The graphical construction method primarily employs human-computer interactive visual interpretation, a technique that combines expert observation with computer-based image processing to enhance accuracy and efficiency in geological hazard identification, particularly for landslides. This approach effectively leverages human expertise alongside the computational power of modern image processing tools. Unlike traditional visual interpretation methods, it integrates real-time analysis software, such as GIS, which provides immediate statistical feedback on identified results. This capability allows operators to monitor overarching landslide trends dynamically during the identification process. Recent advancements in machine learning-based image recognition have further supported landslide detection (Yang and Xu, 2022; Saha et al., 2024; Sharma et al., 2024; Yang et al., 2024). However, compared to these machine learning techniques, human-computer interactive visual interpretation retains a key advantage: the incorporation of expert judgment. This method enables users to interact with the system, guiding it to refine identification parameters for greater accuracy. Additionally, it facilitates deeper insights by allowing experts to interpret and expand on computer-generated data. This iterative feedback loop between expertise and computational analysis significantly enhances both precision and efficiency.

This work primarily utilizes high-resolution, three-dimensional optical remote sensing imagery provided by the Google Earth Pro platform. The satellite imagery is an integration of multisource remote sensing data, including SPOT5 (2.5 m resolution imagery), QuickBird commercial satellite (0.6 m resolution), IKONOS (1 m resolution), Landsat8, WorldView-1 and WorldView-2 satellites (0.5 m resolution), WorldView-3 (0.3 m resolution), WorldView-4 (panchromatic resolution 0.3 m), and GeoEye-1 (0.5 m resolution). Google Earth continuously expands its imagery database and employs advanced data-mining techniques to reduce the effects of cloud cover and atmospheric interference, thereby improving image clarity and usability for analysis. This platform enables multi-angle, comprehensive observation of regional terrain features and landform characteristics (Yu et al., 2024), providing advantageous conditions for human-computer interactive visual interpretation (Yu et al., 2022). In this study, the research area is defined and divided into multiple sub-regions to ensure no areas are missed during the interpretation process. Occasionally, cloud cover obscures some sections; however, Google Earth Pro’s historical imagery function allows us to review these regions over time, enabling more accurate and complete landslide identification across the entire study area.

Landslide identification primarily depends on human visual judgment, requiring personnel to have specialized knowledge of landslide characteristics and assessment criteria. The process relies on identifying discrepancies in color, shape, and texture between the landslide mass and the surrounding geological context, such as landforms and rivers. Key morphological features, including the back scarp, perimeter, and accumulation body, serve as fundamental criteria. Special attention is given to areas with abrupt topographic changes, where regions showing landslide characteristics are accurately delineated using vector polygons. A fully developed landslide should include the following components: the landslide mass, landslide bed, slip surface, back scarp, landslide tongue, landslide steps, and landslide depression, as shown in Table 1. However, not all landslides possess all of these features; nonetheless, the landslide mass and back scarp are present in all cases.

The direct interpretation indicators of landslides primarily focus on the characteristics of the landslide itself in remote sensing images, such as shape, tone, and texture. Shape characteristics: Due to the downward movement of the landslide body, the terrain in the three directions (left, right, and rear) of the landslide tends to be slightly higher, giving it an overall shape resembling a horseshoe, circular chair, bullhorn, or tongue, with the rear wall opening towards the slope base. Tone characteristics: Newly occurred landslides often appear in light tones such as grayish-white or bluish-white due to the destruction of surface vegetation and soil fragmentation. The tone distribution is uneven. Landslide scars tend to appear lighter in tone because they reflect more light, while landslide depressions may appear darker, especially when water accumulates. For older landslides, the recovery of surface vegetation diminishes these color features, but they can still be differentiated from the surrounding tones. Texture characteristics: The original stratigraphic integrity is disrupted, resulting in exposed soil, overturned vegetation, and a fragmented surface. This leads to a rough texture with patchy shadow effects visible in the imagery. The indirect interpretation indicators of landslides primarily focus on environmental factors around the landslide, such as vegetation distribution, topography, geological structure, hydrological information, and ecological landscapes. Vegetation characteristics: For slow-moving or old landslides, the continuous downward movement of the landslide body, combined with the upward growth of trees, results in phenomena like “scythe trees” and “drunken forest” on the landslide surface, which are particularly evident in high-resolution aerial imagery. Hydrological characteristics: Irregular water system patterns on the landslide body, sudden changes in river flow directions at the base of the slope, or narrowing of river channels can indirectly indicate the presence of a landslide. Topographical features: Poor continuity of the landform often results in unique “steep slope + gentle slope” landforms, and the area below the landslide body may exhibit uneven terrain due to the pressure exerted by the sliding mass.

