Mapping Areas Invaded by Pinus sp. from Geographic Object-Based Image Analysis (GEOBIA) Applied on RPAS (Drone) Color Images

Abstract

Invasive alien species reduce biodiversity. In southern Brazil, the genus Pinus is considered invasive, and its dispersal by humans has resulted in this species reaching ecosystems that are more sensitive and less suitable for cultivation, as is the case for the restingas on Santa Catarina Island. Invasion control requires persistent efforts to identify and treat each new invasion case as a priority. In this study, areas invaded by Pinus sp. in restingas were mapped using images taken by a remotely piloted aircraft system (RPAS, or drone) to identify the invasion areas in great detail, enabling management to be planned for the most recently invaded areas, where management is simpler, more effective, and less costly. Geographic object-based image analysis (GEOBIA) was applied on images taken from a conventional RGB camera embedded in an RPAS, which resulted in a global accuracy of 89.56%, a mean kappa index of 0.86, and an F-score of 0.90 for Pinus sp. Processing was conducted with open-source software to reduce operational costs.

1. Introduction

Invasion by exotic species represents one of the greatest threats to biodiversity conservation at a global scale [1] and is indicated by experts as capable of compromising the supply of ecosystem services even in protected areas [2]. As a result, it is important to develop methods for controlling exotic species, especially in natural areas with difficult access.

Experts emphasize that invasion cases are frequent in protected areas, despite their recognized importance in in situ biodiversity conservation. Invasions occur even in conservation units (CUs). They are protected areas under special management regimes in Brazil, created to preserve ecosystems and provide them with a management structure and resources for immediate corrective action.

These invasions in protected areas represent a serious concern throughout the country and the world [3]. In general, protected areas such as CUs in Brazil are extensive territories and usually have reduced staff for management and inspection activities. The total area protected by 334 federal CUs is 1,714,241.92 km2, while the total number of Chico Mendes Institute for Biodiversity Conservation (ICMBio) employees is 3603 [4,5].

Even with all this protecting infrastructure, it should be noted that the dispersion of invasive species is closely associated with human presence, as demonstrated in a global evaluation conducted by Van Kleunen et al. [6], who, through the GloNAF project [7], found that human pressure has modified the geographic composition and global distribution of invasive plants across the continents of the world by transporting and accumulating species, mainly from the Northern Hemisphere.

This scenario has occurred for pine trees of the Pinaceae family, especially the genus Pinus. These pine trees were brought from the Northern Hemisphere to the Southern Hemisphere due to their great economic importance, but they already exhibit invasive behavior in many countries, causing not only ecological but also economic and social damage [8]. The genus Pinus was introduced in Brazil in the 1950s [9] and on the island of Santa Catarina in 1963 [10].

The planting of Pinus sp. on Santa Catarina Island was not properly managed, triggering uncontrolled pine reproduction [10]. Thus, the conservation areas of the island began to suffer from the invasion of this exotic species. The impacts of urban expansion and climate change [11], once combined, represent strong reasons to implement effective and efficient management and controlling strategies for exotic species. In this sense, Ziller et al. [2] proposed a scheme to define priorities for the control of exotic species with invasive potential. The authors demonstrated that the implementation of persistent control represents the path to containing the reproduction of invasive species established in protected areas, as already recommended by Moody and Mack [12]. However, this strategy still depends on human resources to conduct the control, in addition to implementing costs that are generally incompatible with the budgets of CUs.

Given this scenario, in regard to the control of exotic plant species within CUs, Marzialetti et al. [13] pointed out that the use of remote sensing can contribute to locating, detecting, measuring, or even specifically identifying these species, and remote sensing can be used both in planning and in monitoring efforts, especially in areas that are difficult and risky to access [14].

Freely accessible satellite images usually have limited use to monitor this type of species, due to the spatial resolution presented. High-resolution satellite imagery, frequently expensive, can be useful for monitoring critical large areas. As a result, aerophotogrammetric surveys that offer sub-metric spatial resolutions may represent even better alternatives for this kind of monitoring, as demonstrated by Müllerová et al. [15] and confirmed by other studies [13,14], especially with the popularization of unmanned aircraft, formally referred to as a remotely piloted aircraft systems (RPASs) [16].

This technology has had an impact on the spread of uses and applications of this type of equipment, especially in providing high-resolution spatial imagery, which allows for greater detail of objects. However, visual interpretation becomes more time-consuming or even impractical. In this sense, the most recent mapping methods, as reviewed by Osco et al. [17], try to automate the process to achieve high accuracy, using an artificial intelligence (AI) technique known as deep learning (DL). According to LeCun et al. [18], DL “allows computational models that are composed of multiple layers to learn representations of data with multiple levels of abstraction […], and discovers intricate structure in large datasets by using the backpropagation algorithm to indicate how a machine should change its internal parameters that are used to compute the representations in each layer from the representation in the previous layer”. In this sense, a common DNN in the supervised network categories is known as convolutional neural network (CNN). This architecture became widely used in remote sensing mapping since Krizhevsky et al. [19] used it to win an image classification competition by a large margin. Recent studies reviewed by Osco et al. [17] also reported that DL is used to enhance remote sensing observations, combined with other methods, such as super-resolution, denoising, restoration, pan-sharpening, and image fusion techniques. However, the practical application of DL in remote sensing requires the mastery of programming languages, such as Python, through libraries such as TensorFlow and Keras.

RPASs have become more accessible and easier to fly [20], mounted with high-quality RGB sensors, and available in “ready-for-flight” models [21]. Demonstrating the high applicability in vegetation research, this equipment began to be widely used to determine and plan management actions for crops and forest of various species, as reported by Guimarães et al. [22], Onishi and Ise [23], Wang et al. [24], White et al. [25], and Apostol et al. [26].

Guimarães et al. [22] evaluated the use of RPASs in the management process of forests and agricultural resources and established criteria for decision making in terms of the equipment available on the market. Thus, operational costs, the ability to mount sensors, and the management of monitoring missions are variables that must be taken into account when using RPASs.

Regarding the sensors, White et al. [25] assessed the use of an RPAS with an RGB camera to identify individual maritime pine tree seedlings after fire events on the Upper Peninsula of Michigan (USA) in 2012. The authors were successful in individually identifying species using a multispectral camera combined with geographic object-based image analysis (GEOBIA), which allowed successful identification with 90% accuracy for individuals with heights of approximately 0.3 to 1.5 m. In Wang et al. [25], the availability of ultrahigh resolution images by RPAS allowed the classification of an urban forest from spectral information, morphological parameters of the vegetation, texture information, and vegetation indices. In comparison to that determined by images generated by multispectral or hyperspectral specimens, the identification of species determined by ultrahigh-resolution images of RPAs with RGB and segmented by GEOBIA provides more accurate information and is cheaper [26].

Onishi and Ise [23] used RGB images and deep learning mapping in a mixed urban forest in Japan. The authors successfully identified five species (90% accuracy), and a key factor of the ultrahigh-resolution image classification success may be the way in which they separated each class. Thus, greater detail of each class in the GEOBIA results in better classification performance in terms of species recognition, even when compared to species with similar colors [27].

The combination of attributes (features) in GEOBIA, such as texture and tree crown shape, contributes to improving the classification tasks [28]. This characteristic occurs because, unlike traditional pixel-based methods, GEOBIA works with homogeneous and continuous groups of pixels, addressed as objects [28,29]. Therefore, GEOBIA classification considers segmentation one of the main processes [30].

