AC-YOLOv11: A Deep Learning Framework for Automatic Detection of Ancient City Sites in the Northeastern Tibetan Plateau

Highlights

What are the main findings?

• We developed a novel deep learning framework (AC-YOLOv11) integrating dual-path attention and deformable feature calibration for ancient city detection on the Tibetan Plateau.
• The model achieved an F1-score of 94.2% and mAP@0.5 of 82.3% on the newly constructed Qinghai Lake Ancient City Dataset (QHACD), identifying 74 highly probable ancient sites.

What are the implications of the main findings?

• This study demonstrates the feasibility of automated archaeological prospection in high-altitude environments using high-resolution satellite imagery.
• The framework provides a scalable approach for cultural heritage mapping and environmental adaptation analysis across the Tibetan Plateau and beyond.

Abstract

Ancient walled cities represent key material evidence for early state formation and human–environment interaction on the northeastern Tibetan Plateau. However, traditional field surveys are often constrained by the vastness and complexity of the plateau environment. This study proposes an improved deep learning framework, AC-YOLOv11, to achieve automated detection of ancient city remains in the Qinghai Lake Basin using 0.8 m GF-2 satellite imagery. By integrating a dual-path attention residual network (AC-SENet) with multi-scale feature fusion, the model enhances sensitivity to faint geomorphic and structural features under conditions of erosion, vegetation cover, and modern disturbance. Training on the newly constructed Qinghai Lake Ancient City Dataset (QHACD) yielded a mean average precision (mAP@0.5) of 82.3% and F1-score of 94.2%. Model application across 7000 km² identified 309 potential sites, of which 74 were verified as highly probable ancient cities, and field investigations confirmed 3 new sites with typical rammed-earth characteristics. Spatial analysis combining digital elevation models and hydrological data shows that 75.7% of all ancient cities are located within 10 km of major rivers or the lake shoreline, primarily between 3500 and 4000 m a.s.l. These results reveal a clear coupling between settlement distribution and environmental constraints in the high-altitude arid zone. The AC-YOLOv11 model demonstrates strong potential for large-scale archaeological prospection and offers a methodological reference for automated heritage mapping on the Qinghai–Tibet Plateau.

1. Introduction

Accurate identification of ancient city sites on the northeastern Tibetan Plateau is fundamental for understanding regional settlement patterns, human–environment interactions, and cultural development at high altitudes [1]. However, the unique geomorphological conditions of this region—including sparse vegetation, severe erosion, periglacial landforms, and large areas of fragmented gravel surfaces—pose significant challenges to traditional archaeological prospection. Conventional survey methods are highly labor-intensive and often limited by the remote and harsh environment of the plateau. With the increasing availability of high-resolution satellite data, remote sensing has become an indispensable tool for site discovery, landscape reconstruction, and cultural heritage monitoring [2,3,4,5,6,7,8]. Yet, complex topography and the subtle surface expressions of archaeological features still restrict the efficiency and objectivity of manual visual interpretation [9].

In recent years, artificial intelligence has led to major advances in automatic archaeological feature detection [4,10]. Deep learning, especially convolutional neural networks (CNNs), has demonstrated strong potential in recognizing earthworks, burial mounds, qanat systems, charcoal kilns, defensive structures, and other anthropogenic features from LiDAR, multispectral, and very-high-resolution optical imagery [11,12,13,14,15,16]. Studies have shown that automated detection can significantly improve the efficiency, repeatability, and spatial coverage of archaeological prospection, enabling large-scale detection across forested, arid, and mountainous regions [17,18,19,20]. Among various algorithms, the YOLO (You Only Look Once) family of object detectors has become one of the most widely used frameworks because of its balance between detection accuracy and computational efficiency [21]. Improved versions such as YOLOv5, YOLOv7, and YOLOv8 have been successfully applied to cultural heritage mapping and LiDAR-based archaeological detection, providing a promising foundation for automated site-recognition pipelines [22,23,24,25,26].

Despite these advances, the northeastern Tibetan Plateau remains understudied in terms of AI-assisted archaeological detection. Compared with typical low-altitude regions, ancient city sites on the plateau often display weak spectral contrast, irregular outlines, multi-scale degradation, and partial burial by aeolian and fluvial sediments. These characteristics reduce the transferability of existing models and lead to frequent false positives when applying conventional YOLO architectures. Moreover, the lack of annotated training samples from high-altitude archaeological contexts further restricts model generalization. As a result, there is still a need for a task-specific detection framework that can accommodate the unique geomorphological and spectral signatures of mountain-plateau archaeological features [27].

