In the analysis of remote sensing images, the color and texture variations in village features significantly influence the accuracy of deep learning models (Zhang et al., 2020). To systematically capture these variables, we compiled a model development dataset from approximately 51 traditional Li settlements, excluding the nine core case villages. In this study, the UAV DJI Mavic 2 was selected as the main instrument for image acquisition. This UAV is easy to operate and has built-in instruments such as GPS, optical obstacle avoidance lens, and compass, enabling the operator to monitor the vehicle’s position, flight attitude, altitude, and speed in real time. The active stabilization camera gimbal is capable of ensuring clear camera images despite tilting caused by the aircraft vibration and wind.

In the route planning of the flight plan. To meet the requirements of spatial information extraction and analysis of Tangmo villages, the ground resolution G = 30 mm is set. The flight altitude is calculated as 55 m according to the UAV camera specifications as follows: H = P W G f S w where P_w is the number of pixels on the long side of the camera, G is the ground resolution, f is the summed focal length, and S_w is the sensor width.

In order to avoid problems such as blurred images and photo overlap in the shooting, the flight speed, camera parameters, and the settings must also meet the corresponding requirements in the shooting. The course overlap rate and side overlap rate were both set to 85%. The flight speed was 6.8 m/s, calculated from the camera settings, course overlap rate settings as follows: V = H S h 1 − o h f T S where S_h is the sensor height, Oh is the course overlap rate, and T_s is the exposure time interval.

We generated the training dataset through a systematic patch extraction process from the orthophotos of these settlements. High-resolution orthophotos were segmented into 512 × 512 pixel patches with a 30% overlap, and focusing on the core settlement and its immediate productive-peripheral areas (averaging ∼0.18 km² per village), we obtained approximately 16,400 valid, annotated image patches. Each patch underwent pixel-wise annotation in ArcGIS Pro for six landscape elements: mountains, water bodies, woodlands, farmlands, buildings, and roads. This method captured key visual characteristics, such as the hue contrasts between buildings and vegetation, spectral differences between water and cropland (Cao et al., 2024), and distinctive textures such as the centripetal layouts in karst depressions. Real-time data augmentation techniques (random flipping, ±15° rotation, brightness/contrast adjustments) were employed during training to enhance generalization and prevent overfitting, thus ensuring robust feature learning for subsequent analysis in the core case studies (Ko et al., 2020).

In order to precisely locate the plane and elevation coordinates of ground control points, the Trimble R10 global navigation satellite system in the GPS-RTK mode was selected for this study. The horizontal accuracy of the ground control points (GCPs) center coordinates measured by this instrument is 0.013 m–0.033 m, and the vertical accuracy is 0.030 m–0.055 m. To ensure that this accuracy range is reached, online RTK calibration provided by the BeiDou satellite-based system, the FindAUTO location positioning system is used. In addition, the GCPs should be evenly distributed in the study area and ensure that the surrounding area is free of obstructions and the point markings can be clearly identified.

2. Village-Based Spatial Splitting Strategy: Following the construction of the model development dataset, we implemented a stringent village-based spatial splitting strategy to divide the data into training, validation, and test sets. This approach helps avoid the leakage of spatial autocorrelation and ensures that the model is evaluated on completely unseen geographical units. Villages were randomly assigned to three mutually exclusive sets: patches from 40 villages (≈70%) formed the training set, those from 8 villages (≈14%) the validation set (used for hyperparameter tuning and early stopping), and patches from the remaining 9 villages (≈16%) constituted the completely independent test set. This strict partitioning scheme ensures that the model’s generalization capability is assessed in entirely novel settlement contexts, providing a reliable foundation for its subsequent application to the nine core case villages.

The classification system integrates the principles of landscape gene theory with the PLE framework and offers three distinct advantages (Wagner and Fortin, 2013). 1) Characterized by multi-scalar interconnectedness, a distinct hierarchical organization, and quantifiable metrics, this approach significantly enhances the scientific rigor and precision of territorial spatial planning. 2) It provides standardized criteria for the identification of landscape genes and the construction of a corresponding database, enabling a comprehensive characterization of traditional village landscape attributes, precise gene analysis, and integrated multi-plan coordination. 3) The simultaneous consideration of productive, ecological, and living landscape genes facilitates the development of balanced conservation-development strategies that protect cultural and ecological heritage while promoting socioeconomic progress (Huang et al., 2025).

