Wetland Mapping Under Boundary Uncertainty with AlphaEarth Embeddings, LiDAR and Sentinel

1. Introduction

In this connection, wetlands matter for hydrology, biodiversity, carbon storage, flood buffering and climate resilience, yet they remain hard to map at high spatial resolution. Their bound-aries are gradual and fragmented, and in forested low-relief landscapes wetlands look spectrally and structurally similar to upland forest, shrubland and saturated terrain. Field in-ventories are expensive, so remote sensing carries most of the load. Sentinel-2 has been widely used because its visible, near-infrared, red-edge and shortwave-infrared bands track vegeta-tion condition and moisture (Kaplan and Avdan, 2017), multi-temporal Sentinel-2 supports wetland vegetation-community mapping (Villa et al., 2021), and Sentinel-1 + Sentinel-2 time series add radar backscatter and inundation dynamics (Slagter et al., 2020, Mohseni et al., 2023, Wang et al., 2023, Hu et al., 2021). Single-source mapping has its limits, however, because wetlands are controlled by both surface hydrology and veget-ation structure. These multi-sensor workflows are also time-consuming to operate: long optical and radar time series must be gathered, cloud-screened, composited and co-registered be-fore any classifier is trained (Liu et al., 2023), and scaling deep models over such Sentinel-2 stacks is itself computationally heavy (Papoutsis et al., 2023). Even cloud-based Google Earth Engine pipelines remain constrained by processing and memory limits for large areas (Sultan et al., 2025), so the cost of build-ing and updating high-resolution wetland maps is dominated by data preparation rather than by classification.

It is pertinent to mention that multi-source fusion is there-fore the dominant direction. For southern New Brunswick, LaRocque et al. combined Landsat 8 OLI, Sentinel-1, ALOS-

1 PALSAR and light detection and ranging (LiDAR) data (LaRocque et al., 2020b); earlier provincial work used Landsat-5 TM, ALOS-PALSAR and Radarsat-2 with Random Forest (LaRocque et al., 2020a, LaRocque et al., 2014). These ef-forts show that optical + radar + LiDAR fusion improves wet-land classification in the province, but they do not test newer latent geospatial embedding features and they do not explicitly handle polygon-boundary uncertainty in the rasterised labels.

It should be noted that LiDAR has been especially useful be-cause it captures terrain, drainage position and canopy struc-ture. Topographic metrics derived from LiDAR digital eleva-tion models (DEMs) improve forested wetland mapping (Lang and McCarty, 2009), and fine-scale airborne laser-scanning metrics can classify wetland-related habitats (Koma et al., 2021, Maxa and Bolstad, 2009). Even so, LiDAR alone struggles to separate flat upland forest, forested wetlands and shrub wet-lands that share similar elevation and canopy-height profiles. Optical imagery adds complementary moisture and vegetation cues. In Alberta, multi-temporal Sentinel-1/2 with LiDAR-derived topographic wetness improved wetland mapping in the Grassland Natural Region (Onojeghuo and Onojeghuo, 2021), integrating Sentinel-1/2 with Global Ecosystem Dynam-ics Investigation (GEDI) structure data further improves clas-sification (Jafarzadeh et al., 2024), combining high-resolution multispectral imagery with LiDAR raises coastal-wetland clas-sification accuracy (Anokye and Hashemi-Beni, 2025), and cloud-based Random Forest workflows on Google Earth Engine have mapped coastal wetland and mangrove change at regional scale (Nguyen et al., 2024). Such methods also underpin global operational land-cover products at 30 m (Liu and Zhang, 2024).

In the same way, recent work pushes towards deep learning and feature-level fusion. Bakkestuen et al. used U-Net on Sentinel-2 plus LiDAR for large-scale wetland delineation in South Nor-way (Bakkestuen et al., 2023). Martins et al. built a deep net-work for open-water wetlands from WorldView-3 plus airborne LiDAR (Martins et al., 2020), and Guo et al. proposed a multi-source feature-embedding and interaction-fusion network for coastal wetlands using hyperspectral and LiDAR data (Guo et al., 2024). Rapinel and Hubert-Moy combined LiDAR with multispectral imagery for wetland habitat mapping (Rapinel et al., 2015), and Wu and Lane integrated LiDAR with multi-temporal aerial imagery to track wetland inundation dynam-ics on Google Earth Engine (Wu et al., 2019, Seyed Mousavi and Akhoondzadeh, 2023). U-Net and deep neural networks can learn spatial patch structure, neighbourhood context and boundary patterns rather than treating each pixel independently, which is advantageous for wetland mapping where wetlands oc-cur as patches, corridors and landscape-context-dependent fea-tures. These workflows demonstrate the value of richer feature spaces, but typically need specialised data, complex architec-tures or heavy preprocessing.