Several studies evaluated different methods to perform segmentation, either by using different tools or using different criteria from the same tool, resulting in several segmentation results that need to be evaluated and prioritized for classification [28,30]. However, Costa et al. [30] indicated that the analysis of segmentation accuracy is still under development and has several methods proposed to objectively evaluate, using mathematical criteria that consider aspects of geometry and positioning, which may be supervised or unsupervised. Despite this, procedures to evaluate segmentation accuracy have not yet been standardized; thus, subjective evaluation by visual interpretation may be acceptable depending on the application [30].

In studies that evaluate the invasion of exotic plant species by remote sensing, especially in forest environments, it is expected that the target “edges” are more complex compared with anthropic objects such as buildings in urban areas, due to the complex format of each plant or their groupings that usually form transition areas in a gradient between different environments. This results in the definition of boundaries always being imbued with subjectivity, whether for visual delimitation in images or even in field measurements, which makes it acceptable to verify the segmentation quality using subjective visual criteria. One example is the recent study by Modica et al. [31], who reported that they obtained the best segmentation for the LSMS (large-scale mean-shift) algorithm according to their visual evaluation through trial-and-error testing of the parameters.

According to the survey conducted by Modica et al. [31], random forests (RFs) and support vector machines (SVMs) are among the most recommended supervised classifiers for GEOBIA, showing a very satisfactory performance. As explained by the same authors [31], SVM is a nonparametric algorithm based on kernel functions from statistical learning theory. In essence, SVM operates by learning to determine the boundary between training samples of different classes by projecting them into a multidimensional space and finding hyperplanes that maximize the dataset separation, according to the predefined number of classes. RF, on the other hand, is a method that randomly creates many decision trees, independent of each other. They are all trained on the same characteristics but on different sets or on individual datasets derived from the training set, a separation called “bagging”. The algorithm itself generates an unbiased, internal estimate of the generalization error using the “out-of-bag” samples, i.e., data from the training set itself, but which were not used to train the sample under test.

Despite the advances provided by the GEOBIA classification on high-spatial-resolution imagery, the identifying process of a specie should consider the different types of environments where the targets are found. However, even with RPAS images, accurately extracting different thematic classes only from the structural characteristics and colors of RGB images is still a challenge that must be overcome.

Given this information, the present study aimed to evaluate a methodology for mapping Pinus sp. in natural environments using a classification process by geographic regions, known as GEOBIA [32,33]. This study presents and analyzes alternative solutions according to a set of attributes and algorithms classified by machine learning (random forest (RF) and support vector machine (SVM)) applied to determine the best alternative to classify and identify Pinus sp.

Only conventional RGB sensors were chosen, similar to Albuquerque et al. [34], making it possible to perform the study with any good-quality sensor equipment. No ground support points were used to produce the georeferenced mosaics because the positional accuracy obtained from the global navigation satellite system (GNSS) sensor embedded in the equipment was sufficient to reach the purpose.

The main text of this study consists of three parts divided into the introduction, followed by the materials and methods, and the results and discussion The methodological section covers the areas of interest, equipment and software, method flowchart, and accuracy evaluation. The results and discussion segment includes the overall results compared with similar studies and the results obtained in each area of interest, along with confusion matrices, classification color maps, and particularities. The main text is supported by five appendices that complement the textual information with a description of each feature computed, the software parameters and settings, a list of acronyms, and a feature list with the best classifications for each area of interest.

2. Materials and Methods

2.1. Study Areas

Four areas of interest (Figure 1) were selected for the study, all in the Greater Florianópolis region, containing portions of the ecosystem pioneer formations—vegetation with marine influence, according to the Brazilian Institute of Geography and Statistics (IBGE) [35], also known as restinga areas.

Three study areas were located on Santa Catarina Island, municipality of Florianópolis—SC, and one is on the mainland, city of Palhoça-SC. In the northern part of the island is Sapiens Park (Sapiens Parque—SAPIENS). This is a private area that has an extensive natural area and preserves a mosaic of wetlands and restingas. It was proposed as a CU categorized as a Private Natural Heritage Reserve (RPPN), although it has not yet been formally recognized. In the east, Rio Vermelho State Park (Parque Estadual do Rio Vermelho—PAERVE) that was recognized as a CU in 1974 currently has extensive areas dominated by Pinus sp., due to a state a forestry program carried out in the 1960s [9]. In the southern part of the island there is the Municipal Natural Park of Dunes in Lagoa da Conceição (Parque Natural Municipal das Dunas da Lagoa da Conceição—PNMDLC), a public area that, for more than 10 years, has been managed by volunteers from a biodiversity conservation project aiming to contain the invasion of Pinus sp. [10].

On the mainland, Baixada do Maciambu is located in an area of Palhoça in Serra do Tabuleiro State Park (Parque Estadual da Serra do Tabuleiro—PEST) and represents one of the most significant restinga areas conserved in Santa Catarina state, but it is also invaded by Pinus sp. Although it is mostly covered by herbaceous vegetation, this area was chosen because it represents an invasion front, with individuals of Pinus sp. at various ages, from adults to the youngest.

This study aimed to expand the representativeness of diverse environments and, consequently, the challenges of mapping invasion areas. In each area of interest, sampling regions were selected, which covered 300 × 300 m, totaling 9 ha of images per flight.

Regarding the local climate, considering the Köppen classification, this region of the state has a humid mesothermal climate (without dry season) Cf, containing subtype Cfa with a hot summer [36]. According to the world map that was updated for the Köppen–Geiger classification [37], the climatic similarity of this area with that of the southeastern region of the USA is evident, as Cfa encompasses the natural distribution area of the genus Pinus.

2.2. Equipment and Software

The processing was performed on a personal notebook computer with an Intel Core i7-950H processor, 2.60 GHz, and 32 GB of RAM. WebODM 1.8.2 software [38] was used to process the images captured by the RPAS. To conduct tasks and processing in the GIS environment, including the execution of the GEOBIA classification, the software QGIS 3.16.8-Hannover [39] was used mostly with the Orfeo Toolbox (OTB) complement algorithms [40]. The attribute extraction was carried out using the GeoDMA 2.0.3 beta [41] plugin implemented in TerraView 5.6.1 software [42], while attribute selection was performed in Weka software [43], version 3.8.5. The accuracy analysis was performed using spreadsheets in Libreoffice Calc based on the confusion matrix obtained from the processing logs of the TrainVectorClassifier algorithm of the OTB complement implemented in QGIS. The precision, recall, kappa, and F-score values calculated in spreadsheets were confirmed with those provided in the same logs.

The flights used a multirotor equipment model Inspire 1 (T600) manufactured by Da-Jiang Innovations (DJI), equipped with a Model X3 (FC350) camera with 12.4 MP. The execution of the flight plans used the proprietary platform DroneDeploy [44] in its default setting, which, for this equipment, meant 75% frontal and 65% lateral overlap and 12 m/s speed for a flight height set at 100 m (SAPIENS) and 90 m (PAERVE, PNMDLC, and PEST), resulting in approximately 150 photos per flight. The records were made on days with good weather, with sun and few clouds, except in the PEST area, since it was cloudy. After the first flight (SAPIENS), safety reasons led to the height being reduced by 10%. The flights were performed in the SAPIENS area on 06/11/2020, 09:18:21–09:40:10, in the PEST area on 06/04/2021, 10:13:20–10:18:15, in the PNMDLC area on 10/04/2021, 12:10:14–12:15:09, and in the PAERVE area on 11/04/2021, 12:23:38–12:49:59, Brasilia Time (BRT), UTC-3.

2.3. Methodological Flowchart

This methodological flowchart (Figure 2) displays the main steps covering the complete cycle, from the preparatory step for field data acquisition to the post-processing steps involving cartographic product assembly, layer stacking, segmentation, GEOBIA classification, and area estimation.

