Coastal Zone and Mining – Appendix

Collection 3

Version 1

General coordinator Pedro Walfir Martins e Souza Filho Team Cesar Guerreiro Diniz Luiz Cortinhas Ferreira Neto Gilberto Nerino Junior Jhonatan Rodrigues dos Santos

1 Landsat image mosaics

The classification of the cross-cutting themes “Coastal Zone and Mining” used Landsat mosaics that differed from the mosaics used to classify the natural vegetation of the Brazilian biomes. The Coastal mosaics were defined to preserve the maximum of the coastal zone land area while capturing the smallest possible cloud cover. These Landsat mosaics are the third generation of the methodology developed specifically for these cross-cutting themes.

1.1 Definition of the temporal period

Coastal areas are severely affected by atmospheric nebulosity, a condition that is intensified by its proximity to the oceans and its tropical location. On the other hand, the attempt to identify a time interval that covers only the driest season of the year, as alternative to reduce cloud persistency, ends-up severely reducing the number of images available to cover the entire coastal region. Thus, the annual cloud free composites were generated, ranging from 1st of January to 31st of December.

1.2 Mosaic Subsets 1.2.1 Coastal Zone

Since the Brazilian coastal zone (ZCB) is an extensive region, approximately 8,500 kilometers from Oiapoque to Chui (without counting reentrances), and affected by a variety of atmospheric systems, of lesser or greater influence of nebulosity’s, the ZCB were divided into seven different sectors (Figure 1): Sector 1 - Amapa (AP), coastal region of the state of Amapa; Sector 2 - Marajo Island (MAR), coastal region of Marajo Island; Sector 3 - Para / Maranhao (PAMA), a coastal sector of the states of Para and Maranhao; Sector 4 - Piaui / Bahia (PIBA), a coastal sector of the states of Piaui to Bahia; Sector 5 - Espirito Santo / Sao Paulo (ESSP), region that includes the states of Espirito and São Paulo; Sector 6 – Parana/Laguna (PRLA), a coastal region that goes from the state of Parana to the municipality of Laguna in Santa Catarina; Sector 7 - Laguna / Rio Grande so Sul (LARS), a region that ranges from Laguna to the state of Rio Grande do Sul.

Figure 1. The seven sectors delimited in different colors, along the Brazilian Coastal Zone (ZCB) used in the cross-cutting them Coastal Zone.

1.2.2 Mining

The definition of the mining mosaic subset was made based on the Geographic Information System of Mining - SIGMINE, developed by the Coordination of Geoprocessing - CGEO/ CGTIG. The SIGMINE system aims to be a reference system in the search for updated information regarding the areas of mining processes registered in the DNPM - National Department of Mineral Production, associated with other geographic information of interest to the sector produced by public agencies, providing the user with a data query and relational analysis of spatial character. In Collection 3, only the areas registered under the "Mining Concession" status were considered (Figure 2).

Figure 2. Areas registered as "Mining Concession" (in black) by DNPM. MapBiomas Collection 3 locked for mining patterns only inside these black geometries.

1.3 Image selection

The cloud/shadow removal script takes advantage from the Landsat Collection 1 Level-1 QA band and the GEE median reducer. In Collection 1 Tier 1 data, each pixel in the QA band contains unsigned integers values that represent a certain surface, atmospheric, and sensor conditions that may affect the overall usefulness of a given pixel. When effectively used, QA values can improve the data integrity by indicating which pixels might be affected by instrument artifacts or subject to cloud contamination (USGS, 2017). In conjunction to that, GEE can be instructed to pick the median pixel values in a stack of images. By doing so, GEE rejects values that are too high (e.g. clouds) or too low (e.g. shadows) and picks the median pixel value in each band, over time (Figure 3).

Figure 3. Left, Collection 2 “cloud-free composite”. Right, Collection 3 “cloud-free”. 1.4 Final quality

The mosaic quality is related to Landsat cloud-free availability during the image selection periods. The Landsat images availability in Collection 3 period (1985 to 2017) have variations in the number of images available for the yearly mosaics. Throughout this period the Landsat 5 (1985 to 2012), Landsat 7 (1999 to present) and Landsat 8 (2013 to present) provided images for the mosaics compositions. Figure 4 shows the number of image availability of each Landsat for the period covered in Collection 3.

Figure 4. Availability of Landsat images covering the Brazilian territory for the Collection 3 period.

2 Classification

The automatic classification of the Landsat mosaics was completely performed on Google Earth Engine platform. We used the Random Forest classifier proposed by Breiman (2001).

2.1 Classification scheme

We classified each interest class separately. Therefore we had to perform five independent classification processes: 1) Mangrove; 2) Beach and dune; 3) Salt flat (Apicum); 4) Aquaculture and 5) Mining. The classification process was carried out always considering only two possible classes for each pixel, the interest class (Mangrove, Beach and Dun, Salt flat, Aquaculture and Mining) or non-interest class (all that is different from the interest class).

