Shallow benthic and geomorphic habitat maps for Northern and Western Australia (NESP MaC 3.17, UQ, UNSW, AIMS)

This dataset corresponds to high resolution (10 m) raster shallow benthic and geomorphic habitat maps for Northern and Western Australia estimated from Sentinel 2 composite imagery from 2018 – 2023. Benthic classes include sand, rubble, rock, seagrass, coral/algae, microalgal mats and light seagrass. Geomorphic classes include deep, sediment slope, shallow lagoon, deep lagoon, inner reef flat, outer reef flat, reef crest, terrestrial reef flat, sheltered reef slope, plateau, back reef slope, small reef and rocky reef. This dataset covers the area from Houtman Abrolhos Islands in Western Australia through to the northwestern side of Cape York, including both offshore and inshore reef systems. Classifications are limited to shallow regions, just below lowest astronomical tide in turbid areas, and to 10 - 15 m in clear water areas.

These maps were developed by extending the methods used in the Allen Coral Atlas (https://allencoralatlas.org/methods/) and the development of habitat maps for the Great Barrier Reef (GBR10 GBRMP Geomorphic, https://arcg.is/1jfWaa1, GBR10 GBRMP Benthic, https://arcg.is/1GOD4T1). The maps were produced using a semi-automated classification workflow implemented in Google Earth Engine, combining improved low-tide Sentinel-2 satellite imagery composites created by the Australian Institute of Marine Science (AIMS) (https://doi.org/10.26274/2bfv-e921), with Random Forest machine learning classifiers. The classification approach was regionally tailored across five subregions (Shark Bay, West, Northwest, Gulf, and Offshore) to account for differences in water column optical properties and habitat types. Classifications are aligned with updated reef and shallow sediment outlines produced by AIMS as part of this project, and follow conventions from the Allen Coral Atlas and Great Barrier Reef mapping projects, incorporating additional classes to better represent seagrass environments characteristic of this region.

The workflow integrated expert visual interpretation of reference imagery, Object-Based Image Analysis (OBIA) for training data development, and iterative refinement through object-based cleanup rules and regional expert review.

These maps are intended to support regional-scale habitat assessment, marine spatial planning, ecosystem modelling, environmental impact assessment, and prioritisation of monitoring efforts. The dataset provides a consistent and scalable baseline for future reef monitoring and contributes directly to the national reef mapping framework.

This dataset is delivered in three parts: 1. Geomorphic Map (geomorphic/NW_NESP-MaC-3-17_UQ_Shallow-habitat_Geomorphic_2025.tif) High-resolution spatial classification of coral reef, rocky reef, and shallow sediment geomorphic zones across five subregions. The map features 14 classes: Deep Water, Sediment Slope, Shallow Lagoon, Deep Lagoon, Inner Reef Flat, Outer Reef Flat, Reef Crest, Terrestrial Reef Flat, Sheltered Reef Slope, Reef Slope, Plateau, Back Reef Slope, Small Reef, and Rocky Reef. Classifications were produced at 10 m spatial resolution and refined through three stages of cleanup, including object-based rules and expert-guided manual corrections. Geomorphic maps underwent accuracy assessment using validation points generated from expert-interpreted reference segments.

1. Benthic Habitat Map (benthic/NW_NESP-MaC-3-17_UQ_Shallow-habitat_Benthic_2025.tif) High-resolution spatial classification of benthic habitats within shallow reef and sediment environments across five subregions. The map features 7 classes: Sand, Rubble, Rock, Seagrass, Coral/Algae, Microalgal Mats, and Light Seagrass. Classifications were produced at 10 m spatial resolution and limited to areas shallower than 10 m depth. Benthic maps were refined through two stages of cleanup that incorporated both generic ecological rules (enforcing relationships between benthic and geomorphic classes) and region-specific manual corrections guided by expert review. Benthic maps underwent accuracy assessment using validation points generated from expert-interpreted reference segments.
2. Training Data (reference_training_csv/*.csv) Point-based training datasets used to calibrate Random Forest classifiers for geomorphic and benthic habitat mapping. Separate training datasets are provided for each subregion and classification type (geomorphic and benthic). Each dataset contains up to 1,400 calibration points per habitat class, combined with the complete suite of predictor variables including spectral reflectance, bathymetric depth and slope, local variance metrics, GLCM texture measures, and band ratios. Training points were generated through stratified random sampling within expert-interpreted reference segments, with targeted manual enhancement in areas of poor initial classification performance. These datasets support reproducibility, alternative modelling approaches, and extension of mapping methods to adjacent regions with similar environmental conditions.