The first stage (A) consisted of delimiting the polygons for the flights with the RPA and preparing the flight plan and aerial survey carried out first in the SAPIENS area (06/11/2020, 09:18:21–09:40:10), then in the PEST area (06/04/2021, 10:13:20–10:18:15), PNMDLC area (10/04/2021, 12:10:14–12:15:09), and PAERVE area (11/04/2021, 12:23:38–12:49:59).

The second step (B) involved the orthomosaic composition (RGBα), visible atmospherically resistant index green—VARI, vegetation index calculation, and normalized digital elevation model—nDEM generation to better represent the vegetation spectral variability and height.

Raw cartographic products, the orthomosaic (RGBα), the digital surface model (DSM), and the digital terrain model (DTM) were assembled employing the open-source software WebODM 1.8.2 [38]. The settings were adjusted to have “DSM and DTM” options “enabled” and the GSD value set to 5 cm. Images were not resized, and other settings were set to default. The processing time was about 40 min for each area (150 photos).

The derived cartographic products were prepared in QGIS. The raster calculator was used to remove the RGBα alpha (α) band, converting the orthomosaic only to RGB, and calculating the VARI vegetation index using the formula set by Henrich et al. [45].

VARI = Green − Red/Green + Red − Blue.

(1)

According to the method proposed by De Luca et al. [29], to reduce the effect of potential outliers in the subsequent segmentation, this band was discretized and normalized to a usual range of values from 0 to 255, compatible with the RGB band values. This procedure was performed by the QGIS Processing Toolbox, using the SAGA GIS Raster Normalization and GDAL Convert algorithms to normalize and convert the values to 8 bits (byte).

The digital elevation model—DEM was also generated from QGIS by the difference between DSM and DTM, i.e., subtracting DTM values from DSM ones. This procedure, also performed with SAGA GIS, generated the raster elevation values. After being normalized, the nDEM was resampled to 5 cm with SAGA’s resampling algorithm to make the pixel size compatible with other layers. In conclusion, the nDEM raster was also discretized to 8 bits (byte).

The third step (C) involved layer stacking, which was performed with the GDAL Merge algorithm, native to QGIS, observing the “orthomosaicRGB + VARI + nDEM” order, resulting in a raster with five bands, the three RGB bands, plus the VARI vegetation index and elevation model nDEM.

The fourth stage (D) applied GEOBIA while combining zonal, spectral, and shape statistical information. The GEOBIA classification was conducted in the Orfeo Toolbox (OTB) integrated with QGIS, starting with the image segmentation process, whose parameters were set as indicated by Liau [46] in Figure 3.

The interpretation used to identify the appropriate segmentation parameters corresponded to the definition proposed by Modica et al. [31], which was based on the review conducted by Costa et al. [30], assuming that good segmentation corresponds to a relatively high spectral separability between two different land cover (LC) classes and a minimum within polygons of the same LC.

The attribute extraction obtained with OTB was complemented by GeoDMA, a software integrated into TerraView [42] that allows the extraction of zonal, spectral, and textural attributes, important to execute GEOBIA classification through machine learning [32]. The full description of the calculated attributes with OTB and GeoDMA is shown in Appendix A. In addition, all OTB and GeoDMA attributes were also used in the process, alone and together, even if in duplicate, since it was seen that some values calculated for the same attribute did not have a totally identical outcome. There were subtle differences, usually after the sixth or seventh decimal place.

At this stage, a minimum of 2% of the segments were selected as training and validation samples, resulting in the model files being further employed in the classifying process and in the confusion matrices used for the accuracy evaluation step. Global accuracy, kappa index, F-score, precision, and recall were determined to choose the best model files for classification.

The number of classes (

Table 1. Number of classes and typologies by area of interest (AOI).

AOI	Typology
SAPIENS	1—Pinus sp.
	2—Arboreal restinga
	3—Herbaceous vegetation
	4—Marsh
	5—Shade
PAERVE	1—Pinus sp.
	2—Eucalyptus sp.
	3—Casuarina sp.
	4—Herbaceous restinga
	5—Shrub restinga
	6—Shade
	7—Bare soil
	8—Water
	9—Anthropic
PNMDLC	1—Pinus sp.
	2—Herbaceous restinga
	3—Shrub–arboreal restinga
	4—Arboreal restinga
	5—Shade
	6—Bare soil
	7—Herbaceous vegetation
PEST	1—Pinus sp.
	2—Herbaceous restinga
	3—Shrub–arboreal restinga
	4—Arboreal restinga
	5—Shade
	6—Bare soil
	7—Trunks and branches
	8—Marsh

) varied among the study areas, according to the typologies found in each sampled area. The class definition was specified for each area by photointerpretation. For a better understanding, photos taken on the ground and video captured in flight with a camera at an oblique angle, between 45° and 60°, were employed. These images were not processed; they were only used for the interpretation of the vegetation typologies and other classes.

A widely practiced approach was used to run the machine learning algorithms, as described in the study by De Luca et al. [29]. As explained by Frank et al. [43], it is necessary to prepare the data to conduct the machine learning (ML) process toward the training and validation step, which are employed as inputs for the ML algorithms. To do this, it was necessary to first correctly label a certain number of samples for each class to be used in the training and validation process. The determination (labeling) of the regions containing the training and validation samples was performed in QGIS, with the native “selection by location” algorithm, as shown in Figure 4, by visual interpretation, using the same field photographs and videos taken during panoramic flights as described for class definition.

From the total samples indicated, most were employed in the training stage, during which the algorithms identified patterns used to classify the remaining objects. A portion of the labeled samples were not used in the training stage but were employed in the validating stage. The correct classifications and errors were calculated, as recommended by Radoux and Bogaert [33], to enable an accuracy evaluation.

The division of the sample percentages dedicated to training and validation, in tasks involving machine learning, has been defined by researchers. Some opted for a division of 80% training and 20% validation [47], while others prefer a division of 70% training and 30% validation [32]. In this study, 70% of the samples were used for training, and 30% were used for validation to provide a more rigorous evaluation of the results.

Table 2. Number of segments selected for training and validation.

	SAPIENS		PAERVE		PNMDLC		PEST
Total segments	45,445	100.00%	59,589	100.00%	47,531	100.00%	119,190	100.00%
Training	794	1.75%	1078	1.81%	1587	3.34%	2157	1.81%
Validation	339	0.75%	464	0.78%	680	1.43%	926	0.78%
SUM (training + validation)	1133	2.49%	1542	2.59%	2267	4.77%	3083	2.59%

shows the number of segments selected for training and validation in each area in relation to the segment total number in each image. There are no parameters to determine sample size, but De Luca et al. [29] considered it appropriate to select a number of segments no less than 2% of the total number of image segments. In this case, 2% was adopted as the minimum reference value that needed to be exceeded in the sampling.

In the sample selection process, the samples were labeled uniformly across the entire image, trying to maintain the proportion per class. The minimum number of samples needed per class was previously calculated, considering a total of 2% of the polygons resulting from the segmentation. Concentrating too much selection in one class to the detriment of another was avoided, especially in the target class of the study, Pinus sp. However, it is not possible to perform this distribution in a perfectly homogeneous way among the classes, because the area occupied by each class varied, and the total number of polygons per class was unknown before the classification conclusion.

Attributes (features) were selected using tools from Weka, University of Waikato [43]. The vector files with attribute tables that contained the selected samples for training were exported to a spreadsheet to create the datasets used in the attribute selection process. To prevent the model from overfitting, i.e., losing the ability to generalize to new data, no validation polygons were employed in the attribute selection.

