The different approaches used to assess the relationship between land use and cover change (LUCC) and landslides possess unique advantages and significant limitations that should be considered.

First, data-driven methods in the study of LUCC and landslide relationships have undergone significant evolution, from simple linear analysis to complex nonlinear models. This reflects a paradigm shift towards data-intensive and computationally intensive approaches. Before 2019, classical statistical methods dominated the field. For example, univariate analysis (Wasowski et al., 2010), spatial statistical analysis (Xin et al., 2023), and frequency ratio (Sangeeta and Singh, 2023) were commonly used to assign susceptibility weights based on the frequency distribution of landslide occurrences. However, these methods were not strictly based on probability distributions. Although methods like logistic regression and correlation analysis provided statistical explanations, they had limitations in handling non-linearity and data dependencies (Ciurleo et al., 2017; Merghadi et al., 2020).

With the popularization of machine learning, decision trees and neural networks have played a greater role in enhancing nonlinear fitting and geospatial pattern recognition (Ij, 2018). However, it is essential to note some limitations when using machine learning models. For example, in small-sample scenarios, random forests are prone to overfitting due to imbalanced feature space dimensions, while regularization parameter optimization in gradient boosting trees (e.g., XGBoost) relies on manual tuning experience (Merghadi et al., 2020). In terms of interpretability, the "black-box" nature of traditional neural networks makes it difficult to reveal the specific impact pathways of LUCC factors (e.g., vegetation fragmentation, construction land expansion) on landslide-prone environments. Emerging explainability tools in recent years have offered breakthroughs in this regard. For example, SHAP (SHapley Additive exPlanations) quantifies feature contributions through game theory principles, enabling visual representation of the marginal impacts of different LUCC types on landslide susceptibility (Pradhan et al., 2023). Regarding the scientific validity of model validation, studies need to strengthen stratified temporal cross-validation strategies. In small-sample regions, 5-fold stratified sampling (Stratified k-fold CV) should be employed to ensure that the proportion of landslide samples in training and test sets matches the actual geographic distribution (Roy and Saha, 2022). Combined with transfer learning techniques, these methods can effectively alleviate the data scarcity issue. For instance, pre-trained ResNet models can be employed to transfer remote sensing image features, while generative adversarial networks (GANs) can be used to augment simulated landslide samples (Al-Najjar and Pradhan, 2021).

In the context of cutting-edge technology integration, Transformer architecture and reinforcement learning (RL) are driving the field into an intelligent phase. The former captures complex dependency relationships in long-term temporal remote sensing data through self-attention mechanisms, having demonstrated exceptional spatio-temporal feature modeling capabilities in fields such as vegetation dynamics evolution and urban expansion simulation (Li et al., 2024). Its technical logic can be transferred to the analysis of lag effects in LUCC-induced landslides -- for example, decoding the temporal correlation patterns between vegetation coverage decline and slope stability changes. RL optimizes remote sensing image classification strategies through a "state-action-reward" mechanism, enabling adaptive adjustment of spectral feature combinations in LUCC monitoring (Liu et al., 2024b). This provides a methodological reference for multi-source data fusion in LUCC and landslide research.

Secondly, physics-driven methods, based on the physical principles of landslides and mechanical models, assess landslide susceptibility by simulating geological processes. These methods can offer the highest prediction accuracy and are suitable for mapping and analysis in local or small-scale areas. Although physics-driven methods are theoretically reliable, only a small percentage of studies (approximately 14%) have employed them to investigate landslide susceptibility. The limited research scope, difficulty in data acquisition, and computational complexity are key factors hindering the widespread use of physics-based models for large-scale landslide susceptibility mapping (Guo et al., 2023; Yong et al., 2022).

The combination of both physics-driven and data-driven methods is a cutting-edge trend in Earth sciences. However, this hybrid approach is still rare in LUCC-landslide research (approximately 5%) and is typically limited to a simple overlay of physical models with data-driven approaches. For instance, Van Beek and Van Asch (2004) integrated a transient distributed hydrological model with a slope stability model using spatial statistical analysis to evaluate landslide susceptibility. Another approach involves a hierarchical relationship, where landslide susceptibility zones are initially determined using the analytic hierarchy process (AHP), and then a Scoops 3D model is used for finer-scale assessment (Rashid et al., 2020). In theory, hybrid approaches combine the theoretical foundation of physical models with the efficient processing capabilities of data-driven methods, offering more accurate and comprehensive landslide susceptibility assessments (Tehrani et al., 2022; Yang et al., 2024). For example, compared with traditional methods of randomly selecting negative samples, a negative sample extraction method based on physical models significantly improves the accuracy and reliability of the model (Liu et al., 2024a), demonstrating the potential of hybrid-driven methods in landslide susceptibility analysis. However, there is still a long way to go in applying these methods in LUCC-landslide studies. A critical challenge lies in how to incorporate LUCC parameters into physical models. LUCC parameterization faces difficulties related to the complexity of physical models, the dynamic nature of LUCC, differences in temporal and spatial scales, the sensitivity of physical model parameters, model adaptability, and a lack of effective coupling methods (Michetti and Zampieri, 2014). To address these challenges, it is necessary to develop multi-scale, multi-process coupled models, improve model flexibility and adaptability, and refine the acquisition of multi-source heterogeneous data.