Methods:

Regional Framework: The project region was divided into five subregions (Shark Bay, West, Northwest, Gulf, and Offshore) based on differences in water column optical properties and dominant habitat types. Separate classifiers were trained and applied independently for each subregion to account for regional variation in spectral characteristics, though the same suite of predictor variables was used consistently across all subregions.

Data Compilation: Benthic and environmental datasets were compiled from 2021–2024, including georeferenced transect surveys (Reef Life Survey, DBCA monitoring programs), AIMS photo quadrats and video tows, JCU TropWATER seagrass datasets, Seamap Australia records, and high-resolution drone imagery. Data were assessed for spatial accuracy and ecological relevance, with positional uncertainty corrected through visual alignment with satellite imagery where possible. A full list of datasets and attribution is provided in the project final report. Bathymetric layers were collated from Geoscience Australia, AIMS, and state repositories as listed above. Coverage varied by region, with finer resolution available for the Kimberley and Pilbara coastlines.

Reference Dataset Creation: Training data for the habitat classification were created based on the interpretation of the available field data in combination with satellite imagery. An expert selected reference quadrats (a small patch where the benthic and geomorphic classification could be determined with high confidence) across the project region based on clear habitat visibility in Sentinel-2 or Planet imagery, availability of supporting field data, presence of aerial/drone imagery, and geographic coverage across reef types and turbidity zones. Each quadrat was processed using Object-Based Image Analysis (OBIA) in Trimble eCognition software. Separate workspaces were created for benthic (scale parameter = 10) and geomorphic (scale parameter = 50–80) classification. This broke the image into small patches where the texture and colour was similar. RGB bands from Sentinel-2 or Planet Dove served as primary inputs, with blue band emphasis used in deeper or turbid conditions. Segments were assigned habitat class labels only when confident interpretation was supported by field data, high-resolution imagery, Seamap Australia layers, or expert visual interpretation. Two new classes were introduced compared with the Allen Coral Atlas: Sediment Slope and Rocky Reef (using AIMS-supplied outlines). These were needed as the new mapping covers substantial non coral reef areas.

Data Stack Creation in Google Earth Engine:

Satellite Composites:

Two image products were evaluated for region-specific conditions: - UQ Dry Season Sentinel-2 Stack: Generated from cloud-free Sentinel-2 scenes during dry seasons 2021–2023. A dark pixel composite approach estimated and subtracted surface reflectance noise using optically deep regions, enhancing bottom visibility and reducing sun-glint. Statistical percentiles (median, minimum) were computed per pixel for a cloud- and haze-free composite. This was used for classification for offshore reefs with clearer water and stable seasonal conditions. - AIMS Low-Tide Sentinel 2 Composite (Hammerton & Lawrey, 2024): This composite used sun glint correction and was made up of the 10 lowest tide images in each location, from 2018 - 2023.This imagery was selected as primary input for inshore reef and coastal environments due to superior performance in high turbidity zones.

Segmentation and Covariate Stack:

Preprocessing in GEE included land masking (Australian Coastline 50K 2024 - Simplified) rasterised and inverted, and bathymetry standardization by multiplying by -1 and converting to positive integer scales. Slope was calculated from bathymetry in degrees and scaled by 100. Two distinct predictor stacks were generated, segment and pixel based: - Segment Stack (Object-Based): SNIC (Simple Non-Iterative Clustering) segmentation was applied to RGB reflectance bands using hexagonal seed grid spacing, with a scale parameter of 20–80, compactness of 10, neighborhood size of twice the scale factor (40–160), and 4-connectivity. Object-level means were calculated for RGB bands, depth, and slope using connected components reduction (maximum object size 1,500 pixels). Local variance in depth was calculated at the pixel level using a 5-pixel circular kernel (standard deviation ×10). Three band ratios were then computed: Red/Blue, Green/Blue, and Red/Green (multiplied by 1,000 and converted to int16). The final predictor stack contained 8 variables: RGB segment means, depth segment mean, depth pixel variance, and 3 segment-based band ratios. - Pixel Stack (Pixel-Based): No segmentation was applied. Pixel-level metrics included RGB reflectance, depth, and slope. Local variance metrics were calculated using a 5-pixel circular kernel: depth standard deviation (×10) and red band standard deviation (×10). GLCM texture metrics included depth entropy (maximum neighborhood with focal window of 5), blue band sum average (focal median of 5), and red band sum average (focal median of 5). Three band ratios were computed as above. All metrics were converted to int16 for computational efficiency. The final predictor stack contained 12 variables: RGB pixel reflectance, depth, 2 variance metrics, 3 GLCM texture metrics, and 3 pixel-based band ratios.

