The north‒south dichotomy is the most obvious geomorphic feature of Mars. The Southern Hemisphere is highly topographically variable with complex landform types, and includes impact craters, highlands, canyons, dry rivers, sand dunes, yardangs, volcanoes, and other landform types of various sizes. There are three clearly visible impact basins on Mars: the Argyre and Hellas basins in the south and the Isidis basin near the equator. Figure 1 shows the global topography of Mars. On the other hand, the Northern Hemisphere is dominated by low-relief plains. The main northern plain consists of Vastitas Borealis. The Tharsis bulge is near the equator, and its northern edge includes three volcanic regions: Olympus Mons, Alba Mons, and Tempe Terra. The presence of liquid water on Mars has been a mystery pursued by scientists. Orosei et al. (2018), through the Mars Express orbiter mission, found a 20-km-wide body of liquid water at a depth of 1.5 km below the ice cap.

The landform types mentioned above have been influenced by endogenic and exogenic forces. The geological lifetime of Mars ended when the magma activity in the interior of Mars ceased. Subsequently, the external magnetic field of Mars disappeared with the disappearance of nuclear convection. The disappearance of the magnetic field left Mars completely exposed to solar wind and cosmic high-energy rays. The combination of endogenic forces, such as early volcanic eruptions and tectonic movements, and exogenic forces, including meteorite impacts, wind accumulation and erosion, and water scouring, have resulted in the formation of a variety of landforms.

The common landforms included impact craters, volcanic landforms, glacial landforms, valley networks, and sand dunes on the surface of Mars and Earth (Wang, 2018). Typical flowing water features, such as alluvial fans and canyons, were preserved on the Martian surface, indicating the possibility of liquid water on Mars. Atmospheric movement and the high diurnal temperature differences have contributed to the formation of typical aeolian landforms, such as sand dunes and yardangs. The University of Western Ontario (Canada) initiated the Interactive Mapping of Mars (https://imars.uwo.ca/tutorial/) project, which classified Martian landforms into aeolian, water, glacial/periglacial, impact cratering, mass movement and volcanic landforms. Di et al. (2021) classified the landforms according to the causes of formation, including aeolian landforms, fluvial landforms, and tectonic landforms (referring to impact geomorphology and volcanic geomorphology). The above macroscopic landform classifications were further subdivided based on morphology and formation. OuYang and Zou, 2015 proposed classifying Martian landforms into aeolian, fluvial, canyon, impact, volcanic, and glacial landforms. This classification, combined with the traditional method of Martian landform classification, was used in this paper according to the relevant formation mechanisms, as shown in Figure 2.

In addition, the geomorphological classification systems of the Earth and Moon can aid in the geomorphological classification of Mars. The three-level and nine-class classification system of Earth (Institute of Geography Chinese Academy of Sciences, 1987; Geomorphic Map Editorial Committee of the People’s Republicof China, 2009; Zhou et al., 2009) includes the geomorphological category, geomorphological class, geomorphological shape (three levels), and macromorphological type subclass, land elevation and seafloor bathymetry subclass, maincrop force type subclass, maincrop force mode of action subclass, material composition and lithology subclass, geomorphic age subclass, combined morphological subclass, micromorphological subclass, and slope morphological subclass (nine classes). The three-level and eight-class classification system of the Moon (Cheng et al., 2018; Liu et al., 2022) does not include the maincrop force mode of action compared with the three-level and nine-class classification systems of Earth.

In terms of overall classification, Martian landform classification methods can be divided into two categories.

