Monitoring of grassland resources is important to guarantee the regional grassland conservation, management, and sustainable development (Fernandes et al., 2020). Through the development of science and new technologies the grassland ecosystems monitoring has been increasingly enriched (Lyu et al., 2020; 2022) and can make a unique contribution to the ecology and restoration this forgotten ecosystems. Traditional satellite remote sensing technology is a relevant means of monitoring regional grassland ecosystems, but the images obtained by satellite remote sensing tend to have low spatial resolution and the revisit cycle is too long (Peciña et al., 2019; Balasubramanian et al., 2020).

Recent developments for mapping and assessing ecosystem functions and services require information at a very detailed spatial resolution (Villoslada et al., 2020). In this sense, Unmanned Aerial Systems (UAS), also known as Unmanned Aerial Vehicles (UAV), Remotely Piloted Aerial Systems (RPAS), or simply “drones”, have advanced rapidly to satisfy these needs, and have been increasingly used for ecology, environmental and grassland studies (Anderson and Gaston, 2013; Colomina and Molina, 2014; Villoslada et al., 2020).

Modern UAS enables the production of maps and imagery with extremely high spatial resolution and can be a low-cost tool for monitoring vegetation at the landscape scale (Whitehead and Hugenholtz, 2014; Xie, Zhang, and Welsh, 2015; Wang et al., 2019a; Sun et al., 2021). With pixel-size in the order of centimeters (or even millimeters), UAS imagery has been used to extract structural and functional properties of environments, ranging from individuals to ecosystems (Anderson and Gaston, 2013; Singh and Frazier, 2018; Valbuena et al., 2020; Sun et al., 2021). The refined information provides a good opportunity to obtain and quantify the distribution of vegetation across heterogeneous landscapes, capturing variations in plant community cover and revealing an important strategy for biodiversity conservation (Baena, Boyd, and Moat, 2018; Wang et al., 2019a; Díaz-Delgado, Cazacu, and Adamescu, 2019; Prentice et al., 2021).

UAS can also offer very good temporal resolution (Nex and Remondino, 2013; Kampen et al., 2019) and have become more affordable and capable of offering autonomous and flexible functionalities, such as minimization of user intervention, including the ability to plan and conduct surveys to collect aerial photography (Hassanalian and Abdelkefi, 2017; Hamylton et al., 2020). Also, UAS are easy to operate and remotely controlled, allowing the users to go into areas that are inaccessible to humans (Feng et al., 2021). Moreover, UAS remote sensing systems are composed of at least five components: [1] Platform system; [2] Sensor system; [3] Ground control and data transmission system; [4] Data processing system, and [5] Operators (Sun et al., 2017).

Due to advantages in flexibility and high temporal and spatial resolution, UAS remote sensing technology has been applied in many fields, e.g., geography, ecology, and environmental science (Adao et al., 2017; Reis et al., 2019). In the case of grassland studies with UAS imagery it is possible to capture spectral differences between different vegetation species and effectively improve the accuracy of species identification on smaller vegetation units, providing a technical support for the fine monitoring of grassland vegetation species composition (Schmidt et al., 2017). Another important application of UAS imagery is in grassland degradation monitoring, including studies concerning degradation of total vegetation (biomass and productivity), degradation of vegetation structure (reduction of plant height and increase of weeds), processes that pose several challenges to grassland management and sustainable development (Lyu et al., 2020). The hyperspectral and hyperspatial remote sensing technology can further obtain more-in-depth data information and help in monitoring of grassland vegetation structure (Lyu et al., 2022).

