Science & Methods



Planet Satellite Imagery

Learn more about Planet's approach

A high quality global coral reef mosaic from Planet’s PlanetScope satellite imagery is the starting point for creating the Atlas maps. PlanetScope (Dove) imagery exhibits the following technical specifications:
  • Spectral bands
    1. Visual: 3 (Red, Green, Blue)
    2. Analytic: 4 (Blue, Green, Red, and Near Infrared)
  • Pixel Size: 3.125 m
  • Radiometric Resolution
    1. Visual: 8 bit
    2. Analytics: 16 bit
Image processing
Imagery captured by PlanetScope constellation undergo a number of processing steps depending on product delivered. The following steps are taken to transform PlanetScope imagery for use in the Atlas:
  1. Top of Atmosphere Radiance (TOAR)
  2. Flatfield correction
  3. Debayering
  4. Sensor and radiometric calibration
  5. Orthorectification
  6. Surface reflectance
Image mask
The area to be used as the basis for a new global coral map, based on Planet satellite image data, includes coral reefs shallower than 20m deep, between 30° north and 30° south latitude, in clear water (that is, water without high turbidity) and listed as a coral reef in the United Nations Environment Programme (UNEP) 2010 Coral Layer.

To create a global image collection mask for coral reefs for the purpose of tasking acquisition of Planet Dove image data, the following buffer approach is applied:
  1. The existing UNEP 2010 global coral reef map layer is cleaned up to remove small reefs (e.g. bommies, patch reefs) within larger reefs;
  2. Reef areas that enclose non-reef areas are changed to be reef area (e.g. atoll reefs will include the deep lagoon); and
  3. All remaining reef polygons are used to establish a global one kilometer buffer, to conservatively identify global coral reef area.
Mosaicking
Additionally, PlanetScope imagery goes through a mosaicking process. Planet uses “best scene on top” (BOT) techniques for mosaicking PlanetScope imagery. This approach differs from the best-pixel method traditionally used in scientific research projects by stamping the entire scene into the mosaic instead of select pixels.

Carnegie Institution for Science Correction Models

Satellite data are made available to the science team in calibrated at-sensor radiance units (W str-1 m-2 s-1) as spatially contiguous orthorectified mosaics. These data require extensive processing using Carnegie algorithms to generate at-surface, sub-surface, and benthic reflectance data from the Planet radiance imagery. Reaching these three levels of processed data requires modeling of the radiometry of each Planet satellite (Dove, SkySat) used in generating coral reef coverage worldwide. Additionally, the following corrections need to be applied to Planet data to support the UQ mapping component (geomorphic zonation and benthic composition) as well as the Carnegie alert-monitoring component:

Atmospheric correction
The corrections for both the atmospheric effect and water column attenuation derive the benthic reflectance (or bottom reflectance). The derived benthic reflectance is applied to the coral reef classification and bleaching detection with improved accuracy. The method was developed based on the four bands (B, G, R, and near infrared [NIR]) Planet Dove satellite images for deriving the benthic with the assistance of depth data.

Waterbody retrieval
Ocean region is delineated from the corrected satellite images through the normalized difference water index as:

Then the following processing is processed on the water-only region.

Sun glint removal
The removal of sun glint (water surface effect) in the study regions were performed by equation 1 as:
Where Rrs, 0+ is the remote sensing reflectance just above the water surface in blue, green and red bands, Rrs is the water leaving reflectance (R, G and B), and Rrs(NIR) is the reflectance in the NIR band. After the sun glint correction, the below surface reflectance is derived as:


Depth calculation
A band-ratio algorithm is applied for deriving the depth based on B, G, and R bands of the Dove images:
The tunable constant (m0 and m1) is calibrated for the study sites according to the water column attenuation conditions.
this research was supported by The Nature Conservancy

For validation of the water depth product, reference data from field measured water depths is compared with coincident locations on the map product and to calculate regression values. Field measured depth is sourced from previously collected data from existing programs.