One category is mainly based on surface elevation, relief, slope and other features, which can be used to parameterize the surface morphology or extract surface features to classify landforms with clustering or machine learning methods. Bue and Stepinski, 2007 classified Martian landforms into highlands, impact craters, lowlands, high-relief landforms, and channels based on six topographic parameters (elevation, flood, slope, flooded slope, contributing area, and flooded contributing area), using a self-organizing mapping method and Ward clustering. However, a limitation was misclassification for certain features, such as craters. Wang et al. (2017) classified lunar landforms into high-relief, highlands, lowlands, impact craters and other landform types. Comparably, Wang et al. (2017) added relief to topographic parameters and replaced Ward clustering with ISO clustering to achieve an overall accuracy of 83.34% and a kappa coefficient of up to 0.77, despite certain limitations influenced by aggradation, degradation, and complex landform types.

Additionally, with the rapid development of machine learning, deep learning and other methods to segment the content of images at the pixel level, image segmentation has been used in classification tasks. The related steps include creating training sets, designing network model structures, and adjusting parameters, among others, for the binary/multiclass segmentation of landform elements. Shang and Barnes, 2013 used the fuzzy rough feature selection method combined with a support vector machine (SVM) to classify landform types in a single Mars image, and the comparative experimental results showed that the method outperformed decision trees and K-nearest neighbor classification. Jiang et al. (2021) developed a new end-to-end deep learning framework, rotated SSD, to localize and identify different Martian landforms at the same time by using a rotatable-based anchor box mechanism and introducing unsupervised training based on autoencoders. Barrett et al. (2022) used HiRISE images of Oxia Planum and Mawrth Vallis and developed a deep learning terrain classification system (NOAH-H) based on semantic segmentation with deep neural networks. Wright et al. (2022) used NOAH-H to classify four HiRISE images of the terrain near the 2020 Perseverance Rover landing site at Jezero based on impact crater images, and the classification results agreed with the manually created geological maps. Rothrock et al. (2016) developed an algorithm (SPOC) based on deep convolutional neural networks (CNNs) that can identify terrain elements, such as sand, bedrock, wrinkled ridges, and steep slopes, and the algorithm was successfully applied for terrain identification in the Mars 2020 rover landing zone and MSL mission sliding prediction. Wang et al. (2021) designed a new framework for Mars rover image classification with semisupervised contrast learning; it ignores intraclass pairs in labeled data and counterexample pairs in unlabeled data, thus substantially improving the classification model.

With the rapid development of computer vision, image segmentation, deep learning and other fields, Martian landform classification research has also been rapidly developed, and an increasing number of types of landform elements can now be recognized, thus providing the basis for research on landform recognition and the evolution pattern of Mars landings.

Aeolian landforms are the most prevalent and active geomorphic features on Mars, and the ancient Chinese name for Mars “Yinghuo,” is in part due to the sand phenomenon that causes Mars to be bright and pale (Dong et al., 2020b). Li et al. (2022) developed a similar visual degradation process based on the remote sensing images of “Tianwen-1” to synthesize real dust images and used these real dust images to train a deep learning model to identify dust-free images, inspired by the fog formation process on Earth. Yao, 2021 identified 882 dust storms with a diameter of 4,000 km in the southern part of Utopia based on Tianwen-1 landing area image data and analyzed their causes and distribution characteristics.

Sand dunes are the most typical manifestation of aeolian landforms. The methods for identifying sand dunes on Mars can be classified into four types.

1) Visual observation: Hayward et al. (2007) produced a database of medium and large sand dunes with areas larger than 1 km², covering 550 dunes between 65°N and 65°S; however, many dunes are still not included in the database. “Visual observation” could obtain sand dunes intuitively and qualitative information, while it is inefficient for analyzing large-scale datasets.

Machine/deep learning requires a complete training dataset. The diverse shapes (Li, 2018) and sizes (Mandt and LeoneYardang, 2015) of yardang landforms make it difficult to establish a complete training dataset or a unified identification standard. Therefore, the identification of yardang landforms is primarily performed manually. Researchers have focused on the color, morphology, and formation conditions of yardang landforms to characterize the Martian atmospheric environment and prevailing wind direction. Ward (1979), Bridges et al. (2007), and Zimbelman and Griffin, 2010 conducted analyses of the aspect ratio, scale distribution, and material properties of large-scale yardang landforms in the Martian Amazonian plain and Medusa trough layer, respectively. Additionally, some scholars have focused on dating yardang landforms (Kerber et al., 2011; Zimbelman and Scheidt, 2012; Liu et al., 2021a).