As shown in Fig. 6, the impact of different LUCC types on landslide susceptibility varies. The main reasons for this variation include the following.

(1) Uncertainty in the impact of land use conversions on susceptibility. For example, the effects of converting "shrubland to grassland" and "shrubland to forest" on landslide susceptibility are quite different. Shrublands are typically more developed than those of herbaceous plants, allowing them to penetrate deep into the soil and form a complex network that enhances soil shear strength and erosion resistance (Lann et al., 2024; Löbmann et al., 2020). However, this effect is usually intermediate between grassland and forest. When shrubland is converted to grassland, the reduction in shallow roots can decrease the stability of surface soils (Caviezel et al., 2014). In contrast, the conversion of shrubland to forest is generally considered to enhance slope stability, but this process is gradual (Manchado et al., 2022). However, there is currently a lack of in-depth discussion on the relationship between land use transitions and landslide susceptibility. As humans rapidly transform the Earth's surface, land use can change in a short period (Ren et al., 2019). In this context, it is crucial to conduct further research on the impact mechanisms of land use transitions on landslide behavior for accurate susceptibility assessment and prediction.

(2) LUCC's impact on landslide susceptibility is not isolated but interacts with other factors such as topography, geology, climate, and soil (Pham et al., 2022; Sur et al., 2021; Vuillez et al., 2018). For example, wind force applied to tree canopies increases the shear stress on the slope. At the same time, slope steepness increases the downward shear stress on the trees, and when combined with rainfall, it leads to the erosion of the soil, ultimately resulting in landslides (Lann et al., 2024; Pawlik, 2013). As climate change intensifies, the synergistic effects of altered precipitation patterns, extreme weather events, and LUCC on landslide risk are gradually increasing. For example, in regions with alternating extreme drought and heavy rainfall (e.g., Brazil, China, Iran, etc), overgrazing or cropland expansion leads to grassland degradation, reducing the soil-stabilizing capacity of vegetation roots. The development of soil cracks during drought periods and infiltration erosion during subsequent heavy rains creates a "dry-wet cycle" stress, triggering abrupt changes in the mechanical properties of slope soil and initiating large-scale shallow landslides (Lian et al., 2022). In areas where the frequency and intensity of heavy rainfall are increasing, such as the Alps, the Himalayas, and other regions, shallow landslides, rockfalls, debris flows, and avalanches are also gradually increasing (Gariano and Guzzetti, 2016). In the tropical rainforest regions of Southeast Asia, large-scale deforestation has led to increased surface runoff and reduced soil shear strength. When combined with short-duration heavy rainfall brought by tropical cyclones, the landslide risk has significantly increased compared to the original forest areas (Lehmann et al., 2019). Therefore, in the assessment of landslide susceptibility, it is necessary to construct a differentiated analysis framework based on regional characteristics to reduce evaluation biases caused by differences in regional mechanisms.

In landslide susceptibility assessments, it is crucial to consider the interaction between LUCC and other factors -- either through qualitative methods (e.g., sensitive factor models or qualitative reasoning) or quantitative approaches -- because it helps reduce the bias caused by the actual variability of susceptibility patterns.

(3) The effect of LUCC on landslide susceptibility is mediated by human activities. For instance, the abandonment of agricultural land leads to intensified soil erosion and water loss, increasing landslide susceptibility (Dandridge et al., 2023). On the other hand, good agricultural practices can positively impact slope stability (Gariano et al., 2018; Knevels et al., 2021; Pisano et al., 2017). Soil conservation measures in farmlands and the construction of terraced slopes can effectively slow down surface runoff concentration and soil saturation. Furthermore, reasonable urban planning and slope protection technologies, such as vegetation-based slope protection or retaining walls, can significantly reduce the negative impacts of engineering disturbances.

