Mangrove changes over the past decade in South and Southeast Brazil using spaceborne optical and SAR imagery

: Mangroves occur in the tropics and subtropics. This region is constantly covered by clouds and therefore highly challenging to map and monitor. Technological advances in remote sensing have increased the flexibility of performing such analyses. In this study, mapping and change detection were carried out for mangrove areas of the South and Southeast regions of Brazil between 2008 and 2016 using multisensor data and geographical object-based image analysis (GEOBIA). The 823.03 km² mangrove areas in study site in 2008 were reduced to 789.00 km² in 2016, representing a net loss of ~34 km². A change detection analysis of the mangrove areas showed a total gain of 138.21 km², a total loss of 172.24 km² and no change for 650.79 km². The GEOBIA classification accuracy was assessed by performing a statistical analysis of confusion matrix: (2008): global accuracy = 0.92, Kappa index = 0.84 and Tau index = 0.84; and (2016): global accuracy = 0.93, Kappa index = 0.86 and Tau index = 0.86. These results demonstrate the effectiveness of the GEOBIA to map and analyze mangrove dynamics. The results exhibit an excellent accuracy. Furthermore, mangrove areas in the


INTRODUCTION
Remote sensing is an important technique for the monitoring, quantification and detection of changes in mangroves.These ecosystems are coastal environments that are distributed worldwide in tropical and subtropical regions, and are highly impacted by anthropic actions, mainly housing expansion, agriculture and aquaculture activities (Glaser 2003, Alongi 2008).
Global mangrove area is estimated to be 137,760 km² for 118 countries, where 75% of the total area was concentrated in only 15 countries Giri et al. (2011).These data rank Brazil as having the third largest mangrove area of 9,940 km², behind Indonesia and Australia (Diniz et al. 2019).The mapping of mangroves in Brazil is constantly evolving.Initial studies involved visual interpretation of radar images (Terchunian et al. 1986, Herz 1991, Souza-Filho 2005), optical images and multi-sensor data fusion (Souza-Filho & Paradella 2002, Souza-Filho & Paradella 2005).The advent of computation enhanced automatic image classification and optimized the processing time of digital images.Several classification algorithms with specific capabilities were developed to aid visual analysis.The subsequent emergence of pixel value analysis (pixel-pixel classification) has provided more quantitative and replicable information than visual interpretation (Green et al. 1998).
Continuous advances in computational technologies and image segmentation tools that use geographic object-based image analysis (GEOBIA) have enabled extensive areas of the Earth's surface to be classified more rapidly than by a single image analyst (Kamal et al. 2015, Nascimento Jr et al. 2013).Using GEOBIA to map mangrove areas has increased methodological flexibility, precision, and statistical accuracy, while facilitating change detection analysis (Santos 2012, Nascimento Jr et al. 2013, Kamal et al. 2015, Pereira 2015).
In this study, GEOBIA was used to analyze different multisource remote sensing data (Landsat, ALOS PALSAR and digital elevation model) to map and detect changes in mangrove forests in the coastal areas of the South and Southeast regions of Brazil from 2008 to 2016.

Study Area
The study area is located in South and Southeast Brazil and comprises a strip of five coastal states: Espírito Santo -ES, Rio de Janeiro -RJ, São Paulo -SP, Paraná -PR and Santa Catarina -SC (Figure 1).The highest demographic density of these states is located mainly in this regions (IBGE 2011), which may increase anthropic actions in areas near mangrove forests.

Remote sensor data set and digital image processing
The images used in this study include scenes from the satellite Landsat-5 TM ( 2008), Landsat-8 OLI (2016), ALOS-1 PALSAR ( 2008) and ALOS-2 PALSAR (2016).In addition, digital elevation models -DEM (Kartikeyan et al. 1998) were  I shows the main characteristics of the data used.
The Landsat satellite images and the SRTM DEMs were obtained from the "Earth Explorer" portal managed by the United States Geological Survey -USGS (https://earthexplorer.usgs.gov),whose data characteristics and pre-processing steps are described in USGS (2012).The ALOS PALSAR images were acquired through the Japan Aerospace Exploration Agency -JAXA website (https://jda.jaxa.jp/en/).It is a fully polarimetric instrument, which operates in L-band, with 23.6 cm in wavelength (Rosenqvist et al. 2007).Figure 2 shows an example of a set of images used to map the mangrove regions in the study area.
An atmospheric correction was performed on the optical images to reduce the effects of atmospheric brightness in the scenes.Each digital number was converted to ground reflectance using the atmospheric correction module (ATCOOR) of PCI Geomatica software (2017).
The ALOS PALSAR images were preprocessed using Envi 5.5 software.The images were reprojected to WGS84 datum, mosaiced and finally converted to backscatter using eCognition Developer software v. 9.

