A Review of Fine-Scale Land Use and Land Cover Classification in Open-Pit Mining Areas by Remote Sensing Techniques

Over recent decades, fine-scale land use and land cover classification in open-pit mine areas (LCCMA) has become very important for understanding the influence of mining activities on the regional geo-environment, and for environmental impact assessment procedure. This research reviews advances in fine-scale LCCMA from the following aspects. Firstly, it analyzes and proposes classification thematic resolution for LCCMA. Secondly, remote sensing data sources, features, feature selection methods, and classification algorithms for LCCMA are summarized. Thirdly, three major factors that affect LCCMA are discussed: significant three-dimensional terrain features, strong LCCMA feature variability, and homogeneity of spectral-spatial features. Correspondingly, three key scientific issues that limit the accuracy of LCCMA are presented. Finally, several future research directions are discussed: (1) unitization of new sensors, particularly those with stereo survey ability; (2) procurement of sensitive features by new sensors and combinations of sensitive features using novel feature selection methods; (3) development of robust and self-adjusted classification algorithms, such as ensemble learning and deep learning for LCCMA; and (4) application of fine-scale mining information for regularity and management of mines.


Importance of Fine-Scale Land Cover Classification in Open-Pit Mining Areas
During the process of exploration, development, and beneficiation, open-pit mining areas (hereafter MA) may experience denudation, handling, and accumulation.These processes can not only drastically alter the vegetation, soil, and terrain around MA, but also lead to the rapid formation of pits (including waste and in-use), transit sites (including ore piles, ore dressing, and mineral processing pools), solid waste (including dumps, waste piles, closed and in-use tailings, and coal gangue heaps), and other surface mining elements.When these elements are superimposed on the primary geo-environment of a MA, a series of environmental problems occur, such as resource damage, geological disasters (landslide, collapse, debris flow), and environmental pollution, which seriously hinder regional and global sustainable development.
China is currently accelerating construction of green mines and strengthening the ecological construction of mines.The detailed land cover dataset in MA is vital for the national project.Compared with underground mining, open-casting is more damaging to the geo-environment.
In addition, land use and land cover classification in open-pit mine areas (LCCMA) is most often driven by the need for environmental impact assessment procedure and monitoring to assess impact or success of implemented mitigation measures.
Therefore, rapid and accurate LCCMA is key to comprehensively understanding the influence of mining activities on the regional geo-environment.The specific requirements relating to the LCCMA should focus mainly on two aspects: one is more detailed classes, the other is higher accuracy.
Remote sensing is the only practical and cost-effective means of generating such information over large areas [1].Over the past decade, acquisition capabilities of many satellite remote sensing systems have increased rapidly to better than 2.5 m, and comprehensive classification and high-performance computing technology for big data have been successfully developed [2][3][4][5].Accordingly, fine-scale land cover information has become crucial for mapping and managing complex Earth surface environments, such as urban and surface mine environments in MAs, at local and regional scales.Much recent research has focused on fine-scale land cover mapping of complex mining areas [4,6], as it can provide technical benefits for both regional-and national-scale mapping.Moreover, with their rapid development, spaceborne remote sensors of high spatial resolution (HR) have become the dominant data resource for fine-scale land cover interpretation [7].
Mine surface information includes not only mine surface elements around the MA, such as mining kits, transit sites, solid waste, mining construction, and mining roads, but also natural surface elements, such as residential buildings, farming, woodland, and unused land.Conventional methods of interpreting remote sensing data involve visual interpretation based on field investigation.The method (eventually integrated with results by object-oriented classification) of remote sensing data supported by field checks provide land cover/use datasets with high accuracy.Moreover, this integrated approach is straightforward when multitemporal information have to be collected, managed and analyzed.They have been successfully applied at both "regional" and "high detail" scales [8].However, the results depend on the analyst's skill in image interpretation, which is time-consuming and presents difficulty when updating the results [9].Thus, it is necessary to develop automatic classification methodology.