The identification of transverse aeolian ridges (TARs) is largely based on their sediment composition (Fenton et al., 2003; Balme and Bourke, 2005; Balme et al., 2008; Bourke et al., 2003) and the difference in albedo (Bouke et al., 2008; Gou et al., 2022) using visual interpretation. Lu et al. (2022) identified four crescent-shaped lateral sand ridges based on Zhurong rover exploration data and analyzed the erosion process of these beds. The orientation of these beds is related to the angle between the bed crest and the wind direction. Several typical aeolian landforms are shown in Figure 3.

Fluvial landforms have long been regarded as the best evidence for the existence of liquid water on the Martian surface, which is essential for the existence of life and an indication of the warm and humid climate that once existed on Mars. Fluvial landforms can be categorized into narrowly defined fluvial landforms and broadly defined fluvial landforms. Narrowly, Ouyang and Zou (2015) further classified the fluvial landforms of Mars into outflow channels, valley networks, and gullies according to their size scale, as shown in Figure 4. Broadly, Zhao et al. (2021) classified the fluvial landforms of Mars into valley networks, outflow channels, paleolake basins, and alluvial fans and deltas. In this article, fluvial landforms were discussed from a narrow perspective, namely, outflow channels, valley networks, and gullies. These three landforms were distinguished by their sizes, patterns, and formation mechanisms.

The outflow channels on Mars formed during catastrophic floods (Cutts and Blasius, 1981), with a width of 1 km to several hundred km. Consequently, the age of their formation was deduced to be Hesperian (Liu et al., 2021b). Outflow channels are mostly found in fracture zones or near canyons. There are distinct tear-drop islands in the channels. In addition to their size, another obvious difference between valley networks and outflow channels is dendritic branching. Valley networks are mostly found in the old southern hemisphere and are very rare in the younger northern hemisphere, with a width of several kilometers. This suggested that valley network formation occurred on early Mars (Tanaka et al., 2014; Ouyang and Zou, 2015; Liu et al., 2021b). The traditional view holds that valleys resulted from surface runoff on early Mars, which climate was warm and wet (Carr, 2006; Hynek et al., 2010). Another view holds that valleys resulted from groundwater loss or glaciation. Gullies are distributed along slopes, with a width of several meters. Their formation is controversial, and may have included melting snow or CO₂ ice. Generally, the identification of gullies is based on high-resolution data due to their small-scale features.

The methods used to identify valley networks can be divided into 1) Visual interpretation: Carr (2006) mapped the distribution of Martian valley networks and outflow channels. Hynek et al. (2010) mapped a Mars-wide network of canyons based on visible, infrared, and topographic data and analyzed the reasons for the presence of these canyons. Alemanno et al. (2018) modified the approach of Carr (2006), and the results showed that the valley network of Mars is predominantly located in the Southern Hemisphere, with small distributions in the Northern Hemisphere on the rim of the plateau and near Elysium volcano. “Visual interpretation” relies on the level of experts and costs a lot of manpower and time, which is similar with the pros and cons of visual interpretation for aeolian landforms. 2) Hydrological analysis: Stepinski and Collier (2004) proposed a method for extracting drainage networks from lower basins using a contributing area threshold, which was validated for 28 Noachian-age regions on Mars, with high efficiency and accuracy. Gou et al. (2018) extracted valley networks based on the DEM data from the Evros Vallis basin using traditional hydrological analysis methods, such as filling, flow direction analysis, catchment accumulation, and catchment area calculation. “Hydrological analysis” relies on quantitative algorithms and can provide more objective results compared to visual interpretation, reducing the potential for human bias. However, the accuracy of hydrological analysis methods depends on the validity of certain assumptions, which may not hold true in all situations. Furthermore, it is limited by the resolution of DEM. 3) Extraction of elevation change characteristics: Molly and Stepinski (2007) proposed a DEM-based method for the identification of landforms characterized by curvature and separated valleys from other landforms with convex shapes and then reconnected the segments along the drainage networks. The research about identification of gullies is rare due to the data resolution limitations (Li et al., 2015). Li et al. (2015) extracted Martian gullies in six regions using HiRISE images based on mathematical morphology, the bottom-hat transform and path opening and closing, with detection rates reaching 76%–94%. “Extraction of elevation change characteristics” is suitable for regional and global studies, while choosing appropriate parameters for elevation change analysis can be challenging.

