The terms land use and land cover, although used in associated ways, would rather be defined in a dissociated way. Land use refers to the way the biophysical attributes of the land are manipulated and the underlying intention to its manipulation. Land cover refers to the biophysical state of the Earth’s surfaces and the immediate subsoil [46] [47]. Land use causes changes in land cover, and such changes when intense and on a global scale affect the key aspects of the terrestrial systems functions.

According to Briassoulis [47], and Nedd et al. [48] biophysical factors (climate, temperature, topography, soil type, surface water, humidity, vegetation, and fauna) and social (population, technology, socioeconomic, cultural and institutional organization, and political changes) are responsible for such changes and seen interconnected in a space-time perspective. Gómez et al. [46] highlight that distinct types of land cover provide specific habitats and determine the energy and carbon exchange between terrestrial and atmospheric regions. The Knowledge and mapping of land use and land cover are essential to plan and manage natural resources, modeling of environmental variables, and to comprehend the distribution of habitats [48]. Land cover naturally changes over time, also due to the influence and result of anthropogenic activities.

According to Gómez et al. [46], Earth Observation (EO) data provide land use and land cover mapping and monitoring in a consistent and robust manner over large areas, and results are available by different world space agencies at different spatial and temporal scales, matching scientific and political information needs. Geotechnologies have been relevant in the study of land use and land cover as they have enabled observation, identification, mapping, assessment, and monitoring of land cover in spatial, temporal, and thematic scales [46] [49].

Identifying types of land cover provides basic information to generate other thematic maps, and to establish a baseline for monitoring activities. According to Rogan and Chen [49] an effective approach to identify changes for a specific period may maximize exploration in the domains of spatial and spectral resolutions, as using additional data, such as vegetation indexes. On the other hand, 2 significant taxon to separate cover from use changes namely: 1) categorical—known as post-classification comparison, which occurs between a set of thematic categories of land use and land cover (i.e., urban, forest); and 2) continuous—known as pre-classification enhancement, where changes occur in the quantity or concentration of some attribute of the built or natural landscape that may be measured continuously.

Most approaches to monitoring land use and land cover have used traditional image classification algorithms that assume: 1) image data is normally distributed, 2) objects of interest on the surface are larger than the pixel size (H-resolution), and 3) the pixels are composed of a single type of land cover or land use. However, some approaches argue that objects of interest on the surface are smaller than the pixel size (L-resolution), and therefore, they used empirical models to estimate biophysical, demographic and socioeconomic information [49] [50].

Over the years, several studies on land cover have been conducted [3] [23] [51] [52] by using data from various sensors with different resolutions, techniques, and methods of data processing with the purpose to map and monitor land use and land cover. The methods may be grouped into supervised/semi-supervised, and unsupervised classification, and the classifiers grouped into parametric and non-parametric, rigid and flexible (diffuse), or based on pixel/subpixel and object [21] [46].

The parametric such as maximum likelihood, minimum distance and Bayesian classifiers are based on probabilistic theories, modeling the decision boundaries between classes from a fixed number of parameters, independent of the number of samples, using global criteria for the classification [53]. Conversely, the non-parametric, such as support vector machine (SVM), artificial neural network (ANN), guide the grouping of classes based on the digital number (single band/image) or spectral reflectance (multispectral image) and other characteristics such as shape and textural attributes of the scene. The distribution of the image values are independent and its focusing on the local data structure, requiring a high set of samples for the classification process [54] [55] [56] [57].

According to Phiri and Morgenroth [55], advances in object pattern recognition techniques through artificial intelligence and machine learning approaches have contributed significantly to develop advanced non-parametric classifiers, commonly used in GIS and Digital Image Processing commercial software, as well as open source.

The pixel approach, such as Random Forest and spectral matching techniques, is based on the use of the pixel spectral information to find its most likely class, plotting a probability that a given pixel belongs to a certain class or not, i.e., the pixels of a class are more similar from a spectral point of view than the pixels from other classes [55] [58].

The Object-Based Image Analysis (OBIA) approach uses geographic objects as basic units to classify land use and land cover reducing variations in targets within the class and it removes the effects of “salt and pepper” that result from isolated pixels incorrectly classified. It presents the advantage of incorporating several sources of information, such as texture, shape, and position as a basis for the classification [54] [55] [58]. Its main limitation is associated with choosing the appropriate segmentation scale and dealing with different steps, which may be a source of variation if not treated properly.

The sub-pixel-based approach was developed to address divergences in pixel-based classification, such as the separation of land uses and land cover in mixed pixels [50] [51] [59]. The approach proved to be suitable from medium to low spatial resolution sensors, and widely used in regional, continental, or even global mapping [51] [60]. Statistical algorithms, such as Maximum Likelihood (Maxver), Linear Mixture Model (LSMM), and those based on set theories, such as the Possibilistic C-Means (PCM)), and the Fuzzy C-Means (FCM) are some examples of sub-pixel classification.

