A national-scale assessment of land use change in peatlands between 1989 and 2020 using Landsat data and Google Earth Engine—a case study of Ireland

Over the centuries, anthropogenic pressure has severely impacted peatlands on the European continent. Peatlands cover ~ 21% (1.46 Mha) of Ireland’s land surface, but 85% have been degraded due to management activities (land use). Ireland needs to meet its 2030 climate energy framework targets related to greenhouse gas (GHG) emissions from land use, land use change and forestry, including wetlands. Despite Ireland’s voluntary decision to include peatlands in this system in 2020, information on land use activities and associated GHG emissions from peatlands is lacking. This study strives to fill this information gap by using Landsat (5, 8) data with Google Earth Engine and machine learning to examine and quantify land use on Irish peatlands across three time periods: 1990, 2005 and 2019. Four peatland land use classes were mapped and assessed: industrial peat extraction, forestry, grassland and residual peatland. The overall accuracy of the classification was 86% and 85% for the 2005 and 2019 maps, respectively. The accuracy of the 1990 dataset could not be assessed due to the unavailability of high-resolution reference data. The results indicate that extensive management activities have taken place in peatlands over the past three decades, which may have negative impacts on its ecological integrity and the many ecosystem services provided. By utilising cloud computing, temporal mosaicking and Landsat data, this study developed a robust methodology that overcomes cloud contamination and produces the first peatland land use maps of Ireland with wall-to-wall coverage. This has the potential for regional and global applications, providing maps that could help understand unsustainable management practices on peatlands and the impact on GHG emissions.


