Leveraging Deep Learning and Spatial Modeling for Preventive Protection and Sustainable Management of Cultural Heritage: A Case Study of the Liuwan Tombs, Qinghai, China

Abstract

The Liuwan burial complex is the largest known prehistoric clan-based cemetery in the upper Yellow River region, making its preservation vital for Chinese cultural heritage and sustainable local development. To address threats from unregulated agricultural activities and illegal looting, this study proposes a non-invasive preventive protection approach. Surface-visible tombs were identified using low-altitude UAV imagery and deep learning models (YOLOv8n, YOLOv5n, RT-DETR-l, and Hyper-YOLO). By incorporating environmental factors such as elevation, slope, aspect, distance to water, Topographic Wetness Index, and Topographic Position Index, potential tomb distributions were modeled on the Biomod2 platform and key environmental drivers were analyzed. Hyper-YOLO achieved the highest identification accuracy (94.4%). The optimal model, EMwmean (TSS = 0.492, AUC = 0.798), showed that high-potential tomb areas are mainly concentrated in the central region, with tombs preferring elevations of 1964–1978 m, south-facing slopes, and slopes of 13.14–19.19°. This study demonstrates the feasibility of using deep learning to identify surface-visible tombs and predict their potential distributions based on environmental characteristics, thereby providing priority references for heritage protection in Liuwan rather than a definitive inventory of all subsurface remains or cultural phases.

1. Introduction

As important repositories of human history, archaeological sites reveal the development of civilizations and the dynamic relationship between human activities and the natural environment [1,2]. Among various types of ancient sites, tombs are particularly valuable for understanding past social organization due to their distinct morphological features and strong spatial indicators. The Liuwan cemetery, located in the upper reaches of the Yellow River in Qinghai, contains 1730 tombs from multiple cultural traditions, including Banshan, Machang, Qijia, and Xindian [3]. The site is notable for its abundant artifacts, continuous chronological sequence, and well-preserved remains. To date, archaeological research at Liuwan has concentrated mainly on the western section, leaving vast areas in the east insufficiently explored. Furthermore, many unexcavated subsurface features are under serious threat from both agricultural activities and illegal looting [4]. Therefore, accurately identifying the spatial distribution of tombs and quantitatively attributing their environmental determinants have become critical priorities for the preventive conservation of the Liuwan site.

Traditional archaeological survey and excavation can provide crucial and accurate local data and play an important role in understanding the archaeological context of a site [5]. However, because archaeological heritage is fragile and valuable, traditional archaeological work is often constrained by conservation requirements, time, labor, and research scope [6]. As a result, it is difficult to expose archaeological remains over a sufficiently large area, which may, to some extent, limit the understanding of the broader archaeological context and environmental relationships of the community under study [7]. In this context, remote sensing can serve as a complementary tool to traditional archaeology by identifying potential archaeological remains and supporting preventive heritage protection [8]. With the advancement of Earth observation and information technologies, remote sensing has been increasingly applied in archaeological site detection, heritage conservation, and cultural digitalization [9]. For example, Zhu et al. employed aerial imagery to investigate earthen mound tombs in northern Zhejiang, establishing interpretation criteria through a combination of visual interpretation and field surveys and conducting preliminary analyses of their classification and spatial distribution patterns [10]. The Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, utilized color infrared imagery to identify anomalous features of Han Dynasty tombs in Laoshan [11]. Tan et al. demonstrated that thermal infrared and shortwave infrared bands are highly sensitive to temperature anomalies associated with burial remains based on hyperspectral remote sensing data [12]. In recent years, the integration of deep learning and remote sensing has enabled more powerful automated identification of archaeological sites [13]. Caspari et al. combined open-access high-resolution satellite imagery with convolutional neural networks (CNNs) to detect elite burial mounds across the Eurasian Steppe [14]. Berganzo-Besga et al. integrated Random Forest with the YOLOv3 model to develop an automatic tomb detection approach, achieving an accuracy of 97% [15]. To further explore the relationship between archaeological site distribution and environmental factors, Archaeological Predictive Modeling (APM) has gradually evolved from early GIS-based spatial overlay analysis [16] to approaches inspired by Species Distribution Models (SDMs) in ecology [17]. Early localization experiments conducted by Brandt et al. in the Netherlands demonstrated the effectiveness of GISs in archaeological prediction [18]. Banks et al. proposed the Eco-Cultural Niche Modeling (ECNM) framework within the SDM paradigm to investigate prehistoric human–environment interactions [19]. Imen et al. applied the Maximum Entropy (Maxent) model to simulate the spatial distribution of urban heritage sites in the Sola region [20]. Although these approaches have achieved notable progress in site detection and spatial prediction, they are often applied independently in practice, lacking integrated analysis and systematic frameworks.