Development of Land Cover Classification in Open-Pit Mining Areas by Remote Sensing
Since the 1970s, Landsat satellite data have been used for LCCMA.Remote sensing data sources have evolved from low resolution to high resolution.For example, some researchers have focused on identifying contamination in tailings and mine areas using hyperspectral data [10][11][12][13].However, the swath width of a hyperspectral image is usually narrow, and there are fewer satellite data sources.These disadvantages limit application of hyperspectral images.Other studies concentrated on medium and high spatial resolution satellites [14][15][16].The America National Land Cover Database (NLCD) based on Landsat TM was completed in the 1990s by the Multi-Resolution Land Characteristics (MRLC) Consortium, and revealed that the classification accuracy of LCCMA is lower than 60% with 48% omission [17].In addition, maps of mine reclamation are rare, in large part because reclaimed mines are spectrally indistinguishable from grasslands or pastures [18].For instance, bare exposed mine dumps and grasses and trees above it were distinguished from farmland, water, and forest using a single Landsat TM in Pennsylvania, but the total mapping accuracy was only 62.3% [19].Some researchers in China performed LCCMA with Landsat TM and Beijing-1, and also considered mine surface elements as a single class [20].In the coal mining area of Pennsylvania, SPOT was employed to classify land cover with a total accuracy of 70%.However, this study could still not distinguish mine surface elements and reclamation areas from other land types [21].It should be noted that the low classification accuracy of land cover can be improved up to 92% by using HR remote sensing data, such as ALOS fusion images, and the object-oriented method [22,23].
Although land cover interpretation of surface mine areas has been performed using remote sensing techniques for almost 40 years, existing research often merges mining system elements into a single class, or simple subclasses such as developed land or construction land.In this case, the product is of rather weaker value compared to a product with more detail [13,16,21].This could be attributed to the low spatial resolution of the employed remote sensing data, lack of a remote sensing representation model, feature selection method, and classification algorithm used.As the construction of "green mines" continues to be promoted, such approximate classification systems and low classification accuracy cannot meet the national strategic needs of ecological mine construction.Therefore, fine-scale land cover classification of mining areas is vital.

Objectives
This review aims to summarize fine-scale LCCMA studies with a focus on several key issues affecting classification accuracy, such as features, sensitive feature combinations, and classification algorithms.Thus, we checked almost all the available literature about LCCMA based on ISI (web of sciences) and Google scholar.In particular, we analyzed the last 10 years' progress.The used key words were as follows: complex landscapes; mining areas; mountaintop mining; mineral mapping and WorldView-3; machine learning algorithms and classification; remote sensing.First, LCCMA land cover thematic resolution (Section 2) and the three most significant factors (Section 3) affecting LCCMA are presented.Sections 4-6 are dedicated to each respective factor.Then, several scientific issues limiting the development of LCCMA are discussed in Section 7. Further developments and conclusions are presented in Sections 8 and 9.