Introduction
Peatlands are one of the most space-efficient carbon stores in the world and play an important role in the global carbon cycle (Limpens et al. 2008;Yu et al. 2011).These ecosystems cover about 3% or 4.23 million km 2 of global land surface area (Xu et al. 2018) and store approximately 455 to 582 Peta grams of carbon (Gorham 1991;Jackson et al. 2017).For centuries, management-related activities have exerted severe pressure and caused the degradation of peatlands (Tanneberger et al. 2021).Consequently, peatland ecosystems are projected to shift from a carbon sink to a carbon source within this century (Loisel et al. 2021).
Emissions from peatlands where substantial land-use change occurred, e.g.forestry and agriculture, have a greater impact on global land atmospheric Greenhouse Gas (GHG) emissions than previously thought (Ribeiro et al. 2021).Afforestation of natural peatlands and intensive forest harvesting also increase CO 2 emissions and affect water quality (Jovani-Sancho et al. 2021;Palviainen et al. 2021).Furthermore, drained peatlands used for industrial and domestic peat extraction are conduits of fluvial C and GHG emissions.They are also associated with higher risks of carbon loss through fire (Waddington et al. 2001;Turetsky et al. 2015;Wilson et al. 2015;Tubiello et al. 2016;Regan et al. 2019;Peacock et al. 2021).Tiemeyer et al. (2016) analysed GHG emissions from several bog 124 Page 2 of 16 and fen sites dominated by grasslands and reported high emissions.It is evident that these peatland land uses play a significant role in fluvial C and GHG dynamics.
Legislative changes in the EU's 2030 Climate Change Framework mean that wetlands are now included in the quantification of GHG emissions and removals from LULUCF (Land Use, Land Use Change Forestry: EU 2018).Additionally, the EU Nature Restoration law aims to restore the degraded ecosystem across the EU (EU 2022), while the EU biodiversity strategy for 2030 (2020( /2273(INI)(INI)) offers comprehensive guidance on the restoration of wetland ecosystems.On a national scale, the Government of Ireland made two major announcements: (1) the allocation of funding for rehabilitation activities on managed peatlands (Government of Ireland 2020) and (2) to report emissions from managed wetlands, including peatlands (Government of Ireland 2021).Given the increased need to accurately quantify emissions from peatlands, as well as carry out rehabilitation, restoration and rewetting activities, it is necessary to map the spatial and temporal changes in peatland land use.
Despite these requirements, there is a lack of spatial information on land use and land use change within these ecosystems (Fluet-Chouinard et al. 2023).Peatlands are not thoroughly represented in global, regional and national land use and land cover classification maps (Krankina et al. 2008;Pflugmacher et al. 2007;UNEP 2022).This is also evident in the efforts to develop peatland-specific global and regional maps (Xu et al. 2018;Montanarella et al. 2006;Tanneberger et al. 2017), as they are often underrepresented in land cover databases.One of the reasons for this underrepresentation is that peatlands have undergone land use conversion to agriculture and forest, or have been mined and thus the land cover has changed (Connolly 2018).Therefore, it is these land cover types that are represented in land cover maps and not the peatlands which are effectively "hidden" beneath (UNEP 2022).This results in an underestimation of peatland area in land cover maps such as the NLCD (National Land Cover Database: Loveland et al. 2000); the European CORINE (Co-ordinated Information on the Environment: Bossard et al. 2000) land cover dataset and the international land use land cover change dataset by IGBP (Fry et al., 2009).Where they are represented such as in the NLCD or the IGBP land cover maps, it is under the broad classification of wetlands.Similarly, in CORINE land cover, which is one of the principal sources of land cover information in Europe, peatlands are categorised under wetlands and are classified as peat bogs (Bossard et al. 2000).Additionally, the CORINE land cover dataset has low spatial resolution with a minimum mapping unit of 25 ha.In many areas across the globe, peatlands that have undergone land use change still contain a large stock of carbon (Hastie et al. 2022).Consequently, as these land cover maps do not adequately represent peatland extent and thus land use on peatlands, it prompts the need to develop peatland land usespecific maps to track land use change over time.
In Ireland, peatlands account for ~ 21% or 1.46 Mha of the total land surface (Connolly and Holden 2009).However, 85% has already been degraded due to land use activities, such as drainage and subsequent conversion to agriculture, forestry and extraction (domestic/industrial) (Malone and O'Connell 2009;Davies and Forster 2014;Connolly 2018;Fluet-Chouinard et al. 2023).Large areas of peatland are managed by semi-state companies, namely Bord na Móna (BnM) and Coillte.BnM was responsible for industrial peat extraction until 2020 when extraction ceased.Coillte, the semi-state forestry company, has afforested a substantial portion of the peatlands in Ireland (Connolly and Holden 2011).There are also several medium and small private companies still actively involved in peat extraction as well as numerous smaller landowners, who manage peatlands for a variety of land uses such as agriculture, afforestation and peat extraction (Malone and O'Connell 2009).However, the spatial extent of these activities has not been mapped.
The Derived Irish Peat Map version 2 (DIPMv2: Connolly and Holden 2009), was used here to delimit peatland areas and then to track land use change since 1990.Land use maps can be produced using readily available satellite remote sensing data.However, persistent cloud cover is a serious impediment to detecting land use change using remote sensing, especially in Europe and more specifically in Ireland (Tolnai et al. 2016).This is also clear from previous Irish land use and land cover mapping efforts which used remote sensing methods (Walsh et al. 2021;Cawkwell et al. 2018Cawkwell et al. , 2010;;Fealy et al. 2009).Cloud cover was also an issue for the Connolly (2018) peatland land use study, which was the first map of land use on peatlands in Ireland and used imagery from 2005/6.In that map, four major land use classes were mapped: industrial peat extraction, grassland, forestry and residual peatland.However, areas along the western seaboard were omitted due to cloud cover.Given that areas contain extensive blanket bogs which contain some of the largest peatlands soil organic carbon stocks in Ireland, it is essential to track and monitor the land use there (Renou-Wilson and Byrne 2015).
The past decade has seen a paradigm shift in remote sensing from change detection to monitoring and tracking change through time (Woodcock et al. 2020).This has been paralleled and facilitated by the development of planetary-scale cloud-computing platforms such as Google Earth Engine (GEE: Gorelick et al. 2017) and the Geo-Microsoft Planetary Computer Program (Lukacz 2022).These platforms provide readily accessible computing capabilities and integrated remote sensing image archives thus facilitating the processing and analysis of earth observation imagery for large spatial scales.Furthermore, there has been a substantial rise in freely accessible remote sensing data sources like Sentinel-2 and Landsat 8 and 9, contributing to a notable Page 3 of 16 124 increase in remote sensing-based peatland mapping and monitoring (Dronova 2015;Czapiewski and Szumińska 2021).To address the issue of cloud contamination, temporal mosaicking functions within GEE can be employed.Amani et al. (2019) and Mahdianpari et al. (2020) utilised these mosaicking functions in GEE to create cloud-free mosaics using seasonal imagery and successfully performed large-scale wetland mapping with wall-to-wall coverage.
In this study, GEE was used with Landsat data (5 and 8) and a machine learning-based classification algorithm, i.e. random forest to assess, monitor and track land use change on Irish peatlands from 1990, 2005 and 2019.The rationale for choosing these key years is that they relate to baseline data requirements for both the Kyoto Protocol and the Paris Agreement.Additionally, the EU 2030 climate and energy framework aims to reduce (at least 40% cuts) GHG emissions to 1990 levels (UN 1992;EU 2018).The baseline year for quantification of GHG emissions from LULUCF for the Paris Agreement is 2005.Lastly, 2020 had been included for a more recent comparison.To address the issue of cloud cover, a 3-year time period (TP) was selected to facilitate the development of wall-to-wall cloud-free images using the GEE mosaicking functions.These TPs consisted of 1 year before and 1 year after the key years, i.e.TP1 = 1989-1991, TP2 = 2004-2006and TP3 = 2018-2020.The results demonstrate the evolution of land use on peatlands and its current status (TP3), providing accurate spatial data for informed decision-making in sustainable land use management in peatlands.