In addition, besides the affordable RGB cameras, a wide variety of vegetation mapping sensors can be operated from a UAS platform, including multispectral sensors focused on red and near-infrared wavebands, hyperspectral sensors that detect reflectance at many wavelengths, and, the range of LiDAR sensors (light detection and ranging), indicated to study the degree of vegetative ground cover and characterize canopy structure (Colomina and Molina, 2014; Hamylton et al., 2020). RGB cameras are most used due, mainly, to their low cost and ease of use (Pichon et al., 2019), while multispectral sensors can better evaluate plant health and stress status, due to their highest spectral resolution (Wang et al., 2019b). Finally, as another advantage of UAS, ultra-low altitude flying can reduce the effect of clouds on imagery and thereby improve the data quality (Watts, Ambrosia, and Hinkley, 2012; Sun et al., 2021).

One of the applications of Remote Sensing is the use of imagery to classify and delineate different objects and land cover types on the Earth’s surface, a process that involves collecting field data from a series of samples as an input for training a classification model (Zou and Greenberg, 2019; Prentice et al., 2021). Classification in Remote Sensing involves the categorization of response functions recorded in imagery as representations of real-world objects, according to their spectral similarity to the initial values overlapping the samples, which can provide detailed information about land-cover, specifically in a mixed forest-grassland (Corbane et al., 2015; Cullum et al., 2016; Hamylton et al., 2020; Zhang et al., 2021).

In the context of climate change and rapid transformation of grassland environment due to anthropogenic activities, mainly during the lasted four decades (Fernandes et al., 2020; Buisson et al., 2022), the grassland classification is crucial for its management. The identification of grassland classes provides the basis for the protection of grassland resources and for the reconstruction and restoration of a grassland ecological environment (Meng et al., 2022).

Conventional platforms are characterized by lack of spatial detail to solve and classify fine landscape features (individual trees and shrubs), requiring large amounts of cost and resources (Wang et al., 2019a; Meng et al., 2022). In this regard, hyperspatial tools, such as UAS imagery, have been successfully applied in many fine-scale classification studies, allowing the identification of subtle variations in species composition, and capturing detailed vegetation and landscape features, for instance, shadows, stems, and canopy gaps (Laliberte and Rango, 2013). According to a study performed by Christian and Christiane (2014), data collected from UAS can capture more information about environment composition and structure. Thus, using UAS high-resolution cameras is one of the most preferred methods to classify land cover in a mixed savannas-grassland ecosystem (Zhang et al., 2021).

However, there are still several challenges in using this strategy, including the development of appropriate procedures to manage and extract information from high-volume and hyperspatial resolution data. One of the challenges is how to deal with shadows cast by trees, high intra-class spectral variation, and high inter-class spectral heterogeneity (Lu and Weng, 2007; Lu and He, 2017; Berra, Gaulton, and Barr, 2019).

The classification process can be divided into unsupervised and supervised classification algorithms, revealing it as a crucial tool to achieve interpretable results. Unsupervised classification techniques group pixels or segments according to their similarity using a variety of different algorithms. On the other hand, supervised classification requires the use of training sample areas, in other words, spectral signatures of the objects under study (Schafer et al., 2016; Nogueira et al., 2019).

The application of conventional classification techniques on hyperspatial imagery can result in the misclassification of pixels with identical spectral responses to different classes, causing an effect known as “salt and pepper”. Furthermore, traditional supervised classifiers assume the normal distribution of remote sensing datasets (Jensen, 2015), but these datasets can deal with complex and non-linear relations, thus conventional classification algorithms do not fit to classify the ultrahigh remote sensing data (Thessen, 2016; Oddi et al., 2019).

Some strategies have been developed to solve these problems: A. Inclusion of texture features in a classifier; B. Object-Based Image Analysis (OBIA), in which an image is pre-segmented into polygons; C. Using computer vision and machine learning techniques, such as random forest, support vector machine and convolutional neural networks (Nguyen et al., 2019; Zou and Greenberg, 2019). In addition, the accuracy of vegetation identification will depend on four factors: spatial resolution, spectral resolution, habitat complexity, and classification algorithms.