Frequent operations during the identification process may lead to geometric self-intersection issues. Although apparent errors can often be detected manually, smaller discrepancies may evade visual inspection, making algorithmic identification necessary. Unresolved self-intersection issues can hinder the conversion of features into the required GIS format, causing complications in subsequent analyses. Verification is therefore essential after data construction to ensure database integrity and accuracy. This involves using topology checking tools in GIS software to detect self-intersections in polygon features. Identified geometric issues are then corrected to maintain data quality.

Collected information is structured into a database, where each landslide point corresponds to a unique attribute record, ensuring precise matching between graphical and attribute data. This includes details such as location, area size, geometric perimeter, and associated geographical factors like elevation, slope, curvature, lithology, and proximity to faults. After data entry, the attribute data undergoes verification and correction to ensure accuracy and completeness. Different experts independently interpret landslide areas using identical satellite imagery and topographic data, recording key characteristics such as location, area size, and boundaries. Their results are compared to calculate consistency indices that quantify the accuracy and reliability of interpretations. Discrepancies are collectively reviewed to identify error sources and refine unified interpretation standards. The final database is stored in shapefile format, comprising the main (.shp), index (.shx), and attribute (.dbf) files, which enable standardized management of geological disaster data. Statistical analysis of attribute data reveals patterns and characteristics of disaster occurrences. For instance, analyzing disaster frequency and regional distribution helps identify high-risk areas and temporal-spatial patterns, providing critical support for disaster early warning and prevention.

Based on the high-resolution optical remote sensing images provided by the Google Earth Pro platform, a detailed interpretation of landslides in the Zhenxiong County area (covering 3,696 km²) was conducted using a human-computer interaction visual interpretation method. A total of 3,979 landslides were identified, encompassing a combined area of 319.20 km². The largest landslide, measuring 4.55 km², represents a significant ancient slide that diverted a river by filling a valley. In contrast, the smallest landslide covered just 1,779 m², while the average landslide area across the study area was 80,215 m² (Figure 4A). Statistical analysis revealed that there are 72 landslides larger than 0.5 km², accounting for 1.81% of the total number of landslides, with 15 landslides exceeding 1 km². Additionally, there are 758 landslides with areas between 0.10 km² and 0.5 km², 880 landslides between 0.05 km² and 0.10 km², and 2,046 landslides ranging from 0.01 km² to 0.05 km². Only 223 landslides have an area smaller than 0.01 km², as illustrated in Figure 4B. Landslides were categorized by area into small (<10,000 m²), medium (10,000 m²–100,000 m²), and large (>100,000 m²) landslides. It was found that the vast majority (73.54%) of landslides in Zhenxiong County are medium-sized landslides, followed by large landslides, which account for 20.86% of the total number. Small landslides constitute only 5.60% of the total landslide count. Several factors contribute to the prevalence of larger landslides: 1) The morphological features and geomorphology of large landslides are more pronounced, making them easier to identify; 2) Smaller landslides tend to lose their characteristics over time due to erosion, making them difficult or even impossible to recognize; 3) Smaller landslides are more susceptible to vegetation cover, which hampers identification. To gain a deeper understanding of the development of landslide relics in Zhenxiong County, GIS software was utilized to calculate the landslide area percentage (LAP) and landslide point density (LND) across the entire study area. The results showed that LAP and LND were 8.64% and 1.077 landslides per km², respectively, indicating a significant development of both the number and area of landslides in the county.