Classification Thematic Resolution for LCCMA
Different classification thematic resolutions for LCCMA have been established for different purposes.For mining and mine reclamation classification, Maxwell et al. [24][25][26] defined three land cover types related to mining areas.For details, see Table 1.Maxwell and Warner [27] designed two land cover classes to differentiate mine-reclaimed grasslands (Table 1).Guan et al. [28] formed seven classes for land use classification and change detection in a coal mining area: open stope, stripping area, waste-dump area, mine industrial area, farmland, urban area, and the original landscape (grassland) (Table 1).Lechner et al. [29] grouped 22 fine-scale land cover categories into six land cover classes based primarily on disturbance type (Table 1).Karan et al. [30] investigated seven distinct classes with three surface-mining related land cover types (Table 1).Bao et al. [31] focused on the classification of rehabilitated vegetation (Table 1).In Alkan et al. [32], eight classes were generated: sea, artificial lake, forest, agricultural area, coalfield, urban area, open area, and pollution in the sea.However, coalfield represented both coal mining excavation cities and coal dump areas.Townsend et al. [18] examined active and reclaimed mines.Prakash and Gupta [14] recognized 10 classes with two mining land types.Lévesque and Staenz [33] monitored mine tailing revegetation, and focused on bare tailings, exposed lime, and water/wet tailings.Richter et al. [34] classified acid mine drainage pollution zones in tailing areas, into tailings zone (very high), dry vegetation zone (high), transition zone (low to medium), and vegetated island (very low to low).Demirel et al. [35,36] examined three mining related land cover types: mine, dump, and coal stockpile.
In general, there is no universal scheme used among all these studies (Table 1).The datasets are rather inconsistent, and each type is determined based mainly on specific research purposed.For example, in the field of land reclamation in the mining area, it mainly focuses on the types of vegetation in the mining area and the type system of bare soil.Maxwell focused mainly on reclaimed herbaceous vegetation, reclaimed woody vegetation, and barren.However, the barren is determined to include haul roads, active quarries, land disturbed by mining.When more types are proposed, the difficulties of classification are theoretically greater.Thus, it is necessary to set up thematic resolution for the LCCMA according to different application requirements.For instance, it is possible to consider FAO land cover classification system [37] and proposed modifications [38][39][40], a standard land cover classification scheme for remote-sensing applications [40], National Land Use Classification Criteria and mine remote sensing monitoring elements in China as basic thematic resolution.
For fine-scale LCCMA, classification thematic resolution specific to the study area should be constructed such that surface-mined land is classified into multiple second-level classes.For example, based on a in mixed complex surface-mined land in Wuhan City, China, our team developed a thematic resolution with 7 first-level and 20 second-level classes [41].The surface-mined land was classified into three subclasses, i.e., opencast stope, mineral processing land, and dumping site.

Dominant Factors Affecting LCCMA
Classification technologies based on machine-learning algorithms and HR have achieved more accurate results for urban environments, precision agriculture, transportation, forestry surveys, and so on.However, LCCMA classification differs from other fields because specific features of LCCMA increase the difficulty of obtaining high accuracy classification results.These features can be revealed as follows based on our team's experience in the last five years: significant three-dimensional terrain features, stronger feature variability, and homogeneity of spectral-spatial feature.These are described in the following subsections.

Significant Three-Dimensional Terrain Features
In an open-pit area, changes in the ridge and characteristics of soil reconstruct the terrain around the mine area [42].The slope and elevation of the mining area change due to open pit and valley filling.This results in significant terrain features such as pits, transit sites, and solid waste.From a landscape perspective, mining areas can be divided into cliffs, steep slopes, slopes, flattop, slope, mining pits, dry flat ground, wet flat ground, and slope bottoms.The pit and surrounding surface or mountain can form a negative terrain with a height difference of a few meters to several hundred meters.An ore pile with a strip or conical spreading pattern represents a positive terrain.Waste stone piles, pits, and dumps along the valley also show a strip or cone-like distribution on the surface, or a fan-shaped distribution down the hillside, and all of these features differ from ore piles.In addition, the transit site typically contains different sizes of ore products for quarries with conical and strip-like distributions.

Stronger Feature Variability of LCCMA
Opencast mining not only changes terrain characteristics but also drastically changes the type of LCCMA [43].During the survey phase, a large number of forests are cut down and converted into bare land; the topsoil is removed and the overburden material above the ore may be thrown into valleys.With the mining of a large amount ores, accumulations of waste stones, ore piles, and waste rock may be discarded in the valleys, and may be used as a material source for reclamation ridges.Furthermore, during the beneficiation phase, mineral products will form in the transit site.The spectral characteristics of closed and in-use tailings also differ, and the scale varies with the strength of beneficiation activity.In addition, the pit may lie adjacent to the ore piles, waste rock piles, and dumps, and waste piles occasionally lie adjacent or stacked with ore piles, ore dressings, and dumps.Due to the existence of man-made materials (such as asphalt, concrete, roof tiles, and other impermeable surfaces), semi-natural materials (grassland, trees, bare soil, water, etc.), and other natural surface elements, a high degree of heterogeneity may arise, which further leads to strong variability of features derived from remote sensing data.