Given the severe threats currently facing the Liuwan site, there is an urgent need to develop preventive conservation methods. It is essential to acquire information on tomb locations and predict the potential distribution of subsurface remains without causing further disturbance to the site’s topography, thereby identifying priority areas for local heritage protection authorities. Therefore, the integration of automated identification and predictive modeling has become a crucial approach for the sustainable conservation and scientific management of cultural heritage in the Liuwan area.

2. Materials and Methods

2.1. Study Area

The Liuwan cemetery (Figure 1) is located on the second terrace of the north bank of the middle reaches of the Huangshui River in Ledu District, Haidong City, Qinghai Province, northeastern Qinghai–Tibet Plateau, China, at an average elevation of approximately 1900 m. The region experiences a semi-arid plateau continental climate, with a mean annual temperature of 6.9 °C and average annual precipitation of 335.4 mm. The study area generally exhibits higher terrain in the north and lower terrain in the south, extending west to Shagou, east to Liuwan Datanggou, north to Liuwan Datading, and south to Daxi Zhiqu, and is surrounded by terraces on both the eastern and western sides. The Liuwan cemetery dates back approximately 4600 years and represents an important Neolithic to Bronze Age archaeological site in the upper Yellow River region. A total of 1730 tombs have been excavated, including 265 Banshan-type, 1041 Machang-type, 419 Qijia culture-type, and 5 Xindian culture-type tombs, with a total of 37,925 artifacts unearthed, including painted pottery, stone tools, bone artifacts, and jade objects [21]. The site is currently divided into western and eastern sectors. The western sector has been systematically excavated, whereas the eastern sector has not yet undergone formal archaeological investigation. Only occasional discoveries of tombs have been made during agricultural activities by local residents, and the known extent of the cemetery has accordingly expanded eastward by approximately 200 m [4]. The Liuwan cemetery is characterized by the longest occupational duration, the greatest diversity of cultural types within a single site, and the most well-preserved and artifact-rich tomb assemblage in the region. It represents a key archaeological resource for understanding the transition from late prehistoric to early Bronze Age societies in the Gansu–Qinghai region.

2.2. Data Sources and Preprocessing

Deep learning model development was conducted using PyTorch 2.4.1; the computational environment comprised an Intel(R) Xeon(R) Silver 4310 CPU @ 2.10 GHz and an NVIDIA RTX A4000 GPU with CUDA 12.1. During training, all input images were resized to 640 × 640 pixels. The AdamW optimizer was used with an initial learning rate of 0.001. The batch size was set to 16 to maximize GPU utilization, and the number of training epochs was set to 200. Remote sensing image processing was performed using ArcGIS 10.8, while archaeological predictive modeling based on the Biomod2 (4.2-4) framework was implemented in R (RStudio 2026.01.1).

UAV Image Dataset: A custom Liuwan cemetery dataset was constructed using UAV remote sensing imagery acquired with a DJI Mavic 3E (DJI, Shenzhen, China) equipped with a 20-megapixel camera. Following image segmentation and screening, samples were annotated on the MakeSense.ai platform. The dataset was randomly split into training, validation, and test sets at a 6:2:2 ratio. To address the limited sample size, a data augmentation pipeline was implemented in PyTorch, including color jitter, HSV transformation, and Gaussian noise to simulate variations in imaging conditions. Data augmentation was applied stochastically during training to prevent overfitting to specific augmentation patterns.

Tomb Point Data: Tomb point data were derived from the detection results of deep learning models applied to UAV remote sensing imagery; high-confidence tomb detections (confidence > 0.8) were extracted from the full-area Liuwan survey and subsequently converted into vector point datasets using ArcGIS.