Satellite-derived latent geospatial embeddings offer another direction that removes much of this preprocessing burden. AlphaEarth Foundations distils a full year of multi-sensor ob-servations, including optical, radar, thermal and contextual cues, into a single compact, analysis-ready representation of each pixel. It delivers 64 precomputed latent dimensions at 10 m, so users need not gather, clean, composite or co-register raw image time series (Brown et al., 2025, Google and Google DeepMind, 2025). The steps that make Sentinel-2 and other multi-sensor pipelines time-consuming (Liu et al., 2023, Pa-poutsis et al., 2023, Sultan et al., 2025) are performed once, upstream, inside a globally trained model, and a downstream mapping task reduces to reading an off-the-shelf embedding stack. These representations behave as semantically mean-ingful feature spaces well suited to geospatial learning (Dalei et al., 2025), and Ryan et al. reported that AlphaEarth em-beddings can match traditional multi-sensor wetland-vegetation workflows while producing cleaner maps with minimal prepro-cessing (Ryan et al., 2025). That work addressed wetland ve-getation classification. It did not test the value of embeddings as the principal predictor, combined with LiDAR derivatives and Sentinel-2 bands, for boundary-aware wetland/non-wetland prediction at provincial scale, nor did it evaluate AlphaEarth embeddings within a deep-learning segmentation framework.

It may be noted here that a second persistent problem is reference-label uncertainty. Government wetland inventories are distributed as polygons, but converting polygon edges into a 10 m raster produces mixed-class boundary pixels, especially where wetlands grade into upland forest, shrub or saturated ter-rain. If those edge pixels are treated as exact ground truth, both training and validation become biased: a prediction near a mapped edge can be counted as a false positive or false negat-ive even where the ground is ecologically transitional. Previous work has reported persistent wetland/upland confusion for for-ested, small or transitional wetlands (Maxa and Bolstad, 2009, Lang and McCarty, 2009, Koma et al., 2021), and large-area probabilistic mapping has stressed the need for probability out-puts rather than hard class labels alone (Hird et al., 2017, Souza et al., 2025). Explicitly modelling label and prediction uncer-tainty is increasingly recognised as important for reliable seg-mentation (Landgraf et al., 2024).

Despite the progress made by previous studies, three major gaps remain. First, few studies have evaluated the combined use of airborne LiDAR derivatives, Sentinel-2 multispectral bands and AlphaEarth satellite-derived latent embeddings for 10 m wet-land segmentation, and the comparative contribution of each predictor group within both tree-based and deep-learning clas-sifiers has not been quantified in a boundary-aware framework. Second, most wetland mapping workflows rely on government polygon inventories as reference data but convert them dir-ectly into raster labels, treating every labelled pixel as exact ground truth during both training and evaluation. At 10 m res-olution, pixels along polygon edges frequently contain mixed wetland and non-wetland conditions or represent ecologically transitional zones; treating such pixels as unambiguous ground truth introduces bias into both model training and accuracy re-porting. Third, many wetland mapping studies emphasise over-all accuracy without fully evaluating wetland-specific metrics or the proportion of the landscape classified as wetland. Be-cause wetlands occupy a small fraction of most landscapes, a model can achieve high overall accuracy simply by correctly classifying the dominant non-wetland class while still system-atically overestimating or underestimating wetland extent. A map realism check, comparing predicted and reference wetland area proportions, is therefore a necessary complement to pixel-level accuracy metrics because operational maps are used for area-based decisions.

On the basis of these gaps, this study develops an embedding-centred, boundary-aware 10 m wetland prediction framework for the 2024 LiDAR-covered portion of New Brunswick. The contributions are: (1) an embedding-centred predictor design in which precomputed AlphaEarth dimensions act as the principal feature source, augmented by LiDAR derivatives and Sentinel-2 bands, co-registered at 10 m on the EPSG:2953 grid and avoiding bespoke multi-sensor time-series engineering; (2) a 20 m boundary-uncertainty exclusion zone that removes am-biguous edge pixels from both training and evaluation while keeping core wetland and confident non-wetland pixels; (3) an ExtraTrees ensemble classifier with predictor-importance analysis demonstrating the dominance of AlphaEarth embed-dings; (4) an Attention U-Net segmentation model with a con-trolled predictor-group ablation experiment comparing LiDAR-only, Sentinel-2-only, AlphaEarth-only and full-fusion config-urations; and (5) a direct comparison of both classifiers under identical boundary-aware labels and evaluation protocols, with joint evaluation of pixel-level accuracy and mapped-area real-ism. The resulting framework is evaluated against the GeoNB wetland inventory, and its accuracy, map realism and predictor importance are reported in Section 3.

2. Materials and Methods

2.1. Study Area

For the purposes of the present study, the study area is the por-tion of New Brunswick, Canada where 2024 airborne LiDAR coverage from GeoNB is available (Government of New Brun-swick, 2024a). Restricting the analysis to the LiDAR footprint avoided treating missing-LiDAR regions as valid non-wetland samples. The province contains forested, shrub, open, riparian and peatland wetlands as well as low-relief saturated terrain and surrounding mixed forest, so wetland and non-wetland areas are often gradual and partially obscured by canopy. The modelling domain was created by intersecting the 2024 LiDAR valid-area mask with the rasterised GeoNB Wetlands layer. The original 10 m reference grid contained 166,380,000 pixels, of which 97,232,930 had valid LiDAR coverage (Figure 1). Within that valid area the original rasterised reference contained approxim-ately 11.77% wetland and 88.23% non-wetland pixels before boundary-uncertainty treatment (Figure 2).