For the supervised classification of Landsat mosaics, we selected training points for the Random Forest classifier based on reference maps and MapBiomas Collection 2 temporally stable samples. The details of the parameters used in the Random Forest classifier, the reference maps used for each interest class, and the feature space produced for each classification are presented in the sections to follow. 2.2 Coastal Zone Feature space

The tables 1 and 2 show respectively: all spectral indices and all bands, used for classification in the Coastal Zone. After applying reducers in the images collections for the selected periods and highlights the bands used for Random Forest training and classification. Table 1. Spectral indices used for Coastal Zone classification in MapBiomas Collection 3.

Index Expression Reducer Reference

EVI2 2.5 * ((NIR - RED) / (NIR + 2.4*RED + 1))

Median and Standard Deviation Liu & Huete, 1995

NDVI (NIR - RED) / (NIR + RED) Median and Standard Deviation Tucker, 1979

MNDWI (GREEN - SWIR1) / (GREEN + SWIR1)

Median and Standard Deviation Xu, 2006

NDSI (SWIR1 - NIR) / (SWIR1 + NIR) Median and Standard Deviation Rogers, 2004

MMDI Modular Mangrove Detection Index Median and Standard Deviation

Diniz et. al, under review

Table 2. Table of bands used to classify Coastal Zone classes in the MapBiomas Collection 3.

Variable Description Reducer

GREEN Landsat Green band median value Median and Standard Deviation

RED Landsat Red band median value Median and Standard Deviation

NIR Landsat NIR band median value Median and Standard Deviation

SWIR1 Landsat SWIR1 band median value

Median and Standard Deviation

SWIR2 Landsat SWIR2 band median value

Median and Standard Deviation

2.3 Mining Feature Space

A Landsat median composite covering the entire Brazilian territory, from January 1 to December 31, was clipped by the "Concession Mining" vector, obtained through the SIGMINE system. Subsequently, spectral indices developed to identify surface mineral activity (no subsurface mineral activity was here recognized), were added to the set of classification inputs (Table 3). Table 3. Indices and Landsat bands used to Mining classification in the MapBiomas Collection 3.

Band Name Expression Reducer Reference

AFMI Adjusted Ferrous Mineral Index

Median Unpublished

ANMI Adjusted Nitrate Mineral Index

Median Unpublished

SWIR2 Landsat Swir2 band median value

Median USGS

SWIR1 Landsat Swir1 band median value

Median USGS

NIR Landsat Nir band median value

Median USGS

RED Landsat Red band median value

Median USGS

GREEN Landsat Green band median value

Median USGS

2.4 Classification algorithm, training samples and parameters

As lacking reference maps to each one of the years to be classified, reference maps of the closest possible timeframe to the median composites were used. Then the classifications results as a previous year were used for the training subsequent one. This was done for each one of the years without an external reference map available for the classifier training. The Random Forest setup were used to classify the years with and without external reference maps (Figure 4). Table 4. Number of trees, points and variables used in Coastal Zone and Mining classification in the MapBiomas Collection 3.

Parameter Value

Number of trees 100

Number of points 100000

Number of Variables 20 (Coastal Zone) and 14 (Mining)

Classes 2 (binary classification)

2.4.1 Mangrove

As in any supervised method, the Random Forest classifier needs to rely on a training dataset. For mangrove cover recognition the training data was obtained from MapBiomas Collection 2.0 and Dr. Giri Chandra global mangrove cover (Giri et al., 2013) (Figure 5).

Figure 5. Global Mangrove Cover data used as mangrove mapping reference, from 1999 to 2002 (Giri et al., 2013).

2.4.2 Salt flat (Apicum)

In general, the less frequently flooded area of a mangrove swamp, in the transition to topographically elevated lands, is usually devoid of arboreal vegetation. In Brazil, this area is referred to as “Apicum”. In the international scientific literature, this transition zone is usually called salt flat or hypersaline tidal flat. Two different reference maps were here used, the “Atlas Dos Remanescentes Florestais da Mata Atlântica” from 2016/2017, covering the Atlantic Forest coastal region and the “Carta de Sensibilidade Ambiental ao Oleo -Para-Maranhão-Barreirinhas” referent to 2017 and covering most of the Brazilian north coast region (Figure 6).

Figure 6. Apicum (Salt flat) reference maps, from the “Atlas Dos Remanescentes Florestais da Mata Atlântica” from 2016/2017, covering the Atlantic Forest coastal region and the “Carta de Sensibilidade Ambiental ao Oleo -Para-Maranhão-Barreirinhas 2017”, covering most of the Brazilian north coast region.

2.4.3 Beach and Dune

Mapped without distinction between one another, in here the Beach and Dune class refers to sandy strands, bright white in color, where there is no predominance of vegetation of any kind. The training data for this land cover was obtained from MapBiomas Collection 2.0, ranging from 2000 to 2016 and from the “Atlas Dos Remanescentes Florestais da Mata Atlântica” from 2016/2017 (Figure 7).