Environmental Factor Data: Environmental factor data were derived from ASTER GDEM data (30 m spatial resolution) obtained from the Geospatial Data Cloud platform. Considering the characteristics of the wooden tomb structures in the Liuwan cemetery [21], seven environmental variables were selected from topographic and hydrological perspectives, including elevation, slope, aspect, distance to water sources, Topographic Wetness Index (TWI), Topographic Position Index (TPI), and curvature, to investigate their influence on tomb spatial distribution. Topographic factors influence surface drainage and terrain stability, while hydrological factors affect soil moisture; jointly, these conditions control the distribution and preservation of tombs. All environmental variables were classified using the natural breaks (Jenks) method. Aspect was divided into eight classes, while all other variables were classified into five categories. All datasets were unified under the WGS 1984 UTM Zone 47N coordinate reference system.

2.3. Methodology

Focusing on the Liuwan cemetery and its surrounding areas in Qinghai, this study proposes an integrated spatial archaeological detection framework (Figure 2) that couples deep learning-based object detection, ensemble predictive modeling, and spatial statistical analysis. High-resolution low-altitude UAV imagery was used to compare the performance of several state-of-the-art detection architectures—including YOLOv8n, YOLOv5n, RT-DETR-l, and Hyper-YOLO—for obtaining high-confidence tomb locations. Utilizing the Biomod2 platform, the EMwmean ensemble algorithm, optimized based on TSS (true skill statistic, TSS) and AUC (area under ROC curve, AUC) criteria, was applied to simulate the potential distribution of tombs within the study area. In addition, Geodetector software (Excel version, 2015) was employed to quantify the independent contributions of environmental factors to tomb distribution and to reveal nonlinear interactions and synergistic enhancement mechanisms among multiple factors. By integrating automated identification and spatial modeling, this study elucidates the spatial distribution patterns of ancient tombs in the Liuwan area, providing theoretical support for cultural heritage conservation and regional sustainable development.

2.3.1. Deep Learning Object Detection

This study employed four mainstream deep learning models for tomb detection, including YOLOv8n [22], YOLOv5n [23], RT-DETR-l [24], and Hyper-YOLO [25,26]. Among them, YOLOv5 and YOLOv8 are single-stage lightweight object detection models with strong feature extraction and multi-scale feature fusion capabilities. RT-DETR is a Transformer-based object detection model that enhances spatial representation through a global self-attention mechanism. Hyper-YOLO, proposed by Tsinghua University, is an improved variant of YOLOv8 that integrates a Mixed Aggregation Network (MANet) and a Hypergraph-based Cross-layer Representation Network (HyperC2Net), enabling the modeling of higher-order spatial feature relationships and improving detection performance.

Model performance was evaluated using mean Average Precision (mAP), Recall (R), and Precision (P) [27]. Precision reflects the accuracy of model predictions, Recall indicates the ability of the model to detect true targets, while mAP@0.5 and mAP@0.5:0.95 evaluate the overall detection accuracy and localization stability under different IoU thresholds, respectively. The corresponding metrics are defined as follows:

$$\mathrm{m}AP={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^{n}A{P}_i}}{n}}$$

(1)

$$\mathrm{Re}call={\displaystyle \frac{TP}{TP+FN}}$$

(2)

$$\mathrm{Pr}ecision={\displaystyle \frac{TP}{TP+FP}}$$

(3)

where True positives (TPs) refer to instances where positive samples are correctly predicted as positive. False negatives (FNs) occur when positive samples are incorrectly predicted as negative, whereas false positives (FPs) represent cases where negative samples are incorrectly predicted as positive.

2.3.2. Potential Spatial Distribution Modeling

This study employed the Biomod2 platform to construct potential distribution models of the Liuwan cemetery using ten widely used algorithms, including Generalized Linear Model (GLM), Generalized Additive Model (GAM), Multivariate Adaptive Regression Splines (MARS), Random Forest (RF), Generalized Boosted Model (GBM), Classification Tree Analysis (CTA), Flexible Discriminant Analysis (FDA), Artificial Neural Network (ANN), Surface Range Envelope (SRE), and Maximum Entropy Model (MaxEnt) [28].

