This wetland vegetation map for Narran Lakes was produced using a machine learning classification workflow with cluster-guided training, following the approach demonstrated in Wen et al. (2025), and further adapted using the clustering and AlphaEarth embedding-augmented workflow described in Ryan et al. (2026). For this release, the workflow was expanded through the inclusion of ELVIS 1m LiDAR-derived structural predictors, alongside multi-temporal Sentinel-1 radar, Sentinel-2 optical time series, terrain variables and AlphaEarth embeddings.

The product is intended as a landscape-scale baseline for environmental water planning, Murray-Darling Basin (MDB) reporting and monitoring, and conservation management.

Analysis period and map window

This release uses a primary water-year analysis period of June 2024-June 2025, with a spring-summer map window of September 2024-March 2025 used for selected seasonal metrics, wetland-state interpretation, and open water mask. LiDAR-derived structural predictors (ELVIS 1m) were derived from acquisitions collected across the 2023-2024 period, and AlphaEarth embeddings were sourced from the 2024-2025 annual Google Satellite Embedding collection in Google Earth Engine (GEE).

Predictors

Predictors were assembled in GEE. Sentinel-1 and Sentinel-2 metrics were generated in GEE and combined with AlphaEarth embeddings and externally processed ELVIS 1 m LiDAR structural layers.

• Sentinel-2 optical spectral-temporal predictors: vegetation, moisture and wetness indices, seasonal summaries, phenology, texture and hydro-proxy layers.
• Sentinel-1 radar predictors: VV/VH backscatter, angle/orbit-normalised metrics and temporal summaries capturing moisture, inundation and structural variation.
• LiDAR structural predictors: ELVIS 1m height-above-ground (HAG) - derived metrics representing vegetation height, cover, vertical structure and structural heterogeneity.
• Terrain predictor: hydrologically fitted Digital Elevation Model (DEM).
• Embedding-based predictors: AlphaEarth embeddings representing learned spatial, spectral and temporal landscape-contextual features.

Training data

Cluster-guided sampling was used to generate training data across a spatially and temporally variable wetland landscape with limited ground sample coverage. Following Wen et al. (2025) and Ryan et al. (2026), X-means clustering was used to refine cluster structure without requiring a fixed user-defined k. Clusters were reviewed by an expert vegetation ecologist using high-resolution imagery, hydrological/topographic context and existing field observations, then used to guide training-point selection and assign labels, including NSW PCT-aligned classes.

Modelling

Random Forest (RF) modelling followed the general framework applied in Wen et al. (2025) and Ryan et al. (2026), with this release extending the evaluation design through spatially blocked cross-validation (CV) and internal holdout testing. Prior to the vegetation-only classification, a preliminary RF model was run using the full predictor stack to identify and mask broad context/non-target areas, including 'Open water' and 'Cropped/Disturbed' areas. These areas were retained as context classes in the final map product but were excluded from the reported vegetation accuracy assessment metrics. The vegetation classification workflow then included GroupKFold CV using 1 km spatial blocks, an 80/20 train-test split, near-zero variance and correlation-based feature filtering, and RF hyperparameter tuning under blocked 5-fold CV.

The training dataset comprised approximately 3,071 samples organised into 188 spatial blocks of 1 km.

Model accuracy assessment

Model performance for this release was assessed using spatially blocked cross-validation based on 1 km GroupKFold partitions, with additional internal holdout testing. Reported accuracy metrics are based on the mapped vegetation classes only; context/non-target classes were excluded from the summary accuracy calculations. Summary metrics include Overall Accuracy (OA), Cohen’s Kappa (κ), macro-F1 and balanced accuracy.

• MER Functional Groups (8 classes): OA = 0.91, κ = 0.90, macro-F1 = 0.90, balanced accuracy = 0.90;
• NSW Vegetation Formations (8 classes): OA = 0.91, κ = 0.89, macro-F1 = 0.88, balanced accuracy = 0.87;
• NSW Vegetation Classes (10 classes): OA = 0.91, κ = 0.90, macro-F1 = 0.89, balanced accuracy = 0.88;
• NSW PCTs (23 classes): OA = 0.89, κ = 0.88, macro-F1 = 0.88, balanced accuracy = 0.87.

