Occurrence and spatial distribution of native acai groves in high-production areas of the Amazon region

ABSTRACT The floodplain forests of the Amazon estuary have undergone constant change over recent years, where management techniques, especially intensive management, have had an impact on the dynamics of the vegetation and land use. These changes can be monitored using satellite data. With this in mind, the aim of this study was to evaluate the dynamics of ground vegetation on the islands of Jarimbu, Mamangal, Itaboca, Mutirão and Buçu in the district of Igarapé-Miri, Pará, using images from the RapidEye and Planet satellites. The unsupervised ISODATA classification method was used, generating distinct classes of vegetation between each island. To evaluate the efficiency of the classification, an average of 200 random points were used, with another 30 points relating to the type of usage for each class. The Kappa index and overall precision were also analysed, in addition to calculating errors of omission and commission. Monitoring on a seven-year time scale using high-resolution satellites, a more than 50% increase in the Exposed Soil class was seen for the islands of Jarimbu, Mutirão and Itaboca, the latter responsible for an increase of more than 50% in the Urban Area class. On each of the five islands, the Alluvial class, representing the areas of açaí groves, has emerged over the last seven years, increasing in area at the expense of a reduction in the Arboreal class. In this respect, the confusion matrix showed a mean accuracy for the islands of ‘very good’, with a mean overall precision of 77.74%, and a mean Kappa index of 0.73, indicating strong agreement with the reference data and the classification.


INTRODUCTION
The region of the Amazon estuary is formed by a tangle of islands and adjacent areas, where fl oodplain forests occur, and which are infl uenced by the ocean, with two daily fl ood and ebb-tide cycles (PAROLIN et al., 2004).
The Amazon Forest includes the greatest biodiversity of fauna and fl ora on the planet (MOTA et al., 2020).Euterpe oleracea Mart.(the açaí palm) is spread over a large part of the Amazon basin, where consumption of the fruit has required the areas of açaí groves to be expanded.According to Tagore, Canto and Vasconcelos Sobrinho (2018), changes in the natural environment of the floodplains have motivated people who live by the river to manage the açai groves, with the aim of increasing production and productivity.
Recent technological advances in agriculture, as well as the need for conservation and the effi cient use of natural resources, mean that the scale of ground maps and information on land use and land cover need to be refi ned so that these surveys can be interpreted and used for diff erent purposes (COSTA et al., 2016).This scenario highlights the ability of technology to contribute with accurate information in monitoring large areas, and as the study by Dutra, Elmiro and Garcia (2020) demonstrate, there are many classifi cation techniques currently available for use in remote sensing.
In this context, identifying, mapping and planning are necessary to mitigate the impact of these areas, using geoprocessing together with conservation policies.For Souza et al. (2019), the advancement in classifi cation techniques, and improvements in the spatial resolution of sensors have been fundamental for monitoring land use and land cover in the Amazon, corroborating Santos et al. (2017), who included the use of Geographic Information Systems (GIS) in geoprocessing techniques.
From a study by Ponzoni et al. (2015) on the application of remote sensing, it is possible, through the use of mapping, to explore different ways of monitoring vegetation, even of estimating production.The use of classification techniques makes it possible to represent a real-world object, obtaining a thematic map as a result (FLORENZANO, 2011).For this reason, it is important to map and identify classes of land use and land cover through the analysis of spatial databases (ROSA; SOUZA; SÁNCHEZ, 2020).
The aim of this article was to evaluate the dynamics of ground vegetation in fi ve islands of the district of Igarapé-Miri, Pará (PA), using a spatial-analytical approach employing remote sensing techniques; to quantify the changes that occurred, by classifying the landscape of the region during 2013 and 2019, using high-resolution data from the Planet and RapidEye satellites; and to apply validation techniques to the unsupervised ISODATA classifier for the purposes of environmental activities and planning, aiming for the well-being of the community.