Homogeneity of Spectral-Spatial Features
Spectral responses in HR images of the pit, ore pile, waste pile, dump site, and transit site are all similar, and the weathering effect exacerbates the degree of similarity.Post-reclamation mining areas have some spectral characteristics similar to grass or pasture in HR images.Transport routes are mostly dirt roads or gravel roads, and can constitute an independent transport network from non-mine transportation with spectral-spatial characteristics similar to other transport networks.In addition, mining buildings are generally distributed around the waste rock pile in a single row or linear arrangement, and the three base colors of the roof (red, green, and blue) are similar to those of ordinary civil and industrial buildings.Moreover, ore stacks are more likely to be associated with beneficiation sites.In conclusion, spectral differences between landforms in mining areas are small, and the spectral diversity of the class is large, which leads to strong homogeneity in the spectrum space.
A large number of pixel features and object features have been developed, and a series of scale parameter evaluation tools specific to the object method has been promoted; for example, a tool to estimate scale parameter for multiresolution image segmentation of remotely sensed data (ESP) [44], spatial statistics [45], and semi-automatic optimization scheme have been used for object-oriented classification [46].Classification accuracy can be doubled by introducing these features and tools into the classification process for different complex scenes [4,[47][48][49][50][51].
Previous results showed that NDVI and the first PC band had higher importance in terms of classification accuracy than the red and NIR bands [41,52,53].One research team introduced Gaussian low-pass filtering [24], geometric features, a digital surface model [54], textural features [26], and terrain features (slope, roughness, and elevation) [27] with RapidEye and LiDAR to enhance the classification accuracy of land cover in mining areas.The grey-level co-occurrence matrix (GLCM) [54] is the most popular for describing spatial properties, particularly in VHR imagery [55].In addition, other textural measures, such as wavelet transformation [56], have been employed.Few attempts have been made to handle surface-mined landscapes in this manner, even though there are many examples of their use for urban landscapes [49].

Feature Selection Methods
The influence of the feature set embedded in the classification process on classification accuracy is sometimes greater than the effect of the classification algorithm used [58,59].Therefore, the feature selection method is very important for the hundreds of characteristic parameters derived from HR images [41].An effective feature selection method not only can eliminate redundant information in the characteristic parameters but also can obtain the optimal combination of sensitive features [50].Feature selection method has been successfully applied for optimization of features derived from HR images to find the optimal window size through the class separability parameter (coefficient of variance, visualization evaluation, and histogram distance index) [49].Maxwell employed the principal component analysis method to carry out feature selection [24,25,54].The influence of feature selection method on object-oriented classification accuracy has also been analyzed [60].
At present, according to the combination of neutron evaluation criteria and follow-up learning algorithms, the feature selection method can be divided into embedded, filter, and wrapper methods [41,61,62].The embedded method is a component embedded in the machine-learning algorithm, and the most typical is the decision tree algorithm [63][64][65].The filter method eliminates the training steps of the classifier and is therefore suitable for large-scale datasets and as a pre-selector for features [63, 65,66].The wrapper method is based on the performance of the machine-learning algorithm to evaluate the merits of the feature subset [67].Compared with the filter method, the wrapper method typically finds better feature subsets to improve classification performance [68].LiDAR data were used to automatically identify landslides in the Three Gorges reservoir area, and classification accuracy was improved by employing the wrapper method for feature selection [69,70].
A statistically significant increase in accuracy can be achieved by combining the filter feature selection method and the Random Forest (RF) algorithm in the mapping of mining and mine reclamation areas [24].Wrapper feature selection (FS) methods often provide higher performance than filter FS methods in complex landscapes [24,71].Furthermore, wrapper FS methods are advantageous for LCCMA because they reveal the importance of features, unlike other feature extraction methods [41,72].