These are already incorporated into distinct image processing software available on the market, both proprietary and open-source. However, the choice of classifier proves to be complex and challengeable, due to each method presents its strengths and weaknesses as shown in Table 3.

In this regard, Ackom et al. [61], Mohammady et al. [62] suggest hybrid approaches to solve the issues which have become more powerful and diversified, due to the development of powerful and advanced classifiers.

Other strategies can be incorporated such as those which possibility to infer proportions of vegetation cover, commonly known as vegetation indexes. The most used are Normalized Difference Vegetation Index—(NDVI)—Normalized Difference Water Index (NDWI)—Soil-adjusted Vegetation Index (SAVI)—the Normalized Difference Built Index (NDBI)—Spectral Mixture Analysis Modified Soil Adjusted Vegetation Index (MSAVI).

Nevertheless, the success of this approach depends on several factors, such as quality of pre-processing, analyst experience, and classifier performance. However, depending on the complexity of the subject, Gómez et al. [46] point out the following criteria to be considered when choosing the classification algorithm: type of data, statistical distribution of classes, target accuracy, ease of use, speed, scalability and interpretability in order to achieve acceptable accuracy and rationalization of resources (Table 4).

Machine Learning (ML) are algorithms or models built to gain from information and acts appropriately in future circumstances, being grouped in: Lazy (e.g. k-nearest neighbour, Case-based reasoning) and eager (Decision Tree, Naive Bayes, Artificial Neural Networks) and mainly divided into four categories: supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning [81] [82].

According to Galván et al. [81] most of the machine learning algorithms (MLAs)—based on trees, rules, functions, etc.—are eager learning methods, in the sense that the generalization is carried out beyond the training data before observing the new instance. They are a powerful tool for training Artificial Intelligence

Source: Authors (2021).

ML-Maximum Likelihood; MKNN-Modified k-nearest Neighbors; TEP-Thematic Exploitation Platform; SMCF-Superpixel and Multi-Classifier Fusion; MESMA-Multiple Endmember Spectral Mixture Analysis; APEX-Airborne Prism Experiment; LSMA-Linear Spectral Mixture Analysis; NN-Nearest Neighbor; GEE-Google Earth Engine. Source: Authors (2021).

(AI) models that can help increase automation or optimize the operational efficiency of sophisticated systems such as robotics, autonomous driving tasks, manufacturing and supply chain logistics [82].

Classification methods in ML can be Binary, which refers to the classification tasks having two class labels such as “true or yes and false or no”; multiclass— which refers to those classification tasks having more than two class labels; and multi-label—which represent the generalization of multiclass classification, where the classes involved in the problem are hierarchically structured, and each example may simultaneously belong to more than one class in each hierarchical level [82] [83].

Several works on land use and land cover mapping using machine learning classifiers have been carried out [3] [6] [21] [56] [84] —Classification and Regression Trees (CART), Random Forest (RF), kNearest Neighbor (k-NN), Support Vector Machine (SVM), Artificial Neural Network (ANN), Multinomial Logistic Regression (MLR), C5.0, J48 (Decision Tree) have shown more efficient compared to conventional ones.

According to Shetty [21], while some of these classifiers such as SVM find a subset of training data as support vectors by fitting a hyperplane that separates twos classes in the best possible way, CART build simple decision tree from the given training data, ANN follow a neural network pattern and build multiple layer of nodes to passes input observations back and forth during the learning process (Multi-Layer Perceptron) until it reaches a termination condition, RF uses random subset of training data to construct multiple decision trees. Figure 4 and Figure 5 illustrate respectively some Machine Learning Methods highlighting supervised learning and machine learning workflow.

They register improvements of 10% to 20% higher accuracy when confronted with complex data related to large areas [46]. The success of the classifiers for the theme is due to their unlimited assumptions of parametric statistics, and therefore, more suitable for: 1) analyzing multimodal, noisy and/or absent data; 2) analyze combinations of categorical and continuous auxiliary data; 3) reduction of pre-processing steps required in traditional approaches; 4) performance in cloud computing environments, such as Google Earth Engine (GEE).

GEE is a cloud-based platform with multiple petabytes which provides parallel computing and data catalog services for geospatial analysis on a planetary scale [5] [85]. The calculations are automatically parallelized, and data sets ready for public use. The calculations come from several geospatial data development agencies, such as the United States Geological Survey (USGS) and the European Space Agency (ESA), reflection data sets from Landsat surface to Sentinel data sets, various global land cover data, climate data sets, among others. It provides several integrated methods that support pre-processing images, in addition to having a repository of vast functions, such as masks, logical operators, sampling data, etc., and they perform various operations on images and vectors [21].