Characterisation of the st udy area
The study area is located in the Lower Tocantins region, in the district of Igarapé-Miri, part of the mesoregion of north-eastern Pará.It is 78 km from the capital of the state of Pará and has an area of 199,679 ha (IBGE, 2019).The study was carried out on five islands of the PAE (Agroextractivist Settlement Project) which have the greatest occurrence of açaí (Figure 1).
The soils in the area are considered fertile, with a silty clay loam and silty loam texture, high base saturation (greater than 50%), containing high levels of organic matter and significant amounts of potassium and phosphorus (SOARES et al., 2021).
The climate in the region is humid-tropical, corresponding to the Ami megathermal type of the Köppen classification, with an annual rainfall greater than 2,000 mm (ALVARES et al., 2013).The temperature range is small, with the mean annual temperature varying around 27 ºC.Rainfall is abundant from January to June (Figure 2), with more water available during the first three months of the year and a water deficit during September and October (FUNDAÇÃO AMAZÔNIA DE AMPARO A ESTUDOS E PESQUISAS, 2016).
The spatial characteristics of the fi ve islands in the study area are shown in Table 1.

Obtaining and pre-processing the digital image
The Planet and RapidEye high-resolution sensors were used in the present study, which included four RapidEye orbital images and eight from the Planet sensor.The RapidEye images were dated 12 August 2013, and were selected from the geo-catalogue of the Brazilian Ministry of the Environment (MMA), reserved for the Federal Rural University of the Amazon (UFRA) for use in research.The images correspond to location code 2237925 and have a spatial resolution of fi ve metres comprising the following spectral bands: Band 1 (440-510 nm), Band 2 (520-590 nm), Band 3 (630-685 nm), Band 4 (690-730 nm) and Band 5 (760-850 nm).The Planet sensor images were obtained from the State Secretariat for the Environment and Sustainability of Pará (SEMAS) off ered to UFRA for use in research, and dated 9 and 10 August 2019.
Occurrence and spatial distribution of native acai groves in high-production areas of the Amazon region  The images have a spatial resolution of three metres comprising the following spectral bands: Band 1 (455-515 nm), Band 2 (500-590 nm), Band 3 (590-670 nm), Band 4 (590-670 nm).
Each of the images was georeferenced and underwent atmospheric correction.The shapefi les of the islands were downloaded directly from the land collection of INCRA (National Institute of Colonisation and Agrarian Reform).The RapidEye and Planet images were then co-registered, followed by the acquisition of control points, spatial transformation into an adjusted image, and subsequent classification using the ENVI software.Using this process, each scene was represented on a standard geographic coordinate system, allowing their spatial correlation (TULLIO, 2018).
All pre-processing of the satellite images was carried out using the QGis v 3.4.11and ENVI v 5.3 software, generating cropped images showing the boundaries of each island, followed by the mosaic.

Un supervised classifi cation
Th e unsupervised classifi cation was prepared with the ENVI 5.3 software using the ISODATA algorithm (Iterative Self-Organising Data Analysis Technique) developed by Geoff rey and Hall (1965).The ISODATA algorithm requires the operator to have no prior knowledge of the area under study (MORARIU et al., 2018).The minimum spectral distance formula was used for cluster formation and grouping based on the Euclidean distance (SWAIN; DAVIS, 1978) Equation ( 1): ( 1 ) where: η -number of bands; i -band number; cparticular class; X xyi data-file value of pixel X, y in band i; μ ci Mean of the data-fi le values (digital numbers); in i for the sample of class c; SD xyc Spectral distance of pixels X, y, the mean value of class c.
For this classifi cation, a minimum of fi ve and a maximum of 30 classes was adopted, with a maximum of 15 interactions, at the end of which one combination was generated.This procedure includes recognition of the area by the algorithm, which associates pixels to diff erent classes (BLASCHKE, 1954).