Overlaying the identified landslides on the elevation map reveals that most are distributed between 1,000 and 2,000 m in elevation, as shown in Figure 5. Statistical analysis of the landslide distribution based on geographic coordinates indicates that landslide density is significantly higher between 104°17′E and 104°47′E compared to other longitudinal areas. Similarly, in the latitudinal range of 27°40′N to 27°25′N, landslide density is notably higher than in other latitudes. Consequently, landslides in Zhenxiong County are primarily concentrated in the northwestern and southwestern regions. To clearly identify areas with higher landslide densities, we used the kernel density tool in GIS software to calculate landslide point density, setting the search radius to 5 km. As shown in Figure 6, the highest density reaches 3.38 landslides per km². The maximum density is concentrated in four specific regions: the confluence of the Baishui River and Yanxi River in the northwest, the Dahuigou Reservoir, the valleys along both sides of the river from Heitang Village to Hongyan Village, and Xiaogou Village. These areas feature well-developed drainage systems and mid-to-low mountainous terrain, with the western canyons being particularly significant for landslide occurrences. In contrast, the eastern part of Zhenxiong County has relatively flat terrain, with a more uniform landslide distribution. The area around Hongjiayuanzi Village shows a concentrated landslide distribution, where higher elevations and significant topographic variations (with a maximum elevation of 2,300 m) make the terrain more susceptible to geological factors contributing to landslides.

This database includes several typical landslide morphologies, such as “hoop chair,” tongue-like, oval, and “shovel” shapes. Geomorphologically, the features typically exhibit dual ditches with a common source, cracks and cliff faces at the back of the landslide, a distinct boundary between the landslide mass and surrounding mountains, steep steps or benches on the landslide body, and landslide deposits that obstruct rivers, causing unusual river diversions. Most of the landslide relics are ancient, having undergone long-term geological evolution that often modifies them, making identification challenging. However, the boundaries of the landslides are usually clearly visible, and the deposits are distinctly marked, with color differences that set them apart from the surrounding vegetation and terrain. As shown in Figure 7, the thick white dashed line indicates the overall boundary of the landslide, the white arrows denote the rear edge where material has slid down from the highest point, the yellow arrow shows the direction of landslide movement, and the thin white dashed line represents the landslide deposits.