It was necessary to generate three spreadsheets for each area of interest to comprehensively perform the attribute selection: one containing attributes calculated with QGIS/OTB, another with those generated only by TerraView/GeoDMA, and the third containing both. An evaluation with all attributes was performed due to the possible differences between the values produced by the two applications, aiming to discard non-numerical values (not a number—NaN), which can severely degrade the classification. The existence of fields calculated in duplicity with identical values was also verified.

The attribute selection was performed in the Explorer application on the select attributes table. After opening the spreadsheet, the numeric fields “label” and “class” were removed because they were not used in this process. This procedure did not alter the original spreadsheet file but only served to indicate to the software which fields were not to be considered in the attribute selection.

Two tools were used. The first tool, CfsSubsetEvaluation (CfsSE), is a correlation-based feature selection (CFS), does not use an algorithm to mine data, and is considered a “filter” type of tool, because it employs statistical techniques to score the relevance of the relation between the inputs and the target variable and, accordingly, filters the best to be used. According to Nevalainen et al. [48], this tool searches for subsets that are highly correlated with each class and have low internal correlation, making this tool suitable for spectral vegetation data because the spectral characteristics are generally best characterized by a combination of more than one spectral band. The second tool, wrapperSubsetEval, is a wrapper type of tool. As explained by Tiwari and Singh [49], it uses machine learning algorithms to mine the best data subset and, in theory, is able to provide better results than those of the filter tools at the cost of greater computational effort. Since it involves machine learning, the attribute selection was performed with a wrapperSubsetEval (wrapper) combined with a classifier J48 (Decision Trees C 4.5). The CfsSE correlation method was implemented in the Weka software as suggested by Nevalainen et al. [48] using the search methods Genetic Search, BestFirst, and GreedyStepWise. The Wrapper method was implemented according to Frank et al. [43]. All the specific parameters used in Weka for attribute selection are described in Appendix B.

It is important to note that any changes made to the order of the rows or columns in the spreadsheet would result in random changes in the attribute selection. For this reason, the spreadsheets containing the datasets were kept intact for the selection of attributes, without adjustments beyond those already described as necessary.

Regarding attribute selection, it should also be explained that 10-fold two-way cross-validation was employed, resulting in lists of attributes that were sorted according to the frequency with which the attributes were selected. This allowed us to test various combinations of the “best attributes” that were compared with the use of all attributes. Additionally, it was worth observing how much the attribute selection truly improved the classification.

In the CfsSE method, the selected attributes refer to eight of the 10 cross-validation rounds, indicating a cutoff line of 80% frequency, as proposed by Nevalainen et al. [48]. However, instead of testing only the selected attributes in all three search methods (Genetic Search, BestFirst, and GreedyStepWise), the selected attributes were also separately tested. For the wrapper method, the selected attributes ranged from one to five rounds (10% to 50% of frequency), since higher ranges did not select attributes for all areas. All fifty combinations of attribute selection and machine learning algorithm configurations tested are shown in Appendix C, with 25 different feature combinations implemented in the two algorithms (SVM and RF).

The attribute selection by Weka resulted in logs exported in text format (.txt) containing the attributes and their importance in percentages. The logs were converted into spreadsheets to sort and effectively select the lists of attributes to be employed in the classification.

The training and validation step was performed using the TrainVectorClassifier algorithm from the OTB/QGIS Processing Toolbox. Among the available machine learning classification algorithms, two were used in this application, random forest (RF) and support vector machine (SVM), which, according to the OTB Development Team [40], are based on the OpenCV Machine Learning (2.3.1 and later) and LibSVM libraries.

The process inputs were vectors containing the polygons selected for training (70% of the total) and validation (30% of the total), indicating in “field names for training features” the fields selected for training in the attribute selection and in “field containing the class integer label for supervision” the numeric field “class”.

The configuration parameters were kept in default QGIS/OTB configuration for both algorithms. The main settings for LibSVM were SVM kernel type “linear”, SVM model type “csvc”, cost parameter C “1.0” cost parameter Nu “0.5”, and random seed “0”, while other settings were kept disabled or in blank. The settings for RF were maximum depth of the tree “5”, minimum number of samples in each node “10”, terminating criteria for regression tree “0”, cluster possible values of a categorical variable into K <= cat clusters to find a sub-optimal split “10”, size of randomly selected subset features at each tree node “0”, maximum number of trees in the forest “100”, sufficient precision “0.01”, user-defined input centroids “0” (not empty, to avoid error), and random seeds “0”, while other parameters were kept empty or disabled. The parameters employed in QGIS/OTB for classification are described in Appendix D.

Once the training and validation stage was concluded, it was possible to perform the accuracy evaluation on the basis of the confusion matrices generated by QGIS/OTB itself as explained in Section 2.4. The final classification was performed with the four feature sets that reached the best-evaluated accuracy for each one of the four areas of interest. This procedure was performed using the QGIS/OTB VectorClassifier algorithm.

Lastly, the fifth step (E) consisted of fixing the geometry vectors using the native QGIS algorithm “fix geometries” and then converting the vector data from polygon to multipart using the “dissolve” algorithm, also native to QGIS and described in GIS documentation [39]. Once completed, it was possible to calculate the total area values per class, with the QGIS field calculator in square meters using the $area formula, as well as in hectares ($area/10,000). Some development details about this proposed method can be found in Gonçalves [50].

2.4. Accuracy Evaluation

Accuracy estimates were determined on the basis of the count of spatial entities (objects), as indicated by Radoux and Bogaert [33], from the confusion matrix provided by QGIS/OTB. The formulas used in the calculations are shown in

Table 3. Formulas used to evaluate accuracy.

Name	Formula	Objective
User accuracy (precision)	aui = xii/xi+, where xii = objects correctly classified for class i; xi+ = total number of objects classified for class i	It reflects the commission errors that indicate the probability of an object classified in a given class actually belonging to that class [53].
Producer accuracy (recall)	api = xii/x + I, where xi = objects correctly classified for class i; xi = total of reference objects for a class i	It reflects the omission errors, i.e., the probability of an object being excluded (not classified) from the class to which it belongs [53], additionally referred to as “sensitivity”, according to [51].
Global accuracy	ag = xi/n, where xi = total number of correctly classified objects; n = total number of objects considered in the analysis	Measurement used in the global evaluation of the classification [53].
Expected proportion	$Pe={\displaystyle {\displaystyle \sum }_{i=1}^r}\frac{x_{i+}\ast x_{+i}}{n^2}$	Unit proportion for which an agreement is expected by chance [52]. Used to compose the kappa Cohen index.
Kappa Cohen	$K=\frac{Ag-Pe}{1-Pe}$	Kappa employed for the global evaluation of the classification. It differs from the global accuracy by incorporating the elements outside the main diagonal [52].
F-score	$F=\frac{2\ast a{u}_i\ast a{p}_i}{a{u}_i\ast a{p}_i}$	Parameterized measure with equal weights for precision and recall allowing the model performance for each class to be compared, which is the measure most frequently used in machine learning with unbalanced data between precision and recall [51,54]

. The calculated values for producer accuracy or recall, user accuracy or precision, F-score [51], and the kappa Cohen index [52] were confirmed as a function of the values provided by QGIS/OTB. Overall accuracy and Cohen’s Kappa index allowed the overall classification quality to be evaluated, along with precision, recall [53], and especially F-score, which allowed each class to be assessed [51]. These measures could be implemented in spreadsheets according to the formulas presented in Table 3.