Post-processing the digital images
Visual interpretation was used to analyse the images during post-processing within each result.This procedure consists of interpreting the image directly on the computer screen, making use of such basic elements as colour, texture, shape, tonality, size, shadow, pattern, surroundings and geographic location (BARCELLOS et al., 2005;GOMES, 2001;LOCH, 1993;MOREIRA, 2003;TEMBA, 2000).
The image reclassification process was carried out manually, pixel by pixel, using the ClassEdit add-on of the ENVI software.The Flowchart (Figure 3) shows the procedures involved, starting with the acquired images through to the unsupervised classification and data validation.
Eight classes were defi ned for classifi cation: Alluvial -comprising areas with a strong presence of açai groves and the constant presence of water; Arboreal -areas where large and small trees are present; Hydrography -including rivers, holes, lakes and streams; Agriculture -areas consisting of single crops and agroforestry systems; Lowlands -area represented by the presence of açaí groves without the constant presence of water; Exposed Soil -including areas such as sandbanks, bare soil and roads; Urban Areas -consisting of houses, villages and built-up areas, as described by Soares et al. (2021); and Cloud -comprising areas of the image that were unidentifi ed due to cloud cover.

Assessing the accurac y of the classifi cation
A confusion matrix was generated to verify the accuracy of the classifi ed land-use and land-cover data from the Planet sensor.For this purpose, an average of 200 points were randomly collected, with another 30 points resulting from collections in each class for the fi ve islands.Furthermore, parameters of global accuracy and the Occurrence and spatial distribution of native acai groves in high-production areas of the Amazon region Kappa index were used, including errors of omission and commission, as per the methodology used by Congalton (1991).Whereas for the error of omission the number of samples is not classifi ed based on the reference classes, the error of commission refers to the number of samples that are included in a class to which they do not actually belong (FRANCISCO; ALMEIDA, 2012).
Starting with the Kappa analysis as a discrete multivariate technique used in the evaluation of thematic precision, all the elements of the confusion matrix are used, as per Equation 2, proposed by Cohen (1960).
( 2 ) where: r = number of classes; Xij = number of correctly classifi ed elements; Xi+ = total elements classifi ed for category i; X+i = total reference elements sampled for category i; N = total number of samples.
To evaluate the Kappa index, the Landis and Koch criteria (1997) were used, as detailed in Table 2.

Analysing the land use and land cover maps
The classification results for the temporal physical characteristics of 2013 and 2019 for the five islands of Igarapé-Miri (Jarimbu, Mamangal, Itaboca, Mutirão and Buçu), including the classes selected for this study (Hydrography, Exposed Soil, Urban Area, Alluvial, Lowlands, Arboreal and Agriculture) are shown in (Figure 4).
A gradual spatial reduction in the Arboreal class could be seen, due to the increasing dynamics of the Alluvial class and Lowlands, in which the açai grove is becoming predominant in the islands (Figures 4 and 5).The identifi cation of clumps of açai groves, as well as the seven generated classes, is a result of the high resolution of the Planet and RapidEye images.This is corroborated by the work of Asner, Martin and Mascaro (2017)  the precision of the Planet data used for unsupervised classifi cation when detecting the extent of shallow coral reefs, and which gave a mean accuracy of 92%.
The Alluvial and Lowland classes for 2013 add up to 64.49% in areas solely of açai groves.However, in the study by Soares et al. (2021), this total increased to 67.72%, which confirms the formation of homogeneous clumps and management of the açai groves, as described by Homma (2014).
The dynamics of land use and occupation throughout 2013 and 2019 in the area under study i s m o r e s i g n i f i c a n t i n s o m e c l a s s e s , s u c h a s t h e expansion of areas of exposed soil, and urban, alluvial and lowland areas (Figure 5).For Parida and Kumar (2020), multitemporal satellite analysis together with digital image processing are generally employed to monitor changes in vegetation dynamics.
On the islands of Jarimbu and Itaboca (Figure 5A and 5C), the Exposed Soil class more than quadrupled from 2013 to 2019, jumping from 1.82% to 7.44%, and from 1.07% to 5.10% on the islands of Jarimbu and Itaboca, respectively.On Mutirão Island, the area corresponding to this class doubled in size (Figure 5D).This can be attributed to the implementation of new crops and the construction of houses by the community, who cleared the areas for this purpose.Mutirão Island, as it is the export route for all production on the islands, and Buçu, which is small, with its entire area inhabited by the local population.
The areas classifi ed as Alluvial on the fi ve islands, and Lowlands on Itaboca and Mutirão (Figure 5), expanded during 2013 and 2019.This shows that açaí was cultivated on these islands.On the other hand, there is also a reduction of the areas classifi ed as Arboreal in four of the fi ve islands under evaluation, namely, Jarimbu, Mamangal, Itaboca and Mutirão, with a reduction in area of 6.6%, 5.5%, 7, 6% and 5.3%, respectively.With the reduction in the Arboreal class, the Exposed Soil, Urban Area, Alluvial and Lowland classes expanded.
Occurrence and spatial distribution of native acai groves in high-production areas of the Amazon region