Figure 7A illustrates a typical landslide located in Zhangzhai Village, covering an area of 0.31 km². As shown in the figure, the landslide mass has slid down from the southern mountain, forming deposits below. Upon closer inspection, the length of the deposits on the left side of the landslide is notably greater than on the right side. This discrepancy may be attributed to the heterogeneous nature of the rock masses on the left and right sides during the sliding process. Alternatively, it is possible that the left half of the landslide experienced a secondary sliding event after some time, resulting in a larger deposit area on that side. This has led to a noticeable anomaly in the valley’s orientation. Given the long time since the event, settlements have been constructed on the landslide mass. Figure 7B shows another typical landslide located on the southern mountain of Zhongzhai Village, covering 0.09 km². The landslide has an elliptical shape, with a clear boundary between the landslide scarp and the landslide body, as well as a distinct demarcation between the landslide deposits and the surrounding environment. Similar to the previous example, this landslide has caused an abnormal valley orientation. Over time, human activity has resulted in the creation of terraced fields on the landslide mass. Figure 7C depicts a typical landslide situated on the south bank of the Huangshui River, with an area of 0.08 km². This landslide is narrow at the top and widens at the bottom, sliding down the mountain at an inclined angle. Moreover, based on the vegetation on the southern bank of the river below, it can be inferred that a small section on the left side of the landslide deposits has also undergone secondary sliding, indicating the instability of the surrounding mountains and their susceptibility to future landslides. Figure 7D depicts a typical landslide located in Fengyan Village, covering an area of 0.30 km². This landslide slid northwest from the eastern side of Fengyan Village, with its rear edge still visible. Due to the long time elapsed since the landslide event, settlements have been built on the landslide mass. The landslide has created a significant elevation difference from east to west on the mountain, causing the terraces built by humans to display a discontinuous topography stretching from northeast to southwest. The original southwest-northeast oriented valley was disrupted by the landslide deposits, resulting in the valley shifting approximately 180 m to the northwest. Figure 7E features another typical landslide in Sunjiagou Village, covering an area of 1.21 km², classified as a large landslide. Erosion gullies have developed on both sides of the landslide, demonstrating the typical dual-ditch morphology with the same source. Due to the considerable size of the landslide, numerous settlements and terraced fields have developed on and around the landslide mass. Lastly, Figure 7F illustrates a landslide that occurred on 11 January 2013, in Zhaojiagou Village, with an area of 0.91 km². The source, sliding area, and slope morphology of the landslide exhibit a zigzag shape (indicated by the yellow arrows), resembling “boot-shaped terrain.” The overall slope of the rear edge ranges from approximately 50°–90°, with the ridge consisting of steep limestone cliffs at an elevation of about 1800–2000 m. Beneath the cliffs lies a gently sloping “bulge” with an elevation of around 1,690–1800 m. Additionally, a smaller landslide, located southeast of the primary landslide, is enclosed by the blue dashed line in Figure 7F, with a height difference of approximately 151 m between the source and deposit areas.

In the past decade, landslide identification technology has advanced from traditional field geological survey methods (Wei et al., 2010) to semi-automatic recognition through human-computer interaction, and more recently, to fully automated recognition using machine learning algorithms (Van Den Eeckhaut et al., 2012; Moosavi et al., 2014; Wang Y. et al., 2024). These new landslide identification methods offer several advantages over traditional techniques, such as faster processing and lower costs (Xun et al., 2019; Pang et al., 2022). However, despite these advances, landslide relic identification still primarily relies on semi-automated methods, particularly human-computer interaction visual interpretation. Machine learning technologies have become widely applied in landslide identification due to continuous improvements in algorithm performance (Wang et al., 2023; Bhuvaneswari et al., 2024; Yang et al., 2024). Nonetheless, the accuracy of identification remains inconsistent (Moosavi et al., 2014). For example, when geological environment data are used as training samples, machine learning algorithms may mistakenly classify non-landslide areas as landslides due to similarities in environmental features, which reduces identification accuracy. Additionally, automatic recognition technologies based on image or pixel comparison may merge multiple adjacent landslides into a single large landslide, compromising the accuracy of area size and impact assessments. Furthermore, landslide recognition methods based on visible remote sensing imagery can erroneously identify cultivated land or deforested areas as landslides due to color differences (Li C. et al., 2024). This issue is especially problematic for ancient landslides, whose characteristics may have gradually faded due to vegetation changes and human engineering activities, significantly reducing the effectiveness of automatic landslide identification.