Impact landforms include impact craters and impact basins, such as the geomorphic units shown in Figure 5. As tracers of surface processes, impact craters are informative for studying subsurface minerals and useful for dating the planetary surface (Neukum et al., 1975; Wilhelms et al., 1987; Hartmann and Neukum, 2001). Thus, related studies are important for assessing planetary geomorphology. Furthermore, impact basins are often regarded as impact craters when they are identified. Mars and the Moon are largely similar in their morphological characteristics regarding impact crater landforms, and therefore, this paper combines the progression of impact crater studies on Mars and the Moon.

The methods for identifying impact craters can be divided into manual and automatic identification methods. Manual identification aims to identify accurate boundaries based on the visual interpretation of impact craters in images, topographic data and the subsequent derived data. However, manual identification is time-consuming and laborious, relying heavily on expert knowledge of individual identifiers, and the identification standards are inconsistent. Currently, the published Mars impact crater databases include the Barlow (1988) database, MA132843GT (Salamuniccar et al., 2012; Robbins and Hynek, 2010; Robbins and Hynek, 2012) database. The Barlow database manually extracted 25,826 impact craters with diameters greater than or equal to 8 km and relatively young ages based on Viking images, of which 60% were formed by heavy bombardment. The MA132843GT catalog is based on the MA130301GT catalog (Salamunićcar et al., 2011b), which combines manually mapped impact craters and 72,668 impact craters extracted by automatic identification algorithms. The Robbins database contains 384,343 impact craters with diameters greater than or equal to 1 km, which were manually identified from images.

The automatic recognition algorithm saves time and effort compared with manual recognition, and the recognition standard is uniform. Automatic recognition algorithms can be roughly divided into four categories: 1) Traditional edge detection and circle fitting algorithms. Edge extraction is achieved by using grayscale mutation on both sides of the impact crater rim (Kim et al., 2005), such as via the Sobel operator (Lu et al., 2013), the Robert operator (Yuan et al., 2013), the Prewitt operator, the Canny algorithm (Salamunićcar et al., 2010; Jiang et al., 2013), the region growing method (Luo et al., 2014), and the Hough transform. In addition, the bright and dark areas formed by impact craters in optical images (Urbach and Stepinski, 2009) can be paired to identify impact craters. This method can be easily understood, while the results were fitting circles, instead of accurate boundaries. 2) Digital terrain analysis method. Based on the changes in terrain factors, such as elevation and slope at the rim of a crater, watershed analysis, contour lines, and the extraction of terrain features were used to identify impact craters from a three-dimensional perspective (Bue and Stepinski, 2007; Luo et al., 2013; Xie et al., 2013; Liu et al., 2017; Chen et al., 2018). This method obtained changes from elevation and variations in crater boundaries. However, it may provide low efficiency and be influenced easily by minor terrain mutations. 3) Traditional machine learning algorithms. These methods tend to first express the features of impact craters using feature descriptors, such as Haar and PHOG, and then combine them with traditional machine learning models, such as SVM and AdaBoost (Ding et al., 2013; Bandeira et al., 2014), to perform impact crater detection. The quality of the training results depends on the quality of feature selection (Liu et al., 2023). 4) Deep learning algorithms. Deep learning is data-driven and is used to delegate feature selection to a machine learning algorithm. Deep learning can be used to extract features automatically from images, which is comparable to the idea of a “black box”. Nevertheless, it is highly dependent on the quality of the data and uneasily interpretative (Hsu et al., 2021; Silburt et al., 2019; Yang et al., 2020; Zheng et al., 2020; Gao et al., 2022). Combining digital terrain analysis and deep learning may be a further development direction, which could consider geographical information. The results could provide accurate boundaries over large-scale regions.