Analysing the precision of the classifi cation
Mapping precision was assessed by analysing the confusion matrix for the images from 2019, and is shown in Tables 3 to 7.
The results of this study can be considered very good (Table 2), since the average value of the five islands for overall precision was 77.74% and the mean value for the Kappa index was 0.73.Such results indicate a strong match between the reference and classified data.Oliveira et al. (2020) found that the RapidEye sensor, with a resolution of five metres, provided a more accurate classification with rich detailing, in addition to affording greater target differentiation due to the lower spectral mixing of nearby pixels, as seen in the study by Naesset et al. (2016) on mapping and estimating forest areas using RapidEye images, which significantly helped to improve the estimates of forest areas due to their high spatial resolution, resulting in better accuracy.In this study, which validated the classification, the Planet sensor with a resolution of three metres was used; this determines far better accuracy in the results for land use and cover.According to the confusion matrix, the Hydrography class had the most hits in the fi ve islands (Tables 3 to 7), particularly Jarimbu Island, which determined good general Kappa indices of 0.83, and an overall precision of 86.63%, i.e. the pixel distribution of the Hydrography class was generally not confused with the other classes in the image.In the study by Caten, Safanelli and Ruiz (2015), the Hydrography class showed low refl ectance in relation to the other classes, making any changes both constant and of little signifi cance.
In general, for the Exposed Soil class, an average of 79.27% of hits was obtained for the fi ve islands under evaluation.There was confusion between the Exposed Soil and Urban Area classes, which can be attributed to the similarity of the areas and to their size, since the areas are very small and exposed, making it diffi cult for them to be separated by the algorithm.On Itaboca Island, an accuracy of 84.17% was found for the Exposed Soil class, in addition to excellent results for the other classes, resulting in the second-best classifi cation, with a Kappa Index of 0.73 and overall precision of 77.51% (  The mean accuracy of the areas classifi ed as Arboreal was good, reaching 78.31%, but there was confusion with the Alluvial class on all fi ve islands.As these classes include vegetation, such interaction between similar targets is common, and causes uncertainty when mapping.However, the Arboreal reference class was more successfully sampled than the areas characterised as wrongly classifi ed, demonstrating the diff erences between classes.The islands of Itaboca and Mutirão (Tables 5 and 6) have one class in common: Lowlands.This class represents areas of açaí groves without the constant presence of water in their interior; its mean accuracy is the lowest, at 70.65%, being confused with the characteristics of the Arboreal class.
Comparing the classifi cation results for the fi ve islands makes it possible to analyse the accuracy, error of commission and error of omission (Table 8).Jarimbu Island had the best average performance in terms of accuracy (86.2%), with the Hydrography and Urban Area Occurrence and spatial distribution of native acai groves in high-production areas of the Amazon region classes presenting an error of commission of 1% and 0%, respectively.The error of omission in the Hydrography class was 3%, and in the Urban Area class, 14%, which confirms the high level of identification of the areas belonging to this class.However, Buçu Island had the lowest value for accuracy (56%) compared to the other four islands.This can be explained by the error of omission of 44% and of commission of 6%, which implies user error in collecting the data when sampling this class, resulting in pixels left out of their correct class and assigned to another.
Itaboca Island (Table 8) had the second-best mean accuracy (78.5%).Analysing the Alluvial class with 91% accuracy, the error of commission was high at 43%, indicating the incorrect inclusion of several standard samples in other classes.On the other hand, for the same class, the error of omission was lower (9%), showing that few samples were omitted from this class.The same was seen on Mamangal Island for the Arboreal class, where the error of omission was 11% and of commission, 18%, resulting in a good accuracy level of 89%.Olofsson et al. (2014) discuss assessing accuracy as being fundamental to the quality of mapping in both a quantitative and significant way.
Validating the data for Mutirão Island (Table 8), the Exposed Soil and Agriculture classes had an accuracy value of 81% and 80%, respectively; however both the error of commission and the error of omission were greater for the Urban Area class.In the study by Souza et al. (2019), agglomerations, such as smaller patches, also show similarities, which in turn are confused with other targets, e.g.sand, roads and exposed soil.
In the study by Duarte and Silva (2019) on land-use classification by algorithm, these tools can be used to extract complementary information, helping to optimise the processes and reduce errors.2. The overall precision of the method applied in this study was equal to or greater than 71%, and together with the quality of the classifi cation (Kappa index greater than 0.64), demonstrated that the unsupervised ISODATA method can be important for defi ning the landscape, enabling both speed and accuracy when mapping.In this study, it was evident that the results maintained strong agreement between the classified data and its reference in the field, the only exception being Mamangal Island, where the accuracy of the Urban Area class was low (47%), with errors of commission and omission of 26% and 53%, respectively; 3. In seven years, the Exposed Soil class doubled in size on the islands of Jarimbu, Itaboca and Mutirão.Similarly, the urban areas became more widely distributed on each of the islands, which is a cause for concern.On the other hand, the advance and increase in the Alluvial and Lowland classes is an indication that the local communities began managing the areas of açai groves.With the increase in these areas, there was a reduction in the Arboreal class on the five islands, confirming more intense management of the açai groves to the detriment of the native tree cover.