Object-oriented classification and segmentation
This study was developed from geographical object-based image analysis.The remote sensing images were segmented at different levels using the same scale and heterogeneity parameters.However, different weights were assigned to each band/sensor to explore the qualities of different images for certain specific functions.
Next, an unsupervised classification algorithm was used to obtain three main classes (water bodies, upland, and mangrove) and two secondary classes (cloud and mangrove below cloud).To better define the coastline of the study area, the radar images in segmentation level 1 were assigned higher weights than other image types, because radar images are not subject to interference from clouds and are therefore more reliable.
Segmentation level 2 was established below level 1, i.e., level 2 image objects were included in the level 1 image objects to increase the level of detail of the segmentation.The optical sensor images were assigned the highest weights in level 2. The main objective of the level 2 segmentation was to separate the mangrove vegetation from upland vegetation.Radar images were used to classify mangroves below clouds, albeit with lower weights than the other images.Hence, 67.30 km² of mangrove areas under clouds were mapped.This represents approximately 8% of the total area mapped in 2008.In 2016, there were no clouds in the study site.Figure 3 illustrates the differences between the segmentation levels.
In level 3, the classification was refined using the same scale and heterogeneity parameters as level 2. Images of the mangrove class were grouped separately.Border relations between the mangrove and upland classes were used to remove some incorrect mangrove classifications resulting from confusion of the spectral responses of the targets.Thus, polygons classified as mangrove in the continental domain were reclassified as upland class.The expressions used in the classifications and modifications are shown in Table II.
First, the water bodies and continent (coastal plain and upland) classes were classified.The classification parameters were defined mainly based on the backscatter values (in decibels-dB) of the SAR data, as represented by "Expression 1", and were essential to define the coastline (Table II).The mangrove class was classified next.The parameters were chosen based on the mean reflectance of the red, nearinfrared and mid-infrared Landsat-5 TM spectral bands (Mean B3, B5 and B7, respectively) and the Landsat-8 OLI bands 4, 5 and 6 (red, nearinfrared and shortwave infrared, respectively) using "Expression 2" (Table II).Other expressions were also used to optimize the mangrove classification process for 2008 and 2016 (Table II)."Expression 3" consists of an average ratio between SAR and DEM data and was essential in the mapping mangroves below clouds, as was "Expression 4".DEM was also used alone as an alternative to limit the occurrence of mangroves using the forest canopy height.Figure 4 shows the process tree used to classify the three main classes mentioned above and the parameters used for each segmentation level.
The mangroves were classified following the sequence detailed in Figure 4. Initially, the water bodies (rivers, lakes, and ocean) and the continent were classified to define the coastline.The mangroves were then classified by combining the reflectance of the red and infrared bands with the backscatter of the PALSAR images and the elevation of the SRTM and ALSO DEM.The Landsat scenes provided the spectral interval to characterize the mangrove, the PALSAR images were used to identify the forest cover, and the DEM enabled separation of the mangrove vegetation by canopy height.Evaluation of object-oriented classification.
The classification was evaluated using a collection of control points on high-resolution images of the Google Earth Pro platform.These points were selected very close to the limits of what were believed to be mangrove areas or locations likely to be confused with other classes.Thus, both the segmentation and classification were evaluated for areas with close separability thresholds.A total of 600 control points were collected, of which 300 were chosen For the change detection analysis, 1000 control points were distributed between the gain, loss and no change classes, which were also calculated using the indices cited above and the area estimates (Olofsson et al. 2014).The objective of this analysis followed best practices for accuracy assessment of the change classification and estimation of area of change in terms of a classification error matrix.The error matrix was used to crosstabulate the land change class labels allocated by the classification against the reference GCPs collected at sample sites.Accuracy parameters derived from a sample error matrix included overall accuracy, user and producer accuracy of each class (Olofsson et al. 2014).