Bottom Reflectance estimation
In optically shallow waters, the water-leaving reflectance is made up of contributions from both waterbody and bottom sediments. So the below-surface remote sensing reflectance rrs is modeled as:

where rcrs represents the water column contribution. rbrs represents the bottom sediments contribution at below-water surface. H is the estimated depth, and B is the bottom reflectance to be derived. D(at+bb) represents the light attenuation caused by water column absorption and backscattering for water column light components (Dc) or light components from bottom (Db).
Finally, Dc and Db are empirical factors associated with under-water photon path elongation due to scattering and are calculated as below:


rrsdp represents below-surface remote sensing reflectance when the water is infinitely deep and is modeled as:

Then the water inherent optical properties (IOPs) are modeled as different components of water as:


The water IOPs contain the contribution of pure water ( aw(λ ), bbw(λ) ), CDOM ( acdom(λ ) ) and particles ( ap (λ) , bbp(λ) ). Then the bottom reflectance can be derived.
Diagrammatic workflow from Dove reflectance to derive the depth and bottom reflectance. The different components are illustrated below as the methodology sections above. The normalized difference water index (NDWI) is applied to mark water and land regions. For the water regions, the sun glint (or water surface effect) is removed by subtracting the NIR band. The water leaving signals ( Rrs ) are then derived. Next, the subsurface remote sensing reflectance is calculated to remove the sea-air interface effect. Finally, the water column attenuation correction is processed

References

Gao, Bo-Cai. 1996. “NDWI—A Normalized Difference Water Index for Remote Sensing of Vegetation Liquid Water from Space.” Remote Sensing of Environment 58 (3): 257–66.

Lee, ZhongPing, Kendall L. Carder, and Robert A. Arnone. 2002. “Deriving Inherent Optical Properties from Water Color: A Multiband Quasi-Analytical Algorithm for Optically Deep Waters.” Applied Optics 41 (27): 5755–72.

Lee, Zhongping, Kendall L. Carder, Curtis D. Mobley, Robert G. Steward, and Jennifer S. Patch. 1999. “Hyperspectral Remote Sensing for Shallow Waters. 2. Deriving Bottom Depths and Water Properties by Optimization.” Applied Optics38 (18): 3831–43.

Lee, Zhongping, Alan Weidemann, and Robert Arnone. 2013. “Combined Effect of Reduced Band Number and Increased Bandwidth on Shallow Water Remote Sensing: The Case of Worldview 2.” IEEE Transactions on Geoscience and Remote Sensing51 (5): 2577–86.

Li, Jiwei, Qian Yu, Yong Q. Tian, and Brian L. Becker. 2017. “Remote Sensing Estimation of Colored Dissolved Organic Matter (CDOM) in Optically Shallow Waters.” ISPRS Journal of Photogrammetry and Remote Sensing 128: 98–110.

Wabnitz, Colette C., Serge Andréfouët, Damaris Torres-Pulliza, Frank E. Müller-Karger, and Philip A. Kramer. 2008. “Regional-Scale Seagrass Habitat Mapping in the Wider Caribbean Region Using Landsat Sensors: Applications to Conservation and Ecology.” Remote Sensing of Environment 112 (8): 3455–67.

University of Queensland Habitat Mapping
Classification Scheme

A hierarchical, object-based classification approach is applied to map coral reefs and their geomorphic and benthic zones from Planet Dove image data, physical attributes such as bathymetry, slope, and significant wave-height in combination with machine learning and eco-geomorphological driven classification rules.

Four levels of classification, corresponding to the four scales of coral reef environments (Figure 1), each mapped over a set range of depths, (Figure 2) are used:

  1. Reef versus non-reef (0-15 m depth);
  2. Reef Type (0-15 m depth);
  3. Reef geomorphic zonation (0-15 m depth); and
  4. Reef benthic composition (0-5 m)

Figure 1: Minimum mapping units, spatial scale and levels of mapping detail to be used for coral reefs

Figure 2: Example of the hierarchical classification scheme applied in the project, based on work completed in the Capricorn Bunker Group reefs of the Great Barrier Reef.


The classification scheme outlined in Figure 1 and Figure 2 forms the backbone of the Allen Coral Atlas project. The classes and the rules used to define these were revised as the first part of the project and ongoing revision and improvement will take place. The revisions will take into account: Planet Dove data, Planet Dove derived bathymetry, best on offer global wave climatology data, available international community verification and validation data, and most recent regional to global scale coral reef mapping projects, as well as other past global projects and current NOAA UNEP global coral data sets.