The estimated accuracy was calculated for all the performed classifications. By default, the OTB TrainVectorClassifier algorithm provides precision, recall, and F-score for each class and overall classification performance by the kappa index. The obtained classification only with the features calculated by OTB was adopted as a reference to estimate the improvement obtained from the dataset increment with the features calculated by GeoDMA, as well as with successive rounds of attribute selection using Weka. Thus, a baseline was established to estimate the improving classification in relation to the OTB.

The accuracy results were evaluated using the classification performed using the OTB package implemented in QGIS as reference after using this single software rather than another approach involving fewer processing steps and, thus, an easier process. In order to consider the inclusion of other attributes computed and selected in the other software as advantageous, it was necessary to check whether there was an improvement in the classification result.

A widely known reference, suggested by Landis and Koch [55], was employed to interpret the quality of the classification on the basis of the kappa index value, as shown in

Table 4. Reference to interpret the classification performance based on the kappa index value.

Kappa	Strength of Agreement	Performance
<0	Poor	Terrible
0 < k ≤ 0.2	Slight	Poor
0.2 < k ≤ 0.4	Fair	Reasonable
0.4 < k ≤ 0.6	Moderate	Good
0.6 < k ≤ 0.8	Substantial	Very good
0.8 < k ≤ 1.0	Almost perfect	Excellent

Source: Landis and Koch [55].

.

3. Results and Discussion

Table 5. Higher values obtained for global accuracy, kappa, and F-score indices for Class 1 (Pinus sp.) and their respective compositions of software, attributes, and machine learning algorithms.

Global Accuracy
Area	Higher Values	Classifier	Software, Attributes, and Algorithms
SAPIENS	89.97%	libsvm	GEODMA (attributes) Weka (feature selection with CfsSE and GreedyStepWise search method, cutting line 80%)
PAERVE	91.16%	libsvm	OTB and GEODMA (attributes) Weka (feature selection with Wrapper method, cutting line 20%)
PNMDLC	88.68%	libsvm	OTB and GEODMA (attributes) Weka (feature selection with CfsSE in all three search methods combined, cutting line 80%)
PEST	88.44%	libsvm	OTB and GEODMA (attributes) Weka (feature selection with Wrapper method, cutting line 30%)
Mean	%89.56	- -	- -
Kappa
Area	Higher Values	Classifier	Software, Attributes, and Algorithms
SAPIENS	0.87	libsvm	GEODMA (attributes) Weka (feature selection with CfsSE and GreedyStepWise search method, cutting line 80%)
PAERVE	0.90	libsvm	OTB and GEODMA (attributes) Weka (feature selection with Wrapper method, cutting line 20%)
PNMDLC	0.82	libsvm	OTB and GEODMA (attributes) Weka (feature selection with CfsSE in all three search methods combined, cutting line 80%)
PEST	0.85	libsvm	OTB and GEODMA (attributes) Weka (feature selection with Wrapper method, cutting line 30%)
Mean	0.86		- -
F-Score
Area	Higher Values	Classifier	Software, Attributes, and Algorithms
SAPIENS	0.86	libsvm	All OTB features only
PAERVE	0.96	libsvm	GEODMA (attributes) Weka (feature selection with CfsSE and GreedyStepWise search method, cutting line 80%)
PNMDLC	0.83	libsvm	OTB and GEODMA (attributes) Weka (feature selection with CfsSE in all three search methods combined, cutting line 80%)
PEST	0.96	libsvm	OTB and GEODMA (attributes) Weka (feature selection with Wrapper method, cutting line 30%)
Mean	0.90	- -	- -

shows the highest values attained for global accuracy (GA) and kappa indices, representing the classification accuracy considering all classes and those obtained for the F-score for Class 1, which represent the map quality in denoting the invasion areas of Pinus sp., the target of this study. The classifier-based predominance of the SVM is highlighted since it presented the smallest error for the target class in all areas.

According to the reference suggested by Landis and Koch [55] (Table 4), the classifications were excellent in all areas. However, as the resolutions of the orthomosaics generated by RPASs are very high according to Sharma and Müllerová [56], the values considered acceptable according to the literature vary considerably from 70% for global accuracy [57] to 85% for sensitivity (recall) and precision [58]. However, even considering this variation, the achieved results in this study can be considered satisfactory.

Indeed, GeoDMA calculates all metrics generated by OTB with the LargeScaleMeanshift and ZonalStatistics algorithms. It could seem that it would be enough to use only the calculated attributes by GeoDMA. If the calculations were totally identical, one should expect that there would be no differences between the classification results using only GeoDMA or GeoDMA combined with OTB. Nevertheless, for the common attributes between OTB and GeoDMA, subtle differences were observed, usually after the sixth or seventh decimal place, which is why the chosen option was to keep both attributes, leaving the choice up to the algorithm in the attribute selection step. From the attributes generated exclusively by OTB, the number of pixels (nbPixels) and the averages (meanB0, meanB1, meanB2, meanB3, and meanB4) calculated by LargeScaleMeanshift were excluded since the calculation of ZonalStatistics was identical for these fields. Considering only GeoDMA, some attributes had non-numerical values (not a number—NaN), which fatally affected the classification. Therefore, the removal of such attributes, which generally occurred with kurtosis (KURT_BX) and asymmetry (SKEW_BX) associated with bands 3 (VARI) and 4 (DEM), was performed.

The accuracy evaluation analysis showed that the best results were achieved using the GeoDMA attributes only in the SAPIENS area. In the other areas, the best classification result came from the combined OTB and GEODMA features (Appendix A). This showed that, although subtle, there were differences between the calculated attributes values and that the algorithm should make the selection.

Table 6. High values for global accuracy, kappa index, and F-score of Class 1 (Pinus sp.) reached for each classifier compared to those for the classifiers in the OTB software only. The highest values are in bold.