RESULTS
GEOBIA was used with multisensor data to map the mangroves in the coastal region of South and Southeast Brazil.Images and detailed maps are shown in Figure 6.The coastal zone of the study area contained 823.03 km² of mangrove forests in 2008, which decreased to 789.00 km² in 2016.This reduction in the mangrove area of approximately 34 km² is equivalent to a 4.1% loss.
Throughout the study period, a loss of mangrove area was found for almost all coastal states, except RJ, which was characterized by a 3.6% expansion of mangrove areas.The highest mangrove area losses were recorded for ES (~20%), followed by SC (~13%), SP (~3%) and PR (0.45%).

Classification accuracy
For 2008, 278 or 92.67% of the 300 points collected in the mangrove area were correctly classified, whereas 22 points or 7.33% did not match the sampled points.The classification accuracy for mangrove areas in 2016 was the same as for 2008, i.e. 278 or 92.67% of the 300 collected points were correctly classified.
For each year, 300 points were also sampled for the "others" class, of which 274 or 93.6% were classified correctly in 2008, and 26 points or 8.6% did not match the sampled points.The accuracy of the "others" class was higher for 2016 than for 2008: 281 or 93.7% of the 300 collected points were classified correctly for 2016, whereas 19 points or 6.3% did not match the classification.
Several indices were also used to assess the classification accuracy.The following values were obtained for 2008: an overall Kappa index of 0.84; the Kappa index per class reached 0.85 for the mangrove class and 0.82 for the "others" class; an overall accuracy of 92% and a Tau index of 0.84 (Table III).The following values were obtained for 2016: an overall Kappa index of 0.86; a Kappa index per class of 0.85 for the mangrove class and 0.87 for the "others" class; a 93% overall accuracy and a Tau index of 0.86 (Table III).

Geographical object-oriented change detection
The change detection accuracy was evaluated using the random and stratified distributions of 1000 control points.The overall accuracy was 86.50%.Of 13.50% disagreements (Pontius Jr & Millones 2011), 5.60% were quantity disagreements, and 7.90% were allocation disagreements.The largest quantity disagreements (5.60%) were observed for the no change mangrove class, and the largest allocation disagreements were related to the gain mangrove class (5.80%).
The area estimates were adjusted using the methodology of Olofsson et al. (2014).There were more losses than initially quantified, resulting in fewer unchanged areas than initially estimated.The quantity of gains was almost constant.Table IV summarizes the main data obtained from the confusion matrix.
The change detection analysis was used to identify areas of expansion, reduction and no change in mangrove vegetation.The unchanged mangrove areas represented approximately 80% of the study area, for which the largest percentage (~84%) corresponded to RJ, followed by PR (~82%) and SP (80%).There was a 20.90% loss and a 16.80% gain in mangrove areas over the entire period of study.An analysis of the balance between the total gains and losses of the mangrove areas showed that all states had Overall accuracy = 0.92 Kappa index = 0.84 Tau index = 0.84 higher losses of mangrove areas than gains over the period of study, except for RJ, which exhibited more gains in mangrove areas than losses.Figure 7 is a graphical illustration of the balance between losses and gains of mangrove areas by state.Considerable changes were observed near the Piraquê-açu River in ES.Guanabara Bay was representative of gain and loss of mangrove areas in RJ.The most significant changes in SP occurred in the vicinity of Santos Bay.Changes in PR mainly occurred in the Paranaguá Estuarine Complex -CEPII, which comprises four bays (Paranaguá, Pinheiros, Guaraqueçaba and Laranjeiras) with extensive mangrove areas.Most of the mangrove areas in SC are found in the São Francisco Bay or Babitonga.Figure 8 is a change detection map for the entire study area, wherein the areas mentioned above are highlighted.