During model training, a constrained random sampling strategy was used to generate pseudo-absence points equal in number to the observed tomb locations [29]. To ensure robust model evaluation, random cross-validation was conducted with three independent repetitions. In each run, the dataset was randomly split into training (80%) and testing (20%) subsets. Model performance was evaluated using the true skill statistic (TSS) and the area under ROC curve (AUC) [29]. In the ensemble modeling stage, models with TSS > 0.35 and AUC > 0.7 were selected to construct the weighted mean (EMwmean) and committee averaging (EMca) ensemble models. Furthermore, Youden’s index was used to determine the optimal classification threshold for the ensemble models [30].

2.3.3. Geodetector

The Geodetector is a statistical method designed to detect spatial heterogeneity and to identify the driving forces behind it [31]. It does not rely on linear assumptions and is robust to multicollinearity, making it particularly suitable for assessing the contribution of environmental variables to tomb distribution patterns. In this study, factor detection, interaction detection, and risk detection were applied.

(1) Factor Detector: The factor detector is used to quantify the explanatory power of an independent variable X on the spatial distribution of the dependent variable Y, expressed by the q statistic. A higher q value indicates stronger explanatory power. The q statistic is defined as follows:

$$q=1-{\displaystyle \frac{{\displaystyle {\sum }_{h=1}^L{N}_h{\sigma }_h^2}}{N{\sigma }^2}}$$

(4)

where h = 1, 2, …, L denotes the strata of variable Y or X; L is the total number of strata; $N_h$ represents the number of samples in stratum h; N is the total number of samples; ${\sigma }_h^2$ is the variance of Y within stratum h; and ${\sigma }^2$ is the variance of Y in the entire study area.

(2) Interaction Detector: This module evaluates whether the combined effect of two variables enhances or weakens their individual explanatory power, and determines whether their influences are independent or interactively structured.

(3) Risk Detector: This module tests whether statistically significant differences exist between strata, thereby revealing spatial preferences of tomb distribution under different environmental conditions.

3. Results

3.1. Application of Hyper-YOLO for Tomb Detection

3.1.1. Comparative Experiments

Hyper-YOLO was compared with three representative models, including YOLOv8n, YOLOv5n, and RT-DETR-l. The results (

Table 1. Comparison of evaluation metrics for different models.

Model	Precision (%)	Recall (%)	mAP@0.5 (%)	mAP@0.5:0.95 (%)
RT-DETR-l	78.0	82.7	70.8	33.6
YOLOv5n	81.2	82.7	79.2	37.4
YOLOv8n	88.6	83.7	86.0	55.5
Hyper-YOLO	94.4	85.5	88.1	56.2

) show that Hyper-YOLO achieved the best overall performance across all evaluation metrics, including Precision, Recall, mAP@0.5, and mAP@0.5:0.95. Specifically, Hyper-YOLO obtained Precision, Recall, mAP@0.5, and mAP@0.5:0.95 values of 94.4%, 85.5%, 88.1%, and 56.2%, respectively. Compared with the second-best model, YOLOv8n, Hyper-YOLO improved these four metrics by 5.8, 1.8, 2.1, and 0.7 percentage points, respectively. Although the improvement in mAP@0.5:0.95 was relatively small, Hyper-YOLO still ranked first in all four metrics, indicating more balanced and robust detection performance. In contrast, YOLOv5n and RT-DETR-l showed weaker detection performance, especially RT-DETR-l, which obtained the lowest values for all metrics, suggesting limited adaptability to the dataset used in this study.

3.1.2. Detection Results and Feature Analysis

Qualitative comparisons across representative UAV images (Figure 3) show that Hyper-YOLO provides more precise detection results than YOLOv8n. In low-contrast or texture-ambiguous regions, the model demonstrates improved feature extraction capability, resulting in more accurate target localization. Moreover, Hyper-YOLO yields higher confidence scores with fewer false positives and false negatives. These improvements suggest that hypergraph-based feature aggregation enhances the modeling of higher-order relationships, improving robustness in complex backgrounds.