2.2. Data Sources

It is pertinent to mention that four geospatial datasets were used: 2024 airborne LiDAR point-cloud data from GeoNB (Government of New Brunswick, 2024a), 2024 Sentinel-2 multispectral imagery (European Space Agency, 2024), 2024 AlphaEarth satellite-derived latent geospatial embeddings (Brown et al., 2025, Google and Google DeepMind, 2025) and the GeoNB Wetlands reference layer (Government of New Brunswick, 2024b) (Table 1). The LiDAR data carry the stand-ard classification categories (ground, low/medium/high vegeta-tion, buildings, model key points, noise), from which 10 m ras-ter predictors of elevation, canopy structure, intensity, hydro-logical position and curvature were generated. The Sentinel-2 predictor stack used 12 bands as a complement to the LiDAR-derived variables. The AlphaEarth stack provided 64 analysis-ready latent dimensions (A00–A63). Each dimension com-presses a full year of multi-sensor satellite observations into a single 10 m embedding and summarises spatial, spectral and landscape-context information that would otherwise require ex-tensive multi-temporal image acquisition and feature engineer-ing. Because the embeddings are distributed as a ready-to-use annual product, they were ingested directly, with no com-positing, cloud-masking or temporal-statistics computation re-quired. The GeoNB Wetlands polygon dataset was rasterised to the 10 m LiDAR-aligned grid, with wetland set to 1 and non-wetland to 0 before boundary-uncertainty processing.

2.3. LiDAR Derivative Generation

In this relation, LiDAR tiles were processed in ArcGIS Pro (Esri Inc., 2024) using ArcPy and Spatial Analyst. All deriv-atives were produced at 10 m in EPSG:2953 (NAD83(CSRS) / New Brunswick Stereographic). The bare-earth DEM was gen-erated from class 2 ground returns using average elevation bin-ning and served as the master grid. The canopy height model (CHM) was the difference between a 10 m digital surface model (DSM, from first and single returns with maximum binning) and the DEM, with negative values set to zero. A 10 m intensity raster used average binning across all returns. Tile-level rasters were mosaicked using last-overlap (DEM), maximum (CHM) and average (intensity) rules.

Terrain curvature was computed from the DEM. The topo-graphic wetness index (TWI) used filled-DEM flow direction and flow accumulation combined with slope, as in Equation (1):

TWI = ln FA + 1 ,

(1)

tan(β) + 0.001

where FA is flow accumulation and β is slope in radians; the constant 0.001 prevents division by zero on very flat ground. TWI is a physically grounded predictor for wetland occur-rence because it captures the tendency for flow to accumulate in poorly drained, low-gradient terrain, the landscape position most commonly occupied by wetlands. The final LiDAR pre-dictor stack contains five 10 m rasters: elevation, TWI, canopy height, intensity and curvature.

2.4. Raster Alignment and Preprocessing

It should be noted that all predictors and reference rasters were co-registered to the LiDAR-derived DEM master grid in EPSG:2953. Continuous predictors (elevation, TWI, canopy height, intensity, curvature, Sentinel-2 bands, AlphaEarth di-mensions) used bilinear resampling; categorical rasters (wet-land reference mask, LiDAR valid-area mask, buffered labels) used nearest-neighbour resampling. This guarantees that each training sample represents the same ground location across all predictor groups and the label. A one- or two-pixel shift near a wetland boundary can otherwise change the apparent class of a 10 m pixel.

A valid-area mask flagged the 97,232,930 pixels with com-plete LiDAR coverage; only these pixels entered sampling and evaluation. For the ExtraTrees model, outlier clipping was applied to the LiDAR-derived predictors using percent-ile bounds, because extreme intensity or curvature values can dominate tree-based splits. For the Attention U-Net, isolated artefacts in LiDAR predictors were additionally repaired us-ing local neighbourhood median values calculated from non-outlier neighbouring pixels, to preserve local spatial structure rather than simply truncating values. Final cleaned ranges were: elevation 247.52–523.69 m, TWI 3.82–13.60, canopy height 0.00–27.13 m, intensity 33,192.73–61,536.71 (raw re-turn units), curvature −2.44–2.41. The canopy-height min-imum was left at 0 m because open, shrub and non-forested saturated wetlands legitimately carry little canopy. Sentinel-2 and AlphaEarth tiles extended beyond the LiDAR-covered study area; all training, validation, prediction and evaluation were constrained to pixels with valid LiDAR coverage.

2.5. Boundary-Uncertainty Treatment

As is often the case with polygon inventories, polygon-to-raster conversion produces uncertain wetland-edge pixels. A 20 m ex-clusion zone (two 10 m pixels) was created around each raster-ised wetland boundary. Pixels inside that zone were assigned class 255 and excluded from sampling, training, threshold se-lection, accuracy assessment and final map evaluation. The res-ulting boundary-aware reference carries three classes: 1 (core wetland), 0 (confident non-wetland) and 255 (uncertain bound-ary). Before the exclusion, 96,854,976 pixels had a com-plete predictor stack across all sources (slightly fewer than the 97,232,930 LiDAR-valid pixels, as cells lacking full Sentinel-2 or AlphaEarth coverage were dropped), comprising 11,395,605 wetland and 85,459,371 non-wetland pixels (11.77% wetland). After the exclusion, 85,299,502 pixels remained, comprising 6,216,473 core wetland and 79,083,029 confident non-wetland pixels (7.29% wetland), and 11,555,474 boundary pixels were removed (Table 2, Figure 3).