An example of machine learning applications can be found in Li et al. [86], article published in Remote Sensing Journal. In this article, it’s proposed to generate a land cover map of the whole African continent at 10 m resolution utilizing machine learning algorithms, and multi-source data in the GEE platform. A workflow designed for on it is shown in Figure 6 where it’s highlighted the 5 steps of machine learning workflow.

Through this study it was possible to generate a map of land cover of all Africa, obtaining an accuracy of 81% for 5 classes, which is relatively superior to the existing 10 m land cover product (e.g., FROM-GLC10) in detecting urban class in city areas and identifying the boundaries between trees and low plants in rural areas. Part of the results of this study are shown in Figure 7 (more details can be found in the original article).

Time series of medium spatial resolution optical data present significant results compared to a single scene. It presents high capacity to characterize environmental phenomena describing trends, as well as discrete events of change in characterization and identification of nature changes in land use and land cover [51] [87] [88].

Landsat data is considered the appropriate/standard to classify land use and land cover changes [55], due to its spatial resolution (30 m), temporal (16/8 days), covered area (185 × 185 km), rigorous calibration, and consistency in the radiometry of sensors (TM/ETM+/OLI).

For Chi et al. [20] traditionally, data fusion can be carried out in terms of pixel-level fusion, feature-level fusion, and decision-level fusion. However, big data in remote sensing usually comprise different scales and/or formats.

According to Huang and Wang [11] Big Spatial Data (BSD) can integrate data from various sources, providing a more comprehensive picture and when doing so, a huge amount of data is pulled from different formats, devices or systems and given a geographic context to facilitate building a complete picture or analysis, but it’s important to wonder how to integrate the data from various sources, where data features are significantly different (e.g., spectral signatures in optical remote sensing data, electromagnetic radiation in microwave data, structural features of texts, unstructured features of images by a digital camera, etc.).

The use of multi-source data, also provides land use and land cover mapping, and improve classification accuracy [6] [23] [54] [86] by collecting samples with high resolution sensors and fusing different sensor products (optical/optical or optical/radar) allowing clear target differentiation.

In this regard, Häme et al. [70] used the Hierarchical Clustering method to detect and identify changes in land cover using paired data Sentinel-2/Sentinel-2, Landsat-8/Sentinel-2, and Sentinel-2/ALOS 2 PALSAR in an area of 12,372 km² in Finland. Joshi et al. [89] reviewed 112 studies on fusing optical and radar data, which offer unique spectral and structural information, for land cover and use assessments were they assessed advantages of fusion for land use analysis in 32 studies, and a large majority (28 studies) concluded that fusion improved results compared to using single data sources. MateoGarcia et al. [90] proposed and implemented a methodology to mask clouds (Cloud Mask) using GEE to map a type of biome based on data from OLI/Landsat-8. The algorithms used (FMask and ACCA), showed relevant quantitative performance, improving from 4% to 5% in classification accuracy, and 3% to 10% in commission errors. Adamo et al. [80], Samal and Gedam [75] present others applications.

Monitoring and managing the territory requires accurate information on land cover. Attempts to obtain accurate land use and land cover maps will always accompany professionals in the area [21]. The validation of land cover products is essential to show the quality of remote sensing products for decision making. Evaluating and reporting with appropriate information metrics is essential for the community of users [46]. Thus, factors such as size and quality of training samples, thematic accuracy, choice of classifier, and the size of study area affect the accuracy of classified maps [13] [14] [21] [91].

Comprehending these factors aids to find the appropriate accuracy classification for a given problem studied [21]. Selecting samples must comply with statistical criteria, such as the type and method of sampling. Mastella and Vieira [56], Shetty [21] states that in remote sensing, simple and stratified random sampling are largely used, with most validation indexes based on simple random. However, authors who employ and recommend systematic sampling methods to study land use and land cover, due to their accurate results, despite the absence of an unbiased estimate of the variance [21].

Accuracy assessment is the key component to have maps with remote sensing data, as it supports evaluating the performance of various classifiers and the effect of sampling [6]. The literature recommends the inclusion of an error or confusion matrix [13] [14] to assist the identification of confusion between classes, as well as potential sources of error [21] [46] [56]. Furthermore, quantitative metrics derived from confusion matrix provide significant support, such as global accuracy, which expresses how close the classified map is to the original, as well as weighted metrics (producer accuracy and user accuracy, Kappa index, Tau, statistical Z, among others) by area and confidence intervals.