To make a map of the landscape remap needs data to train a model called a classifier. Remap uses data points that identify all of different classes present in the landscape, which are then used to predict across the rest of the landscape. The result is a map where all pixels have been classified into the classes originally presented to remap.
We suggest making maps of ecosystems where each class is a single ecosystem type, such as mangrove, salt marsh or a forest type. However, remap will attempt to classify the landscape into any set of classes that you provide as training data, which means you can make maps of deforestation (2 classes: forest and non-forest), land use (3 classes: agriculture, bare land and urban area) or ecosystems (For example, mangrove, salt marsh and sandy beach). In many cases it will be sufficient to simply train the class of interest and have an "other" class to distinguish it from.
Remap is subject to a range of computational limits when making calls to the Google Earth Engine. Therefore, to ensure we can successfully classify an area we have set a limit to the focal region of 100,000 km^2. If you require maps that are larger than this, we suggest implementing your workflow directly in the Google Earth Engine. We are actively working to increase this limit, so please let us know if you require a larger focal region using the feedback link.
As with all remote sensing classifications, there are a number of choices that can be made to produce the most accurate map classifications. We suggest trying these three steps if you want to produce the most accurate map possible in remap:
Remap uses a machine learning classification method called Random Forests. This method requires training data (which we call the training set, a set of observations of class occurrence) that are used to train, or teach, the classifier the most likely class that a pixel belongs to. Therefore, we recommend you develop a fairly large training set (>50 point per class) with the aim of independently sampling all of the map classes you wish the Random Forest to identify across the landscape.
When you click classify the classifier aims to predict all of the classes from the set of training observations provided to it. To acheive this, it utilises information from a set of predictors (data layers) to classify each pixel in the focal region. We have provided a default set of biophysical (slope, elevation), spectral (Landsat Archive imagery) and climatic (precipitation, temperature) predictors that work well across a variety of landscapes. However, it is possible to select predictors that are expected to strongly influence the distribution of particular ecosystems, and tuning remap to work for your ecosystem of interest will improve the classification. For instance, if you are working with a coastal ecosystem sugh as mangroves, slope is likely to be important predictor of their distribution. You can use Select Predictors to ensure slope is included in the classification.
Remap will classify the landscape only into the classes you provide it. Therefore, in cases where there are many contrasting map classes, it may be necessary to carefully think through the best set of classes that you wish remap to distinguish. For example, if you are interested in mapping mangrove ecosystems, you will get a better classification if you map mangroves, water and land, rather than simply mangrove and other. This is due to the water and land classes being very different classes for the classifier to assign to one single class. This process is also important if you wish to map ecosystems adjacent to urban areas: it will be necessary to make urban a map class so that it can be effectively distinguished from the natural ecosystems in your region.
From Murray et al (2017):
Extent of occurrence (EOO) and area of occupancy (AOO) are two standardized measures of geographic range size used in both IUCN red list criteria. Neither AOO nor EOO is intended to be a precise estimate of total area in which a species or ecosystem occurs. Rather, in risk assessment protocols such as the IUCN red lists they function as standardized, complementary and widely applicable measures of risk-spreading against spatially explicit threats.
EOO is measured by a minimum convex polygon encompassing all known occurrences of a species or ecosystem.
AOO is an index related to total occupied area, estimated by summing the areas of grid cells of a standard size (~2×2 km for species and 10×10 km for ecosystems.
Refer to the IUCN Red List of Ecosystems guidelines for further information about EOO and AOO.
If your .csv file with training points is not being recognised by remap, you can try a couple of different things:
latitude, longitude, labelin the same order (
latitude, longitude, label).
Wet Schlerophyll Forests
,as a decimal point. For example, a single point should have the format
-34.1524, 151.0259, Wet Schlerophyll Forests.
Remap allows you to save your current workspace so that you can quickly and easily return to work:
7. Export Dataand select
2. Build Training Set, select
Upload JSON. Remap will automatically zoom to your region, import your previous training points and you can resume your workflow.
Satellites have been imaging the earth for around 50 years. In that time they have been instrumental in identifying large, detrimental changes to ecosystems. Indeed, the trajectories of many of Earth’s major ecosystem types, such as the –3.7% annual decrease in the extent of tropical peatlands, –2% annual decline of coastal wetlands in East Asia and the >1% net annual change of forests globally were quantified with large-scale remote sensing analyses of satellite data.
The basis of all of these observations of ecosystem loss and change has been accurate maps of ecosystem distributions. Maps of ecosystems may be developed from a range of sources, but the use of satellites such as Landsat have allowed long-term changes (nearly up to 50 years) to be detected.
In remap, we use Landsat satellites and a range of other datasets to develop accurate maps of ecosystem distributions. Our classification methods, which utilise machine learning to rapidly and accurately classify satellite image pixels, are proven to work for a range of ecosystem types.
The IUCN Red List of Ecosystems identifies ecosystems as Critically Endangered, Endangered, Vulnerable or Least Concern. Together, the Red Lists for species and ecosystems will provide a more comprehensive view of the status of the environment and its biodiversity than either can on its own. Find out more about the IUCN Red List of Ecosystems: www.iucnrle.org
Remap was developed with funding from a Google Earth Engine research award. We appreciate feedback to help further develop Remap into a tool that helps assess and conserve ecosystems across the world.