Building upon the scholarship of landscape genes and field investigations in Hainan’s Li settlements, we developed a PLES gene classification system (Table 1) (Liu S. et al., 2024; Liu Y. et al., 2024). This system employs a multi-scalar analytical approach to examine village landscape characteristics at macro-, meso-, and microscales. Ecological landscape genes integrate macroscale landscape patterns, such as mountain-water networks, with microscale elemental interrelationships, thereby constructing comprehensive ecological networks that reflect the composition and interactions within ecosystems. Production landscape genes encompass the geometries of farmland, adhering to morpho-adaptive principles, and capture the functional interdependencies between settlements and croplands, thereby reflecting agro-ecological synergy (Ren and Buyandelger, 2024). This framework delineates the ecological matrices surrounding villages and operationalizes the units of the productive landscape. Living landscape genes integrate macroscale settlement forms and boundary morphologies, mesoscale road network circuitry and connectivity patterns, and microscale spatial articulations. Collectively, these elements capture the configurations of settlements, transportation networks, and organizational details (Wang L. et al., 2024). Compared to traditional binary classification systems, our system demonstrates superior conceptual comprehensiveness, methodological systematicity, and academic rigor (Zhuang et al., 2022).

To objectively assess the configurational characteristics of living landscape genes, which include settlement morphology, transportation networks, and spatial articulation, we utilized space syntax theory (Hillier and Hanson, 1984). We computed four primary metrics from axial-segment models, which were derived from 1:500 scale cadastral maps and drone orthophotos, using DepthMapX software version 0.8.0:

Circuity (C): This metric quantifies the deviation level of movement paths, defined as the ratio of the actual network distance to the Euclidean straight-line distance between two points. It serves as an indicator of path efficiency and network directness.

Where Pactual denotes the shortest-path distance along the network, and PEuclidean represents the straight-line distance.

2. Connectivity (Con): This measures the number of direct connections a node (e.g., street junction) has with adjacent nodes, reflecting local permeability.

Nodes with high connectivity often function as access hubs within settlement layouts.

3. Intelligibility (I): This metric assesses how well local connectivity can predict global integration through linear regression, indicating the legibility of the spatial structure.

Where R represents the coefficient of determination for all nodes.

4. Integration (Rn): This quantifies a node’s accessibility to all other nodes based on topological depth. It evaluates centrality and movement potential.

Where MDn(i) indicates the mean topological depth within radius N. We adopted both R (local integration) and Rtotal (global integration).

5. Shape Index (S): This index measures the compactness of a settlement’s boundary shape in comparison to a circle. Lower values indicate higher compactness.

Where P denotes the perimeter of the settlement boundary, and A indicates the area enclosed by the boundary.

6. Aspect Ratio (A): This metric quantifies the elongation of a settlement by comparing the lengths of its major and minor axes. Where Lmajor is the length of the major axis of the Minimum Bounding Rectangle (MBR) and Lminor denotes the length of the minor axis of the MBR

Where Lmajor is the length of the major axis of the Minimum Bounding Rectangle (MBR) and Lminor denotes the length of the minor axis of the MBR.

7. Fractal Dimension (FD): This parameter characterizes the complexity and self-similarity of settlement boundaries. Higher values suggest greater irregularity.

Where P is the perimeter, and A indicates the area.

Drawing upon the identified landscape genes as foundational reference points, this study utilized a multi-level geographic coding theory. It employed a combination of alphabetic codes and numerical identifiers to encode landscape gene data at various levels, culminating in the creation of a PLES-Gene Database (Fan et al., 2023). To deepen our understanding of the spatial characteristics of traditional settlements, representative village samples were selected through classification and merging processes within this database. The analysis of these settlements covered three dimensions: ecological spatial structure, characteristics of production activities, and the layout of living spaces. The study subsequently summarized and extracted the constituent elements and organizational patterns of traditional village landscape genes.