Study area
This research focuses on Irish peatlands, found within Ireland's temperate maritime climate, which is characterised by mild winters, cool summers and abundant rainfall throughout the year.The spatial extent of peatlands was delineated using DIPMv2 (Connolly and Holden 2009).This map classifies peatlands into three distinct classes: (i) high-level montane blanket bog, (ii) low-level Atlantic blanket bog and (iii) raised bog.Blanket bogs are characterised by their extensive coverage with a "blanket-like" continuous layer of peat.They represent most of the Irish peatlands; 926,700 ha (65 %) and are present mostly, near low-lying coastal plains and across gentle slopes and mountain plateau.In contrast, raised bogs, known for their dome-shaped peat accumulation, cover ~ 533,300 ha (35%) of the peatlands in Ireland and are situated mostly in the inland areas (midlands) (Feehan and O'Donovan 1996;Connolly and Holden 2009).There is also the presence of minerotrophic peatland, fens which cover approximately 22,180 ha (Foss 2007).These were excluded from Connolly and Holden (2009) work on DIPMv2 and therefore are not considered in this study.

Satellite image data
Landsat-5 and 8 multitemporal and multispectral optical remote sensing satellite images were used in this study.Six spectral bands were selected for the land use classification: red, green, blue, near infrared and both short-wave infrared.The swath width of the sensor is 185 km and has a spatial resolution of 30 m (USGS 2016).Atmospherically corrected Landsat-5 and Landsat-8 products were used including Tier 1 Surface Reflectance from the TM (Thematic Mapper) and OLI (Operational Land Imager).The atmospheric correction of TM and OLI products is based on the Landsat Ecosystem Disturbance Adaptive Processing System and Landsat-8 Surface Reflectance Code respectively (Masek et al. 2006;Vermote et al. 2016;Foga et al. 2017).It includes snow, water, cloud and shadow masks that are based on the per-pixel saturation mask and CFMask (C Function of Mask) algorithm (Foga et al. 2017).The QA (quality assessment) band is used to represent these masks within the Level-1 data.The CFMask has been validated by Foga et al. (2017) and shown to have the best overall accuracy among different cloud-masking algorithms.Topographic effects can also have an impact on the classification results (Álvarez-Martínez et al. 2018a, b).However, most of the study area did not have significant terrain variations, so topographic corrections were not explicitly considered for this study.The entire study area was covered by 9450 columns and 14,341 rows of Landsat imagery for each TP.

Methodology
A remote sensing-based image classification approach was used here (Lu and Weng 2007).This included the acquisition and processing of satellite imagery which was conducted in GEE.The imagery was processed to mask out the clouds using the pixel_qa band.It was then constrained to the peatland areas in Ireland using the DIPMv2.After defining training areas using heads-up digitisation, a classification of land use was performed using the Statistical Machine Intelligence and Learning Engine (SMILE) random forest algorithm.Accuracy assessment of the results was performed using System for Earth Observation Data Access, Processing and Analysis for Land Monitoring (SEPAL) (Fig. 1).

Training data
The spatial resolution of the Landsat data (30 m) means that only certain features on the ground can be discriminated.In this case, only four classes could be identified: industrial peat extraction, forestry, grassland and residual peatland (Table 1).These classes are also consistent with the land-use map developed by Connolly (2018) which used a relatively higher spatial resolution (23 m) imagery but did not provide wall-to-wall coverage due to cloud cover contamination along the western seaboard.
The training data was derived from randomly distributed polygons digitised by an expert operator using the visual interpretation (Richards 2013)