Mangrove extent in South and Southeast Brazil
The analysis of remote sensing images to quantify the mangrove area in the coastal zone of South and Southeast Brazil is important for conservation.The mangrove forest area in South and Southeast Brazil was estimated at 823.03 and 789.00 km 2 for 2008 and 2016, respectively, in this study.Different methods have previously been used to estimate the mangrove area: for example, the SOS Mata Atlântica (Atlantic Forest) project (SOS Mata Atlântica 2018) and the Mapbiomas project (2019), conducted by Diniz et al. (2019), quantified the mangrove area for South and Southeast Brazil at 892.9 km 2 and 636.02 km 2 , respectively, for 2008 and at 953.8 km 2 and 615.14 km 2 , respectively, for 2016.Figure 9 shows the mangrove areas estimated by different authors for each of the coastal states of South and Southeast Brazil (Diniz et al. 2019, SOS Mata Atlântica 2009, 2018).
The estimates obtained in the present study are generally similar to those obtained by the Mapbiomas project, given the similar nature of the data analysed.That is, Landsat images, with 30-m spatial resolution, were used to map mangrove forests.ALOS PALSAR images were also used in this study to minimize the effect of cloud cover that was around 8% in 2008, whereas a series of cloud-free images (Diniz et al. 2019) were used in the Mapbiomas project.In both studies, a loss of mangrove areas was observed, with the exception of RJ, for which there was a mangrove gain.Relatively larger areas were estimated by the Atlântica project (2009,2018), because the mangrove ecosystem was mapped as a whole, considering both mangrove forests and associated hypersaline fields, locally known as "apicum".Our estimate can be considered more accurate than previous estimates because of the inclusion of ALOS PALSAR images, with a 10-m spatial resolution, which are associated with Landsat images and classified using GEOBIA instead of isolated pixels.
Table IV.Summary of data obtained from confusion matrix for change detection.Loss, gain and no change were calculated from methodological approach proposed by Olofsson et al. (2014).Overall accuracy (OA), overall disagreement (OD), quantity disagreement (QD), and allocation disagreement (AD) were estimated from Pontius Jr & Millones' (2011).mangrove and Florianópolis Island (Trindade 2009), whereas expansion is associated with the progradation of mudflats subsequently colonized by mangrove vegetation in protected areas within coastal bays, such as Santos and Guanabara bays (Figure 8), where mangrove forest presents higher structural development (Cavalcanti et al. 2009) Comparing this dynamic with other regions of the country shows that the highest loss of mangrove area in Brazil is in the South and Southeast, because the mangrove area remained practically stable over the period analysed in the Northeast region (MapBiomas 2018), and there was an increase of approximately 10% in the North (Nascimento Jr et al. 2013).However, we found that the mangroves of South and Southeast Brazil (the most densely populated region of Brazil) are still well preserved in comparison to other developing regions with high population densities, such as China, the Philippines and Vietnam.For example, China has lost 48% of its mangrove area over the last 50 years (Jia et al. 2018), whereas the Philippines (Primavera 2000) and Vietnam (Valiela et al. 2001) lost 73% and 62%, respectively.Furthermore, in response to global changes associated to warming process, Soares et al. ( 2012) expect that mangrove forest will expand southward of their present latitudinal limit at Laguna, Santa Catarina State.

CLASS
Brazilian mangroves have been successfully conserved in response to public policies, that is, mangrove areas have been permanently protected since the publication of the Brazilian Constitution of 1988.In addition, a series of sustainable-use conservation units, including extractive and restricted-use reserves, such as national parks and biological reserves (Gerhardinger et al. 2009, Tenório et al. 2015), were created along the Brazilian coast, making it difficult to annex mangrove areas for different use purposes.
Major drivers of mangrove changes in the study site are related to port activities (Ferreira & Lacerda 2016), urbanization (Lacerda et al.  -(Brasil 2012, Borges et al. 2017).Due to lack of proper evaluation of mangrove functioning along the Brazilian coast and environmental inconsistencies of the new Brazilian Forest Code that excluded salt flats from mangrove protection areas, anthropogenic drivers have the potential to increase threats and reduce the effectiveness of conservation of this important ecosystem (Oliveira-Filho et al. 2016).At least, an annual assessment of mangrove extension has been carried out from MapBiomas project (Souza et al. 2020, Diniz et al. 2019), hence will be possible to assess the status and sustainable use of mangrove forest and salt flats along the Brazilian coast to break negative anthropogenic impacts on mangroves.Restoration efforts need to be increased to minimized losses and expand the use of natural services given by mangrove forests (Ferreira & Lacerda 2016).

CONCLUSIONS
Satisfactory results were obtained by using optical and radar remote sensing data in conjunction with GEOBIA techniques to map mangrove forest areas in South and Southeast Brazil.This alternative mapping scheme exploits the advantages of different sensor types, irrespective of their characteristics.There has been an overall 4% reduction in the mangrove areas of South and Southeast Brazil over the last decade.This reduction is reflected in all states except in Rio de Janeiro.The mangrove area loss of 4% is considered low in the most densely populated region of Brazil.Hence, we have an important opportunity to tell success stories that have helped protect mangroves, making then an important and positive case study for the conservation optimism movement for now (Friess et al. 2020).
The results obtained in this study can be used as a reference for future studies and to monitor the development and dynamics of mangroves to evaluate the main natural and anthropogenic factors affecting change in the coastal landscape of South and Southeast Brazil.