Martian glacial landforms are closely related to the presence of water, the stable presence of which is necessary to support life. The topography of Mars has a distinct north‒south dichotomy, with high south and low north topography. The large topographic height difference redistributes water between the poles, resulting in glacial landforms in the mid-latitudes (Hepburn et al., 2020), as shown in Figure 6.

Souness and Hubbard, 2012 identified and cataloged Martian glacier-like landforms from 8058 CTX images, 1,309 of which were concentrated at mid-latitudes of 39.3°N and 40.7°S, and inferred the response mechanisms of Martian glacial landforms based on latitude and altitude. There are various types of Martian glacial landforms, including concentric crater fills (CCFs), lobed valley fills (LVFs), lobate debris aprons (LDAs), and viscous flow features (VFFs) (Levy et al., 2010; Pedersen and Head, 2010; Liu et al., 2021c). Milliken et al. (2003) used 13,000 Mars orbiting camera images distributed globally to identify 146 images containing VFFs and marked the locations of these VFFs as points. Galofre et al. (2022) combined data from CaSSIS, SHARAD, CTX, and HiRISE to explore the deformation history of tongue-shaped rocky debris slopes, their internal structure, and their influence on regional climate change from surface morphology, which can be used to infer past glaciation on Mars. Petersen et al. (2018) used SHARAD radar data to estimate the water ice content (greater than 80%) in LDAs. Furthermore, only a small fraction of cases that matched typical surface water erosion indicated extensive subglacial erosion on the Martian surface after analyzing more than 10,000 Martian valleys.

In addition to Martian glacial landforms (beyond CCFs, LVFs, LDAs, and VFFs), periglacial landforms, such as pingos (Dundas et al., 2008; Richard et al., 2021), scallop-shaped depressions (Lefort et al., 2010), polygonal terrains (Soare et al., 2021), and thermokarst depressions (Bernhardt et al., 2016), are also present. Especially, some glacial-relevant landforms were found in the mid-latitudes of the Utopia Planitia. Wang et al. (2021) applied 2D-PCA and identified potential water ice in Utopia Planitia using FP-SPR data from the Zhurong rover. Bina and Osinski, 2021 found a new landform “decameter-scale rimmed depressions” in Utopia Plantita using SHARAD, which represents a further marker for the presence of ground ice in the northern plains of Mars.

Although glacial landforms, such as VFFs, CCFs, LVFs, and LDAs, have been discussed, studies have focused more on the formation, evolution, material composition, and structure of the above landforms and less on their identification. Additionally, the relationship of periglacial and glacial landforms between Mars climate evolution has been a focus of glacial landform research.

Topographic data are characterized without the limitation of optical light and darkness and reflect the real terrain. Global-scale Martian data are usually characterized by low resolution, and localized data have a high resolution but a limited range. In low-resolution data, some fine-textured geomorphic units, such as gullies, miniature yardangs, and secondary impact craters, cannot be identified. Data resolution is an important factor affecting the recognition accuracy. However, the current Mars global DEM data resolution is 463 m/pixel (the blended DEM from MOLA and HRSC data resolution has a resolution of 200 m/pixel), which is not sufficient for the recognition and analysis of fine features. The production of higher-resolution global-scale Mars topographic data would be beneficial for the identification of finer geomorphic units.