Satellite image processing
To address the issue of persistent cloud cover in Landsat imagery, an efficient cloud removal methodology is crucial in regions where cloudy conditions are prevalent (Amani et al. 2019).Recent studies have shown that a period of 3 years (for a particular season) is optimal to create a cloud-free mosaic for national-scale mapping (Amani et al. 2019;Mahdianpari et al. 2020).Therefore, in this study, GEE was utilised to access imagery with less than 10% cloud cover over each TP to create cloud-free mosaics.It is also pertinent to mention that most peatlands investigated in this research do not experience land use change on short temporal scales.Hence, it was also suitable to use imagery over the entire year as opposed to being constrained to a specific season.This also allowed for the acquisition of wall-to-wall coverage of cloud-free imagery without any gaps created due to the application of cloud masks.The table shows the temporal coverage of images on each TP (Table 2).
Based on the selection criterion of temporal (Table 2) and cloud coverage (less than 10%), several hundred images were available for each TP.Two cloud-masking functions in GEE were used to mask out the clouds: cloudMaskL457 and maskL8sr, which use a QA band denoted as pixel_qa (Fig. 1).These were implemented for the Landsat-5 and Landsat-8 data, respectively.Finally, a single cloud-free composite image was generated for each TP using the GEE median (ee.reducer) composite function.The function calculates the median value of all the pixels in matching bands across the image stacks and eliminated pixels with extreme value.It also removes any remaining artefacts in the imagery.Hence, each TP image represented pixels with no or minimum cloud cover (or shadow).

Image classification
A pixel-based classification approach was used in this study.The Random Forest classification model was used (Breiman 2001).
It is one of the most widely used classification algorithms in remote sensing and has also shown reliable results in wetland mapping (Amani et al. 2019;Mahdianpari et al. 2020).It is an ensemble classifier that uses a large number of decision trees that combine decisions from a set of classifiers using a voting system (Breiman 2001;Pal 2005).GEE has implemented a slightly modified version of the algorithm known as the SMILE random forest (Li 2016), which was used here.It closely resembles the original algorithm in structure but includes decision tree size regularisation through the maxNode and nodeSize parameters.
The optimum value for the two most important parameters in the random forest classifier, namely the number of trees and number of variables, was adjusted using a trial-and-error method and the square root of the number of features, respectively.A value of 20 was finally used for the number of trees.The model was trained for each TP using the training samples.Due to spectral similarities between the industrial peatlands and many residual peatland sites, most of the areas not identified as industrial extraction sites (containing bare peat on the surface and mostly non-industrial) were classified as such.
To resolve this issue, boundary data from BnM was used to delineate the spatial extent of the industrial peat extraction sites on raised bogs and blanket bogs.Any bare (exposed) peat areas, outside the BnM industrial extraction sites that were classified as industrial peatlands, were merged with residual peatlands.This is due to the coarse spatial resolution of the imagery and could be improved in future studies with higher spatial resolution datasets.National-scale peatland land use maps at 30-m spatial resolution were thus created for each TP.The output maps from this analysis depict the spatial extent and temporal trend of four major land use categories on Irish peatlands over a 30-year period.

Accuracy assessment using validation data
An overall accuracy assessment was conducted for the output maps using independent data.The maps were uploaded to the SEPAL platform (FAO 2020) which is a cloud-computing-based platform developed for the automated mapping, monitoring and assessment of land cover.There are several tools available on the platform for land cover analysis.In this study, the following tools were used; stratified area estimator-design (SAED) and stratified area estimator-analysis (SAEA).SAED was used to create a sampling point.The tool uses Eq. ( 1) (Cochran 2007) to calculate the overall sample size (n).It ensures that enough sample points (validation data) are created for each stratum (class).
where n is the total number of validation points, N is the number of pixels for each class in the classified raster  representing the spatial unit, S(Ô) represents the standard error of the estimated overall accuracy, S i is the standard deviation of each class (stratum), , where U i is the conjectured value of the user's accuracy for a land use class and W i is the proportional area of each class i that is mapped (Olofsson et al. 2014).One thousand seven hundred four sample points were generated which were used for accuracy assessment (Table 3).
The labelling of validation data was carried out using Collect Earth, a desktop-based tool combining Google Earth and Google Earth Engine (Bey et al. 2016).This was conducted employing visual interpretation of high-resolution imagery in the Google Earth desktop application in both 2005 and 2019.Each point was carefully assessed by a single expert operator.For areas with low spatial resolution imagery, the temporal archive and GEE were utilised to ascertain the specific class; however, these points were less than 10% of the total.Upon completion of the labelling process, the points were further analysed in the SEPAL platform using the SAEA tool.This was used to obtain a confusion matrix and estimate the overall, producer and user accuracies.The simplest statistic is the overall accuracy which defines the proportion of locations that are correctly predicted by the model.The producer's accuracy indicates the probability of a correctly classified reference pixel, whereas the user's accuracy defines the probability of the pixel classified on the map agreeing with the class on the ground (Story and Congalton 1986;Congalton 1991).