Land Cover Classification Algorithms
We summarized the relationship between the number of samples used to train and validate the machine-learning classification algorithms and overall accuracies for different algorithms in related to literatures (Table 4).It indicated that classification accuracies were not only related to the number of training samples, but also related to the number of classes, the number of validation samples, classification algorithms, feature parameters, and remote sensing data sources.
In fact, the used classification scheme usually significantly affected the classification accuracies.Based on a preliminary classification of woody vegetation, herbaceous vegetation, barren areas, and water using NAIP imagery and SVM algorithm, the classification of mine-reclaimed grasslands and non-mining grasslands achieved an overall accuracy range of 97.2% to 97.9% using different terrain variable combinations, but with only 200 objects for training.(100 objects for each class) and 198 pixels for validation (115 and 83 pixels for those two classes) [27].Thus, it is necessary to review classification algorithms in here.
In recent years, the advantages of machine-learning classification algorithms, such as Support Vector Machine (SVM), RF, Artificial Neural Network (ANN), boosted classification and regression trees, and k-nearest neighbor (KNN) algorithms, to handle complex high-dimensional heterogeneous features have been revealed [62,68,[73][74][75][76].In addition, they boast efficient operation and high classification accuracy, and have been widely used in remote sensing classification using HR images at a regional scale [77,78].
SVM is more sensitive to feature selection and requires a large number of parameters, similar to ANN, which reduces the degree of automation in the classification process [79].RF is an integrated learning classification method and features fewer parameters and higher efficiency than ANN and SVM.RF is more capable of processing highly correlated terrain features [27,80].Typically, RF obtains better classification results as well as determines the importance of features [68,81].However, the classification accuracy of SVM was found to be higher in a comparison of the performance of object-oriented machine-learning algorithms at extracting mining land reclamation grassland information [24].Nevertheless, RF and SVM often outperform the ANN algorithm.
These machine-learning classification algorithms have been used in complex landscapes [54,82]; notably, the performances of RF, SVM, boosted classification and regression trees, and KNN have been compared in classifying surface-mined and reclamation land [24,26,27,54].Although RF and SVM often showed similar classification abilities and vary in different classification tasks, it was reported that SVM can provide more accurate classifications of surface mining and mine reclamation areas than the RF algorithm in complex surface-mined landscapes.Li et al. [41] compared RF, SVM, and ANN.They concluded that, for mapping of surface-mined and agricultural landscapes, RF achieved the greatest overall accuracy, followed by SVM and ANN.However, for classification of surface-mined land, SVM had the highest accuracy, followed by RF and ANN.Other previous studies have only used one algorithm, such as SVM [30,36], decision tree [18], or maximum likelihood [35] algorithms.
In fact, there is no consensus on the best classifier.However, research into finer-scale, more complex land cover classification has revealed how to best take advantage of the different classifiers and use integrated learning methods [68].

Some Key Scientific Issues Limiting the Development of LCCMA
The three land cover factors discussed in Section 3 restrict the development of fine-scale remote sensing information models and technical methods, and increase the difficulty of classification.Therefore, higher resolution images, texture, filtering, terrain, and other characteristics and object-oriented classification have been employed, but the results show significant room for improvement.The main issues involve the following three aspects.

The Existing Classification Representation Model of Remote Sensing Information Cannot Accurately Describe LCCMA Characteristics
The three-dimensional features of mining areas involve both positive and negative local terrain.Here, the negative local terrain mainly refers to pits.While positive local terrain mainly refers to transit sites and solid waste.They clearly differ from the positive and regular terrain features of buildings in the urban environment.These positive and negative terrain features cannot be characterized only by the geometric and spectral characteristics of HR images, and existing terrain features can only describe the positive terrain features of the mining area.In addition, most research adopted single-phase remote sensing data without any emphasis on the ability of multi-temporal data to characterize LCCMA variability.The ability of input features to distinguish LCCMA from natural surface elements is almost equivalent, which is a major reason why fine-scale classification cannot be achieved effectively.

Lack of Sensitive Feature Combinations for LCCMA
High-dimensional feature sets of remote sensing can be generated using HR images, but not all features are beneficial for improving classification accuracy, i.e., some information is "redundant".Those that can effectively improve the classification accuracy are individual features and combinations of features known as sensitive features and combinations of sensitive features, respectively in this study.However, LCCMA research focus on the classification algorithm, and no sensitive feature combination that will improve classification accuracy can be proposed, especially for mining area characteristics.