To further analyze the effect of the model on feature extraction for tomb targets, three representative patches were selected for Grad-CAM++ visualization. Unlike Figure 3, which presents the detection results of different models, Figure 4 focuses on the attention responses of the models. The three patches represent three typical recognition scenarios. The first corresponds to tomb targets with clear morphological features and is used to examine whether the model focuses on the tomb bodies and their boundary features. The second represents areas with relatively pronounced terrain undulation and is used to evaluate the model’s attention to targets under local terrain variation and shadow conditions. The third represents small-scale and weak-feature targets and is used to examine the model’s ability to attend to less obvious tomb targets. By comparing the feature responses of tomb targets under these three scenarios, the compatibility of the model with the tomb detection task can be further evaluated.

Based on these results, Hyper-YOLO was used for full-area inference, yielding 560 high-confidence detections (confidence > 0.8). To evaluate the reliability of the detection results, 100 samples were randomly selected from the high-confidence detections and manually interpreted using high-resolution UAV imagery with a spatial resolution of 0.035 m. This spatial resolution supports detailed interpretation of the morphological features of surface-visible tombs. The validation results showed that 92 of the 100 randomly selected samples exhibited morphological features consistent with tombs in the imagery, corresponding to an image-based visual interpretation accuracy of 92%.

To evaluate the spatial distribution pattern of the tomb locations identified by Hyper-YOLO, this study employed the Average Nearest Neighbor (ANN) method for spatial statistical testing. The Average Nearest Neighbor (ANN) analysis (

Table 2. Results of the Average Nearest Neighbor (ANN) analysis for detected tomb sites.

Metric	Observed Mean Distance (m)	Expected Mean Distance (m)	Nearest Neighbor Ratio (ANN)	Z-Score	p-Value
Value	8.54	16.39	0.52	−21.68	<0.01

) indicates that the observed mean distance (8.54 m) is lower than the expected mean distance (16.39 m), with a nearest neighbor ratio of 0.52 (<1) and a z-score of −21.68 (p < 0.01), indicating a statistically significant clustered pattern.

The nearest-neighbor connection map of the tombs (Figure 5) further shows that the central and eastern parts of the study area are characterized by dense connections and short inter-point distances, indicating a clustered distribution pattern. In contrast, the marginal areas show sparser connections and longer inter-point distances, suggesting clear spatial heterogeneity. The spatial statistical test and spatial visualization results consistently indicate that the Liuwan cemetery exhibits a clustered spatial pattern, which may reflect the influence of potential environmental preferences and sociocultural factors on burial site selection among the ancient inhabitants of Liuwan.

3.2. Potential Spatial Distribution of Tombs

This study used Pearson correlation analysis to examine the degree of linear correlation among the environmental variables, with the aim of identifying highly correlated variables and reducing variable redundancy and the potential effects of multicollinearity. Variables with an absolute correlation coefficient greater than 0.7 (|r| > 0.7) were considered highly correlated. As shown in Figure 6, elevation (dem), slope (podu), aspect (poxiang), distance to water sources (water_dist), and the topographic wetness index (TWI) exhibited weak-to-moderate correlations, with all correlation coefficients below 0.7. This indicates that there was no strong linear correlation observed among these variables, and they could therefore be jointly used in the subsequent analysis. Pearson correlation analysis indicates a strong correlation between the Topographic Position Index (TPI) and curvature (r = 0.97, |r| > 0.7); given TPI’s greater archaeological relevance, curvature was excluded and elevation, slope, aspect, distance to water sources, TWI, and TPI were retained for subsequent analysis.

3.2.1. Model Performance Evaluation

The TSS and AUC values of the selected individual and ensemble models in Biomod2 are presented in

Table 3. TSS and AUC values of each model.

Models	RF	CTA	GBM	GAM	MaxEnt	EMca	EMwmean
TSS	0.429	0.411	0.426	0.353	0.352	0.453	0.492
AUC	0.792	0.734	0.784	0.735	0.746	0.768	0.798

. Among the individual models, the Random Forest (RF) model achieved the highest TSS and AUC values, indicating the best predictive performance. Based on the criteria of TSS > 0.35 and AUC > 0.7, five individual models were selected to construct ensemble models. The committee averaging (EMca) ensemble model achieved a TSS of 0.453 and an AUC of 0.768, while the weighted mean (EMwmean) ensemble model achieved a TSS of 0.492 and an AUC of 0.798. Overall, the EMwmean model exhibited the highest predictive performance, showing a significant improvement over the individual models.