Geomorphic and Benthic Mapping Principles

The geomorphic and benthic mapping approach currently implemented is combining machine learning with Object Based Analysis (OBA). The satellite image and physical attribute data are first segmented into ‘objects’ following an OBA paradigm. Using a training data set, a machine learning classifier is then used to make a preliminary classification of geomorphic and benthic classes. Based on an established framework for OBA on the Great Barrier Reef (Roelfsema et al. 2018), the preliminary classification is then improved and refined using the relational and contextual principles of OBA.

OBA Image Segmentation

The OBA paradigm is based on the image and other spatial data to be first segmented into groups of pixels with similar characteristics (e.g. colour or texture, or a physical property such as water depth). This is akin to how we use our eyes to segment images into objects for interpretation. Each image ‘object’ is then given a set of metrics based on its constituent pixels, which could be the mean or standard deviation, or something more complex like gray-level co-occurrence matrix (GLCM), texture metrics.

Curation of training and validation data sets

Training and validation data sets are created for individual reefs or small groups of reefs within a mapping region, based on the segmented image and physical attribute data. These training and validation data sets have two main origins:

  • An OBA mapping methodology developed and adjusted based on past studies is used to create high-detail and high-accuracy maps, from which training and validation data can be drawn. The five initial geographically-disparate sites mapped for the Allen Coral Atlas followed the same mapping approach. The training reefs within a mapping region as representation of the reef present in that region and vary in: reef type, geomorphic zonation and benthic composition but are representative for the region. A mapping region can be a geographic entity (e.g. Indonesia) or a group of reefs significant in size (e.g. Great Barrier Reef). This mapping approach may include manual editing and selection of areas to ensure only high confidence areas are included in the final training and validation data set
  • Existing high-resolution and high-accuracy maps and/or high-accuracy field data are ingested and used either in the mapping approach above (1) or are used solely to manually annotate the segmented image objects

Machine Learning: Random Forest

The Random Forest classifier is a well-established machine learning algorithm. It is an ‘ensemble learning’ methods, which means it classifies the input data based on a number of constructed ‘decision trees’ that each have some component of random variation in the parameterisation. The output classification is the mode or mean of all the decision trees, and they are particularly useful because they balance predictive performance with overfitting, and are also robust to redundant predictor variables (Breimann et al. XXX). The classifier is trained used the curated training data set, with the input variables included both the pixel and segmented objects from the image and physical environment data.

OBA Refinement

The output classification from the machine leaning classifier is then processed using a number of automated OBA membership rules. Membership rules form the typology of a mapping class defined by different attributes. These attributes not only include the brightness of an object but also the texture, depth, slope, waves or location in relation to other objects. In marine environments, seafloor features are especially challenging to distinguish due to submerged characteristics by variation in water from tides, water column composition from water movement, and surface roughness.

Software Used

To create geomorphic and benthic maps for the training set the rule sets are  adjusted, developed and tested using the commercial software Trimble eCognition 9.3. This software enables highly efficient and accurate OBA mapping, but the curated training and validation data set is in no way limited to this software – any source of high-resolution and high-accuracy map or labelled segment data can be used as training and validation in the Allen Coral Atlas workflow. The machine learning mapping and the  For the mapping regions, the machine learning and OBA refinement stages are implemented in Google Earth Engine, and open source and free cloud-based processing environment that provides capability to access and process the Planet Dove imagery, along with a range of other satellite image archives (e.g. Sentinel 2, Landsat). The image segmentation and OBA refinement workflow was not previously available in Google Earth Engine, so that software capability is a major output of the Allen Coral Atlas project (Lyons et al. in prep).

Supporting Environmental Data - Global Wave Climate Model and Data

Wave exposure is the dominant force influencing the ecological makeup and physical structure of coral reefs. Changes in the benthic ecological community as well as some crucial metabolic and biological functions of coral reefs have been linked to variations in wave energy. Long-term geomorphic development of coral reefs is also driven by the relative exposure of coral reefs to wave processes. A thorough understanding of wave exposure is now an important component of benthic ecological surveying in coral reefs. Wave exposure on coral reefs has typically been determined using a suite or computationally onerous models which limits wave modelling to a local or regional basis which. To calculate the wave exposure for every reef in the world a wave model that is flexible, computationally fast and links with global wave models. This model uses principles of wave refraction and diffraction to determine the dissipation of wave energy from deepwater sources to shallow reef environments and through often complex coral reef regions. This provides the local wave height for every reef prior to wave breakpoint and hence the wave exposure index for each coral reef.