Human-computer interaction visual interpretation technology can partially compensate for the limitations of automatic identification methods, offering significant advantages in the accuracy and completeness of landslide relic identification (Li et al., 2021; Huang et al., 2023a; Wang W. et al., 2024). However, this technology also faces several challenges. For example, optical remote sensing relies on favorable optical conditions, making it difficult to capture clear surface images in foggy or cloudy weather. Additionally, this method requires human experts to have substantial geological and geomorphological knowledge to effectively guide the system’s analysis and decision-making. Furthermore, current technology cannot identify landslides in the initial sliding stage or those experiencing minor deformations, requiring the integration of other techniques, such as InSAR, for more comprehensive identification and analysis (Antonielli et al., 2019; Li N. et al., 2024). Moreover, this identification method still incurs significant labor and time costs. In terms of objectivity, past experiences often necessitate field surveys for validation. However, the areas accessible to humans and the perspectives available during on-site investigations are frequently limited. To address this, researchers often use small devices like drones for observation (Yavuz et al., 2023), which offers advantages similar to satellite imagery. As a result, the application of human-computer interaction visual interpretation on satellite images is nearly indistinguishable from field surveys, and this method has been validated in other studies (Li et al., 2021; Liu et al., 2023; Wang W. et al., 2024; Zhang et al., 2024), fully meeting the requirements for identifying landslide relics. However, enhancing the precision of landslide recognition while maximizing automation remains an area of ongoing research. With the advancement of deep learning technology, future landslide identification techniques will likely increasingly rely on artificial intelligence algorithms, such as Convolutional Neural Networks (CNN) and Generative Adversarial Networks (GAN), to improve both the automation and accuracy of identification. As the quality of landslide data continues to improve, the accuracy of automatic landslide recognition will also increase. This highlights the importance of high-quality basic landslide data, suggesting that future automatic landslide recognition technologies and the quality of existing data will be mutually reinforcing.

A highly accurate, complete, and detailed landslide data inventory is playing an increasingly important role in the field of landslide geological hazard research. Firstly, the establishment of the landslide inventory will fill the gap in the basic data on landslide disasters in the study area, providing solid data support for disaster prediction and risk assessment. The landslide inventory is a core foundational dataset for landslide disaster management. It includes key information such as the location and size of landslides, providing a reliable basis for governments and relevant departments to develop precise disaster prevention and reduction strategies. For example, the inventory can help identify high-frequency landslide areas and potential hazard zones, supporting disaster risk zonation and management.

Secondly, the landslide inventory provides essential parameter inputs for landslide susceptibility assessment. Based on the landslide inventory, regional landslide susceptibility models can be developed, especially as recent studies increasingly focus on using landslide databases to establish regional landslide susceptibility, hazard assessment, and risk evaluation (Miao et al., 2023; Abdo and Richi, 2024; Chicas et al., 2024; Guo et al., 2024; Kassa, 2024; Kaur et al., 2024). Such studies require substantial data as the foundational basis for model development (Huang et al., 2024; Ma et al., 2024a; Ma et al., 2024b; Shao et al., 2024a; Sun et al., 2024a; Wu et al., 2024), particularly for training samples in machine learning algorithms (Tang et al., 2023; Zhuo et al., 2023; Singh et al., 2024). These data are crucial for determining the reliability and accuracy of the models. For instance, due to the inability to obtain a complete landslide database for the high-altitude regions of the Himalayas, Du et al. developed a quantitative method that combines heuristic and multi-class statistical models to assess landslide susceptibility in areas with incomplete inventory data and high uncertainty in landslide interpretation (Du et al., 2020). While this method somewhat mitigates the issue of sparse landslide data, it still faces challenges in verifying model accuracy. With a relatively complete landslide inventory, there would be enough samples to validate the accuracy of model methods and further enhance model precision. Sahrane et al. found that studying landslide susceptibility in homogeneous and heterogeneous environments requires the use of different datasets (Sahrane et al., 2023). Landslide inventories with limited data may be reliable in monotonous and repetitive areas, but they often prove unreliable in regions with significant geological and geomorphological diversity (Fu et al., 2020). In contrast, this study effectively addresses the issue of inaccurate risk assessment models by conducting detailed identification of landslide geological hazards in Zhenxiong County.

Finally, the establishment of the landslide inventory provides a data foundation for optimizing monitoring and early warning systems. The inventory data enables the identification of key monitoring areas, optimization of monitoring point layouts, and improvement in the accuracy and efficiency of disaster monitoring. For example, deploying comprehensive monitoring equipment such as surface displacement sensors, rain gauges, and groundwater level meters in high-risk landslide zones can significantly enhance early warning capabilities. Moreover, the inventory can serve as calibration data for landslide simulations. By analyzing historical landslide events, it helps improve the accuracy and reliability of numerical simulations, supporting research on dynamic evolution and triggering mechanisms of landslides.