2.6. Sampling and Spatial Split

Needless to say, two complementary sampling strategies were used, one tailored to each classifier. For the ExtraTrees en-semble, a class-prevalence-aware pixel sampling strategy was adopted because the core-wetland share was only 7.29%. An earlier balanced 50:50 experiment gave high sample-based ac-curacy but severely overpredicted wetlands across the full area, confirming that balanced sampling distorts map-level outputs in this landscape. The selection of a realistic class ratio for training follows from the observation that balanced sampling overrepresents wetlands by a factor of roughly four to six relat-ive to their true landscape prevalence, which causes the model to assign elevated wetland probabilities to non-wetland terrain features in the full map. The final ExtraTrees training sample drew approximately 40,000 core wetland and 460,000 confident non-wetland pixels from the buffered reference (about 8% wet-land), preserving a realistic class imbalance. Samples were par-titioned using spatial block-based validation, in which contigu-ous blocks, rather than individual pixels, are assigned wholly to training or validation, to reduce spatial autocorrelation between training and evaluation sets.

For the Attention U-Net, training samples were extracted as 128×128 pixel patches from the aligned raster stack. Each full-fusion patch contained 81 input channels: five LiDAR-derived predictors, 12 Sentinel-2 bands and 64 AlphaEarth embedding dimensions. Each patch was accompanied by a corresponding label raster and a valid-pixel mask. Patches were accepted only when a sufficient proportion of pixels contained valid LiDAR coverage and valid boundary-aware labels. Training, valida-tion and test sets were separated spatially using row-based split-ting, which reduces the risk of evaluating the model on pixels spatially adjacent to training samples. Patch-based sampling without a spatial split can allow a model to memorise the neigh-bourhood structure of specific locations rather than generalising across the landscape, which is particularly problematic at 10 m resolution where adjacent patches share geographically similar terrain and vegetation conditions.

2.7. ExtraTrees Classifier

In this connection, wetland prediction with the ExtraTrees ap-proach used an Extremely Randomised Trees ensemble (Geurts et al., 2006) implemented in scikit-learn (Pedregosa et al., 2011), on top of NumPy (Harris et al., 2020) and Rasterio (Gil-lies et al., 2013–) for I/O, with Matplotlib (Hunter, 2007) for figures. ExtraTrees was chosen for its ability to handle nonlin-ear relationships, mixed predictor types and high-dimensional feature spaces without normality or linearity assumptions, in line with the strong performance of tree-based ensembles for remote-sensing land-cover classification (Sharma et al., 2025, Du et al., 2015). The ensemble used 500 trees with the de-fault scikit-learn settings (Gini splitting, square-root feature sampling per split, no depth limit) and a fixed random seed for reproducibility. Because ExtraTrees uses fully random rather than optimised split thresholds at each node, the resulting en-semble is generally less prone to overfitting on individual pre-dictor values than standard Random Forests, which is advant-ageous in a high-dimensional predictor space such as the 81-channel fusion stack. Wetland was the positive class. Three stages were compared: (i) a LiDAR-only baseline using the five LiDAR predictors; (ii) full LiDAR + Sentinel-2 + AlphaEarth fusion using the original hard rasterised labels; and (iii) the same fusion model trained and evaluated on the 20 m boundary-aware labels. The end-to-end ExtraTrees workflow is shown in Figure 4.

2.8. Attention U-Net Model Development

It may be argued that wetland segmentation was additionally performed using Attention U-Net, a convolutional semantic segmentation architecture that extends the standard encoder–decoder U-Net with soft spatial attention gates (Aburaed et al., 2023). Attention U-Net was selected because wetland mapping is a spatial segmentation problem: wetlands occur as patches, corridors, margins and landscape-context-dependent features rather than as independent pixels. Unlike the pixel-wise Ex-traTrees classifier, which treats each pixel independently based on its local predictor values, the Attention U-Net processes spatial patches and implicitly encodes neighbourhood relation-ships, spatial gradients and contextual information across tens to hundreds of metres. This is relevant for wetland mapping be-cause wetland patches, edges and corridors carry characteristic spatial signatures, such as elongated riparian forms, clustered peatland complexes and gradual ecotonal fringes, that pixel-wise classifiers cannot capture without explicit spatial feature engineering.

The attention gate mechanism applies soft spatial weights to the encoder-to-decoder skip connections, allowing the network to selectively emphasise feature responses from wetland-likely regions while suppressing irrelevant background features. This is particularly valuable in the high-dimensional predictor stack used here, where 64 latent embedding dimensions are present alongside terrain and spectral inputs, and the attention mechan-ism can allocate representational capacity to the most wetland-discriminative spatial patterns without manual feature selection.