Object-Based Image Analysis (OBIA), more specifically Geographic Object-Based Image Analysis (GEOBIA), has provided important methodological refinements for high-resolution image classification, having advantages over traditional pixel-based methods, and is considered a superior classification technique (Liu and Abd-Elrahman, 2018; Dujon and Schofield, 2019). GEOBIA methods are based on pre-clustering the image pixel into homogeneous objects (regions, clusters), according to specific spectral characteristics (features, color, texture), and shape characteristics (Blaschke, 2010).

Machine Learning (ML) approach, one of the most popular latest technologies in the fourth industrial revolution (Sarker, 2021), has been developed as a response to the rigidity of many other computer programs in comparison with the world’s infinite versatility. Emerged within the field of artificial intelligence, ML aims to deal with complex data and to learn without being explicitly programmed (Mahesh, 2020; Prentice et al., 2021). The data and desired result are provided to a learning algorithm (a “learner”), which then generates the algorithm that turns one into the other (Dujon and Schofield, 2019; Hamylton et al., 2020). ML algorithms are divided into four categories: Supervised learning, Unsupervised learning, Semi-supervised learning, and Reinforcement learning. In the case of Supervised learning (the category focused on the research), the most common supervised tasks are the “classification algorithms”, which separate the data, and the “regression models”, which fit the data, but do not offer predictive capability (Sarker, 2021). The supervised learning technique, specifically the classification process, has been successfully applied to sequential RGB images from ground digital cameras focused on species recognition, and in studies that involve ecology and Earth sciences applications (Almeida et al., 2014; 2016; Thessen, 2016; Maxwell, Warner, and Fang, 2018).

In ecology, the Random Forests algorithm is commonly used to perform the supervised classification process, followed by maximum likelihood classification, support vector machine, K-means clustering, convolution neural networks, and thresholding (Dujon and Schofield, 2019). Random Forest (RF) is an ensemble learning technique based on decision trees, which has higher accuracy when compared to traditional methods, enabling the simultaneous classification of features based on a set of training samples, making possible the determination of the best performing explanatory variables (Breiman, 2001).

Considering the advantages and challenges explained, the present work aimed to answer the following questions: Are the machine learning techniques effective to process the ultrahigh-resolution data generated by UAS images? Is the OBIA an effective method for classifying heterogeneous vegetation? Using a Random Forest classifier, what is the better combination of parameters for proposing a high quality of ultrahigh spatial resolution classification? We focus our fieldwork on a highly diverse, seasonally dry, complex mountaintop vegetation system, the campo rupestre or rupestrian grassland, considered a vegetation mosaic or a continental archipelago (Morellato and Silveira, 2018; Mattos et al., 2019; Vasconcelos et al., 2020). The mosaic formed by the vegetation types, from grasslands, and rocky outcrops to cerrado and forest patches, poses a challenge to accurately classify campo rupestre vegetation diversity at a landscape scale, a prime measure for the conservation of this threatened ecosystem (Fernandes et al., 2018; 2020). This high diversity of vegetation types cannot be distinguished in conventional imagery, like Sentinel-2 or even Planetscope imagery, but in UAV imagery this is possible.

The results of orthomosaic registration revealed the importance of complete alignment of the time series to obtaining concrete, effective and accurate results in the classification process. Figure 3 shows the comparison between the Euclidean distance displacement before and after the registration process, revealing the improvement in the quality and accuracy of images generated by UAS.

Regarding the segmentation process, according to enumeration in Table 3, segmentation 10 showed the best parameter combination, producing results that properly covered the homogeneity areas of the image. We used photointerpretation and visual techniques to choose the best parameter combination.

In addition, concerning the classification processes, all values extracted during the validation process performed in R (KAPPA index, overall accuracy, and overall error) are shown in Table 5. The classification to the date 05 of September 2016 received 0.987 for KAPPA Index and 0.989 for general accuracy. Moreover, the classification to the date 05 of January 2017 received 0.985 for KAPPA Index and 0.988 for general accuracy. Finally, the classification using the stack with two dates received 0.969 for KAPPA Index and 0.974 for general accuracy.