Research on regional landslide disasters has been increasingly prevalent, leading to the establishment of numerous landslide inventories across various regions (Shen et al., 2023; Rüther et al., 2024; Shi et al., 2024). Most of these inventories have been created for the purpose of training machine-learning models or investigating landslide disasters in specific scenarios (Gao et al., 2024; Li et al., 2024c; Yingze et al., 2024). As a result, the completeness of the landslide data in these inventories may not be fully representative of the areas in question. Table 2 presents previous studies related to landslide disasters in the vicinity of Zhenxiong County, all of which include landslide relic inventories. To evaluate the completeness and detail of these inventories, the authors used landslide density (i.e., the number of landslides per unit area) as a metric. Since Zhenxiong County is located in the northeastern part of Yunnan Province, the sources of these studies were selectively drawn from this region whenever possible. In total, the authors reviewed ten research outcomes, nine of which were conducted within Yunnan Province, with two covering the entire province. Additionally, one study was from Bijie City in neighboring Guizhou Province, which borders Zhenxiong. This approach ensures the comparability of the landslide inventories.

A comparative analysis of landslide inventories from Zhenxiong County and surrounding areas reveals deficiencies in detail and completeness in inventories from other regions. These deficiencies are mainly reflected in the following aspects: (1) Differences in the purposes of landslide inventory compilation have led to varying levels of data completeness. Some landslide inventories were created primarily for machine learning training or specific geological phenomena studies (Gao and Wang, 2016; Ji et al., 2020; Cheng et al., 2021). In such cases, data collection often emphasizes the representativeness of landslide features rather than the comprehensiveness of landslide events. (2) The scope of the study area influences the detail of landslide records. Certain studies encompass broad areas, which limits detailed records of landslide disasters within smaller, specific areas. Compared with Wang et al. and Wu’s research, inventories covering the entire Yunnan Province provide broad coverage but often overlook landslide events in localized areas like Zhenxiong County, thus failing to fully capture landslide distribution and frequency in such regions (Wang et al., 2014; Wu, 2015). (3) The methods of landslide data collection also impact inventory detail. Studies that incorporate high-resolution satellite imagery and drone data tend to achieve more comprehensive landslide information compared to those relying solely on field surveys or historical records. (4) Variations in inventory standards and data processing approaches lead to discrepancies. Some studies apply differing landslide definitions or data filtering methods, resulting in biases in landslide density calculations. For instance, certain inventories record only large-scale or high-impact landslides, omitting smaller or non-lethal events. These landslide lists cannot fully reflect the actual situation of regional landslides, especially in mountainous areas with frequent landslides but small scale. This study aims to achieve a comprehensive identification of historical landslide relics to accurately reflect the landslide hazards in the study area. Consequently, the landslide inventory presented here shows a higher density (1.077 landslides per km²) than those in previous studies.

Zhenxiong County, located in northeastern Yunnan Province, is characterized by a complex geological environment and frequently experiences landslide disasters. The severe landslide event in Liangshui Village, Tangfang Town, on 22 January 2024, has underscored the urgent need for landslide research and early warning systems in the region. This study primarily focuses on establishing a comprehensive and accurate inventory of landslide relics within Zhenxiong County and provides a preliminary analysis of landslide size and spatial distribution. Moving forward, we plan to conduct a more detailed analysis of landslide distribution in relation to various environmental factors, including elevation, slope, aspect, proximity to rivers, and lithology. Based on this understanding of landslide spatial distribution, the study will then assess landslide susceptibility across Zhenxiong County. Additionally, by incorporating local rainfall and seismic activity data, we will conduct an analysis of landslide hazards to develop a comprehensive risk assessment model. This model aims to evaluate the potential risks of landslides in the study area, providing critical technical support for the prevention and mitigation of regional landslide disasters.