MapBiomas Water Method
WATER SURFACE MAPPING : Method Summary
Presentation
The objective of MapBiomas Water is to provide monthly and annual data about the surface water dynamic, water bodies and and their discrimination as natural or anthropic (small and large dams and water in mining areas) for the whole national territory with data since 1985. In addition, the water surface mapping will provide more detail for the annual land use and land cover map of the Brazilian biomes, adding information of wetland occurrence and more detail about small water bodies (i.e. > 0.5 ha).
The water surface mapping in Brazil used all Landsat satellite scenes with less than 70% cloud cover, at 30 meters spatial resolution. Mapping was conducted at the sub-pixel scale, with spectral mixture model (MEM), and empirical classification rules based on fuzzy logic. The mapping covered the period from 1985 to 2020, at the monthly scale, with a total of 184,558 Landsat scenes processed (average of 5,126 per year) and it has been analyzed on the Google Earth Engine platform.
Organization and Database
The general coordination of MapBiomas Water is conducted by Imazon and WWF-Brazil, and the technical and operational coordination is conducted by Geokarten. The reconstruction of the monthly historical series of water surface area was conducted by experts from all biomes with leadership from the following institutions: Imazon (Amazon and Pantanal biomes), Geodatin (Caatinga), Solved (Atlantic Forest), Geokarten (Pampa) and IPAM (Cerrado). We also have valuable inputs for the mapping and validation of the ArcPlan results (Pantanal and Atlantic Forest). The algorithm for mapping water surface and water bodies was originally developed by Imazon, and improved in this first stage of the MapBiomas Water work (see details in the ATBD).
The development of the MapBiomas Water dashboard counted on relevant contributions from users in a design thinking process. The dashboard development was conducted by Geodatin with significant contributions from the MapBiomas Water working group and inputs from the design thinking process.
Four types of products were produced by MapBiomas Water:
1) Maps of monthly water surface occurrence frequency;
2) Maps of transitions from water surface to other land use and land cover classes;
3) Maps of trend (increase and decrease) of water surface area; and
4) Maps of water body types.
The dashboard consists of maps, statistics and tools for visualization, analysis and data access.
Method
1. Pre-processing: consists of the selection of Landsat scenes from sensors: Landsat 5, Thematic Mapper (TM) Landsat 7, Enhanced Thematic Mapper Plus (ETM+) Landsat 8, Operational Land Imager (OLI); the application of cloud and shadow masking to each scene, and the exclusion of scenes with more than 70% cloud cover. The visible and near and mid-infrared spectral bands were selected for the application of the Mixture Spectral Model (MEM). The result of the MEM is a set of compositional bands, from each pixel of the Landsat image, for the components Vegetation, Non-Phototosynthetic Vegetation (NPV), Soil, Shade and Cloud. Water behaves as a dark body (i.e. low reflectance) in Landsat images and therefore shows a high percentage of the Shadow component in the pixel. Edge of lakes, rivers and wet environments (floodplains present a mixture of Shadow (water), Vegetation and Soil in the pixel, which allows to detect water in environments with these types of materials.
2. Water Surface Classification: It was based on the compositional bands of the MME on fuzzy decision rules to obtain maps of probabilities of water occurrence in the Landsat pixel. The monthly maps were obtained by combining the average probability map of water occurrence in a given month with decadal probability of the month and average to year probability. Pixels with low mean annual probability (i.e., >0.35) are excluded, with probability between 0.35 and >0.5 are considered as ephemeral (to be considered as water only if the probability in the month is >0.67), and with values >0.5 of annual and decadal probability are considered with water to fill data gaps.
3. Water body classification. Water body mapping was used on information extracted from the annual mapping, including: the first and last occurrence of the water body in the year, the total water surface frequency (in the historical series), and the annual frequency. This information was organized into matrix data and used in an object segmentation algorithm. The next step was to extract attributes from auxiliary maps of dams from the National Water Agency (ANA) database and Mining (source: MapBiomas Collection 6 - Mining). The water body segments were classified with the random forest algorithm into the following water body classes: Natural, Hydropower, Small Dams and Mining Water. We also included the class "False positives" as a product of the classification of the water surface segments. This information was used to remove the cases of false positives that remained in the monthly and annual water surface maps.
4. Mapping Accuracy Analysis
Accuracy analysis was conducted for the annual water classification data using the MapBiomas reference data collected by LAPIG/UFG (also at annual scale). The class "River, Lake or Ocean" was considered with water surface. We applied a sample stratification method based on annual water frequency classes in an attempt to reduce the sampling error of the producer's accuracy. The water frequency classes were:
Permanent (appears + 90% in the annual product and at least once in the last three years).
Intermittent (appears at 50% to 89% frequency in the annual product).
Infrequent: (appears between 10% and 49% frequency in the annual product).
Ground: less than 10% frequency in the time series.
250 meters distance to the next water body.
500 meters distance to next water body
5,000 meters distance to next water body.