The processes using machine learning techniques require attention to their parameter choice. In the case of Random Forests, the main parameter tested is n-estimators, referring to the number of decision trees generated by the classifier. We tested three different numbers of decision trees, as exposed in Materials and Methods, and 500 trees were chosen as the better number for classifying the rupestrian grassland.

In general, the greater number of decision trees means a greater number of analyzes that the algorithm will perform to arrive at a result. This number of decision trees was chosen due to the peculiarities present in the type of image (hyperspatial resolution) and in the type of vegetation (rupestrian grassland with high vegetation heterogeneity). In this sense, for effective classification of the rupestrian grassland, more decision trees were needed. Thus, all data was processed using 500 decision trees and these results are presented in Figure 4.

We classified each image individually and, also, the stack image (a merged between the image dated 25 September 2016 and 05 January 2017). According to Table 6, which shows the proportion of each spectral class according to the total area, in each classification performed, the Rocky Outcrop class received the highest percentage of area classified (25.398% in September 2016, 24.107% in January 2017 and 25.114% in stack), followed by Sandy Grassland (20.806% in September 2016, 21.108% in January 2017 and 18.689% in stack). The lowest percentage of area classified focus on Bare Soils (1.243% in September 2016, 0.994% in January 2017 and 1.408% in stack), followed by Bare Rocks (1.924% in September 2016, 2.431% in January 2017 and 2.325% in stack). The stacking process helped the algorithm to properly classify the spectral classes present in rupestrian grassland according to reality, furthering the capture of different phenological stages and patterns of each vegetation type.

In Tables 7–9 we presented the results of analyzes performed with the error matrix of each classification process. An error matrix is a square array of numbers set out in rows and columns, in which the columns represent the reference data (validation data) while the rows represent the classification generated from the UAS data. The major diagonal of the matrix indicates the percent of pixels classified correctly, while the off-diagonal cell values represent the percent of pixels classified incorrectly in each spectral class.

In all results, we detected an uncertainty between two spectral classes: sandy grassland and wet grassland. In Table 7, 18.380% of pixels of sandy grassland were classified correctly, but 0.322% of these pixels were misclassified as wet grassland. In Table 8, 19.320% of pixels were classified correctly and 0.283% misclassified. In Table 9, 17.601% were classified correctly and 1.410% misclassified.

The other two relevant analyzes performed in the error matrix are presented in Table 8, where 0.493% of pixels of bare soils were misclassified as stony grassland. Also, in Table 9, 0.490% of pixels of stony grassland were misclassified as bare soils. This analysis showed some uncertainty between bare soils and stony grassland.

Multi-temporal UAS images associated with RGB cameras, and the machine learning methods applied in our study produced consistent results, as far as concerns the ecology analysis and are adequate for mapping vegetation, at least at the life-form level (i.e., tree, shrub, and herbaceous species). The combination between the OBIA method and the Random Forest classifier reduced the misclassified pixels and the effect known as ‘salt and pepper’, generating extremely high general accuracy and kappa index. In general, machine learning approaches have high potential to capture the non-linear relationship between remote sensing data and vegetation parameters and have the capability of integrating multisource information at different levels (Yao, Qin, and Chen, 2019). In addition, the results acquired confirmed the importance of using machine learning algorithms in remote sensing vegetation classification, due to their powerful adaptation, self-learning, and parallel processing capabilities (Navin and Agilandeeswari, 2020; Meng et al., 2021; 2022).