Efficient Integrated Classification Algorithm for LCCMA is Inadequate
LCCMA involves a large dataset, and a suitable classification algorithm is more capable of dealing with large amounts of data and data variability.Therefore, reasonable and efficient classification algorithm is guaranteed to improve the accuracy of LCCMA.However, while many studies have been conducted in the fields of urban environment, forest, agriculture, etc., there is a lack of algorithms specific to LCCMA.
In LCCMA research, the feature set derived from HR images often has diverse characteristics, which can be expressed in three ways.First, the feature set is different when derived from different sensors.Second, for the same remote sensing data, the multi-temporal feature set varies due to differences in phenology and mining activities.Thirdly, the combination of sensitive features is the fusion and reconstruction of multi-source and multi-temporal features, and there are obvious differences and complementarity with the above two feature sets.
According to our knowledge, there are obvious differences and complementarities for the combination of multi-source and multi-temporal feature parameters set and sensitive feature parameter combination used for LCCMA classification.Due to the diversity of these feature sets can result in different classification results, this difference needs to be considered.Therefore, it is better to develop new classification algorithms.Around this issue, we can work from two aspects.On the one hand, constructing the "diversity" in integrated learning theory is taken as the starting point, and the complementary information of multi-source multi-temporal and multi-temporal feature parameter sets and sensitive feature parameter combination are fully used to enhance the diversity through multi-feature sets, and a classification framework based on diversity of feature sets is constructed.On the other hand, an excellent heterogeneous multi-classifier is used for integrated learning, and a feature set diversity-heterogeneous multi-classifier integrated classification algorithm is established.
In general, designing a classification algorithm that can fully cope with feature set diversity is a key scientific issue.The scientific issues of LCCMA are presented in Figure 1.

Unitization of New Satellite Sensors
Over the past two decades, the spatial resolution of civil HR images has increased rapidly from 30 m to 0.3 m, and it is clearly possible to improve land cover accuracy with higher resolution data (Table 5).Moreover, over the past five years, new remote sensors have been introduced that can provide more features for users from different perspectives.For example, survey satellites Ziyuan-3 (ZY-3) and Tianhui-2 can provide simultaneous HR images and stereo images, and WorldView-3 and Sentinel-2 have a finer resolution.These satellite data, which can provide sensitive characteristic parameters for fine-scale LCCMA, can be divided into two types.Firstly, some sensors not only feature increased spatial resolution but also the number of spectral bands, such as RapidEye, Worldview-3, and WorldView-4.The red-edge band of RapidEye satellite imagery can be used to identify the type and growth state of vegetation [53,[82][83][84], which may be useful for fine-scale LCCMA.Some studies have also assessed the effect of the red-edge band on some special classifications [53,82,84].WorldView-3 (WV-3) is a newly launched (August 2014) commercial satellite sensor with high spatial resolution, eight visible to near-infrared bands (0.42 to 1.04 µm), and eight shortwave infrared (SWIR) bands (1.2 to 2.33 µm) [85].These characteristics may further benefit vegetation analysis.WV-3 significantly outperforms Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) for mineral mapping because of additional SWIR bands and the improved spatial resolution of 7.5 m from ASTER's 30 m pixel size [85].The limited release (government only) 3.7 m spatial resolution data will be even more capable.Unlike other multispectral sensors, the WV-3 instrument is suitable for detecting hydrocarbons [86].Accordingly, the WV-3 has great potential to improve the classification accuracy of LCCMA.
Secondly, some stereo surveying satellites have been successfully launched.The terrain in a mine area is extremely rugged, and topography also determines the spectral value of the remote sensing imagery, resulting in strong variability in reflectance due to direct shadows, as well as cast shadows [87].The inclusion of texture in the classification can help overcome this problem.Texture quantifies the spatial variation in image tone values of neighboring pixels [88].The results of several studies over the past seven years have revealed that integrating HR images and topographic data is indispensable.However, airborne LIDAR data, Shuttle Radar Topography Mission digital elevation models (DEM), and ASTER global DEM (GDEM) have the disadvantages of coarse spatial resolution, and earlier generation time.This is a particular issue because land cover changes significantly in time and space in surface-mined landscapes [41].These drawbacks result in an inability to meet the mapping requirements of LCCMA.As a result, several new stereo and HR satellite remote sensing sensors, such as the successfully launched ZY-3 and TianHui (TH) -1/2/3, have been developed to provide simultaneous HR multispectral (MS) bands and topographic data.These new data are beneficial for reducing the limitations of topographic data.
ZY-3 is China's first domestic civil three-line-array stereo mapping satellite.It is equipped with two front and back view TDI CCD cameras with a resolution higher than 3.5 m, and a width larger than 50 km, one TDI CCD camera with resolution higher than 2.1 m and width larger than 50 km, and one multispectral camera with a resolution higher than 5.8 m.TH-1 has a three-dimensional survey camera and a CCD camera with a ground resolution of 5 m and a camera angle of 25 • .The two stereo satellites aim to enhance the effectiveness of scientific research, mapping, and disaster relief using stereo-topographic techniques.In fact, they could provide spectral and spatial features for LCCMA.Airborne LiDAR is a new generation of aeronautical remote sensing system, integrating GPS and IMU, and can use the flight to directly generate original DEM and orthoimage data.Over the past two decades, airborne LiDAR has been widely recognized as an accurate and rapid three-dimensional measurement of the Earth's surface [4].In addition, it is now straightforward to obtain DEM by unmanned aerial vehicle data processing.Thus, it is another new option to improve classification accuracy with multitemporal, high spatial resolution information to be assimilated to optical remote sensing data [89].