Results
This study is the first to quantify land use on raised and blanket bogs separately across Ireland over three decades.The resulting maps (depicted in Fig. 3a, b, c) provide insights into the intensity, extent and trends of land use on peatlands during this period.The prevalence of industrial peat extraction sites can be seen in the midland region, where most of the raised bogs are situated.These peatlands are transitioning into grassland and forestry, as evident in the 2019 map (Fig. 3c).Within these midland areas, grassland primarily occupies the peripheries of industrial peat extraction sites.
Additionally, grassland emerges as a predominant land use category across both peatland types.Along the western seaboard's blanket bog regions, forest and residual peatland remain prevalent.
The spatial extent of each peatland land use category was assessed for each TP.The industrial peatland extraction sites managed by BnM have undergone a significant transition.
One key indicator of this change is the steady reduction in exposed bare peat areas, which have decreased by almost 40%, from 69,600 to 41,700 ha between 1990 and 2019 (Fig. 4).Afforestation increased by 17% between 2005 and 2019 from 191,600 to 224,700 ha, respectively.Over the past 30 years, the grassland area has decreased from 436,700 to about 357,900 ha (Fig. 4).
Land use change dynamics between 1990 to 2005 and 2005 to 2019 are depicted in Tables 4 and 5, respectively.These tables show the major land use change between the four mapped land use classes.The bold values along the diagonal line in the tables indicate the land use categories that remained constant and were consequently subjected to the same land use.The cumulative area represented by these constant values is estimated to be approximately 1 Mha between 1990 and 2005, and 1.1 Mha between 2005 and 2019.This constitutes ~ 71% of the total peatland area, identified at this spatial resolution.
The other ~ 30% of the peatland area has experienced a change in land use.The largest change between 1990 and 2005 was a 27% reduction in the grassland sites due to afforestation, i.e. 40,700 ha.Afforestation showed a 5% increase in 2005 compared to 1990.The conversion of grassland to residual peatland, in some cases, could be attributed to differences in the spectral signature and is discussed in detail later.Meanwhile, the industrial peat extraction sites underwent substantial land use change between 2005 and 2019 (Fig. 3b, c).Almost one-third (36%) of these sites were converted to forest, grassland and residual peatlands.The increase in afforestation is also apparent as 34,700 ha of grasslands and 5100 ha of industrial peat extraction sites converted to forestry (Table 5).

Raised bogs
Grassland (agriculture) has been the dominant land use in raised bogs, accounting for ~ 43% of the total area of raised bogs in 1990 and 47% in 2019 (Table 6).A smaller but significant area has been afforested, increasing from ~ 8% in 1990 to ~ 14% in 2019.Peat soil exposed through industrial peat extraction constituted ~ 12% in 1990 reducing to ~ 7% in 2019.Approximately 90% of the BnM industrial peat extraction sites occur on raised bogs, in the midlands.However, there has been a notable decrease in industrial bare peat sites, as they transition to other land use types.

Blanket bogs
Much less land use change can be detected in blanket bogs (Table 7) at this spatial resolution.Grasslands on peatlands have decreased from 20% in 1990 to 14% in 2019.Afforested areas have increased marginally from 16 to 17%.The residual peat area increased slightly between 1990 and 2019 from 64 to 70%.However, as with the raised bogs, the residual peat category is a catchall for many peatland land use categories that cannot be detected at this spatial resolution.The BnM industrial areas at Bellacorrick, Co. Mayo, are now captured under the residual peatland category.These industrial peat extraction sites on blanket bogs account for between 0.2 and 0.8% of the total area of BnM landholdings.However, peat extraction ceased there in 2003 and there has been a transition from industrial to forest and residual peatlands.This may depict revegetation through rehabilitation activities after the cessation of extraction in 2003 (Farrell and Doyle 2003).

Accuracy assessment
The accuracy of the map is calculated by comparing the mapped classes against the reference data.The overall accuracy was 86% and 85% for the 2005 and 2019 maps respectively (Table 8).In both the 2005 and 2019 maps, the forest class exhibited the highest user's accuracy, with values of 95% and 88%, respectively.The producer's accuracy, on the other hand, was highest for residual peatland class in both 2005 and 2019 maps, i.e. 95% and 93% respectively.The accuracy assessment of the 1990 map was not carried out due to the unavailability of high-resolution reference data.