Figure 1 .
Figure 1.Distribution map of mangroves for coastal states in South and Southeast Brazil

Figure 2 .
Figure 2. Example of a set of space-borne images used to map mangroves.a) Landsat-5 TM image in color composite 5R4G3B; b) Landsat-8 OLI in color composite 6R5G4B; c) 2008 and d) 2016 ALOS PALSAR monochromatic L-band in HH-polarization images; e) 2000 SRTM digital elevation model; and f) 2014 ALOS PALSAR digital elevation model.

Figure 3 .
Figure 3. Multiresolution segmentation.a) and b) Segmentation level 2. c) and d) Segmentation level 1. Observe that segmentation level 1 presents larger and fewer objects in comparison to level 2. Note in detail (yellow box in figure a and c) differences between segmentation levels (figure b and d).
Table II.Expressions used to classify mangrove area.B5 -Mean B3)/(Mean B5 + Mean B3)] + Mean B7 [(Mean B5 -Mean B4)/(Mean B5 + Mean B4)] + Mean B6 Expression 3 Mean HV/Mean DEM Mean HV/Mean DEM Expression 4 10 x log (Mean HV)/DEM 10 x log (Mean HV)/DEM Object-oriented change detection An object-oriented change detection analysis was adapted from the methodology proposed by Nascimento Jr et al. (2013) The mangrove 2008 and mangrove 2016 classes were initially generated.Arithmetic operations were used to generate the classes of gain, loss and no change for the mangrove area (Schlosser & Pfirman 2012).The gain class represented areas that were present in the mangrove 2016 class but not the mangrove 2008 class.The loss class represented areas that were not present in the mangrove 2016 class; that is, areas that were only classified in the mangrove 2008 class.Finally, the no change class represented unchanged areas, i.e. those areas that remained constant for the two classifications.Figure 5 is a flowchart summarizing all the change detection stages.

Figure 4 .
Figure 4. Flowchart showing processing steps for mangrove classification: SAR, optical and DEM images were subjected to three segmentation levels, and mangrove classification was obtained after combination of segmentation levels 2 and 3.

Figure 5 .
Figure 5. Example of change detection process using detail of area in Espírito Santo state, near the mouth of the Piraquê-Açu river.First: Input images of Landsat mosaics 2008 and 2016.Second: Segmentation of Landsat mosaics 2008 and 2016.Third: Classification for respective years from geographical object-based image analysis.Finally, object-oriented change detection analysis was performed to identify areas of gain, loss, and no change of mangrove forests.

Figure 6 .
Figure 6.Detailed map of mangrove areas for most representative regions of investigated states in 2016.
.50 OD = 13.50QD = 5.60 AD = 7.90 Mangrove losses in South and Southeast Brazil and elsewhere Over 56% of the Brazilian population reside in South and Southeast Brazil (IBGE 2011).These regions have the highest demographic densities, ranging from 49 to 87 inhabitants per km 2 , but surprisingly lost only 4% of their mangrove area over the last decade.Furthermore, the fringed mangroves of the southeast and south of Brazil are bounded by highland areas ("Serra do Mar") and are the ones most threatened by sealevel rise (Souza-Filho et al. 2023).The loss of mangrove area is mainly related to processes of occupation of the coastal zone as observed in Piraquê-Açu river, Espírito Santo State (Figure 8) due to impact of upland use in the adjacent

Figure 7 .
Figure 7. Balance of gain and loss of mangrove areas for each state from 2008 to 2016.

Figure 8 .
Figure 8. Change detection map for study area: regions with most significant changes for each state are demarcated by red boxes.

Figure 9 .
Figure 9. Estimates of mangrove area by different authors for the coastal states of South and Southeast Brazil (MAPBIOMAS -Diniz et al. 2019, SOS Mata Atlântica 2009, 2018).

Table I .
Main characteristics of remote sensing data.

Table III .
Classification confusion matrix for 2008 (A) and 2016 (B).