	INDEX PER CLASSIFIER	OTB	GEODMA+weka_CfsSE_GreedyStepWise_cv10x1_80%	GEODMA			MAX	PROCESSING SCHEME SETUP	IMPROVEMENT	%
SAPIENS	Global Accuracy rf.	84.66%		89.68%			89.68%	GEODMA	0.050147	5.92%
	Global Accuracy libsvm	87.32%	89.97%				89.97%	GEODMA+weka_CfsSE_GreedyStepWise_cv10x1_80%	0.026549	3.04%
	KAPPA rf.	0.805405		0.868927			0.868927	GEODMA	0.063522	7.89%
	KAPPA libsvm.	0.839244	0.872983				0.872983	GEODMA+weka_CfsSE_GreedyStepWise_cv10x1_80%	0.033739	4.02%
	FSCORE PINUS rf.	0.797814		0.851064			0.851064	GEODMA	0.053250	6.67%
	FSCORE PINUS libsvm.	0.855556					0.855556	OTB	0.000000	0.00%
PAERVE	INDEX PER CLASSIFIER	OTB	OTB_and_GEODMA+weka_CfsSE_GreedyStepWise_cv10x1_80%	GEODMA+weka_CfsSE_BestFirst_cv10x1_80%	GEODMA+weka_CfsSE_GreedyStepWise_cv10x1_80%	OTB_and_GEODMA+weka_Wrapper_bd_cv_20%	MAX	PROCESSING SCHEME	IMPROVEMENT	%
	Global Accuracy rf.	77.37%	81.25%				81.25%	OTB_and_GEODMA+weka_CfsSE_GreedyStepWise_cv10x1_80%	0.038793	5.01%
	Global Accuracy libsvm	82.76%				91.16%	91.16%	OTB_and_GEODMA+weka_Wrapper_bd_cv_20%	0.084052	10.16%
	KAPPA rf.	0.729318	0.776774				0.776774	OTB_and_GEODMA+weka_CfsSE_GreedyStepWise_cv10x1_80%	0.047456	6.51%
	KAPPA libsvm.	0.797338				0.895904	0.895904	OTB_and_GEODMA+weka_Wrapper_bd_cv_20%	0.098566	12.36%
	FSCORE PINUS - rf.	0.826087		0.894942			0.894942	GEODMA+weka_CfsSE_BestFirst_cv10x1_80%	0.068855	8.34%
	FSCORE PINUS libsvm.	0.850467			0.957627		0.957627	GEODMA+weka_CfsSE_GreedyStepWise_cv10x1_80%	0.107160	12.60%
PNMDLC	INDEX PER CLASSIFIER	OTB	OTB_and_GEODMA+weka_CfsSE_COMBIN_cv10x1_80%	OTB+wekaWrapper_bd_cv_30%	OTB+weka_Wrapper_bd_cv_40%		MAX	PROCESSING SCHEME	IMPROVEMENT	%
	Global Accuracy rf.	84.12%		86.32%			86.32%	OTB+wekaWrapper_bd_cv_30%	0.022059	2.62%
	Global Accuracy libsvm	84.85%	88.68%				88.68%	OTB_and_GEODMA+weka_CfsSE_COMBIN_cv10x1_80%	0.038235	4.51%
	KAPPA rf.	0.745134		0.782718			0.782718	OTB+wekaWrapper_bd_cv_30%	0.037584	5.04%
	KAPPA libsvm.	0.771837	0.824535				0.824535	OTB_and_GEODMA+weka_CfsSE_COMBIN_cv10x1_80%	0.052698	6.83%
	FSCORE PINUS rf.	0.646617			0.706767		0.706767	OTB+weka_Wrapper_bd_cv_40%	0.060150	9.30%
	FSCORE PINUS libsvm.	0.717647	0.828947				0.828947	OTB_and_GEODMA+weka_CfsSE_COMBIN_cv10x1_80%	0.111300	15.51%
PEST	INDEX PER CLASSIFIER	OTB	OTB+weka_Wapper_bd_cv_30%	GEODMA	OTB_and_GEODMA		MAX	PROCESSING SCHEME	IMPROVEMENT	%
	Global Accuracy rf.	84.99%		86.18%			86.18%	GEODMA	0.011879	1.40%
	Global Accuracy libsvm	87.47%	88.44%				88.44%	OTB+weka_Wapper_bd_cv_30%	0.009719	1.11%
	KAPPA rf.	0.802609		0.819808			0.819808	GEODMA	0.017199	2.14%
	KAPPA libsvm.	0.839762	0.850954				0.850954	OTB+weka_Wapper_bd_cv_30%	0.011192	1.33%
	FSCORE PINUS rf.	0.923636			0.930403		0.930403	OTB_and_GEODMA	0.006767	0.73%
	FSCORE PINUS libsvm.	0.949495	0.962457				0.962457	OTB+weka_Wapper_bd_cv_30%	0.012962	1.37%

shows the highest results obtained in the accuracy evaluation in each AOI for the GA, kappa index, and F-score of the target class (Pinus sp.), for both classifiers (SVM and RF). The values are presented in comparison with the reference value obtained from processing the classifications using only the features calculated by the OTB, shown in the first column. The acronyms in the column headings indicate the setup used in the processing that originated features employed in the classification. The colors applied to the kappa index values correspond to the quality assigned, according to Table 4. The list of acronyms and their meanings are exhibited in Appendix C, while Appendix E shows the feature list.

In all four areas, the SVM classifier surpassed the RF classifier, with an average performance 6.24% higher, as shown in

Table 7. Classifier performance comparison.

Area	Kappa RF	Kappa SVM	Difference	%
SAPIENS	0.868927	0.872983	0.004056	0.47%
PAERVE	0.776774	0.895904	0.11913	15.34%
PNMDLC	0.782718	0.824535	0.041817	5.34%
PEST	0.819808	0.850954	0.031146	3.80%
Average				6.24%

.

According to the results, the SVM performance was consistently superior in three areas (PAERVE, PNMDLC, and PEST) and slightly superior in the SAPIENS area. Regarding processing time, which reflects computational effort, SVM, on average, consumed approximately 30 times more effort than RF, taking an average of 33 s to complete the training and validation steps, while RF took an of 1 s.

The accuracy evaluation indicated an improvement given the additional procedures for the GEOBIA flow implemented in this study, such as increasing the number of attributes or features computed by GeoDMA and the attribute selection performed in Weka through filtering methods by statistical calculations (CfsSE) and by data mining with machine learning (wrapper). If the GEOBIA classification had been conducted only in the most basic way, using only the attributes computed in QGIS/OTB, without submitting them to selection, no improvement would be possible.

Table 8. Mean improvement obtained in the kappa and F-score indices of Class 1 (Pinus sp.) with the use of the techniques employed to improve classification.

Classifier	Percentage Improvement
	Global Accuracy	Kappa	F-Score of Class 1 (Pinus sp.)
RF	3.74%	5.40%	6.26%
libsvm	4.70%	6.14%	7.37%
Mean	4.22%	5.77%	6.81%

shows the average percentage quantification of these improvements reached in the overall accuracy and in the kappa and F-score indices.

Compared to the results procured when using only the OTB package for GEOBIA classification, a consistent classification improvement was reached, with an average increase of 4.22% in GA, 5.77% in kappa index, and 6.81% in F-score for the target class (Pinus sp.).

The improvement in classification with the inclusion of additional attributes calculated with GeoDMA and then with the attribute selection conducted with Weka was not immediate, showing randomness, as displayed in Figure 5.

These results showed that no feature selection method was superior to another. One could also observe that, in machine learning, the inclusion of more features did not necessarily result in improvement. Similarly, selection by the simplest method (CfSE), referenced by Nevalainen et al. [48] as suitable for vegetation, proved successful for some classifications but not others. The same scenario occurred for the most computationally expensive method (wrapper).

According to Table 6, the F-score value of the class referring to Pinus sp. increased for the SAPIENS area when only the set of attributes of the OTB was used in the classification. For the other cases, the selection of attributes consistently improved the classifications quality. This result shows that selecting only the best set of attributes may not be effective, as observed by Nevalainen et al. [48], who selected the method that employs 80% of the most frequent attributes. However, this set of features did not improve the classification accuracy because, in their study, the best performance, GA 95.5% and kappa 0.92, was achieved when all attributes were used. It is also worth mentioning that the data were captured by a hyperspectral camera and processed by deep learning in their research. The classification was performed by including the point cloud and not the orthomosaic, achieving the best results with random forest and multilayer perceptron (MLP).

Unlike Nevalainen et al. [48], who obtained the best results using all attributes, this study had random results because of the rising number of attributes not necessarily reflected in the improved classification. Similarly, the choice of attributes indicated as the most relevant attributes did not always improve accuracy. The results from this study indicate that testing possible attribute combinations according to different attribute selection methods and ordinating classification results, based on accuracy values, can improve the classification result. In addition, the automation of this process will enable the testing of a wide range of possibilities.

Despite having used fixed-wing equipment, the approach used by De Luca et al. [29] is the closest to the methodology employed here. Their results of GA 94% and kappa 0.92 were higher than those in this study, but the near-infrared (NIR) band method was employed in a less environmentally diverse context.

Nascente et al. [59], on the other hand, also used a ready-to-fly multirotor quadricopter equipped with an RGB sensor. They mapped with GEOBIA the cover of invasive vegetation in vereda environments in the Cerrado biome. Despite similar methodology and objectives, the technique is different since it employs nearest neighbor classifier combined with feature space optimization. The results also achieved GA and kappa indices higher than 80% and 0.8, respectively. This demonstrates that more than one method may perform accurate mapping with equipment limited to the RGB spectrum.