Processed stacks were exported to GEE assets at 10 m resolution for classification.

Calibration and Validation Point Creation:

The calibration and validation datasets were derived from the previously created reference segments. These segments, developed through OBIA in eCognition and tagged with high-confidence benthic or geomorphic labels, were exported into ArcGIS Pro and separated by region. Each subregion was treated independently to preserve regional contextual accuracy during sampling and classification. Labelled reference segments were uploaded into GEE as assets. Within the labelled reference segment polygons, two independent sets of 1,000 random points per class were generated: one for calibration (random seed 42) and one for validation (random seed 666). Both sets were sampled from the same reference polygon library, with different random seeds ensuring spatial separation and independence between training and validation datasets. This stratified random sampling ensured even representation of each class.

Targeted Training Enhancement:

Following initial classification runs, visual inspection identified areas where class discrimination was poor, typically due to highly variable water quality and environmental conditions along the northern Australian coast, or insufficient training point coverage in specific areas. To address these deficiencies, additional training points with known class labels were manually placed in problem areas. These manual points were stratified-sampled to a maximum of 400 points per class to prevent over-representation, then merged with the original 1,000 auto-generated calibration points. This process resulted in final calibration datasets of up to 1,400 points per class (1,000 auto-generated + up to 400 manual). The validation dataset remained unchanged at 1,000 points per class to ensure independent performance evaluation. This targeted enhancement was applied to all regions for both geomorphic and benthic classifications. Once generated, all calibration points were drilled through the data stack within GEE to extract predictor variable values at each location. Training points with null values in any predictor band were filtered out prior to model training. The resulting training datasets contain the complete suite of predictor variables from both pixel-based and segment-based data stacks, including RGB reflectance, depth, slope, variance metrics, GLCM texture metrics, and band ratios. While the classifier selected specific subsets of these variables (8 for geomorphic, 12 for benthic), the full predictor stack was retained in the published training datasets to support alternative analyses and model development.

Classifier Development:

The classification process was conducted using a Random Forest classifier implemented within GEE. Random Forest is a robust ensemble learning method that constructs multiple decision trees and aggregates their outputs to produce a more accurate and stable classification result (James et al., 2013). The classifier was configured with 1,000 trees, a minimum leaf population of 2, and variables per split set to the default square root of the number of predictor variables. A random seed (42) was set to ensure reproducibility across runs. Training data for the classifier came from the calibration points created in the previous step, each enriched with predictor variables derived from the composite satellite imagery and bathymetry stack.

To improve classification accuracy and ensure ecological and geographic consistency, three separate classifiers were developed and applied independently to three major habitat categories defined by the AIMS reef outlines (v0-3): Coral Reefs, Rocky Reefs, and Shallow Sediment environments. This approach aligns with the NESP Common Language for Marine Habitat Classification and recognizes that the broader environmental context of each habitat type differs substantially. By separating the classification into three discrete habitat contexts, the model was provided with more coherent environmental backdrops, reducing misclassification between spectrally similar but ecologically distinct features. For each region, separate Random Forest classifiers were trained and run within each AIMS outline category, using only the relevant calibration points and predictor variables.

Special Treatment of Rocky Reefs: Rocky reefs presented a unique challenge for geomorphic classification. Because rocky reefs are formed by non-carbonate geological processes rather than by reef-building organisms, the standard coral reef geomorphic zonation scheme (reef flat, reef crest, reef slope, etc.) is not ecologically appropriate. Therefore, for geomorphic mapping, all areas within rocky reef outlines were assigned a single class "Rocky Reef" (class 27) rather than being classified into coral reef-derived geomorphic zones. Benthic classification within rocky reef areas proceeded normally using the standard benthic class scheme. This helped to ensure that fine scale sediment areas were separated out from the relatively coarsely mapped rocky reef boundaries. After classification, the three outputs—coral, rocky, and sediment—were merged into a single benthic or geomorphic map prior to the cleanup and refinement stage.

Post-Classification Refinement:

The refinement process followed a structured, iterative approach for improving both the geomorphic and benthic classification maps through object-based image analysis and manual correction. Geomorphic maps underwent three cleanup stages, while benthic maps underwent two stages.