These interwoven factors further complicate the relationship between LUCC and landslide susceptibility. This complexity not only increases the difficulty of modeling landslide susceptibility but also raises higher requirements for related decision support systems. Encouragingly, recent research has gradually advanced in this direction: the coupling of LUCC with other factors is becoming better understood (Pacheco Quevedo et al., 2023), and multi-factor landslide susceptibility prediction models based on machine learning are gradually maturing (Agboola et al., 2024).

Based on the analysis of the research field above, we summarize the following limitations and future challenges:

To effectively reduce landslide risk, it is essential to establish a sustainable land management framework (Fig. 8).

Protecting lands with high ecological value, such as pristine forests, wetlands, and grasslands, is a key strategy for reducing landslide risk and enhancing ecosystem services (Wang et al., 2023). These areas play a crucial role in soil and water conservation, as well as slope stability. By restoring vegetation and stabilizing soil structure, their natural buffering capacity can be improved, thus reducing the risk of landslides and other natural disasters (Paudel et al., 2024). Based on this, governments should develop relevant policies and laws to promote ecological compensation mechanisms, encouraging local governments and communities to actively engage in the protection and restoration of vulnerable ecological zones (Liu et al., 2023). Ecological compensation funds can be raised through diversified financing channels such as green bonds and ecological protection funds, attracting investment from the private sector and international organizations. The United Nations' 2030 Agenda for Sustainable Development, Goal 15, explicitly calls for the "protection, restoration, and sustainable use of terrestrial ecosystems", offering a strong policy framework and international support (UN, 2015). China's projects, such as the Saihanba Desertification Control and North China Ecological Water Supplementation, provide practical experience for this strategy and have demonstrated the dual benefits of ecological protection and disaster risk reduction (Fu et al., 2023).

For damaged land types such as mining areas, abandoned farmlands, and degraded grasslands, a series of improvement and restoration strategies should be implemented. First, soil and water conservation measures, such as terracing and vegetation restoration, should be applied in degraded farmlands to reduce soil erosion and enhance soil stability. For mining areas and post-landslide regions, specific land restoration plans should be formulated, including slope reshaping and surface vegetation recovery, to reduce landslide risks at the source (Chatterjee et al., 2024). Similar to the Program of Central America Dry Corridor, restoring ecosystems and integrating traditional farming practices to enhance the productivity of natural landscapes can provide biodiversity and soil and water conservation benefits for these regions (Gotlieb et al., 2019). To achieve these goals, restoration projects should be piloted in areas with frequent economic activities or dense populations, in line with national disaster risk reduction strategies, to ensure the effective implementation of restoration measures. Additionally, international funding support, such as the Global Environment Facility (GEF) and other international environmental protection funds, can provide the necessary financial backing for restoration projects, ensuring smooth execution and long-term results (GEF, 1990).

To effectively optimize land use and reduce the exacerbation of landslide risks due to human activities, it is necessary to focus on land types such as cultivated land, artificial forests, and agroforestry land, and adopt corresponding management strategies. First, precision agriculture technologies should be promoted in cultivated land to minimize soil disturbance and protect surface structure (Oliver et al., 2013). In artificial forest areas, mixed-species planting should be encouraged to avoid the vulnerability of monocultures under extreme climate conditions (Seddon et al., 2021). Furthermore, for low-density development areas, zoning planning should be implemented, with reasonable restrictions on construction density in steep slope regions to reduce the contribution of improper construction to landslide risks. Similar to land consolidation measures in Bavaria, Germany, land merging and rehabilitation can improve agricultural production conditions, enhance rural landscape layouts, reduce soil erosion risks, and strengthen soil stability (Jiang et al., 2022). This experience can be applied to other regions, especially in countries with well-developed agriculture and forestry. These strategies should be integrated with national disaster risk reduction policies to ensure the long-term sustainability of land management.

To address the landslide threats in high-risk and rapidly developing areas, particularly in densely populated regions with high economic value, monitoring and early warning strategies should be implemented. A high-density real-time monitoring network should be established in high-risk areas, focusing on rainfall, soil moisture, and surface displacement, to promptly identify potential landslide risks (Pecoraro et al., 2019). For densely built-up areas and critical transportation corridors, a landslide risk early warning system should be developed and integrated with community response mechanisms to ensure a rapid emergency response in the event of a disaster (Fathani et al., 2016). Given that land management in many high-risk areas is difficult to improve in the short term, early warning systems become a critical tool. The United Nations has set a goal to ensure that everyone globally will have access to early warning protection by 2027, guiding countries to establish and enhance landslide early warning systems (UNDRR, 2022). Furthermore, these monitoring and early warning systems should be aligned with national disaster risk reduction technical standards, promoting technical cooperation and information sharing among countries in emergency management, particularly strengthening technical support for developing and high-risk countries to ensure timely prediction and response to landslide risks worldwide.