Datasets: National Oceanic and Atmospheric Associate (NOAA) Wave Watch III global wave model hindcast reanalysis (1979-present). Planet derived bathymetry.

Geomorphic Zonation

The OBA protocol incorporates additional attributes: water depth (derived from satellite imagery), slope (calculated from water depth), historical significant wave height and surface reflectance. In this project’s methodology, surface reflectance is considered a proxy for consolidated (dark e.g. reef matrix, coral, algae) or unconsolidated material (bright e.g. sand). Geomorphic mapping requires first the data sets to be divided for shallow reef area and reef type is mapped at final stage.


For Reef and Reef Type level, a rule set was developed that extracted deep water areas within surface reflectance image and depth mosaic in deeper than 15 m based on depth, and extracted land features using a global land mask (described in Planet’s satellite imagery methodology). The remaining area was thus considered reef area. Within each reef area a division was made between rocky reefs and carbonate reefs, where carbonate reefs were assumed to have a relatively large horizontal platform and rocky reefs did not have a platform. Reef Type was further divided in reef top (above 3 m), and reef slope (3-10 m).

For each geomorphic zone category, a ruleset was developed to assign a label to each segment based on a set of biophysical attributes such as water depth (depth shallower than 3 m = Reef Top), color (e.g. brightest = sand), slope derived from water depth (>10 degree = slope) and neighbourhood relationships, e.g. Fore Reef Exposed is adjacent to Reef Rim, includes reef crest. Based on the ruleset, areas with a depth of >15 m are labelled Deep Water. Areas with a depth of >10 m are labelled Deep Slope or Deep Plateau, dependent upon slope and water depth attributes. Areas between 0.75 -10 m and neighbouring Reef Top are labelled Not Reef Top. Not Reef Top exposed to historically high significant wave heights (Hs95 > 2 m) are labelled as Fore Reef Exposed, while areas with low significant wave heights (Hs95 < 2 m) are labelled as Fore Reef Sheltered. Reef Top is re-segmented into smaller segments and assigned Reef Crest, Reef Flat, Outer Reef Flat, Inner Reef Flat and Shallow Lagoon based on the following criteria: water depth, neighbourhood relationship, slope and brightness level of individual bands. Brightness was used as a surrogate for consolidation where, objects area contain in general more sand and are therefore unconsolidated, darker objects when found close to the reef slope are in general hard consolidated. (a surrogate for sand). Shallow Lagoon as an example is predominantly a sandy area, hence it will have high brightness level. Next to that Reef Rim (Reef Crest) is shallower than outer reef flat, followed by inner reef flat and shallow lagoon.
Benthic Cover Type

Membership rulesets to assign a dominant benthic cover type label to a segment are based on the brightness of the segments, band ratios, segment location within each of the geomorphic zones, and with visual assessment and guidance of expert reef knowledge and/or field data. Rules vary between geomorphic zones dependent upon the type of ecological relationship and/or threshold value for a dominant benthic cover type. Dominant benthic cover type labels include: Coral, Algae, Benthic Microalgae Mats (BMA), Seagrass, Rock, Rubble, and Sand. In this segmentation, Algae is dominated by macroalgae (> 2 cm), and Rock includes turf algae (< 2 cm) and crustose coralline algae. Patch reef categories represent small patches that include coral and algae (approximately 10 m - 50 m diameter) that inhabit sandy areas.

Due to increased bottom reflectance attenuation with increasing water depth, in some cases the only differentiation made is between bright and dark objects. Bright objects are assumed to represent unconsolidated material (e.g. bright = sand), and dark objects consolidated material (e.g. dark = coral, rock or algae).

To provide more confidence in differentiating between coral and algae, historical impacts could be incorporated together with local knowledge. Coral and algae have similar visual characteristics within a high spatial resolution multi-spectral satellite image, making them harder to differentiate from each other. Therefore, historical knowledge of impacts such as cyclones, bleaching, Crown of Thorns, or decline in water quality are helpful in verifying presence of either benthic type. For instance, recent severe bleaching would most likely turn areas (objects) assigned as coral to algae.