Development of Effective Feature Sets
In the future, it is necessary to pay more attention to developing classification representation models of remote sensing information to expand the LCCMA theory.The advantages of unmanned aerial vehicle data and stereo mapping satellite images and hyperspectral images, such as those from ZY-3, and WV2/3, should be fully utilized.Based on the analysis of the major characteristics of LCCMA, other novel feature models for LCCMA are urgently required.

Obtaining Sensitive Feature Combinations Based on Feature Selection Method
Overall, effective feature selection procedures should have the following characteristics: they should be based on heterogeneous and complete feature sets; they must be robust (i.e., capable of overcoming data-dependence by, for example, using multiple random runs and a final feature determination method with a stable threshold); and they must not be prone to under-fitting (i.e., do not excessively exclude features).
Employed feature selection methods often discard "redundant" features that are highly correlated with the selected features, but are interdependent.This is mainly because the indicators used ignore the intrinsic nature of these feature parameters.In fact, the individual "redundant" features are not "redundant" and they can form a set of sensitive features that are more capable of identifying LCCMA when combined with other interdependent features.Thus, these feature selection methods do not provide combinations of sensitive features and are not suitable for fine-scale land cover classification in mine areas.Thus, a key scientific problem is obtaining combinations of sensitive features for LCCMA, which is also an important guarantee of improved classification algorithm performance.The existing feature selection method ignores the intrinsic structure among features.
Therefore, when a novel feature selection method is designed, especially for LCCMA, two aspects should be considered.Firstly, preference should be given to evaluation indexes such as integrating mutual information (MI) and conditional mutual information (CMI).Secondly, a feature selection method based on dynamic weight updating should be designed to ensure that all features in the combination of sensitive feature parameters are related to each other and are interdependent from each other.
Under the information theory framework, the feature selection method based on dynamic weight renewal technology should be adopted to break down the technical bottleneck of existing methods, which could result in a loss of characteristic parameters with benefits for LCCMA.From the theoretical point of view, LCCMA is useful for exploring theoretical models and technology based on HR images in complex geological environments.From the perspective of application prospects, LCCMA can improve the availability of domestic HR satellite images to obtain more detailed land cover information in mine areas.This is significant for mine management, geographical condition monitoring.