In Mars geomorphology mapping research, some specific geomorphological units and Mars geological maps have been produced, but there is a lack of global geomorphological regionalization maps for Mars. Therefore, the digital geomorphological regionalization methods of the Earth and Moon could be used as references to draw the Mars geomorphological regionalization classification map based on the topographic indexes of Mars, such as relief, slope and elevation. This approach could aid in exploring the spatial differentiation characteristics of the Mars surface evolution process.

Mars has landforms that are typical of both the Earth and the Moon. The “three classes and nine levels” classification system for Earth’s landforms and the “three classes and eight levels” classification system for lunar landforms are useful for establishing a complete classification system of Martian landforms. However, the applicability of these classification systems to Martian landforms needs to be further explored. At present, studies on impact craters (Lei, 2017), sand dunes (Li et al., 2020), and yardangs (Liu, 2021) have established a detailed morphological index system, but index systems for describing other geomorphic types has been less studied. Most of the existing geomorphological classification studies (Bue and Stepinski, 2006; Wang et al., 2017; Deng et al., 2022; Liu et al., 2022) have divided geomorphological units into subtypes based on morphology. However, there have been few studies of descriptive indexes reflecting the formation or material composition. Combining the cause of formation and morphology is the key to geomorphological classification (Shen et al., 1982; Cheng et al., 2011). Combined with multisource remote sensing data, including visible images, thermal infrared images and topographic data, descriptive indexes considering the formation, material composition and geological age was introduced to build a complete index system and a database of corresponding geomorphic units. The construction of an index system for Martian geomorphic classification needs to be further explored.

In terms of recognizable objects, automatic recognition has been rapidly developed in the field of impact crater recognition; however, to recognize more complex geomorphic units, such as sand dunes, yardangs, canyons, valley networks, gullies and degraded impact craters, manual recognition is still the main identification method. Manual identification could represent the real morphology accurately, but it is time-consuming and laborious and depends on the professional knowledge level of the experts. That results in heterogeneous mapping standards and heterogeneous mapping quality. Therefore, in research on the recognition of geomorphological units with complex morphology, automatic recognition methods need to be studied further.

In terms of the applicable scope of the recognition methods, most existing terrain classification methods are applicable to small-scale or local-scale terrain classification, while there are fewer automatic recognition methods applicable to large-scale or even global-scale classification. Although large-scale catalogs of impact craters (Robbins and Hynek, 2012), sand dunes (Hayward et al., 2007), yardangs (Liu, 2021), valley networks (Alemanno et al., 2018), and glacial-like landforms (Souness and Hubbard, 2012) were proposed, these features were identified manually rather than automatically. Because the accuracy, efficiency and robustness of these automatic methods for the global-scale regions were unsatisfactory, the automatic methods were typically applied in local regions. The methods that are applicable only for local small-scale target recognition lose the essential advantages of automatic recognition methods. Therefore, automatic recognition methods that are applicable at large and global scales require further research.

In terms of the requirements of recognition methods based on the users’ level of knowledge, the current Mars morphology recognition are mostly based on traditional machine learning, with a lack of end-to-end deep learning methods. On the one hand, the effect of machine learning depends on feature selection (Guyon and Elisseeff, 2003; Domingos, 2012), which requires a high level of expertise for the user. Deep learning is data-driven with the ability of automatic representation learning (Qian et al., 2022; Chen et al., 2023). On the other hand, traditional machine learning provides interpretable models, while deep learning is often seen as a “black box”. It is difficult to interpret the modeling process. Furthermore, deep learning requires more computational resources and larger training datasets than machine learning. Currently, end-to-end deep learning models are widely used in various professional fields, and it is hoped that they can be applied to intelligent terrain type recognition in the case of Mars. Broadly applicable training datasets and well-performing models are expected to be developed to identify multiple Martian terrain units.