Utilizing the PLES-Gene Database and insights derived from typological analysis, this study developed a Landscape Gene Atlas for traditional Li settlements. This atlas, inspired by the structure of biological gene sequences and the organizational patterns of genetic coding, provides a structured visual representation of the combinatorial patterns characteristic of traditional village landscapes. By employing graphical and symbolic expressions, the atlas transforms abstract landscape gene data into intuitive and organized visual information. This representation not only facilitates the standardized storage and efficient management of landscape genes but also significantly enhances the capacity for multi-dimensional data analysis and comparative studies among researchers. As a result, the atlas proves to be an invaluable tool for revealing the intrinsic characteristics of these landscape genes. Moreover, the encapsulated data within this atlas lay a scientific foundation and offer evidence-based support for practical applications such as conservation planning, cultural heritage transmission, and the spatial optimization and regeneration of traditional settlements (Zhang and Yang, 2024).

The spatial distribution and morphology of traditional Li settlements on Hainan Island are fundamentally influenced by terrain and mountain-water configurations. Our key findings include:

There was a significant spatial adaptation observed between village patches and topographic parameters: Patch area and density decreased exponentially with increasing slope gradients. The FD of patch shapes exhibited a strong positive correlation with relief amplitude. The Landscape Shape Index progressively declined as slopes steepened, indicating a trend towards more squared patch geometries.

Genes related to elemental relationships govern the integration of settlements with cardinal landscape elements, forests, mountains, and waterways, which are vital to the environmental landscape of Hainan’s traditional settlements. These genes are expressed through adaptations that are specific to various geomorphic features, such as Mountain Massifs, Hilly Terrains, Alluvial Plains, and Erosional Terraces, thereby generating distinct spatial genotypes. For instance, in the Alluvial Valley Plains, settlements exhibit a linear clustering pattern closely aligned with river corridors (integration > 1.8), which places them in direct proximity to waterways. Conversely, mountain forests maintain a significant buffer zone, retreating at least 500 m from the edges of villages, defining a hydro-proximal adaptation. Within the Tectonic Highland Bench, settlements align with contours and feature sinuous access roads, encircled by mixed forest buffers along both upslope and downslope peripheries. Rivers consistently mark the lower boundaries of settlements, creating a vertically nested forest-road adaptation. In the Karst Depression Cluster Settlements, dwellings are dispersed within solution dolines, with access routes displaying high sinuosity to navigate the complex terrain. Perimeter rivers typically follow tectonic fractures, and forests integrate seamlessly into the rugged topography, contributing to a topographic-concealed adaptation within ravines. These genotypes fundamentally encode the integration of settlements with varying environmental landscape elements.

The traditional farmland morphology of the Li ethnic group on Hainan Island exhibits significant spatial heterogeneity across different altitudinal gradients. In karst terrains, such as Maona Village, dissolution-induced surface fragmentation, coupled with shallow soils and poor water retention, necessitates the utilization of micro-topography within karst fissures for a dispersed distribution, resulting in a discrete patch mosaic pattern. In non-karst high-altitude areas, such as Echa Village, gully dissection leads to strip-based cultivation adapted to steep slopes, manifesting as segmented ribbon and terraced ribbon patterns. Conversely, in non-karst low-altitude areas, such as Luoshuai Village, the presence of extensive alluvial plains with deep soils and dense river networks supports intensive cultivation on flat terrain, forming a continuous grid pattern.

The spatial coupling pattern between Li settlements and farmlands demonstrates a distinct response to topographic gradients. In karst terrains, dissolution fissures fragment the landscape into micro-topographies, necessitating a dispersed layout of farmlands adjacent to settlements. This results in an “Interwoven farmland-settlement mosaic” (e.g., Maona Village), where agricultural plots and village structures are intricately interspersed. In non-karst high-altitude areas (e.g., Chubao and Zahan villages), steep slopes necessitate strip-based cultivation adapted to the terrain, leading to the emergence of the “Settlement surrounded by farmland” and “Farmlands adjacent to settlements” patterns. Conversely, in non-karst low-altitude areas (e.g., Luoshuai Village), expansive alluvial plains and dense river networks facilitate the development of contiguous farmland on flat terrain, also forming a “Farmlands adjacent to settlements” pattern.