Four configurations were trained in a controlled ablation ex-periment: (i) LiDAR-only: five LiDAR-derived predictors (5 channels); (ii) Sentinel-2-only: 12 spectral bands (12 chan-nels); (iii) AlphaEarth-only: 64 embedding dimensions (64 channels); (iv) full fusion: LiDAR + Sentinel-2 + AlphaEarth (81 channels). All four configurations used the same patch locations, training/validation/test spatial split, boundary-aware labels and evaluation metrics, making observed performance differences attributable to predictor-group contribution rather than sampling differences. The model was trained using a combined masked binary cross-entropy and Dice loss. Bin-ary cross-entropy alone tends to be dominated by the majority non-wetland class, which can suppress gradient signals related to wetland pixels; the Dice loss component directly optimises the overlap between predicted and reference wetland masks, providing a complementary objective that is more sensitive to wetland recall. Pixels labelled 255 (uncertain boundary) were masked and ignored during loss computation. Wetland was the positive class. Model outputs were continuous wetland prob-ability maps; binary wetland classification maps were derived through threshold optimisation (Section 2.9). The Attention U-Net workflow is shown in Figure 5.

cross-entropy and Dice loss while ignoring boundary-uncertain pixels (class 255). Threshold optimisation produces the final probability and binary wetland maps.

2.9. Threshold Optimisation and Accuracy Assessment

On the basis of the validation design, because the default prob-ability cutoff of 0.50 is not appropriate for a minority class, probability thresholds were optimised separately for each clas-sifier. For the ExtraTrees models, thresholds from 0.10 to

0.75 in steps of 0.05 were tested on the validation split, each scored on wetland precision, recall, F1-score, IoU, overall ac-curacy and predicted wetland proportion. For the hard-label ExtraTrees fusion model, 0.30 gave the best balance, producing a predicted wetland proportion of 12.66% against the 11.77% original reference. For the boundary-aware ExtraTrees fusion model the optimum was 0.35. On the validation split, this threshold gave 79.39% precision, 70.45% recall, 74.65% F1-score and 59.56% IoU, with a 7.04% predicted wetland propor-tion against the 7.29% buffered reference (Figure 6).

For the Attention U-Net models, thresholds from 0.20 to 0.80 were tested on the validation split. The full-fusion Attention U-Net selected a threshold of 0.80, which is notably higher than the ExtraTrees optimum and reflects a fundamental dif-ference in the probability distributions produced by the two model architectures. Tree-based ensemble classifiers typically produce probability estimates spread across a wide range, in-cluding moderate values for uncertain pixels. Deep neural net-works trained with binary cross-entropy and Dice loss tend to produce more polarised outputs, concentrating high wetland probability only in pixels where the model is highly confid-ent. In the validation sweep, the stricter threshold of 0.80 gave the best F1/IoU balance for the full-fusion network, because it removed many moderate-probability false positives while re-taining enough high-confidence wetland pixels to keep the pre-dicted extent close to the buffered reference.

boundary-aware fusion model (optimum near 0.35), the configuration adopted for the final ExtraTrees map.

Accuracy used confusion-matrix metrics with wetland as the positive class, defined in Equation (3):

P =   T P  ,        R =   T P  ,

(2)

F1 = 2 P R , IoU =    TP    .

(3)

  P+R     TP +FP +FN

3. Results

3.1. Boundary-Aware Reference-Label Availability

It may be noted here that the 20 m boundary exclusion sub-stantially changed the available reference labels and the ap-parent wetland prevalence (Table 2). Before exclusion, wet-lands represented 11.77% of the valid LiDAR-covered refer-ence area; after exclusion the retained core-wetland proportion was 7.29%. A total of 11,555,474 pixels were assigned to the uncertain boundary class and removed. Of these, 5,179,132 pixels were originally labelled as wetland (the wetland edge pixels just inside polygon boundaries) and 6,376,342 were ori-ginally labelled as non-wetland (the upland pixels immediately adjacent to polygon edges). The reduction in wetland percent-age from 11.77% to 7.29% reflects the high spatial density of wetland boundaries in the GeoNB inventory, which contains many small, fragmented patches and linear riparian corridors that are disproportionately affected by a two-pixel buffer. These boundary-aware labels were used for all subsequent sampling, training, threshold optimisation and accuracy assessment.

3.2. Effect of Multi-Source Fusion (ExtraTrees)

In spite of this limitation, the LiDAR-only ExtraTrees baseline reached a wetland F1-score of 0.425 and an IoU of 0.270 at threshold 0.25 (Table 3), so terrain and canopy variables on their own captured useful but limited wetland information. The hard-label LiDAR + Sentinel-2 + AlphaEarth fusion model reached 0.920 overall accuracy, 0.648 precision, 0.698 recall,

0.672 F1-score and 0.506 IoU at threshold 0.30, with 12.66% predicted wetland against an 11.77% reference share. Under the boundary-aware labels, the same fusion model scored 0.965 overall accuracy, 0.781 precision, 0.724 recall, 0.752 F1-score and 0.602 IoU at threshold 0.35 (Figure 7). Together, predictor fusion and boundary-uncertainty treatment produced the most reliable and spatially realistic ExtraTrees wetland prediction.

  TP +FP    TP +FN  

F1 = ^{2 P R}, IoU =   ^T P   .

(3)

  P+R    TP +FP +FN

Reporting precision, recall and F1-score, rather than overall accuracy alone, follows recommended practice for evaluating sampled remote-sensing classifications (Rangongo et al., 2024). Because non-wetland pixels dominated the landscape, overall accuracy alone was not interpreted as the primary performance indicator. Map-level realism was additionally checked by com-paring the predicted wetland proportion with the buffered refer-ence proportion.

