Benchmark maps of 33 years of secondary forest age for Brazil

Our method was implemented in the Google Earth Engine (GEE) platform¹⁹. We divided it into four steps. Figure 2 summarizes our approach, including the input of the raw data (land-use and land-cover from 1985 to 2018 and the water surface), and the output data (from 1986 to 2018), which included maps of the annual secondary forest increment (Product 1), annual secondary forest extent (Product 2), annual secondary forest loss (Product 3; from 1987 to 2018), and annual secondary forest age maps (Product 4).

Input data

We used the land-use and land-cover data from the Brazilian Annual Land-Use and Land-Cover Mapping Project (MapBiomas Collection 4.1; https://mapbiomas.org/en/colecoes-mapbiomas-1)¹ as input data. This dataset was obtained through the classification of images from the Landsat satellite series (30-m spatial resolution) using a theoretical algorithm implemented in the GEE platform¹⁹. Details about the processing of the dataset can be found in the Algorithm Theoretical Basis Document²⁰. More detail about the land-use and land-cover classes can be found in the MapBiomas website (https://mapbiomas.org/en/codigos-de-legenda?cama_set_language=en).

Moreover, we used the maximum water surface extent data (from 1984 to 2018) developed by Pekel et al.²¹ (https://global-surface-water.appspot.com) to avoid the inclusion of false detection within wetland areas in our products. This dataset contains a map of the spatial distribution of the water surface cover from 1984 to 2018, globally²¹. These data were obtained from 3,865,618 Landsat 5, 7, and 8 scenes acquired between 16 March 1984 and 31 December 2018. Each pixel was individually classified into water or non-water cover using an expert system²¹ implemented in the GEE platform¹⁹.

Step 1 – Reclassifying MapBiomas data

All MapBiomas land-use and land-cover maps from 1985 to 2018 (34 maps) were reclassified into binary maps. We assigned the value “1” for all pixels in the Forest formation class of the MapBiomas product (Legend ID: 3) and “0” for the other land-use and land-cover classes. In our reclassified maps, pixels with value of “1” were, then, associated to the class “Forest”, which includes only forests classified as old-growth and secondary (before 1985). Mangrove and forest plantation classes were excluded from our secondary forest map.

Step 2 – Mapping the Annual Increment of Secondary Forests

We mapped the annual increment of secondary forests using the forest maps produced in Step 1. This process was carried out pixel-by-pixel, where every pixel classified as Forest (value 1) in the analysed year (y_i; between 1986 to 2018) and classified as non-forest (value 0) in the previous year (y_i-1; i = 1985, 1986… 2017) was mapped as secondary forest. As forest cover maps before 1985 were not available in the MapBiomas product, maps of secondary forest increment start in 1986, when it was possible to detect the first transition (1985 to 1986). Thus, 33 binary maps were obtained, where the secondary forest increments (non-forest to forest) have a value of 1 and the other transitions a value of 0 (forest to forest, non-forest to non-forest, and forest to non-forest). Here, we only considered secondary forest growth in pixels that had previously an anthropic cover (forest plantation, pasture, agriculture, mosaic of agriculture and pasture, urban infrastructure, and mining) and did not overlap wetland areas.

Step 3 – Mapping the Annual Extent of Secondary Forests

We generated 33 maps of the annual extent of secondary forests. To produce the map of secondary forest extent in 1987, we summed the map of the total secondary forest extent in 1986, which is the same map as the secondary forest increment in 1986 from step 2, with the 1987 increment map, resulting in a map containing all secondary forest pixels from 1986 and 1987. Knowing that the sequential sum of these maps results in pixels with values higher than 1, to create annual binary maps of secondary forest extent, we reclassified the map produced for each year by assigning the value 1 to pixels with values between 2 and 33 (secondary forest extent) and pixels with a value 0 were kept unchanged. Finally, to remove all secondary forest pixels that were deforested in 1987, keeping in the map only pixels with the extent of stand secondary forests, we multiplied the resulting map by the annual forest cover map of 1987, produced in step 1 (Fig. 3). This procedure was applied year-by-year from 1986 to 2018 to produce the maps of annual secondary forest extent. The removal of deforested pixels provides a product depicting the extent of secondary forest deforested in each specific year and they were also included as complimentary maps (from 1987 to 2018) in our dataset.

Step 4 – Calculating the Age of Secondary Forest

Finally, we calculated the age of the secondary forests (Fig. 3). First, we summed the 1986 map of annual secondary forest extent (from Step 3) with the 1987 map to obtain the age of secondary forests in 1987 (Fig. 4). We continued this summation year-by-year until the secondary forest age map of 2018 was obtained (Fig. 4). The values of each pixel in 2018 correspond to the age of the secondary forest. To ensure the elimination of deforested secondary forests from each age map, we executed a similar procedure as described in step 3 by removing all forest pixels overlaying non-forest areas (Fig. 4). As our analyses started in 1986, it was not possible to identify secondary forests before this year. The 1986 age map, therefore, only shows one-year old secondary forests, and the 2018 map shows ages of secondary forest varying between 1 and 33 years (Fig. 4). If a secondary forest pixel with any age is cleared in a given year, it is then removed and a value of zero is attributed to the pixel. The age of this pixel, subsequently, will only be computed again if the algorithm detects a new non-forest to forest transition in the forest cover map (Step 1), which depends on the MapBiomas project classification method.

Source: Ecology - nature.com