The spatial morphology of traditional Li settlements shows significant differentiation influenced by landforms. In karst terrain, settlements conform to the terrain, adopting linear layouts along ridges or waterways to capitalize on topographic fluctuations and hydrological flows. In non-karst terrain, settlement morphology diversifies: in plains, settlements predominantly develop into compact clusters, leveraging the open terrain. Along riparian corridors, settlements typically form ribbon-like patterns, constrained by the radius of farming areas. In mountainous landscapes, settlements exhibit paddy-interlocked configurations, where their boundaries intricately interlock with paddy field textures. Such geomorphic adaptations exemplify the Li’s ecological wisdom in a “terrain-responsive settlement with integrated farmstead clusters,” illustrating how traditional settlements spatially evolve in synergy with natural systems.

This study examines the boundary morphology of traditional Li settlements through the application of FD. The results reveal diverse characteristics across different FD thresholds: Settlements with an FD below 1.3 display a simple morphology, as exemplified by Jinmiaolang and Zajin villages. These villages feature geometrically regular boundaries with aligned edges, reflecting a spatially structured expansion governed by orderly construction logic. Settlements with FD values ranging from 1.3 to less than 1.5 demonstrate a more complex morphology, as observed in Luoshuai and Chubao villages, which exhibit intricate boundaries. These boundaries represent a negotiated balance between environmental constraints and deliberate spatial organization. For FD values of 1.5 or higher, settlements such as Hongshui and Maona villages exhibit complex boundary morphologies. Some of these settlements develop highly convoluted perimeters shaped by terrain and hydrography, while others manifest atomized building distributions that result in extreme dispersal, effectively negating clearly definable boundaries. This spectrum of forms, correlated with FD thresholds, substantiates the Li’s construction philosophy of ‘working with natural potentials’ and illustrates how geomorphic and socio-cultural factors jointly shape settlement boundaries.

The road network configuration of Li settlements on Hainan Island exhibits distinct typological characteristics, primarily based on the connectivity degree and circuitness number. These settlements are classified into four types: high-connectivity outward-oriented, high-connectivity inward-oriented, low-connectivity outward-oriented, and low-connectivity inward-oriented. The findings indicate that the connectivity degree exceeds 30% in the majority of settlements. However, constrained by water systems and topography, their road networks demonstrate low density and an irrational hierarchical structure. Influenced by the traditional Li concepts of reclusiveness and defense, these settlements exhibit strong internal connectivity but weak external accessibility. This pattern is in stark contrast to the road network configurations observed in local Han settlements. A minority of settlements, such as Hongshui and Maona Village, exhibit low-connectivity inward-oriented characteristics, with a connectivity degree below 30%. These settlements show limited connectivity between primary external roads and village entrances, further emphasizing the relatively enclosed road network morphology shaped by geographical constraints and the cultural ideology of the Li ethnic group.

The overall circuitness of road networks in Li settlements across Hainan Island remains relatively low, with most settlements exhibiting a circuitness typically below 30%. This limited formation of street network circuits reflects a state of simplified connectivity. Structures with medium-to-low circuitness are predominant, often manifesting as pectinate or dendritic patterns. In contrast, networks with high circuitness feature fewer nodes and connections but demonstrate heightened network utilization efficiency, typically forming peripheral loops. This configuration enhances external transportation accessibility, thereby facilitating mobility for both villagers and external persons. Collectively, the road network characteristics of Hainan’s Li settlements embody a distinctive morphology, shaped by geographical constraints and historical development factors.