Discussion
This study seeks to address the lack of robust spatial data on land use and land use change on peatlands in Ireland.It uses freely available Landsat satellite imagery and the GEE cloud-computing platform.A novel and robust method was developed in GEE using the random forest machine learning approach to identify four peatland land use classes and to track land use dynamics over 30 years from 1989 to 2020.The use of these methods in GEE enabled the study to overcome the challenges posed by persistent cloud cover and obtain accurate wall-to-wall land use maps for Irish peatlands.Several studies have found cloud cover to be a major issue with using remote sensing in Ireland (Cawkwell et al. 2018;Connolly 2018;Walsh et al. 2021).The findings demonstrate significant changes across all four peatland land use categories examined here.The accuracy assessment results from the 2005 and 2019 datasets show a particularly good agreement between the land use maps and the reference data; however, high-resolution reference data was not available to conduct the accuracy assessment of 1990 land use results.Table 4 provides information on the direction and magnitude of changes in land use class that facilitates the evaluation of the consistency of the land use classification for that particular TP.The transition from forest to grassland observed may be linked to policy incentives for agricultural production (Donnellan et al. 2015;Renou-Wilson and Byrne 2015), whereas small changes in bare peat in industrial areas between 1990 and 2005 are consistent with the policy regarding industrial peat extraction (Bullock et al. 2012).
In TP2: 2005, grassland covers 23% (~ 342,800 ha) of the peatland area, forestry covers 13% (~ 188,800 ha) and industrial peat extraction covers 4% (~ 65,500 ha).However, a previous study by Connolly (2018) reports higher estimates for grassland and forestry areas: 437,000 ha and 337,000 ha, respectively.Similarly, the industrial peat extraction area was estimated to be ~ 50,000 ha but did not include BnM landholdings on blanket bogs.Therefore, our study estimated an accurate BnM industrial peat extraction area of 63,900 ha.The remaining 59% (~ 864,500 ha) consists of degraded remnant and cutaway peatland areas, which were not detectable in either study due to the coarse spatial resolution of the imagery.We believe that the utilisation of robust machine learning-based pixel classification techniques, coupled with cloud-free wall-to-wall coverage, as employed in our study, enables more accurate estimations of the four land use categories when compared to the previous study (Connolly 2018).
Peatlands have been under considerable land use pressure in Ireland, particularly over the past several decades with the introduction of mechanised peat extraction (Smith and Crowley 2020).There are varying spatial and temporal changes across raised and blanket bogs.Industrial peat extraction areas are being converted to grassland, forestry and residual peatlands (revegetated and abandoned), whereas grasslands and forestry, located on peatlands, can be replaced with each other.According to this pixel-based approach, approximately 30% or 400,000 ha of Ireland's peatlands have experienced a change from one type of peatland land use to another in the past 30 years.Of the remaining 70%, about 47% were classed as residual peatlands.The land use changes that occur in those areas are relatively small scale, e.g.domestic peat extraction and are not detectable using Landsat data.However, they are extensive and should be examined using higher-resolution imagery in future studies.
The remaining 23% represent land use on peatlands including industrial, forested or grassland, that have remained stable over the study period.The driving factors for the land use change detected in this study could be linked to land use policy changes related to peat mining, afforestation and agricultural production over the last thirty years.

Industrial
The assessment of industrial land use was limited to the BnM landholding.The increase of afforestation on these lands has mainly occurred in the past 15 years.This could be attributed to the afforestation of former industrial peat extraction sites by BnM in collaboration with Coillte as depicted in Fig. 5.It also shows the transition from industrial peat extraction sites to grassland and residual peatland.Between 1990 and 2005, there were minor changes in the east of the midlands in the areas (outlined by the blue box in Fig. 5).However, between 2005 and 2019, there is a marked increase in forestry (yellow) and residual peat (orange) in this area.Given the relatively rapid land use change detected between 2005 and 2019, it would be prudent to track and monitor this on all industrial peatland sites using this methodology particularly, as all extraction activities on these sites officially ended in 2020 (BnM 2021).

Forestry
While there is an overall increase in forestry on peatlands over the past three decades, there was also a sub-trend where some afforested areas were converted to residual peatlands.For example, several EU LIFE-Nature projects have the aim of the "restoration of afforested bogs to their former status" including 1976 ha of blanket bogs on the west coast and more recently the raised bogs sites in the midlands (Bullock et al. 2012;Griffin 2016).This increasing trend may be related to afforestation policy incentives offered on a national scale.These incentives began with the establishment of Coillte in 1989 (Ward 1989), a semi-governmental organisation for the management of state-owned forestry.The introduction of EU and national afforestation schemes and grants such as the farm compensatory allowance scheme in 1987, the forest premium in 1990 and more recently, the afforestation scheme 2014-2020 incentivised forestry for farmers/landowners (Government of Ireland 2022).In this study, a substantial increase in forestry on peatlands was detected in the past 15 years, mainly on raised bogs.There are further plans by BnM and Coillte, e.g. to convert 1500 ha of the former industrial area into native woodlands (BnM 2019).Therefore, this land use change trend will continue over the coming years.