3.3. Final Map Production

It may be noted here that the boundary-aware fusion model for each classifier was applied to the full valid 2024 LiDAR-covered domain, producing a continuous 10 m wetland prob-ability raster and, after thresholding, a binary wetland classi-fication raster. Pixels outside the LiDAR footprint or inside the 20 m boundary zone were assigned NoData in the evalu-ation step. Both probability and binary products were produced for each classifier, giving users the option to apply alternative thresholds according to their application-specific requirements.

LiDAR + Sentinel-2 + AlphaEarth fusion (hard labels) and the same fusion under 20 m boundary-aware labels, for overall accuracy, wetland precision, recall, F1-score and IoU. Wetland-specific metrics rise sharply once the embeddings are added and again under boundary-aware evaluation.

3.4. Effect of Boundary-Uncertainty Treatment

The hard-label ExtraTrees fusion model achieved a 0.672 F1-score and a 0.506 IoU; on the boundary-aware reference the same model scored 0.752 and 0.602. These values are not dir-ectly comparable because the evaluation domain itself changed when the 20 m boundary zone was removed. The point is not that every wetland edge was mapped more accurately, but that the model is assessed more defensibly when the most ambigu-ous pixels are set aside. The boundary-aware evaluation focuses on the model’s ability to correctly classify pixels that are con-fidently labelled as either core wetland or definite non-wetland, which represents a cleaner test of classifier performance. Map-level realism also held: the predicted wetland proportion under the boundary-aware evaluation was 6.75% against the 7.29% buffered reference, a difference of −4,548.9 ha.

Table 3. Accuracy and mapped-area comparison across the three ExtraTrees modelling stages, each at its optimised threshold (Thr.). Metrics are overall accuracy (Acc.), wetland precision (P), recall (R), F1-score and IoU; ground-truth (GT) and predicted (Pred) wetland percentages and the wetland area difference (∆ ha) summarise map realism. “Fusion” denotes the.

LiDAR + Sentinel-2 + AlphaEarth predictor stack.
Model	Labels	Thr.	Acc.	P	R	F1	IoU	GT/Pred %	∆ ha
LiDAR-only	Hard	0.25	0.863	0.420	0.431	0.425	0.270	11.77 / 12.07	+2,987.7
Fusion	Hard	0.30	0.920	0.648	0.698	0.672	0.506	11.77 / 12.66	+8,689.7
Fusion	20 m buffered	0.35	0.965	0.781	0.724	0.752	0.602	7.29 / 6.75	−4,548.9

Table 3. Accuracy and mapped-area comparison across the three ExtraTrees modelling stages, each at its optimised threshold (Thr.). Metrics are overall accuracy (Acc.), wetland precision (P), recall (R), F1-score and IoU; ground-truth (GT) and predicted (Pred) wetland percentages and the wetland area difference (∆ ha) summarise map realism. “Fusion” denotes the.

LiDAR + Sentinel-2 + AlphaEarth predictor stack.
Model	Labels	Thr.	Acc.	P	R	F1	IoU	GT/Pred %	∆ ha
LiDAR-only	Hard	0.25	0.863	0.420	0.431	0.425	0.270	11.77 / 12.07	+2,987.7
Fusion	Hard	0.30	0.920	0.648	0.698	0.672	0.506	11.77 / 12.66	+8,689.7
Fusion	20 m buffered	0.35	0.965	0.781	0.724	0.752	0.602	7.29 / 6.75	−4,548.9

3.5. ExtraTrees Final Accuracy and Map

On the basis of the threshold search, the final boundary-aware ExtraTrees model used threshold 0.35 and was evaluated on 85,299,502 pixels (Table 4). The confusion matrix gives TP = 4,502,056, TN = 77,823,497, FP = 1,259,532 and FN =

1,714,417. Overall accuracy is 96.51%, wetland precision 78.14%, recall 72.42%, F1-score 75.17% and IoU 60.22%. The predicted wetland area is 57,615.88 ha against a buffered refer-ence area of 62,164.73 ha, a difference of −4,548.85 ha (6.75% vs 7.29% reference). The model maintains relatively low com-mission while detecting most core wetland pixels, indicating that the embedding-dominated predictor stack drove spatially realistic classification. The probability and binary products are shown in Figure 8.

Table 4. Summary of the boundary-aware ExtraTrees fusion model at threshold 0.35, with its confusion-matrix counts, wetland-specific accuracy metrics and.

predicted-versus-reference wetland area
Metric	Value
Selected model	ExtraTrees ensemble
Predictors	LiDAR + S2 + AlphaEarth
Selected threshold	0.35
Valid evaluated pixels	85,299,502
Buffered ref. wetland %	7.29
Predicted wetland %	6.75
Buffered ref. area (ha)	62,164.73 Predicted wetland area (ha) 57,615.88 Area
difference (ha)	−4,548.85
Overall accuracy	96.51%
Wetland precision	78.14%
Wetland recall	72.42%
Wetland F1-score	75.17%
Wetland IoU	60.22%
TP / TN	4,502,056 / 77,823,497
FP / FN	1,259,532 / 1,714,417