Reference data for calibration and validation
Geomorphic zonation map reference data is created by assigning a mapping category through visual interpretation to points randomly generated across the full extent of the geomorphic map and overlaid on existing satellite imagery. The assignment is conducted by unbiased experts who are not involved in producing the geomorphic zonation map, but are familiar with the reef region assessed.

Reference data for the dominant benthic cover type map 

is derived from previously collected field data through existing programs, from newly collected field data as part of the project, and through citizen science data. A protocol for data collection is being created and will include a suggested capacity building process.
Field data are required for training and verification of the mapping approach and map products such as: water depth, Geomorphic, Benthic community, benthic change.

Reference data for the benthic mapping will be sourced predominantly through georeferenced photo quadrates that could be collected from boat, on snorkel or scuba or using a Remote Operated Vehicle (ROV) or Autonomous Underwater Vehicle (AUV). Photo quadrates could be acquired at random points or along transects and set intervals, located around and on top of the reef, and its position synchronized with standard hand held GPS. Photos will be analyzed using machine learning to provide a consistent output of benthic mapping categories.

Additionally, field verification will be used for contextual editing in regards to the habitat information but also in regards to administrative information e.g. such as addition of local reef names. Verification data will be sourced at different levels of detail and accuracy, and at different times throughout the project (Figure 3).


Figure 3: Options for different levels of knowledge able to be collected by the communities: (top left) detailed surveys through geolocated benthic photo quadrates:, (top right) basic surveys, through descriptive characterisation in the water; (lower left) local knowledge, where the fisherman or other are asked to provide their input on what is where; and (lower right) remote surveys, where technology is deployed to gather autonomous information about the seafloor.


Quantitative Validation
Accuracy measures have been determined for the geomorphic zonation and dominant benthic cover type by comparing map outputs with reference data using an error matrix to produce overall, producer and user accuracies.

Qualitative Validation
qualitative assessment has been conducted to take into consideration the factors providing confidence for the producers and users of the habitat maps. This is be done through assessment of the mapping process itself. This approach uses resampling framework to test the stability or ‘confidence’ of the classification of the image objects.

For more information on mapping and monitoring through remote sensing from the University of Queensland, check out this Remote Sensing Toolkit


The National Geographic Society: Field Engagement

The National Geographic Society, in partnership with the Allen Coral Atlas and, prior to her passing, Dr. Ruth Gates, is developing a strategy for field engagement with the wider coral reef science, monitoring and management community. As the Allen Coral Atlas continues to expand, key elements will include building coalitions with networked institutions, including NGOs and research and government monitoring and mapping programs. With innovative data visualization, the team hopes to build awareness and understanding of the Atlas tool so can be leveraged globally by users, who can also contribute data and provide feedback to help improve the Atlas over time.


Field teams are collecting geo-referenced data in selected representative regions to help test, develop and implement the mapping algorithms. The Society is also developing a larger program focused on local engagement and capacity development, and aligning with efforts to use machine learning to analyze georeferenced photo quadrats.


Ultimately, the goals of the habitat mapping and monitoring of the Atlas can help report on progress toward achieving international targets such as the Sustainable Development Goals and Convention on Biological Diversity Aichi targets. We anticipate alignment with existing efforts (e.g., the International Coral Reef Initiative and the Global Coral Reef Monitoring Network) to facilitate planners, managers and policymakers using the findings and data from the Allen Coral Atlas to achieve conservation impact.

Coral Reef Watch

Coral Reef Watch (NOAA NESDIS)

The Coral Reef Watch near real-time 5km global products on the Allen Coral Atlas site are the most recent day's published sea surface temperature (SST), SST Anomaly, Coral Bleaching HotSpot, Degree Heating Week (DHW), a 7-day maximum Bleaching Alert Area, and 7-day SST Trend data from NOAA's Coral Reef Watch program. Please see their website for more information about the program. For more technical details about the 5-km products, see Liu et al. 2017and 2014, and Heron et al. 2016 and 2015. If these products are used in any way, please follow the citation guidance.