WATER SURFACE MAPPING: METHOD SUMMARY

Here we present a summary of the MapBiomas Water Method. Access the ATBD (Algorithm Theoretical Basis Document ) in this LINK for more methodological details. .

Presentation

The main objective of MapBiomas Water is to map the dynamics of surface water throughout the territory of the Amazonian countries (Pan-Amazonia), on a monthly and annual basis from 1985 to 2023. The dataset is publicly available on a web platform to improve the management and use of water resources across Pan-Amazonia. 

The surface water mapping in the Amazonian countries utilized all Landsat satellite scenes with cloud coverage of less than or equal to 70% and a spatial resolution of 30 meters. The mapping was conducted at a sub-pixel scale (SWSC) using Spectral Mixture Analysis (SMA) and empirical classification rules based on fuzzy logic. The mapping covered the period from 1985 to 2023 on a monthly scale, with a total of 396,000 Landsat scenes processed and analyzed on the Google Earth Engine platform. 

Organization and database 

The overall coordination of MapBiomas Water is led by Imazon and RAISG, while technical and operational coordination is directed by Geokarten. The reconstruction of the monthly historical series of surface water was carried out by specialists from all biomes of the Amazonian countries, under the leadership of the following institutions: Fundación Amigos de la Naturaleza -FAN- (Bolivia), Fundación Gaia Amazonas -FGA- (Colombia), EcoCiencia (Ecuador), Instituto del Bien Común -IBC- (Perú), Provita y Wataniba (Venezuela), Alliance of Bioversity International y CIAT (Guianas y Suriname). The surface water mapping algorithm was developed by Imazon and adapted by MapBiomas Water in this initial phase of work. 

The development of the MapBiomas Water control panel (dashboard) was conducted by Geodatin and includes significant contributions from the MapBiomas Water working group and platform users in the design thinking process. 

Three types of products were produced by MapBiomas Water: 

1. Monthly and annual surface water maps;
2. Surface water transition maps between “Water” and “Non-water” classes. This product was processed using the annual surface water database;
3. Trend maps (increase and decrease) in surface water. This product was calculated from monthly surface water data in 5 km x 5 km grids. 

The dashboard (link) consists of maps, statistics, and visualization, analysis, and data access tools. It is possible to view the data on an annual and monthly scale, as well as obtain it in different territorial units. Finally, the dashboard also provides a link to access the MapBiomas Water data API. 

• Method

The following diagram illustrates the main stages in the process of classifying surface water in the Amazonian countries, involving a surface water sub-pixel classifier (SWSC), decision tree, and post-classification procedures to generate annual and monthly surface water datasets. 

Figure 1 – Stages of surface water classification.

Description of classification steps:

1. Pre-processing: 

Pre-processing: Consists in the selection of Landsat scenes from the sensors: Landsat 5 Thematic Mapper (TM), Landsat 7 Enhanced Thematic Mapper Plus (ETM+), and Landsat 8 Operational Land Imager (OLI); applying cloud and shadow masking to each scene and excluding scenes with more than 70% cloud cover. The visible, 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 for each pixel of the Landsat image, for the Vegetation, Non-Photosynthetically Active Vegetation (NPV), Soil, Shade, and Cloud components. Water behaves as a dark body (i.e. low reflectance) in Landsat images and therefore has a high percentage of the Shadow component in the pixel. The edges of lakes, rivers, and humid environments, such as floodplains, present a mixture of Shadow (water), Vegetation, and Soil, which allows the detection of water in environments with these types of materials. 

2. Classification of Water Surface:

The original Surface Water Sub-Pixel Classifier (SWSC) algorithm employs three hierarchical binary decision rules (e.g., true, false). Since water absorbs most of the electromagnetic radiation, an image with Shade fraction, the combination of GV and Soil, and Cloud is used to classify pixels as surface water. Additionally, a classification based on fuzzy logic (independent fuzzy rules) is applied, where the degree of truth/certainty (memberships) de que un pixel Landsat es clasificado como agua. Luego se calculó el grado de verdad promedio para obtener un mapa continuo de memebership that a Landsat pixel is classified as water is determined. The average degree of truth was then calculated to generate a continuous membership map with values ranging from 0 to 1. Based on these memberships, pixels are classified to produce monthly surface water layers. 

By calculating the median of the memberships of pixels across the available Landsat scenes for each month, pixels were classified as water based on defined thresholds. Procedures were then applied to restore false negatives and remove false positives, using temporal metrics. Subsequently, gap-filling techniques were implemented to reclassify as water those pixels that were either cloud-covered or located in areas without Landsat scenes during a given month. This was achieved using a combination of two rules: the annual median probability and the decadal median of the corresponding month. Finally, the presence of cloud shadows or other dark objects in the Landsat scene can produce false positives in water classification. To address this, a removal filter was applied to reclassify those pixels as non-water. 

The annual surface water maps include a distinction between permanent and seasonal water. This classification is based on thresholds corresponding to the number of months a pixel is classified as water. For the first case, a frequency of ≥ 6 months is considered, while for the second case, a frequency of 1 to 5 months is used. 

Figure 2 – Monthly classification process.

3. Classification of Water Bodies

Figure 3 – Process of Water Body Classification.

For the classification of water bodies, the following information extracted from the annual mapping of water surface (permanent) was used: i) the first and last occurrence of the water body in the year, ii) the total frequency of the water surface in the historical series, and iii) the annual frequency. This information was organized into raster data and used in an object segmentation algorithm.

Subsequently, attributes were extracted from auxiliary maps of hydroelectric plants and mining from the following entities:

• Mapa de coberturas de la tierra 2018 - IDEAM (2018) (Colombia)2018) (Colombia)
• Cartografía básica 1:25.000 and 1:100.000 – IGAC (2018) (Colombia)
• MapBiomas Amazonía Collection 6 
• MapBiomas Colombia Colección 2

Water body segments were classified using the Random Forest algorithm into five categories: natural, other artificial, hydroelectric, mining, and aquaculture. Additionally, a "false positives" class was included to eliminate persistent overestimations in the annual and monthly surface maps. Spatial and frequency filters were then applied; further samples were taken, and manual polygon delineations were conducted to improve the results.