3.2.2. Potential Tomb Distribution Areas

The optimal EMwmean model was used to predict the potential distribution of tombs, and the results are shown in Figure 7b. The results indicate that the potential tomb distribution in the Liuwan region is primarily concentrated in three spatial zones—eastern, central, and western—forming a generally east–west-oriented pattern across the terrace landscape. The central region is dominated by high-suitability areas, whereas the eastern and western regions are primarily characterized by moderate suitability, with only small patches of high suitability scattered within them. The optimal classification threshold for the ensemble model was determined to be 522 using Youden’s index. Based on this threshold, the study area was categorized into four suitability classes: low (1), moderate (2), high (3), and very high (4) (Figure 7a).

To further assess model reliability, 50 tombs with clearly identifiable morphological features were manually extracted from high-resolution UAV imagery and used as independent validation samples. Overlay analysis reveals that 94% of these samples fall within the predicted high-suitability areas, indicating a strong spatial agreement between the observed tomb locations and the model predictions.

3.3. Analysis of Environmental Driving Mechanisms

The factor detector results (Figure 8a) indicate notable differences in the explanatory power of the selected environmental variables. Ranked by q-values, the six factors are ordered as follows: elevation (0.0499) > aspect (0.0498) > slope (0.0413) > TWI (0.0162) > distance to water sources (0.0088) > TPI (0.0023). Elevation, aspect, and slope show relatively higher q-values and pass the significance test (p < 0.05), suggesting that topographic conditions may contribute more to the spatial variation in tomb distribution compared with other variables. TWI and distance to water sources exhibit lower but still statistically significant explanatory power. In contrast, TPI shows minimal explanatory ability (q = 0.0023) and does not pass the significance test (p > 0.05). Overall, all q-values remain relatively low, indicating that no single environmental variable can independently explain the spatial distribution of tombs in the study area, and that the pattern is likely influenced by multiple interacting factors.

The interaction detector results (Figure 8b) show that the explanatory power of all pairwise combinations of variables is higher than that of individual factors, indicating the presence of interaction effects among environmental variables. Most factor pairs exhibit nonlinear enhancement, while only the interaction between aspect and TWI shows a bivariate enhancement pattern. This suggests that the combined influence of environmental variables is generally non-additive, with interaction effects contributing to improved explanatory performance. Among all combinations, interactions involving topographic variables (elevation, aspect, and slope) tend to show relatively higher q-values. The interaction between elevation and aspect yields the highest q-value (q = 0.1345), exceeding the explanatory power of either variable alone. This result suggests that the joint configuration of elevation and aspect may be more closely associated with spatial variations in tomb distribution than either factor alone, although its explanatory power remains moderate.

The risk detector results indicate that the probability of tomb occurrence varies across the strata of different environmental factors.

Table 4. Interval classification of tomb presence under different environmental factors.

Graded Classification	Elevation	Aspect	Slope	Distance to Water Sources	TPI	TWI
1	1924~1940	N	1.01~4.76	0~80.77	−2.56~−1.22	4.59~5.85
2	1940~1951	NE	4.76~7.26	80.77~159.63	−1.22~−0.33	5.85~6.77
3	1951~1964	E	7.26~9.84	159.63~238.24	−0.33~0.33	6.77~7.90
4	1964~1978	SE	9.84~13.14	238.24~322.60	0.33~1.22	7.90~9.90
5	1978~2005	S	13.14~19.19	322.60~432.96	1.22~2.78	9.90~13.92
6	—	SW	—	—	—	—
7	—	W	—	—	—	—
8	—	NW	—	—	—	—

“—“ indicates that no stratification is available. Based on the results of the risk detector, six environmental factors were classified using the natural breaks method. Aspect was divided into eight categories, while the remaining factors were divided into five categories each.