Geomorphic Refinement (3 stages): - Stage 1 (Initial Noise Reduction): A small-object filter removed features below 196 pixels to reduce noise and complexity. Areas deeper than 20 m were reclassified as Deep (2) based on bathymetry data. Terrestrial Reef Flat beyond 2,000 m from land was converted to Outer Reef Flat. Focal mode smoothing (radius 1 pixel, 2 iterations) was applied, and small objects were replaced with the smoothed values. - Stage 2 (Fast OBIA Cleanup): A broader focal mode smoothing (radius 3 pixels, 2 iterations) captured neighborhood context. Objects were identified using connected components analysis, and small objects (based on size threshold) were replaced with the underlying smoothed classification to remove spatial inconsistencies and refine boundaries. - Stage 3 (Manual Refinement): Region-specific manual adjustments were applied using hand-drawn polygon masks and custom rules. Corrections included: • Lagoon-based rules (converting classes within or around lagoon boundaries, with different treatments for oceanic vs. closed lagoons) • Mid-zone masking to remove misclassified areas • Distance-based buffering and cleaning (e.g., 1,000 m buffer around lagoons) • Class-specific conversions guided by geomorphic context • Depth-based corrections where applicable • Rocky reef delineation using externally supplied masks

Benthic Refinement (2 stages): - Stage 1 (Initial Refinement and OBIA Rules): Small-object filter removed features below 10 pixels. Shallow no-data areas were reclaimed using focal mode smoothing (radius 3 pixels, 2 iterations). Benthic mapping was limited to areas shallower than 7.5 m depth. Generic OBIA rules were applied to enforce ecological relationships between benthic and geomorphic classes. - Stage 2 (Manual Refinement): Region-specific manual adjustments were applied using hand-drawn polygon masks. The final geomorphic map (Stage 3) was used to guide benthic corrections. Manual rules included class conversions based on local context (e.g., Sand to Rock, Rubble to Coral/Algae, Seagrass to Coral in specific zones) and application of benthic masks to remove misclassified areas.

Expert Review and Final QA:

Following Stage 3 (geomorphic) and Stage 2 (benthic) cleanup, maps were uploaded to a custom GEE viewer app and circulated to regional experts familiar with the target environments: Dr Mitch Lyons (UNSW), Dr Emma Kennedy (AIMS), Dr Alex Ordoñes Alvarez (Southern Cross University), Dr Tom Holmes and Dr Claire Ross (DBCA Perth), Dr Zoe Richards (WA Museum/Curtin University), Dr Katie Chartrand (TropWATER JCU), and Assoc Prof Chris Roelfsema (UQ). Feedback on key habitats and problem areas informed localized corrections applied prior to final export. These expert-guided refinements are not reflected in the accuracy assessment statistics to preserve validation independence.

Final quality assurance involved detailed visual inspection comparing maps with field data, bathymetry, and satellite imagery to ensure spatial distribution aligned with ecological expectations. Refined maps were depth-masked and exported to the final product archive.

Accuracy Assessment:

Accuracy was assessed using an independent set of validation points withheld from the training process. Validation datasets consisted of 1,000 randomly generated points per class (random seed 666). Assessments were conducted separately for benthic and geomorphic classifications in each region.

Validation points were sampled through the classified maps at 10 m resolution within GEE. Confusion matrices were generated comparing the mapped class at each validation point with the known reference class. Standard accuracy metrics were calculated following Congalton & Green (2008), including: • Overall Accuracy – the percentage of total validation points correctly classified • User's Accuracy – the probability that a map pixel labelled as a given class is actually that class on the ground (commission error) • Producer's Accuracy – the probability that a reference point of a given class is correctly mapped as that class (omission error) Classes with no training or validation data were excluded from the confusion matrix and accuracy statistics. Additionally, externally derived classes, such as deep water, deep lagoon, and rocky reef outlines contributed by AIMS, were excluded to ensure that the reported metrics reflect only those classifications produced within the UQ mapping workflow. Full validation results are summarised for each region and published in the final report, with full confusion matrices and per-class accuracy tables provided.

