Fig 1.
Geographical locations of the Atlantic Forest biome in green and of our region of interest in red (a); region of interest with 2017 land use/cover classes from the MapBiomas project, the WorldView-2 and WorldView-3 tiles used in this study, and the area covered by the aerial photography of 1962 called further “Rio do chapeu” region (b).
Fig 2.
Example of images (256 × 256 pixels) with natural (a-b) forest and planted eucalyptus forest (c-d).
Satellite image courtesy of the DigitalGlobe Foundation.
Fig 3.
Example of images (256 × 256 pixels) with the species of interest, Tibouchina pulchra (a-d).
Satellite image courtesy of the DigitalGlobe Foundation.
Fig 4.
U-net architecture for the forest types segmentation, adapted from [26] and reproduced from [14].
The number of channels is indicated above the cuboids and the vertical numbers indicate the row and column size in pixels.
Table 1.
Numerical evaluation of the models, training and validation sample size and convergence details.
Table 2.
F1-scores of the segmentation for natural and eucalyptus forests as well as for T. pulchra and C. hololeuca trees.
Fig 5.
Example of validation images (128 × 128 pixels) for T. pulchra trees with manual delineation in orange and the U-net delineation in green (a-d).
Satellite image courtesy of the DigitalGlobe Foundation.
Fig 6.
Distribution and relative age of the “Rio do Chapeu” forests in 2017.
Two types of forests, planted or natural, were identified with the U-net method. And, among the natural forests, new forests indicate forests that were pasture in 1962 (from the registered 1962 aerial photos) while old forests indicate forests that were already there in 1962. Note that all eucalyptus forest have been planted after 1962. All the pixels which are not planted or natural forests are in white.
Fig 7.
Distribution of Tibouchina pulchra (transparent magenta) and Cecropia hololeuca (transparent blue) dominances in the natural forests of the “Rio do Chapeu” region.
Dominance was defined as pixels with more than 2.84% of C. hololeuca or more than 17.5% of T. pulchra within a radius of 25 m, for C. hololeuca and T. pulchra, respectively. New forests indicate forests that were pasture in 1962 (from the registered 1962 aerial photos) while old forests indicate forests that were already there in 1962. All the pixels which are not natural forests are in white.
Fig 8.
Map of the natural forests in 2017 with additional information on Tibouchina pulchra and Cecropia hololeuca dominances.
Dominance was defined as pixels with more than 2.84% of C. hololeuca or more than 17.5% of T. pulchra within a radius of 25 m, for C. hololeuca and T. pulchra, respectively. Natural forest, T. pulchra and C. hololeuca were mapped on the extends cover by the Worldview-3 and -2 image.
Fig 9.
Distribution of the natural forests and non forest (drawn from a random distribution) in relation to elevation and five elevation-related variables at the regional scale.
The five elevation-related indices consider the 8 neighbor pixels and are slope (°); aspect (compass direction facing slope in °, where 0 is the North); topographic position index (elevation of the pixel in relation to the mean elevation of the neighbor pixels); terrain ruggedness index (sum of the change in elevation between the pixels and its 8 neighbors cells); and finally roughness (difference between the maximum and the minimum value of a cell and its 8 surrounding cells) [31, 32].
Fig 10.
Comparison of the distribution of the forests with or without T. pulchra dominance in relation with annual mean of climate variables over the Worldview2 image extent.
Climate variables were extracted for T. pulchra dominance polygon centroids (n = 461355) and from randomly distributed natural forests points without T. pulchra dominance (n = 50000).