Therefore, some challenges need to be emphasized and both the hardware and software of UAS remote sensing technology require improvements: [1] The endurance of UAS is relatively limited, the flight stability is not strong enough in areas with large terrain fluctuation and the lack of flight altitude limits the image size; [2] Although more lightweight and smaller sensor systems have become available, such as hyperspectral and LiDAR sensors, but they are still expensive; [3] The integration between UAS platforms and sensors requires improvement, e.g., most of the multispectral, hyperspectral, and thermal sensors are built independent of the UAV platform, so, need an extra GPS module and, also, UAS are often equipped with a single sensor, multisensor integration is beneficial to improve monitoring accuracy and efficiency; [4] The mosaic workload is significantly higher than satellite imagery, which takes up more time for image processing, resulting in the need to develop more robust algorithms, like deep learning techniques, in addition, the technology of mass data processing needs to be improved due to the richness and variety of data obtained; [5] The most vegetation classifications via UAS require human participation and interpretation, indicating that the combination between UAS remote sensing with ground data and satellite data needs to be strengthened; if the dataset used for training is extent, computer learning techniques would generate a satisfactory classification outcome; [6] The use of UAS images to monitor tropical savannas leaf phenology is a challenge due to difficult methods to extract accurate quantitative phenology estimates under variable lighting and viewing conditions; [7] The application scenarios of UAS remote sensing in grassland ecosystem monitoring need to be expanded and deepened, mainly in animal investigation and soil physical and chemical monitoring; and also, the correlation between the scientific research of UAS remote sensing monitoring and practical decision making of grassland management is still insufficient (Neumann et al., 2019; Park et al., 2019; Lyu et al., 2020; 2022; Sun et al., 2021).

The hyperspatial imagery, like UAS technology, associated with machine learning classification techniques enabled the identification of even fine grassland vegetation types such as wet, stony, and sandy grasslands. Although, some adjustments were necessary for improving the accuracy of orthomosaic classification, such as [1] precise sensor position; [2] orientation data, and [3] several GCPs during the orthomosaic registration process (Tsai and Lin, 2017).

This study examined the use of Unmanned Aerial Systems remote sensing technology associated with machine learning techniques (OBIA method and Random Forest algorithm) to classify and understand the vegetation distribution across a grassland landscape. Through the results obtained with the validation process all classifications performed received general accuracy and KAPPA index above than 0.96, indicating a high efficiency of machine learning techniques to process the ultrahigh spatial resolution data generated by UAS images. Using the UAS to explore and analyze vegetation phenomena is becoming a new efficient remote sensing technique, providing vegetation recognition, the extraction of structural and functional properties of ecosystems and the quantification from individual plants down to the leaf scale. A big challenge of applying a robust algorithm is the choice of the better combination of parameters for proposing a high quality of vegetation classification. Our results demonstrated that the better combination of parameters, for segmentation process, were 20 for the number of K-means clusters, which controls the general homogeneity and the number of resulting clusters, and 600 pixels for the minimum size of the resulting cluster, which produced results that properly covered the homogeneity areas of the image. In addition, for the Random Forest classifier the number of decision trees chosen were 500 decision trees, where the greater number of decision trees means a greater number of analyzes that the algorithm will perform to arrive at a result.

In recent years, machine learning techniques associated with ultrahigh spatial resolution imagery has been widely used in grassland classification due to its high accuracy and powerful processing ability, however, still there are set of challenges and limitations, such as high time-consuming, because processing this big dataset require a great computational configuration, making it impossible processing the data with conventional techniques and traditional GIS tools; therefore, automatic identification of vegetation classes requires further exploration. In future studies, we suggest explore the method of combining machine learning algorithms, multitype indices (NDVI, DEM, temperature, precipitation, and so on) or multisensor integration to improve the accuracy of grassland classification. Considering future directions the UAS remote sensing technology has been rapidly improved and developed along with of precision and intelligence; the endurance, stability and flight height and other performance parameters of UAV platforms will be significantly improved; the cost of sensors can be reduced; machine learning techniques will become an important technical means to provide a technical solution for automatic processing and analysis of massive monitoring data, and finally, the decision-making support of UAS remote sensing for grassland management will be enhanced.