Limitations of the Data: - Optical Depth Constraints: Benthic mapping is limited to approximately 10 m depth due to the penetration limits of satellite imagery. Features deeper than this may be misclassified or omitted. - Spectral Confusion: Coral and macroalgae cannot be reliably distinguished using Sentinel-2 imagery due to similar spectral signatures. These are grouped into a combined class. - Validation Limitations: Validation data were derived from expert interpretation of satellite imagery, not from purpose-collected in-situ surveys. Spatial correlation between training and validation samples may inflate accuracy estimates. - Bathymetry Resolution: Bathymetric data quality varies across regions. Coarse resolution in areas like the Gulf of Carpentaria limits the precision of slope and depth-based classification. - Sparse Benthic Cover: Features with low benthic cover (> 10%) are difficult to detect at 10 m resolution and may be underrepresented. - Data Availability: Some regions lacked sufficient field data, requiring reliance on expert interpretation and external outlines. This may affect classification reliability in data-poor zones. - External Class Dependencies: Classes such as rocky reef and deep lagoon were informed by external outlines and not generated directly from classification, limiting reproducibility. - Sand bank geomorphic confusion: Mapped areas include shallow sand banks. The geomorphic classification scheme treats these sandy areas as though they were part of a coral reef, leading to potentially confusing results. Sand banks are typically represented as Reef Slope, Back Reef Slope, Sheltered Reef Slope, with the specific classification primarily related to the depth of the sand bank.

Format of the Data: - NW_NESP-MaC-3-17_UQ_Shallow-habitat_Benthic_2025.tif: GeoTIFF raster, internal tiles, overview pyramids, 10 m resolution, EPSG:4326. Each pixel represents the dominant benthic cover class. - NW_NESP-MaC-3-17_UQ_Shallow-habitat_Geomorphic_2025.tif: GeoTIFF raster, internal tiles, overview pyramids, 10 m resolution, EPSG:4326. Each pixel represents the dominant geomorphic zone. - Calibration/validation points: CSV

Spatial reference: Geographic coordinate system: WGS 1984, EPSG: 4326, Units: degrees, Spatial resolution: 10 m

The classes used for the geomorphic and benthic maps are described as per Lyons et al (2024). They are in line with the classes used for the Allen Coral Atlas and the Great Barrier Reef habitat maps, with the addition of two geomorphic categories – Sediment Slope and Rocky Reefs, and the benthic category Sparse Seagrass. These additional classes were required to describe habitats not covered in the previous mapping projects.

Geomorphic Class Descriptions: Geomorphic zones describe the physical structure and formation of reef features. These zones guide understanding of energy regimes, substrate type, and biological potential. The classification schema is based on the Reef Cover standard (Kennedy et al., 2021). Class Code | Geomorphic Zone | Full Description 2 | Deep Water | Water deeper than mapping threshold (~20 m) 10 | Sediment Slope | A shallow sloping area of unconsolidated sediments - sand, silt or mud 11 | Shallow Lagoon | Any fully- to semi-enclosed, sheltered, flat-bottomed, sediment-dominated lagoon area shallower than approximately 5 m 12 | Deep Lagoon | Any sheltered broad body of water, fully- to semi-enclosed by reef, with a variable depth (but deeper than approximately 5 m and shallower than the surrounding ocean) and a soft bottom dominated by reef-derived sediment 13 | Inner Reef Flat | A low-energy, sediment-dominated, horizontal to gently sloping platform behind the outer reef flat 14 | Outer Reef Flat | Adjacent to the seaward edge of the reef, outer reef flat is a levelled (near horizontal), broad, and shallow carbonate platform, displaying distinct wave-driven zonation 15 | Reef Crest | A zone marking the boundary between the reef flat and the reef slope, generally shallow and characterized by highest wave energy absorbance 16 | Terrestrial Reef Flat | A broad, flat, shallow-to-semi-exposed area fringing reef flat, found directly attached to land at one side; it is subject to freshwater run-off, nutrients, and sedimentation 21 | Sheltered Reef Slope | Any submerged, sloping area extending into deep water but protected from strong directional prevailing wind or current, either by land or by opposing reef structures 22 | Reef Slope | A submerged, sloping area extending seaward from the reef crest (or flat) toward the shelf break; windward-facing or any direction if no dominant prevailing wind or current exists 23 | Plateau | Any deep, submerged (> approximately 5 m), hard-bottomed, horizontal to gently sloping (angle shallower than approximately 10°), seaward-facing reef platform 24 | Back Reef Slope | A complex, interior—often gently sloping—reef zone occurring behind the reef flat; of variable depth (but deeper than reef flat and more sloped), it is sheltered, sediment-dominated, and often punctuated by coral outcrops 26 | Small Reef | Small isolated reef features 27 | Rocky Reef | A rocky outcrop or reef not created by carbonate organisms - determined by the AIMS outlines