The aims of Feduck et al. [60] were similar to the objectives of this study in identifying a detection method capable of mapping the smallest individuals. However, their study used single photographs taken in an automated flight plan, hovering 15 m above ground level. No orthomosaic was generated, and the sampling-based method reached a GA of 86% and kappa of 0.81. These photographs can be used to estimate samples inside vegetation clearings, but it may be impractical to execute orthomosaic flight plans in forested environments.

Considering studies that proposed mapping invasive alien species, it is also worth mentioning the results achieved by Lehmann et al. [14], who also employed open-source platforms to map invasions of Acacia mangium in a mussunanga environment in northeastern Brazil, and reached a GA of 85% and kappa of 0.82, and by Samiappan et al. [60], who worked on the mapping of Phragmites australis, Poaceae, using RPAS images in marshes in the south USA, attaining a GA of 85% and kappa of 0.70. Both studies had similar accuracy results, slightly below those of this study, despite the use of cameras capable of capturing the infrared spectrum.

The next section presents the confusion matrices, maps, and interpretations of the most accurate classification models for each area of interest according to the experimental results. Appendix E lists all the features used to create these classification models.

3.1. SAPIENS

Table 9. Confusion classification matrix for the SAPIENS area. The rows represent the reference map, and the columns represent the classified map. The classes are 1—Pinus sp. (PI), 2—arboreal vegetation (AV), 3—herbaceous vegetation (HV), 4—marsh (BA), and 5—shade (SB). The table also includes the measures of producer accuracy or recall (PA), user accuracy or precision (UA), F-score, global accuracy (GA), and kappa Cohen index (K).

Reference		Classified					Sum
		1	2	3	4	5		PA (Recall)
		PI	AV	VH	BA	SB
1	PI	75	7	0	0	8	90	0.83
2	AV	8	71	0	2	0	81	0.88
3	HV	0	0	66	0	0	66	1.00
4	BA	2	3	0	49	0	54	0.91
5	SB	4	0	0	0	44	48	0.92
Sum		89	81	66	51	52	339
UA (precision)		0.84	0.88	1.00	0.96	0.85
F-score		0.84	0.88	1.00	0.93	0.87
GA		89.97%
K		0.872983

shows the confusion matrix for the best classification obtained for the SAPIENS Park area. This classification was achieved using the SVM classifier with the attributes calculated by GeoDMA with Weka using CfsSubsetEvaluation executed with the GreedyStepWise search method. Compared to OTB, this improved the kappa index by 4.02% (Table 6).

Figure 6 shows the SAPIENS area image and the best-procured classification. The numerical results of the accuracy evaluation were corroborated by the consistent results from the visual analysis.

Overall, the classifier distinguished the class Pinus sp. in the tree vegetation in a very challenging context. The best-represented class was herbaceous vegetation, which was expected due to its homogeneity and difference in hues. The shade category, despite not representing a feature in the field itself, was treated as a class due to the difficulty in defining covert objects. Despite this, shade was more abundant where there were individuals of Pinus sp. due to their larger size compared to other vegetation types. This feature can also serve as an indicator of its occurrence.

3.2. PAERVE

PAERVE had the highest class number (Table 1) among the evaluated areas, representing an additional challenge. In addition to Pinus sp., Eucalyptus sp. and Casuarina sp. were also identified as exotic species. The confusion matrix (

Table 10. Confusion classification matrix for the PAERVE area. The rows represent the reference map, and the columns represent the classified map. The classes are 1—Pinus sp. (PI), 2—Eucalyptus sp. (EU), 3—Casuarina sp. (CA), 4—herbaceous restinga (HR), 5—shrub restinga (SR), 6—shade (SB), 7—bare soil (BS), 8—water (A), and 9—anthropic area (AN). The table also includes the measures of producer accuracy or recall (PA), user accuracy or precision (UA), F-score, global accuracy (GA), and kappa Cohen index (K).

Reference		Classified									Sum
		1	2	3	4	5	6	7	8	9		PA (Recall)
		PI	EU	CA	HR	SR	SB	BS	A	AN
1	PI	110	5	0	1	1	2	0	0	0	119	0.92
2	EU	1	69	0	0	4	0	0	0	0	74	0.93
3	CA	0	0	9	1	0	0	1	0	0	11	0.82
4	HR	0	1	1	17	4	0	2	0	0	25	0.68
5	SR	0	4	0	0	68	0	0	0	0	72	0.94
6	SB	0	1	0	0	0	34	0	0	0	35	0.97
7	BS	0	0	1	1	0	0	32	0	0	34	0.94
8	A	0	0	0	0	0	0	0	59	2	61	0.97
9	AN	0	0	2	0	0	0	3	3	25	33	0.76
Sum		111	80	13	20	77	36	38	62	27	464
UA (precision)		0.99	0.86	0.69	0.85	0.88	0.94	0.84	0.95	0.93
F-score		0.96	0.90	0.75	0.76	0.91	0.96	0.89	0.96	0.83
GA		91.16%
K		0.895904

) indicated few errors, reflecting the higher accuracy achieved in this area when compared to other classified areas. This result shows that the class diversification did not hinder the classification process, which worked even for less representative classes such as Casuarina sp., anthropic areas, and water.

Figure 7 shows the PAERVE area image and the best-obtained classification. This area also showed visual coherence, as a good distinction between Pinus sp. and Eucalyptus sp. was observed. However, some segments were classified as restinga shrubs amid Eucalyptus sp. vegetation, which does not correspond to what is actually present in the area. The same occurred with the Casuarina sp. class, which represented a further challenge to the classifier given the few samples in the image, in addition to the low crown density that made the soil and objects visible, such as people and vehicles below the trees.

3.3. PNMDLC

Overall, the classification result in this area was also satisfactory, with a kappa index of 0.82 (

Table 11. Confusion classification matrix for the PNMDLC area. The rows represent the reference map, and the columns represent the classified map. The classes are 1—Pinus sp. (PI), 2—herbaceous restinga (HR), 3—shrub–arboreal restinga (SRA), 4—arboreal restinga (AR), 5—shade (SB), 6—bare soil (BS), and 7—herbaceous vegetation (HV). The table also includes the measures of producer accuracy or recall (AP), user accuracy or precision (AU), F-score, global accuracy (GA), and kappa Cohen index (K).

Reference		Classified							Sum
		1	2	3	4	5	6	7		PA (Recall)
		PI	HR	SRA	AR	SB	BS	HV
1	PI	63	0	0	13	1	0	0	77	0.82
2	HR	0	31	0	0	0	1	0	32	0.97
3	SRA	1	0	65	27	0	0	0	93	0.70
4	AR	10	1	10	345	5	0	0	371	0.93
5	SB	1	0	0	6	38	0	0	45	0.84
6	BS	0	0	0	0	0	55	1	56	0.98
7	HV	0	0	0	0	0	0	6	6	1.00
Sum		75	32	75	391	44	56	7	680
UA (precision)		0.84	0.97	0.87	0.88	0.86	0.98	0.86
F-score		0.83	0.97	0.77	0.91	0.85	0.98	0.92
GA		88.68%
K		0.824535

), despite the complexity presented by the environment mosaic represented by the various classes. Again, the best-represented class was herbaceous vegetation, which had higher differentiation, in terms of both spectral and textural aspects and height.

Figure 8 shows the classification results attained for the PNMDLC, which were generally consistent with the image according to a visual evaluation.