Table 4. Summary of the boundary-aware ExtraTrees fusion model at threshold 0.35, with its confusion-matrix counts, wetland-specific accuracy metrics and.

predicted-versus-reference wetland area
Metric	Value
Selected model	ExtraTrees ensemble
Predictors	LiDAR + S2 + AlphaEarth
Selected threshold	0.35
Valid evaluated pixels	85,299,502
Buffered ref. wetland %	7.29
Predicted wetland %	6.75
Buffered ref. area (ha)	62,164.73 Predicted wetland area (ha) 57,615.88 Area
difference (ha)	−4,548.85
Overall accuracy	96.51%
Wetland precision	78.14%
Wetland recall	72.42%
Wetland F1-score	75.17%
Wetland IoU	60.22%
TP / TN	4,502,056 / 77,823,497
FP / FN	1,259,532 / 1,714,417

3.6. Predictor Importance: Dominance of AlphaEarth Embeddings

It is in this context that feature importance in the final Ex-traTrees model was concentrated in the AlphaEarth embeddings

LiDAR + Sentinel-2 + AlphaEarth model, in EPSG:2953 coordinates. Left: continuous wetland probability (colour bar 0–1). Right: binary wetland classification (wetland in green) after thresholding at 0.35. Pixels outside the LiDAR footprint are shown as NoData.

(Figure 9). The five highest-ranked predictors were all embed-ding dimensions (A22, A08, A03, A04 and A17), and embed-dings accounted for 19 of the top 20 positions; LiDAR intensity was the only non-embedding predictor to appear (rank 20), and no Sentinel-2 band reached the top 20. Although the LiDAR and Sentinel-2 layers remained useful, the embeddings sup-plied most of the model’s discriminative power. These dimen-sions are latent features and should not be read as individual physical variables; their collective importance indicates that the embeddings already encode much of the multi-sensor and con-textual landscape information that wetland discrimination re-quires. Impurity-based importance tends to favour numerous continuous features such as the 64 embedding dimensions, so the ranking reflects relative influence within this model rather than a controlled measure of marginal accuracy gain per pre-dictor group. For context, the LiDAR-only model ranked TWI, canopy height, intensity and curvature as its main contributors, consistent with the role of hydrological position and canopy structure in wetland occurrence.

3.7. Attention U-Net Ablation Performance

The ablation experiment demonstrated that the full multi-source fusion model produced the strongest wetland segmentation across all evaluated metrics (Table 5, Figure 10). The Sentinel-2-only configuration had the weakest performance, achieving 58.14% F1-score and 40.98% IoU, confirming that spectral re-flectance alone was insufficient for reliable wetland detection in this forested and topographically complex landscape where canopy cover and spectral similarity between wetland and up-land vegetation limit class separability. The LiDAR-only con-figuration achieved 70.22% F1-score and 54.10% IoU, a sub-stantial improvement that reflects the strong physical controls of terrain, drainage and canopy structure on wetland occur-rence. The AlphaEarth-only configuration performed simil-arly to the LiDAR-only model, achieving 69.68% F1-score and non-embedding feature to enter the list (rank 20), and no Sentinel-2 band reaches the top 20.53.45% IoU, which indicates that the latent embedding dimen-sions encode landscape information closely related to, though not identical to, the terrain and canopy structure captured by LiDAR. The full LiDAR–Sentinel-2–AlphaEarth fusion model achieved the strongest result with 96.92% accuracy, 84.19% precision, 68.90% recall, 75.78% F1-score and 61.01% IoU, confirming that the three predictor groups contribute comple-mentary information that is not fully captured by any single source.

3.8. Attention U-Net Threshold Sensitivity

Threshold testing showed that all Attention U-Net configura-tions were sensitive to the chosen probability cutoff, and that the correct threshold differed substantially from both the con-ventional 0.50 default and the ExtraTrees optimum of 0.35. The full-fusion model maintained the highest IoU across the tested range and peaked near the selected validation threshold of 0.80. The LiDAR-only and AlphaEarth-only configurations produced intermediate performance, while the Sentinel-2-only config-uration remained consistently lower at all thresholds tested. The precision–recall analysis showed that the full-fusion model achieved the best overall trade-off between wetland omission and commission errors, maintaining higher precision at com-parable recall levels relative to all single-source configurations (Figure 11). The threshold sensitivity curves for the Attention U-Net showed steeper IoU-versus-threshold relationships than those observed for ExtraTrees, consistent with the more po-larised probability distribution produced by the deep-learning model.

multi-source fusion model achieves the highest performance on both metrics, while the Sentinel-2-only configuration is consistently the weakest. LiDAR-only and AlphaEarth-only models perform similarly, each substantially below the full-fusion level.

Figure 10. Attention U-Net ablation performance across the four predictor configurations for F1-score and IoU. The full.

Figure 10. Attention U-Net ablation performance across the four predictor configurations for F1-score and IoU. The full.

Figure 11. Threshold sensitivity and precision–recall trade-off of the four Attention U-Net predictor configurations. The.

full-fusion model consistently achieves the highest IoU across thresholds and exhibits the strongest precision–recall balance.