Benthic Class Descriptions: Benthic classes represent the dominant biological or substrate cover in shallow water. These were derived using expert interpretation of segmented satellite imagery. Class Code | Benthic Habitat | Full Description 11 | Sand | Any soft-bottom area dominated by fine unconsolidated sediments. Can include sand, mud or silt in this class 12 | Rubble | Any habitat featuring loose, rough fragments of broken reef material 13 | Rock | Any exposed hard-bottom area with uncommon-to-scarce corals and fleshy macroalgae—it encompasses limestone reef matrix but also underlying non-reefal bedrock and "beach rock" 14 | Seagrass | Any habitat where seagrass is the dominant biota 15 | Coral/Algae | Any hard-bottom area supporting living coral and/or algae 18 | Microalgal Mats | Visible accumulations of microscopic algae in sandy sediments 19 | Light Seagrass | Any habitat covered in sparse seagrass

Benthic and Geomorphic Training Data - Data Dictionary These datasets contain the calibration point data used to train Random Forest classifiers for geomorphic and benthic habitat mapping across five subregions of Northern and Western Australia (Shark Bay, West, Northwest, Gulf, and Offshore). Separate training datasets were created for geomorphic zonation and benthic habitat classification in each region. Both geomorphic and benthic training datasets share identical column structure. The class_num field contains different class codes depending on the dataset type (see class reference tables above). All predictor variables are present in both datasets, though different subsets were used in classification: geomorphic used 8 segment-based variables, while benthic used 12 pixel-based variables. Column Name | Description | Units/Type system:index | Unique identifier for each training point | String b1_p | Blue band mean reflectance (pixel-based) | Integer (scaled) b1_s | Blue band mean reflectance (segment-based) | Integer (scaled) b1_savg_p | Blue band GLCM sum average texture metric (pixel-based) | Integer b2_p | Green band mean reflectance (pixel-based) | Integer (scaled) b2_s | Green band mean reflectance (segment-based) | Integer (scaled) b2_savg_p | Green band GLCM sum average texture metric (pixel-based) | Integer b3_p | Red band mean reflectance (pixel-based) | Integer (scaled) b3_s | Red band mean reflectance (segment-based) | Integer (scaled) class_num | Benthic habitat class code (11=Sand, 12=Rubble, 13=Rock, 14=Seagrass, 15=Coral/Algae, 18=Microalgal mats, 19=Light seagrass)

Geomorphic habitat class code (10=Sediment Slope, 11=Shallow Lagoon, 12=Deep Lagoon, 13=Inner Reef Flat, 14=Outer Reef Flat, 15=Reef Crest, 16=Terrestrial Reef Flat, 21=Sheltered Reef Slope, 22=Reef Slope, 23=Plateau, 24=Back Reef Slope, 26=Small Reef) | Integer depth_maxent_p | Depth entropy GLCM texture metric (pixel-based, focal window 5, ×100) | Integer (scaled) depth_p | Bathymetric depth (pixel-based) | metres depth_s | Bathymetric depth (segment mean) | metres depth_stdDev_p | Local standard deviation of depth in 5-pixel circular kernel (×10) | Integer (scaled) gb_p | Green/Blue band ratio (pixel based, ×1000) | Integer (scaled) gb_s | Green/Blue band ratio (segment-based, ×1000) | Integer (scaled) rb_p | Red/Blue band ratio (pixel-based, ×1000) | Integer (scaled) rb_s | Red/Blue band ratio (segment-based, ×1000) | Integer (scaled) red_stdDev_p | Local standard deviation of red band in 5-pixel circular kernel (×10) | Integer (scaled) rg_p | Red/Green band ratio (pixel-based, ×1000) | Integer (scaled) rg_s | Red/Green band ratio (segment-based, ×1000) | Integer (scaled) slope_p | Terrain slope calculated from bathymetry (pixel-based, ×100) | Integer (degrees scaled) slope_s | Terrain slope calculated from bathymetry (segment mean, ×100) | Integer (degrees scaled) .geo | Point geometry (coordinates) | GeoJSON

eAtlas Processing:

The original data were provided as tiled mosaics corresponding to each of the five model areas. These were merged to GeoTIFF images covering the whole study area. This processing also included some small cleanup of spurious classification values. The full processing code is available from https://github.com/eatlas/NW_NESP-MaC-3-17_UQ_Shallow-habitat.

Location of the data: This dataset is filed in the eAtlas enduring data repository at: data\custodian\2023-2026-NESP-MaC-3\3.17_Northern-Aus-reef-mapping\data\NW_NESP-MaC-3-17_UQ_Shallow-habitat