Images to Maps
Remote sensing images are a grid of pixels with numerical values whereas maps use X and Y coordinates to deﬁne the location of points, lines, and areas (polygons) that correspond to map features. Transforming a satellite image into a map that has high accuracy is a process that requires experts and fieldwork.
To create a map from an image, classiﬁcation techniques are applied. One of the more popular approaches is an object-based image analysis, a recent and powerful image classiﬁcation technique.
An object-based image analysis first segments the image into objects. The segmentation is based on the spectral signatures of the pixels, the shape and size of the object, or the texture of the pixels within the object. Once the image is segmented, the user matches each land cover class to a few sample objects. The image is then classiﬁed.
Classification schemes are used to assign a class to an object on the image. For coral reefs, having prior knowledge of the area will help to identify potential classes to be mapped. These classes are used to form a classiﬁcation scheme which is used to set up ﬁeld surveys and target representative areas for training data collection.
Below is an example of a hierarchical classiﬁcation used for the maps in the Allen Coral Atlas that includes geomorphic and benthic classiﬁcation and is general enough to be used on a global scale.
Field data is a key component of creating a map of high quality. Field data builds a link between the image and the mapped features.
Field survey objectives:
- Identify the habitat classes to be mapped and create a classiﬁcation scheme
- Locate representative areas of each habitat class in order to collect training data
- Generate data needed for image calibration, such as water clarity and depth
- Collect suﬃcient stratiﬁed random points to evaluate the accuracy of the classiﬁcation (predicted vs. observed classes)
There are a variety of methods that can be used to collect ﬁeld data for classiﬁcation validation besides SCUBA and snorkel surveys. Data can be collected using aerial and surface drones, underwater instruments such as remotely operated vehicles (ROVs), spot checks from the boat using a glass-bottom viewer, or even local knowledge.
Training data are any type of georeferenced ﬁeld data collected for image classiﬁcation. They are representative “ground referenced” samples of each class in the classiﬁcation scheme. They are used to “train” the classiﬁcation algorithm to recognize the unique spectral reﬂectance patterns in each of the classes. These can be collected in several ways such as SCUBA and snorkel surveys, aerial and surface drones, remotely operated vehicles (ROVs), boats, or even local knowledge.
An accuracy assessment is an important part of any classiﬁcation project and provides a measure of how accurate the map product is. It uses independent “ground referenced” data to calculate a statistically-based accuracy score based on the comparison of the predicted (mapped) class verses observed class in the ﬁeld. In other words, the classiﬁed pixels or classiﬁed objects are compared to reality, or what actually exists at that location. While independent ﬁeld data collection can be time consuming and expensive, these data can also be derived from interpreting high-resolution imagery, existing classiﬁed imagery, or local experts.