3.9. Attention U-Net Final Accuracy and Map

The final selected Attention U-Net model was the boundary-aware full-fusion configuration at threshold 0.80. On the test set, the model correctly classified 5,310,694 non-wetland pixels and 277,980 wetland pixels, with 52,197 false-positive and 125,469 false-negative wetland pixels (Table 6). The model achieved 96.92% accuracy, 84.19% wetland precision, 68.90% recall, 75.78% F1-score and 61.01% IoU. The model was there-fore conservative: it reduced false wetland predictions but missed some true wetland pixels, with a false-negative rate con-siderably higher than the false-positive rate.

When applied to the full valid 2024 LiDAR-covered area, the Attention U-Net predicted 5,444,230 wetland pixels (79,855,272 non-wetland) from 85,299,502 valid evaluated pixels, corresponding to a predicted wetland proportion of 6.38% and a predicted wetland area of 54,442.30 ha (Table 7). This is slightly conservative relative to the 7.29% buffered ref-erence proportion, consistent with the model’s high-precision, lower-recall behaviour at threshold 0.80. The final probability and binary maps are shown in Figure 12.

3.10. Comparison of ExtraTrees and Attention U-Net

The Attention U-Net full-fusion model slightly outperformed the ExtraTrees fusion model in both F1-score and IoU (Table 8).

Figure 12. Final 10 m Attention U-Net wetland products for the 2024 LiDAR-covered study area generated by the.

Figure 12. Final 10 m Attention U-Net wetland products for the 2024 LiDAR-covered study area generated by the.

boundary-aware full-fusion model. Left: Continuous wetland probability map (0–1). Right: Binary wetland classification obtained using the optimal threshold of 0.80, with predicted wetland areas shown in green. Areas outside the LiDAR footprint are displayed as NoData.

ExtraTrees achieved 75.17% F1-score and 60.22% IoU evalu-ated on the full 85.3 M-pixel domain, while Attention U-Net achieved 75.78% F1-score and 61.01% IoU on its test set (ap-proximately 5.8 M pixels). Direct metric comparison should be interpreted with care because the evaluation domains dif-fer in size; full-raster wetland area statistics, derived from ap-plying each model to the complete valid domain, are more directly comparable. Both classifiers produced conservative full-raster maps relative to the 7.29% buffered reference: Ex-traTrees predicted 6.75% (57,615.88 ha) and Attention U-Net predicted 6.38% (54,442.30 ha). The Attention U-Net achieved higher wetland precision (84.19% vs 78.14%) but lower recall (68.90% vs 72.42%), indicating a more conservative detection behaviour that generates fewer false positives at the cost of missing more true wetland pixels.

4. Conclusions

On the basis of the results, a boundary-aware 10 m wet-land mapping framework was developed for the 2024 GeoNB LiDAR-covered portion of New Brunswick, fusing LiDAR-derived terrain and canopy structure with Sentinel-2 multis-pectral bands and 64-dimensional AlphaEarth latent geospa-tial embeddings, on top of a 20 m exclusion zone that removes ambiguous wetland-edge pixels from training and evaluation. Two classifiers were developed and directly compared under identical boundary-aware labels and evaluation protocols: an ExtraTrees ensemble and an Attention U-Net segmentation model. The Attention U-Net was further tested in a predictor-group ablation experiment across LiDAR-only, Sentinel-2-only, AlphaEarth-only and full-fusion configurations.

The LiDAR-only ExtraTrees baseline gave moderate wetland discrimination (F1 = 0.425, IoU = 0.270), establishing the contribution of terrain, hydrological position and canopy structure as a baseline. The full ExtraTrees fusion model reached F1 = 0.672 and IoU = 0.506 under hard labels, and F1 = 0.752 and IoU = 0.602 under boundary-aware labels, with 96.51% overall accuracy and a predicted wetland share of 6.75% against a buffered reference of 7.29%. The Attention U-Net ablation independently confirmed the value of multi-source fusion: Sentinel-2-only (58.14% F1), LiDAR-only (70.22% F1) and AlphaEarth-only (69.68% F1) configurations all un-derperformed the full-fusion model, which achieved 96.92% accuracy, 84.19% precision, 68.90% recall, 75.78% F1-score

and 61.01% IoU, mapping 54,442.30 ha (6.38%) of wetland.

Feature importance in the ExtraTrees model was dominated by the AlphaEarth dimensions, which accounted for 19 of the 20 most important predictors, ahead of all Sentinel-2 and nearly all LiDAR predictors. The comparable performance of the AlphaEarth-only and LiDAR-only Attention U-Net configur-ations further confirms that precomputed embeddings encode landscape-relevant information sufficient to approach terrain-based discrimination, and that combining all three predictor groups yields complementary gains. The Attention U-Net slightly outperformed ExtraTrees in F1-score and IoU but was more conservative, achieving higher precision at the cost of lower recall; the overall performance difference was modest, indicating that both tree-based and deep-learning classifiers can effectively use the AlphaEarth embedding stack.

These results demonstrate that precomputed satellite embed-dings can serve as the principal predictor for high-resolution wetland mapping, that a 20 m boundary-uncertainty exclusion zone provides a more defensible evaluation framework for polygon-derived reference data, and that both ExtraTrees and Attention U-Net classifiers can produce accurate, spatially real-istic wetland maps from public geospatial data without the multi-temporal image acquisition and preprocessing burden of conventional Sentinel-2 workflows.