presents the environmental factor strata classified using the natural breaks method, while Figure 9 illustrates the variation in tomb occurrence probability within each stratum. Overall, elevation, slope, and aspect show relatively pronounced stratified differences in tomb occurrence probability. For elevation, the probability of tomb occurrence first increases and then decreases with increasing altitude, reaching its highest value in the 1964–1978 m interval (0.442). For slope, tomb density generally increases with slope gradient, reaching a maximum probability (0.476) within the 13.14–19.19° interval, which is the highest among all classes. Aspect shows directional variability, with relatively higher tomb densities on east- and south-facing slopes, while northwest-facing slopes exhibit lower values. In contrast, distance to water sources and TPI show relatively weak variation across classes, indicating limited stratification effects. The TWI shows an inverse relationship, with lower values associated with higher tomb densities. Overall, tomb occurrences in the Liuwan area tend to cluster within specific environmental ranges, particularly at elevations of 1964–1978 m, on slopes of 13.14–19.19°, and on sun-exposed aspects. These patterns suggest that topographic conditions may be associated with tomb placement.

4. Discussion

This study identifies surface-visible tombs using high-resolution UAV imagery and deep learning, and further predicts potential tomb distribution areas by integrating environmental variables with spatial ensemble modeling. This technical workflow uses the spatial distribution patterns of exposed or surface-visible tombs to infer areas where tombs may remain unexposed or buried underground. The results do not constitute a final confirmation of subsurface archaeological remains, but rather provide priority references for subsequent field survey, non-invasive prospection, and heritage management.

4.1. Tomb Detection

In tomb detection, the performance differences among models are primarily reflected in their ability to extract environmental context and texture features of the targets. After the excavation of the Liuwan site, most trenches and tombs were backfilled as required for heritage protection, and the land was returned to local residents for cultivation. As a result, only a limited number of representative tombs remain visible on the surface. However, due to prolonged unregulated human and livestock activity, as well as illegal looting, the cemetery has suffered severe damage. Under these challenging detection conditions, Hyper-YOLO demonstrated significantly higher detection accuracy than YOLOv5n, YOLOv8n, and RT-DETR-l, largely attributable to its innovative architectural modules. Compared with the YOLO series, which rely on PANet for local multi-scale feature fusion, Hyper-YOLO integrates a Mixed Aggregation Network (MANet) in the backbone and a hypergraph-based neck network (HyperC2Net), enabling the modeling of higher-order correlations among object features. This synergy preserves more spatial structural details and improves detection accuracy for archaeological targets with weak salience and high deformation. The RT-DETR-l model, based on the Transformer architecture, typically requires large-scale datasets for adequate training; however, the limited study area and sample size in this research do not meet such requirements. Additionally, its real-time design may compromise precise localization of targets in complex backgrounds, limiting its performance advantages. To ensure the reliability of spatial analysis, only high-confidence tomb detections (confidence > 0.8) were retained for further study. This approach of extracting tomb locations via object detection models enables non-destructive acquisition of cultural heritage information and provides high-quality samples for spatial analysis.

4.2. Distribution Prediction

In terms of spatial prediction, the ensemble model EMwmean outperforms individual models (TSS = 0.492; AUC = 0.798), demonstrating that ensemble approaches can effectively integrate the strengths of selected base models to improve predictive performance. However, the filtering criteria applied during ensemble construction may also constrain model performance to some extent [32,33]. Iterative ensemble strategies based on model performance could further optimize prediction results. The protection of tomb remains has long been a priority in the Liuwan area; however, the unknown nature of subsurface relics makes comprehensive preventive conservation challenging [4]. In this study, spatial ensemble modeling was used to predict the potential distribution of tombs, providing a scientific basis for site protection. Based on model predictions, protection zones can be clearly delineated within areas of high potential tomb distribution, and interpretive signage can be installed along major roads or in areas with high foot traffic. This approach not only safeguards invisible subsurface remains but also promotes public education about the Liuwan site. Among the potential tomb distribution zones, areas with high and very high potential should be prioritized by cultural heritage authorities. Measures such as protective fencing, regular patrols and monitoring, anti-looting prevention, and the installation of protection markers and explanatory signs can be implemented first in these areas. For areas with medium and low predicted tomb potential, the model results can provide auxiliary references for subsequent archaeological surveys and land-use coordination; however, their specific management should still be determined in combination with heritage protection planning, field investigation, and necessary archaeological prospection. Since the model in this study mainly targets surface-visible tomb features, its prediction results cannot replace the identification and assessment of other types of buried archaeological remains. Therefore, before agricultural production, land-use adjustment, or construction activities are carried out, necessary investigation and assessment should still be conducted in accordance with cultural heritage protection requirements.