Grassland
Irish peatlands have been drained and reclaimed for agricultural development for centuries (Hammond 1981;Connolly 2018).This grassland category includes several types of grassland land uses including silage, hay, grazing and pasture (Fealy et al. 2009).While the other land use categories exhibit either consistent downward or upward trends since 1990, the grassland area on peatlands in Ireland has experienced an overall decline between 1990 and 2019.However, between 2005 and 2019, an increase is observed.A more detailed examination of this phenomenon in Fig. 6 highlights a continuous decrease in grassland on blanket bogs, while grasslands on raised bogs display a decrease between 1990 and 2005, followed by an increase in 2019.
To be precise, the land use conversion analysis (Table 5) shows that between 1990 and 2005, ~ 40,700 ha of grassland on peatland have been converted to forestry.This could be attributed to afforestation incentives discussed earlier.The increase in grasslands on raised bogs between 2005 and 2019 could be explained by a national drive towards agricultural production with incentives such as the abolition of the milk quota (Donnellan et al. 2015).BnM also converted 2500 ha of industrial cutaway land to agricultural grasslands (Renou- Wilson et al. 2015).Furthermore, government grant aids in the 1980/90 and payments through the EU Common Agricultural Policy (CAP), encouraged the conversion of large areas of peatland (mostly upland bogs) to grazing lands (Bullock et al. 2012).

Residual peatland
The residual peatland class comprises a range of small-scale land use categories that make up a spatially complex and heterogenous landscape (Connolly 2018).These land use categories include domestic peat extraction, rewetted and rehabilitated areas, natural degradation and bare peat which can be difficult to discriminate within the Landsat imagery.Consequently, the spectral signature variations are likely the result of land use and revegetation associated with the rehabilitation or abandonment of these lands.This leads to a shift from one class to another within a particular TP, followed by a return to the same class.An example of this is a change from grassland to residual peatland and then Mayo have predominantly reverted to poor fen and scrub with embryonic peat-forming conditions and thus converted to the residual peatland category (Farrell and Doyle 2003).
The cutover and remnant peatland (high bog) areas present a dynamic landscape and detailed habitat (area of occupancy) mapping using high-resolution imagery is essential (Álvarez-Martínez et al. 2018b;Ingle et al. 2023).COPERNICUS Sentinel-2 or other higher-resolution imagery could be used for this purpose, but this was beyond the scope of this study.Future work could also include additional TPs and the analysis of the spectral signature over time.
Overall, the general spatial distribution and temporal trend of the four main land use classes show that peatland industrial areas, which are present mostly on raised bogs, are converting to residual peatland, forestry and grassland.Although residual peatlands possibly show the rehabilitation, restoration and rewetting efforts being implemented.Furthermore, the use of blanket bogs for afforestation activities will also lead to ecosystem degradation and increasing peatland emissions (Jovani-Sancho et al. 2021).Degradation could be further exacerbated by climate change.For instance, periods of dry weather might render these ecosystems more susceptible to wildfires, and episodes of drought followed by heavy precipitation could increase their vulnerability to landslides (Connolly et al. 2021).
The overall accuracy for the maps in both 2005 and 2019 is 86% and 85%.However, more detailed information can be obtained by looking at the user and producer accuracy.The user accuracy for forestry and grassland is the highest in the 2005 (86%) and 2019 (85%) maps.The accuracy of the residual peat class is the lowest, primarily because this category is an amalgamation of multiple small-scale land uses.This is a limitation of using medium spatial resolution data (30 m).The industrial class also has lower accuracy.This class represents the exposed peat areas within the BnM landholdings.However, the peripheries of BnM sites still have domestic peat extraction activities ongoing which result in exposed peat.This is one cause of confusion between industrial peat extraction sites with residual peatlands, resulting in lower accuracy for both classes.Future work could use highresolution satellite imagery and very high-resolution aerial imagery to improve the land use maps, and map residual peatland class in further detail.This includes domestic peat  (Bradley and Mustard 2005).This could introduce errors in classification results and eventually the confusion matrix.The elimination of these uncertainties using the methods proposed by Álvarez-Martínez et al. ( 2010) could lead to a more refined accuracy assessment and a more efficient analysis of the temporal evolution of changes.Additionally, this study uses an existing peatland map (DIPMv2) for the delineation of peatland areas, although it is the latest and most reliable map of peatlands in Ireland with an overall accuracy of 88%.It underestimates the extent of peatlands with areas of less than 7ha (Connolly and Holden 2009).Therefore, this study can be expanded to these missing areas if that peat map is updated.Nevertheless, our study is an essential first step towards understanding peatland land use and offers a significant perspective and valuable insights into the dynamics of land use change on peatlands in Ireland.While higher spatial resolution remote sensing data is available, this methodology outputs maps of the major land use categories on peatlands providing valuable information on how they are changing through space and time.It could be a useful dataset for the recently announced "Land Use Review" by the Government of Ireland (2023).This methodology could be used to identify hotspots of land use change to support policy-related decision-making, aid Nation Inventory Reporting, promote sustainable land use practices and implement the principles and actions outlined in the National Peatland Strategy of Ireland.Additionally, these maps could also be used to refine estimates of GHG emissions from peatlands using Intergovernmental Panel on Climate Change (IPCC 2003) methodologies.The method used here is robust and could be applied to track and monitor peatland land use and land use change over time both in Ireland and across the globe.