Traditional Li settlements exhibit a high degree of spatial integration, with the majority of these settlements achieving an integration value exceeding 0.7. This high level of integration indicates robust internal connectivity and accessibility, facilitating ample space for residents’ daily activities. Settlements characterized by such high integration typically reside in low-altitude valley plains, where the flat terrain supports a closely interlinked configuration of farmland, habitation, and roads. This configuration is distinguished by high road density and circuitous routes. In contrast, settlements with low integration are predominantly found in high-altitude, mountainous areas. The challenging topography of steep slopes and ravines dictates that road networks in these areas adhere to contour lines, creating sinuous patterns and forming dendritic topological structures. A significant negative correlation exists between village integration and terrain complexity, underscoring an inherent spatial pattern: lowland settlements tend to feature simplified boundaries and high spatial coherence, whereas mountainous settlements display complex boundaries with fragmented connectivity. These observations elucidate the geographical logic and adaptive characteristics that have shaped the morphological evolution of Li settlements.

Traditional Li settlements exhibit a dichotomy in spatial intelligibility, characterized by strong associations in lowland areas versus weak linkages in highland areas. Settlements with high intelligibility are clustered in low-altitude valley plains, which are marked by open terrain, well-developed water systems, and fertile soils that are favorable for human habitation and agriculture. These settlements demonstrate strong local connectivity with closely interconnected components. Conversely, settlements with low intelligibility are situated in complex, high-altitude terrains. Hindered by steep slopes and deep ravines, their road networks are characterized by low topological density. This results in the development of locally isolated subsystems, which exhibit weak integration with the broader global structure and fragmented spatial coherence. This phenomenon underscores the profound adaptation of Li settlement morphology to varying natural environments, highlighting geography’s deterministic role in shaping spatial configurations.

Living landscape genes are predominantly influenced by social structures, ethnic traditions, and historical legacies:

Social Structures: The kinship groups and Hemu system within Li settlements significantly shape the street network patterns and other aspects of living landscape genes through mechanisms of resource allocation and defense. For example, the land co-cultivation model promotes modular settlement distributions, while the need for defense encourages the integration of bamboo barriers and stilt architecture.

These historically convergent forces foster the formation of topological connectivity and other living landscape genes, sustaining cultural identity and community cohesion. However, the pressures of accelerating modernity, such as increased mobility, the erosion of traditional institutions, and the shifting values of youth, pose threats to their socio-cultural foundations.

Ecological landscape genes are influenced by resource systems, terrain, and climate:

Resource Systems: On Hainan Island, the composite natural resource system provides the essential life-support substrate for traditional settlements. With central mountainous rainforests at its core, the island features interconnected river networks that function as hydrological corridors, and is bordered by coastal plains forming alluvial basins. This configuration is crucial in shaping the expression of Elemental Relationship genes and other ecological landscape genes. High-elevation settlements, such as Hongshui Village, develop configurations that are buffered by hills and fronted by rivers, featuring clustered habitation on fluvial terraces near rivers. Low-elevation settlements like Zajin Village are strategically developed along fluvial corridors within alluvial oasis areas, characterized by level terrain and optimal loamy soils, resulting in compact, medium to small-scale habitation units that have clear visibility of hills and accessibility to rivers.

These adaptations represent ancestral ecological intelligence that ensures survival in marginal environments. However, the effectiveness of traditional strategies is increasingly challenged by climate change, conservation policies, and the introduction of modern materials.

Productive landscape genes evolve from land-use systems and economic transitions:

Land-Use Adaptations: The development of Farmland Morphology and other productive landscape genes in traditional Li settlements on Hainan Island is shaped by their agricultural practices. This is illustrated by: Swidden fallow-cycle constrained dispersal at Maona Village in Wuzhishan, where settlements are distributed as discrete altitudinal nodes along ridge crests. This configuration generates low-connectivity production gene units with fragmented mosaic configurations. Contour-cascade intensification in Dali Village’s valley terraces, which are engineered through elevation-parallel cultivation systems. These systems facilitate the creation of high-density production gene units with terraced ribbon morphotypes.

These co-evolved systems exemplify circular resource utilization and eco-economic balance. However, quadruple pressures, agricultural scaling, tourism expansion, land intensification policies, and labor outmigration, are undermining production systems, posing threats to economic resilience and cultural landscape integrity. Critically, these local pressures are often embedded within the regional dynamics of urban growth and land use deconcentration, processes known to reconfigure landscape patterns and modulate ecological risks (Zeng et al., 2025). The conservation of landscape genes, therefore, must be analyzed and addressed within this larger, dynamic context of spatial development.