Figure 7. The training data obtained from MapBiomas Collection 2.0, ranging from 2000 to 2016 and from the “Atlas Dos Remanescentes Florestais da Mata Atlântica 2016/2017”.

2.4.4 Aquaculture/Saliculture

Area related to artificial lakes/lagoons, where aquaculture and/or saliculture activities predominate. In here no samples were required, once this class recognition does not rely on the spectral domain of the target. Aquaculture/Saliculture areas are usually occupied by water filled tanks, a characteristic that complicates its distinction from natural lakes, rivers oceans and other water related targets. In here, the context domain, rather than the spectral, dominates the distinction of this class from other natural water reservoirs. The process happens in 2 steps. First, based on MNDWI index, a water mask was created, then over the water class only, context parameters were extracted (Figure 8), such as:

● Area, ● Perimeter, ● Size/perimeter ratio ● Polsby Pooper index ● Distance between class center

Figure 8. Over the water mask, context parameters were extracted, and used to distinguish between Aquaculture/Saliculture from other natural water reservoirs. The context parameters used are; area, perimeter, size/perimeter ratio, Polsby Pooper index, distance between class centers.

3 Post-classification

The post-classification process includes the application of temporal and spatial filters.

3.1 Temporal filter

This step uses the temporal information to identify not-allowed class transitions between consecutive years. We used a temporal filter with a total coverage of 3 years, 1 year before and 1 year after the central year, which is year to be filtered. Applied over the classified products, the temporal filter sets off a group of transition rules, which identifies and reclassifies not allowed temporal transitions of any given class. In this way, it is possible to remove cloud residues, reduce the

effects of stripes for the operation of the SLC sensor, as much as correcting the impossible transitions.

3.2 Spatial filter

Spatial filters are traditionally used in Remote Sensing (RS) and Digital Image Processing (DIP) areas, although most of the traditional filter approaches end-up removing isolated classified pixels by using traditional “Moving Window” approach. Thus, isolated pixels were removed, but the removal contaminates the adjacent class codes. To void such adjacent contamination, the filter applied over Collection 3 takes into consideration the GEE connectedComponents function. Such function finds connected components with the same value of the first band of the input and labels them with a globally unique value. Connectedness is specified by the given kernel. Objects larger than maxSize are considered background, and are masked, which ends-up removing isolated pixels without contaminating adjacent class codes (Figure 8).

Figure 9. GEE connectedComponents function results.

3.3 Integration with biomes themes

After the application of the temporal filter, for each one of the 33 years (1985-2017), Coastal Zone and Mining mappings were then integrated with others cross-cutting themes and biomes, by applying a set of specific hierarchical prevalence rules (e.g. Table 5). As output of this step, a final land use and land cover (LULC) was obtained for each annual mapping.

Table 5. Prevalence rules for combining the output of digital classification with the crosscutting themes in the MapBiomas Collection 3.

4 Validation strategies 4.1 Use of reference maps

The LULC maps used as reference were the “Atlas Dos Remanescentes Florestais da Mata Atlântica - 2016/2017” coordinated by SOS-Mata Atlantica, the “Global Mangrove Cover Data - 2013” coordinated by Dr. Giri Chandra (Giri et al., 2013), and the “Atlas dos Manguezais do Brasil - 2018”, produced by Chico Mendes Institute for Biodiversity Conservation – ICMBIo, (ICMBio, 2017). The comparison of the maps will be included in the next version of ATBD. 5 References Breiman, L. Random forests. Machine learning, v. 45, n. 1, p. 5-32, 2001.

Diniz et. al, under review. Giri, C., Ochieng, E., Tieszen, L. L., Zhu, Z., Singh, A., Loveland, T., … Duke, N. (2013). Global Mangrove Forests Distribution, 2000. Palisades, NY: NASA Socioeconomic Data and Applications Center (SEDAC). ICMBio. (2017). Atlas dos manguezais do Brasil. (ICMBio, Ed.) (1st ed.). Retrieved from http://www.icmbio.gov.br/portal/images/stories/manguezais/atlas_dos_manguezais_do_brasil.pdf. HUETE, A. R. A soil-adjusted vegetation index (SAVI). Remote Sensing of

Environment, v. 25, n. 3, p. 295–309, 1988.

ROGERS, A. S.; KEARNEY, M. S. Reducing signature variability in unmixing coastal marsh Thematic Mapper scenes using spectral indices. International Journal of Remote Sensing, 2004. TUCKER, C. J. Red and photographic infrared linear combinations for monitoring

vegetation. Remote Sensing of Environment, v. 8, n. 2, p. 127–150, 1979.

USGS. LANDSAT COLLECTION 1 LEVEL 1 PRODUCT DEFINITION. p. 26, 2017.

XU, H. Modification of normalised difference water index (NDWI) to enhance open

water features in remotely sensed imagery. International Journal of Remote

Sensing, v. 27, n. 14, p. 3025–3033, 20 jul. 2006.