Post-processing and manual edits

Post-processing included edge-aware smoothing, gap filling, local majority filtering, morphological refinement and class-specific minimum mapping units (MMU): ~0.2 ha. Final outputs were reviewed and manually refined by an expert vegetation ecologist to address residual artefacts, boundary inconsistencies and selected low-confidence areas, particularly disturbed lake margins and residual open water or cropped areas not removed by masking.

Class hierarchy

Vegetation classes were assigned at PCT level using the NSW BioNet Vegetation Classification (https://vegetation.bionet.nsw.gov.au/) and aggregated to Vegetation Class and Vegetation Formation levels within the NSW nested hierarchy (https://www.environment.nsw.gov.au/topics/animals-and-plants/biodiversity/nsw-bionet/the-nsw-vegetation-classification-framework). To support broader MDB reporting, wetland PCTs were also aligned to indicative Monitoring, Evaluation and Reporting (MER) functional groups, using analogous groupings from comparable wetlands and long-term water planning frameworks.

Key fields dictionary

‘PCT_ID’ (PCT Code); ‘PCT_Desc’ (PCT Name); ‘Veg_Class’ (NSW Vegetation Class); ‘Veg_Form’ (NSW Vegetation Formation); ‘MER_FG’ (MER Functional Group for LTWP reporting); ‘Hectares’ (polygon area). Context classes (‘Cropped/Disturbed’, ‘Open water’, ‘Dam') are included in the final map for completeness but were excluded from the reported vegetation accuracy assessment metrics.

Intended use

This product is intended as a landscape-scale baseline for environmental water planning, monitoring and reporting, conservation management, and long-term wetland condition assessment and change detection. MER functional group alignments are indicative and intended to support broader reporting consistency. The product is not intended for statutory or site-scale assessment without targeted field verification.

Input data limitations

• Satellite predictors: Sentinel-1 and Sentinel-2 predictors may be affected by cloud, inundation timing/state, sensor geometry, and radar backscatter variability in dynamic wetland conditions.
• Temporal alignment: Temporal mismatch among Sentinel-1, Sentinel-2, LiDAR, and ancillary layers can introduce predictor inconsistency when vegetation structure/hydrology changes between acquisition dates.
• LiDAR-derived structural metrics: LiDAR predictors are sensitive to point-cloud density, ground/vegetation classification, HAG normalisation, and rasterisation choices, with errors in ground modelling, point classification, or interpolation potentially propagating into derived structural predictors.
• Class separability: Spectral and structural similarity in transitional or mixed vegetation communities, particularly under variable inundation or disturbance conditions, may result in class confusion in some areas.

Validation scope

Accuracy metrics should be interpreted as internal estimates from blocked spatial cross-validation and internal holdout testing. They represent raw model performance prior to post-processing and expert editing. A withheld, independent ground-based validation dataset collected separately from model training will be used to validate the final map product in a future release. Users requiring statutory or site-scale confidence should undertake targeted field validation.

Versioning

This version is ‘Version: v1.0’ (release date: 2026-04-XX). Results are version controlled. Updates will include improvements to training data, predictors, and post-processing

Acknowledgements

This mapping project was supported by the NSW Water for the Environment Program and developed within the Department of Climate Change, Energy, the Environment and Water (DCCEEW).

Related publications

Wen, L., Ryan, S., Powell, M., and Ling, J.E. (2025). From Clusters to Communities: Enhancing Wetland Vegetation Mapping Using Unsupervised and Supervised Synergy. Remote Sensing, 17(13): 2279. https://doi.org/10.3390/rs17132279

Ryan, S., Powell, M., Ling, J.E., and Wen, L. (2026). Streamlining Wetland Vegetation Mapping with AlphaEarth Embeddings: Comparable Accuracy to Traditional Methods with Cleaner Maps and Minimal Preprocessing. Remote Sensing, 18(2): 293. https://doi.org/10.3390/rs18020293