Figure 3 -
Figure 3 -Methodological fl owchart of the main stages of classifi cation

Figure 5 -
Figure 5 -Classes of land use with their respective areas (ha) for 2013 and 2019, for the islands under study in Igarapé-Miri, Pa

Table 1 -
Characteristics of the islands of the Lower Tocantins River in the district of Igarape-Miri

Table 2 -
, studying Classification criteria based on the bands of the Kappa index

Table 3 -
Confusion matrix for 2019 for Jarimbu Island, district of Igarapé-Miri PA

Table 4 -
Confusion matrix for 2019 for Mamangal Island, district of Igarapé-Miri PA

Table 5 -
Confusion matrix for 2019 for Itaboca Island, district of Igarapé-Miri PA

Table 7 -
Confusion matrix for 2019 for Buçu Island, district of Igarapé-Miri PA

Table 6 -
Confusion matrix for 2019 for Mutirão Island, district of Igarapé-Miri PA to fl oods.It was, however, the class with the least pixels mistakenly distributed in other similar classes, such as the Arboreal, or even Exposed Soil or Urban Areas, which are in direct contact with the Alluvial class in question and where the tonality of the targets is very close.

Table 8 -
Accuracy (AC) and errors of commission (C) and omission (O) related to the fi ve classifi ed islands Adapted from Souza et al. (2019) Space-time analysis by the classification of orbital images from 2013 to 2019 demonstrated a great variation in the pattern of land use and cover on the five islands under study;