4.3. Driving Factors Analysis

Although the built-in functions of Biomod2 can assess variable importance, they primarily evaluate the independent effects of environmental factors and cannot directly capture interactions among variables. Therefore, the Geodetector method was employed to analyze environmental driving mechanisms. The results confirm that tomb distribution is non-random and closely associated with specific environmental conditions [34]. Among the selected environmental factors, elevation, slope, and aspect showed relatively higher explanatory power, suggesting that they may be important environmental factors associated with tomb distribution in the Liuwan area. This finding is generally consistent with previous studies on the importance of environmental variables [35]. However, the overall q-values of individual environmental factors were relatively low, indicating that a single natural environmental variable has limited explanatory power for tomb distribution. This may be related to the limited environmental gradient within the study area and the combined influence of multiple natural and cultural factors on tomb distribution. The interaction detector results showed that some factor combinations had higher explanatory power than individual factors, suggesting that tomb distribution is more likely influenced by the combined effects of multiple environmental conditions. Therefore, these results can provide a reference for prioritizing subsequent archaeological surveys and heritage protection management, but should not be used directly as deterministic management boundaries.

4.4. Limitations and Future Work

Despite these findings, several limitations remain. Existing excavation reports of the Liuwan site provide only general distribution information, lacking precise spatial coordinates. Additionally, due to backfilling, the tomb locations identified through deep learning cannot be linked to specific cultural phases (e.g., Banshan, Machang, Qijia, and Xindian), limiting the ability to analyze temporal variations in site selection preferences. Moreover, tomb distribution is influenced not only by environmental conditions but also by human factors such as settlement structure, social stratification, and belief systems, which are difficult to quantify [36,37]. Consequently, reliance on environmental variables alone may introduce bias into the prediction results. However, given the relatively small spatial extent of the study area and its simple topographic conditions, the currently observable surface features are considered sufficiently representative to support the analysis of environmental preferences. Future research should incorporate precise tomb location data and classify sites by cultural type to investigate spatiotemporal variations in site selection. In addition, integrating technologies such as microwave remote sensing and ground-penetrating radar could further improve tomb detection accuracy and provide stronger technical support for archaeological exploration and heritage conservation in the Liuwan area.

5. Conclusions

Focusing on the Liuwan cemetery and its surrounding areas, this study proposes an integrated framework combining deep learning-based object detection with site distribution modeling to identify tombs and simulate their potential spatial distribution. The main conclusions are as follows: (1) In tomb detection, the Hyper-YOLO model achieved the best overall performance, with precision and mAP50 improved by 5.8% and 2.1%, respectively, compared to the baseline YOLOv8n model, demonstrating superior adaptability to tomb detection tasks. (2) For spatial distribution prediction, the EMwmean ensemble model based on the Biomod2 platform exhibited strong predictive capability (TSS = 0.492, AUC = 0.798), effectively simulating the potential distribution of tombs within the study area. (3) Geodetector analysis indicated that the spatial differentiation of tombs in the Liuwan area is the result of synergistic effects among multiple factors. Elevation, aspect, and slope showed the highest explanatory power and are the dominant drivers influencing tomb distribution, with tombs preferentially located at elevations of 1964–1978 m, on south-facing slopes, and with slopes of 13.14–19.19°. The proposed approach, which integrates deep learning-based object detection, ensemble predictive modeling, and spatial statistical analysis, provides an effective technical means for detecting partially visible archaeological remains in small areas. By inferring potential distribution zones based on environmental characteristics, this method offers valuable theoretical and practical support for future archaeological excavations and the sustainable conservation of cultural heritage in the Liuwan region.