3.4. PEST

The PEST area also had a very challenging context due to diversity of species and environments in the mosaic, as reflected in the variation of classes needed for classification. The target class Pinus sp. in this area also achieved an excellent F-score (

Table 12. Confusion PEST area matrix of classification. The lines represent the reference map. The classes are 1—Pinus sp. (PI), 2—herbaceous restinga (HR), 3—shrub–arboreal restinga (SRA), 4—arboral restinga (AR), 5—shade (SB), 6—bare soil (SL), 7—trunks and branches (TB), and 8—marsh (BA). The table also includes the measures of producer accuracy or recall (AP), user accuracy or precision (AU), F-score, global accuracy (GA), and kappa Cohen index (K).

Reference		Classified
		1	2	3	4	5	6	7	8	SUM	PA (Recall)
		PI	HR	SRA	AR	SB	SL	TB	BA
1	PI	141	0	3	1	0	0	0	0	145	0.97
2	HR	0	62	11	0	0	0	0	7	80	0.78
3	SRA	1	4	303	22	1	0	0	0	331	0.92
4	AR	6	0	40	36	1	0	1	0	84	0.43
5	SB	0	0	1	0	30	0	0	0	31	0.97
6	SL	0	0	0	0	0	17	0	0	17	1.00
7	TB	0	0	0	0	0	0	41	0	41	1.00
8	BA	0	7	1	0	0	0	0	189	197	0.96
sum		148	73	359	59	32	17	42	196	926
UA (precision)		0.95	0.85	0.84	0.61	0.94	1.00	0.98	0.96
F-score		0.96	0.81	0.88	0.50	0.95	1.00	0.99	0.96
GA		88.44%
K		0.850954

). The most confusion occurred among the classes arboreal restinga and shrub–arboreal restinga, which are naturally difficult to discern even in the field.

Figure 9 shows the classification results. Notably, the mosaic diversity of environments formed by mostly herbaceous species was well represented in the classification, demonstrating the possibility of using the method in this context.

This area, mostly covered by herbaceous vegetation, was specifically chosen because it represents a region of invasion fronts around the matrix, close to different types of environments found in the restinga of the Baixada do Maciambu. By selecting this region, the challenge was to map the typical behavior of invasion fronts in the park, with differently aged individuals, from seedlings to adults, increasing the chance of testing the detection method in a more diverse condition than in consolidated invasion areas that tend to be further homogeneous. In addition, invasion fronts tend to be prioritized in biological invasion control projects; hence, it was considered important to represent them in the classification.

Nevertheless, during the visual evaluations, some segments classified as Pinus sp. were actually jerivá palms, Syagrus romanzoffiana (Figure 10). Like the Pinus sp., they usually stand out from the surrounding vegetation due to their height. Miyoshi et al. [61] used a deep learning method with hyperspectral images representing the state of the art in remote sensing to detect and classify individuals of S. romanzoffiana since the species occur mainly in early secondary forests, serving as an ecological indicator of regenerating forests.

Since an altimetry raster was one of the attributes used in the classification, this factor likely helped the confusion. Unlike the multispectral images used by Miyoshi et al. [58], the RGB spectrum may not have sufficient amplitude to allow an adequate distinction.

However, in the study areas, the occurrence frequency of S. romanzoffiana was low. According to a manual count over 9 ha, 32 individuals were found in PNMDLC, eight were found in PEST, seven were found in SAPIENS, and none were found in PEST. Thus, manual counts can be employed to adjust classifications if necessary. Overall, the classification results were considered satisfactory on the basis of the used technology.

3.5. Area Estimates

Area estimates per class were obtained from the classification, as shown in

Table 13. Area calculations by classified typology.

AOI	Typology	Area (m2)	Area (ha)	%
SAPIENS	1—Pinus sp.	40,273.377	4.027	44.76%
	2—Arboreal restinga	22,914.420	2.291	25.47%
	3—Herbaceous vegetation	10,817.774	1.082	12.02%
	4—Marsh	2919.212	0.292	3.24%
	5—Shade	13,042.063	1.304	14.50%
	Total	89,966.846	8.996
PAERVE	1—Pinus sp.	42,583.091	4.258	47.33%
	2—Eucalyptus sp.	14,589.845	1.459	16.21%
	3—Casuarina sp.	1868.936	0.187	2.08%
	4—Herbaceous restinga	8295.598	0.830	9.22%
	5—Shrub restinga	7256.411	0.726	8.06%
	6—Shade	10,289.606	1.029	11.44%
	7—Bare soil	4320.588	0.432	4.80%
	8—Water	264.970	0.026	0.29%
	9—Anthropic	510.185	0.051	0.57%
	Total	89,979.230	8.998
PNMDLC	1—Pinus sp.	6,444.565	0.644	7.16%
	2—Herbaceous restinga	5481.469	0.548	6.09%
	3—Shrub–arboreal restinga	14,529.664	1.453	16.15%
	4—Arboreal restinga	56,572.321	5.657	62.87%
	5—Shade	4781.252	0.478	5.31%
	6—Bare soil	1774.236	0.177	1.97%
	7—Herbaceous vegetation	399.328	0.040	0.44%
	Total	89,982.835	8.997
PEST	1—Pinus sp.	4654.108	0.465	5.17%
	2—Herbaceous restinga	18,393.586	1.839	20.44%
	3—Shrub–aboreal restinga	35,376.910	3.538	39.32%
	4—Arboreal restinga	8173.890	0.817	9.09%
	5—Shade	1528.963	0.153	1.70%
	6—Bare soil	221.288	0.022	0.25%
	7—Trunks and branches	5299.039	0.530	5.89%
	8—Marsh	16,323.355	1.632	18.14%
	Total	89,971.139	8.996

. The measurement of these areas using vectors was aimed at validating the representativeness of each class in the landscape represented by the image.

This result shows that the proposed methodological flow can be employed both for a quick and detailed diagnosis of the study area and for studies that involve monitoring on the basis of repeating the method application.

4. Conclusions

From the classification results, it was possible to map the areas invaded by Pinus sp. in the areas of interest. Thus, the understanding is that the used method can help both the invasion interpretation and action plans needed to manage and control Pinus sp. as an invasive species.

In addition to mapping the target species in this study, the method was also helpful as a classifier of images in general, and it can be employed in managing other exotic plant species or even in other environmental applications.

Compared to the RF, the SVM classifier obtained better results in all areas studied considering the measures used in the accuracy evaluation. According to these results, SVM behavior for this application is apparently more consistent. Nevertheless, further studies involving the method’s application with similar equipment in other areas should be encouraged to assess whether such behavior can be confirmed as a trend. Regarding the used attributes, the classifiers differed significantly across the four areas (Appendix E).

The main methodological difference in this study from others was the effort to increase the set of attributes and select them by applying distinct methods, with subsequent ranking according to the accuracy assessment. This approach allowed improving the classification substantially in all four areas evaluated. Thus, although a single set of attributes, which is ideal for carrying out a classification in multiple areas, has not been found, it was possible to identify which set of attributes best fit to classify each area and achieve an excellent kappa index.

Given the results achieved, it is likely that further accurate classifications can be reached by adjusting the used technique. Thus, the suggestion is to investigate other vegetation indices as a substitute or in conjunction with VARI, as well as other algorithms, to extract textural attributes and other attribute selection techniques.

The recommendation is also to apply the technique in larger areas, since computational capacity can represent a bottleneck. An alternative is to divide an area into subareas with adequate dimensions, dividing the segmentation flow into stages and eliminating the first segmentation, which is optional, as proposed by De Luca et al. [29]. Another possibility is to develop the model in a restricted area that sufficiently includes all classes before application to the entire segmented area.