Conclusion
This study presents the first long-term spatiotemporal assessment of peatland land use and land use change for Ireland.It is conducted over a 30-year period from 1989 to 2020.The spatial extent of four major peatland land use categories (grassland, forestry, industrial peat extraction and residual peatlands) are quantified, for both blanket and raised bogs.The methodology integrates the DIPMv2 to delimit peatlands within the GEE cloud-computing platform.The random forest machine learning algorithm is combined with Landsat data to classify the major land use classes.
The results indicate that at least 30% of peatlands in Ireland have undergone major land use change in the last 30 years.The results clearly show the conversion of industrial peatlands (BnM owned) as extraction ceases and the land is converted to several other land uses including rewetted areas.The forestry area has also increased, and grassland is one of the largest land use classes on peatlands in Ireland.The methodology also showcases how GEE can be used to overcome issues such as persistent cloud cover to deliver a national-scale assessment of land use change on peatlands and identify hotspots of change.Future studies should use higher-resolution imagery that can then be applied to examine these areas in detail, particularly the residual peatland class where change is occurring but cannot be detected with Landsat data.This robust method could be used to generate peatland land use-specific maps for peatland areas across the globe using the available global and regional peatland extent maps.Area (ha) Raised Bogs Blanket Bogs

124
Page 4 of 16 of Landsat imagery available for each TP.Training samples based on these polygons were created for each land use class and each TP in GEE enabling image classification.Figure 2 depicts the spectral reflectance values of the training samples at different wavelengths.There is a significant difference in reflectance values between grassland, forestry and industrial classes at different wavelengths and more noticeable in the NIR region.Residual peatland class, on the other hand, may exhibit some similarities with other land use classes at specific wavelengths, given it includes cutover and dynamic vegetation environment.Nevertheless, the differences in spectral reflectance between all classes are prominently discernible.Hence, the variation observed indicates potential effectiveness in distinguishing the mapped land use types in this study.

Fig. 2
Fig. 2 Spectral reflectance of the land use types across the six spectral bands

Fig. 3 a
Fig. 3 a Peatland Land use map 1990.b Peatland land use map 2005.c Peatland land use map 2019 Fig. 3 (continued)

Fig
Fig.6Grassland area on raised and blanket bogs

Table 1
Description of each land use class mapped in the study Forest Land covered with trees; afforested areas are mostly covered by evergreen tree species Grassland Natural and/or agricultural grassland that is used for pasture or hay, silage and grazing Residual peatland Mosaics of revegetated areas (heath, scrub, etc.), exposed peat through domestic peat extraction, remnant peatlands (high bog)

Table 2
Temporal coverage of the imagery used for each TP.

Table 3
Distribution of validation sample points for each land use derived using the SEPAL

Table 4
Land use (LU) change

Table 6
Sites that are going through rehabilitation such as the former BnM industrial site at Bellacorrick, Co.

Table 8
Accuracy assessment results for 2005 and 2019 results Peatland land use change in the midlands extraction and rehabilitation, restoration and rewetting sites.The use of Landsat (medium resolution) for the hard classification of large areas may introduce uncertainties (Álvarez-Martínez et al. 2010).For example, in heterogeneous land use classes such as the residual peatlands and grassland in this study, the allocation of pixels to specific land use classes may cause uncertainty