other studies that use spectral differences to discriminate between species.Michez et al. (2016) and Hill et al. (2010) found that the late season was a more appropriate period for image acquisition for forest areas because phenological differences between forest species are more enhanced during the senescence period.In other words, different habitats may exhibit greater spectral variations at different times of the year, and these can be used to devise an optimal period for monitoring.Despite this, misclassifications still occurred in the July model when evaluated using data from other times of the year.The water table in dune slacks (2190) remains at or close to the surface for most of the year (Fossitt, 2000).However, in 2020, they were flooded in May and dry in July and October.This water table fluctuation could account for the misclassifications in this area.
The use of UAV surveys to acquire centimeterresolution images is currently limited to small areastypically a few square kilometers-due to factors such as the flight speed, legally required proximity of operators to UAVs and battery life.This limitation might not be practical for operational use in regular habitat or vegetation mapping at regional or national scales.Our method is a practical approach for monitoring areas locally, and it could be extended to a larger scale when combined with freely available satellite imagery, such as that from the Sentinel-2 mission (e.g.Kattenborn, Lopatin, et al., 2019;Riihimäki et al., 2019).While the UAV sensor used in this study operates in five spectral bands that are also present in Sentinel-2 sensors, understanding sensors' spectral characteristics (e.g.bandwidth, signal-to-noise ratio) is crucial for optimizing multi-sensor synergies (Alvarez-Vanhard et al., 2020;J. Jiang et al., 2022).Moreover, aside from having lower spatial resolution compared to UAV imagery, the issue of cloud cover is another notable limitation when using satellite imagery.This means that countries with high levels of cloud cover, such as Ireland, can have less benefit from using optical satellite imagery for habitat monitoring (Connolly, 2018).Nevertheless, UAV technology is continuously developing and, in particular, the flight range of UAVs is increasing.This could address the current challenges in terms of spatial areal coverage without compromising the spatial image resolution.
With increased technological advances, in both computer processing power and remote sensing platforms, there are opportunities to explore the application of deep learning algorithms in habitat mapping.Moreover, the calculation of different texture features from the imagery could be used to represent vegetation structure within habitats.These data, combined with image bands, vegetation indices, and threedimensional information, could further improve habitat classification accuracy.The addition of texture features in the classification was demonstrated by previous habitat and vegetation mapping studies (E.M. Wood et al., 2012;Gonçalves et al., 2016).It is also worth emphasizing that future analysis in a coastal dune environment can take advantage of the UAVderived data in other applications, such as in examining high-resolution morphodynamic processes (Laporte-Fauret et al., 2019).
The robust methodology presented here uses a lowcost multi-spectral UAV system to produce a highly accurate (>94%) high-resolution Annex I dune habitat map.It also highlights degradation caused by both natural and human disturbance (e.g.frequently used tracks) and the presence of invasive species.This methodology is transferable and could be used for the accurate mapping of other habitats.Accurate maps are invaluable resources that provide information about habitat extent and condition.The maps produced from this methodology are useful for policymakers as well as site managers and can be used to monitor changes from baseline data as well as progress in restoration activities.Moreover, ecologists and other monitoring experts can integrate UAV remote sensing with the traditional field methods for the targeted surveillance and monitoring of habitats (e.g.detection of the extent and direction of change of the habitat location and status) (Anderson & Gaston, 2013;Cruzan et al., 2016).

Conclusion
Under the Habitats Directive, one key requirement for each EU Member State is the spatiotemporal monitoring of its natural habitats, including coastal dunes.The emergence of mapping technologies and UAV platforms providing high-resolution spatial data is a tool that can support habitat monitoring.
This study aimed to improve the method for monitoring complex dune habitats by integrating UAV data and using the Random Forest classifier to automatically map these habitats with a high level of accuracy.The result was a series of models and maps with accuracy ranging from 84% to 94%.The integration of both spectral and topographic data derived from three sets of images gave the highest accuracy map (94.26%).The classification results also suggest that the middle period of the growing season is better than the early or late seasons for image acquisitions for coastal dune systems in Ireland.The high accuracies obtained in this study highlight the application of this robust methodology to aid the spatiotemporal monitoring of Annex I habitats.

Figure 1 .
Figure 1.Inset map showing the location of the study site in County Kerry, Ireland (left); the boundary of the Tralee Bay and Magharees Peninsula, West to Cloghane SAC (upper right); and the boundary of the study site (lower right).

Figure 2 .
Figure 2. A typical dune profile (modified from Delaney et al., 2013) and sample field photos of Annex I and Fossitt habitats present within the study site.

Figure 3 .
Figure 3. Methodology workflow in using multi-temporal high-resolution UAV imagery to characterize different habitat types.

Figure 4 .
Figure 4. Diagram of the reference data collection and processing steps.

Figure 5 .
Figure 5. Preparation of Test 1 and Test 2 datasets to be used for modeling scenarios.May (M), July (J) and October (O) were the three field campaign months in this study.

Figure 6 .
Figure 6.Confusion matrices obtained for each RF model.'Truth' refers to the number of habitat samples from the reference data.'Predicted' refers to the number of habitat samples identified from the classification.The color scale along the major diagonal line indicates the proportion of reference samples classified correctly as a function of overall accuracy rounded between 0.0 and 1.0.

Figure 7 .
Figure 7. Supervised Random Forest classification map of coastal dune habitats in the Magharees, County Kerry, and the computed area (in ha) for each habitat type classified using UAV images acquired at three different times within the growing season.The classified map is based on the July data.

Table 1 .
Habitat types found within Irish dune systems and corresponding habitat codes (following European

Table 2 .
List of variables used in the study and their descriptions.

Table 3 .
Datasets used in training models.

Table 4 .
5-fold cross-validation mean accuracy and standard error (SE) values for models obtained for each dataset.The datasets considered are described in Table3.Values are in percentage (%).

Table 5 .
Per-class precision (P) and recall (R) values, and F1-scores (F1) obtained for models from each dataset.The habitat codes are described in Table1.

Table 7 .
Summary table showing the mean and standard error (SE) values for each model evaluated on the Test 2 dataset.

Table 6 .
Summary table showing the mean and standard error (SE) values for each model evaluated on the Test 1 dataset.