Robust and Self-Adjusted Classification Algorithms
Classification rules are mainly based on the statistical values of the spectral bands and textural features of images, and lack reasonable remote sensing classification models.Therefore, it is obviously not possible to improve the classification accuracy of LCCMA at the regional scale.Thus, some key issues should be researched when HR techniques are employed to classify LCCMA [90].For instance, it is necessary to develop and summarize the mechanism and regularity of LCCMA distribution and formation with remote sensing.
For LCCMA, there are class imbalance issues, which have been discussed in land cover classification studies across complex landscapes [91], i.e., classes relating to mining were minority samples and target classes.The class imbalance issue has resulted in low accuracies for minority classes when using custom classification algorithms.Moreover, the more advanced classification algorithms such as ensemble learning [92] and deep learning [93] have proved to be very effective.As a result, those methods should be explored and analyzed in the future.

Applications for Mine Regulation and Management
Some studies have explored the effects of mining using mining information [42,[94][95][96][97].For example, Palmer et al. [98] systematically discussed the consequences of mountaintop mining, such as ecological losses, downstream impacts, and threats to human health.Maxwell and Strager [43] assessed landform alterations induced by mountaintop mining.Becker et al. [99] examined the impacts of mountaintop mining on terrestrial ecosystem integrity.Ross et al. [42] investigated the effects of mountaintop mining on surface topography, bedrock structure, and downstream waters.Miller and Zégre [100] provided insights into the complexity of catchment hydrology in a mountaintop removal-mining region.Nippgen et al. [101] examined saline baseflows of the Appalachian Watersheds affected by mountaintop removal coal mining.Brooks [102] evaluated the impact of mountaintop mining on the watershed and regional scale nitrogen exports.Surbera and Simonton [103] investigated the disparate impacts of coal mining and reclamation in West Virginia and central Appalachia.
Nevertheless, fine scale mining information would be more effective for determining the effects of mining.For this purpose, using fine scale mining information for the assessment of various impacts should be promoted to provide more effective information for mine regulation and management.In addition, the application of fine-scale mining information can help develop the remote sensing of mining areas around the world.

Conclusions
Fine-scale LCCMA is important for a comprehensive understanding of the influence of mining activities on the regional geo-environment and for environmental impact assessment procedure at local scales around the world.This paper reviews the latest developments in fine-scale LCCMA using remote sensing techniques, focusing mainly on the proposed land cover thematic resolution, significant characteristics of MA, and the employed remote sensing data sources, features, feature selection methods, and classification algorithms.Then, some key scientific issues limiting the development of LCCMA are discussed.The conclusions are as follows: (1) the existing classification representation model of remote sensing information cannot accurately describe LCCMA characteristics; (2) there is a lack of sensitive feature combinations for LCCMA; and (3) there is a lack of efficient classification algorithms for the LCCMA.
Finally, some future research directions are proposed as follows: (1) new satellite sensors, especially those with stereo survey ability and unmanned aerial vehicle data processing, should be utilized to provide more effective sensitive features; (2) other novel feature models should be proposed to promote LCCMA based on the analysis of its major characteristics; (3) it is important to determine how to obtain sensitive feature combinations based on the feature selection method.Effective feature selection procedures should be based on heterogeneous and complete feature sets, robust, and not prone to under-fitting.In addition, a novel feature selection method was planned, especially for LCCMA characteristics of high dimensions and large data, with a preference for evaluation indexes and dynamic weight updating; (4) limiting class imbalance issues is important because classes related to mining are minority samples and target classes.Furthermore, robust and self-adjusted classification algorithms should be employed, such as ensemble learning and deep learning.(5) Lastly, the application of fine-scale mining information is key for mine regulation and management.

Figure 1 .
Figure 1.Scientific issues for the LCCMA and proposed research directions.

Table 1 .
Land cover sub-classes in mining areas.

Table 2 .
Remote sensing data sources.

Table 3 .
Main features derived from the HR images and employed for LCCMA over the past two decades.

Table 4 .
The relationship between the number of samples used to train and validate the machine-learning classification algorithms and overall accuracies for different algorithms in related to literatures.

Table 5 .
Optical remote sensors launched successfully in recent decades with a spatial resolution higher than 10 m.