To address these challenges, this study proposes AC-YOLOv11, an enhanced deep learning framework designed for automatic detection of ancient city sites using GF-2 and Sentinel-2 imagery. The model integrates improved feature-fusion modules, deformable convolution, and attention mechanisms to enhance sensitivity to subtle linear and polygonal archaeological structures. A multi-source remote sensing dataset was constructed, incorporating field-verified ancient city sites, multi-temporal samples, and negative samples representing complex backgrounds typical of the Qinghai Lake basin. We further optimized anchor assignments and introduced adaptive spatial-context encoding to reduce misclassification in gravel-desert and eroded-terrace environments. Experimental results demonstrate that AC-YOLOv11 significantly outperforms baseline models in precision, recall, and robustness, especially for small and degraded archaeological features. The integration of deep learning with high-resolution imagery provides a practical pathway for large-scale archaeological prospection and heritage protection strategies on the Tibetan Plateau.

This study not only contributes a new methodological framework for site detection in high-altitude regions but also provides a reproducible workflow integrating remote sensing, AI modeling, and archaeological validation [28]. The results highlight the potential of advanced deep learning tools for discovering and monitoring cultural heritage in challenging environments and offer an important reference for future archaeological research and heritage management on the Tibetan Plateau and beyond.

2. Study Area Overview

The Qinghai Lake Basin lies on the northeastern margin of the Qinghai–Tibet Plateau (97°50′–101°20′E, 35°19′–38°20′N), covering approximately 5.6 × 10⁴ km² and representing China’s largest inland saline lake basin. The terrain exhibits a northwest–southeast gradient in elevation, ranging from about 2600 to 5700 m a.s.l., and includes diverse geomorphic units such as mountains, river valleys, lakeshore plains and terraces. The basin is characterized by a typical cold, high-altitude semi-arid climate: mean annual temperature is approximately −4 °C, and precipitation ranges from 350 to 450 mm, while evaporation greatly exceeds precipitation, forming a classic inland closed dry environment [29]. Vegetation is dominated by alpine meadows and shrublands, which play a vital role in soil retention and constrain the development of ancient agro-pastoral activities. Geographically, the basin sits at the intersection of the East Asian monsoon region, the northwestern arid belt and the high-altitude cold zone, serving as a key node between the Yellow River source region and the Hexi Corridor. This unique setting has endowed the basin with central significance for Holocene human migration and civilizational exchange [30].

Archaeological and chronological evidence indicate that since the Neolithic period, the basin has functioned as an important agro-pastoral transition zone, with settlements mainly distributed along rivers and lakeshores, reflecting strong human dependence on water resources and pastures [31,32]. During historical times, the region evolved into a frontier interaction zone for the Qiang, Xianbei, Tuyuhun, Tubo and Central Plains dynasties, where numerous ancient city sites bear witness to the development of the northern route of the Silk Road [33,34]. Under the combined influence of natural environment and human activities, the distribution of ancient cities in the Qinghai Lake Basin presents a distinct “lakeshore–riparian” pattern: lakeshore plains and river valleys provided favorable conditions for agriculture and animal husbandry, while at the same time offering obvious advantages in defense and transportation.

Climatic and palaeo-environmental records further show that water-level fluctuations of Qinghai Lake responded closely to regional precipitation and temperature variability, highlighting its sensitivity to climate change [35,36]. Landscape-pattern analyses and remote sensing reconstructions reveal that human settlement expansion generally correlated with environmental stability and resource accessibility [37,38]. Overall, the Qinghai Lake Basin serves as a critical window for understanding human–environment interactions on the Tibetan Plateau and for deciphering the spatial patterns of Silk Road traffic networks in the highlands [39]. In this study, a distribution map of ancient city sites in the Qinghai Lake Basin is compiled, marking the main known sites and their geographical locations (Figure 1).

3. Materials and Methods

3.1. Sources and Processing of Ancient City Site Data

The ancient city inventory used in this study was primarily derived from published archaeological survey reports, historical–geographical records and regional gazetteers, and was cross-checked with modern toponyms and geomorphic clues [40]. Data compilation was based on open-access sources, including Atlas of Chinese Cultural Relics: Qinghai Volume and Studies on Ancient Cities in Qinghai (Qinghai Gucheng Kaobian), from which the names, spatial descriptions and coordinate information of sites within the Qinghai Lake Basin were systematically extracted. In total, 37 ancient city sites were identified. To ensure spatial accuracy and consistency, a multi-stage verification process was implemented. (1) Text–image cross-validation: directional semantics (e.g., “near the river,” “on the terrace,” “by the mountain pass”) described in the literature were matched to high-resolution remote sensing imagery. (2) Topographic consistency testing: DEM-derived slope, curvature and shaded-relief layers were analyzed to confirm whether each site was situated in a geomorphically plausible context, such as valley terraces, lacustrine plains or corridor nodes. All spatial data were unified under the WGS 1984 coordinate reference system and converted to standardized vector layers containing fields for site ID, Chinese and English names, cultural period, longitude, latitude, city shape and area. This integration of archaeological literature and remote sensing-based geospatial verification provides a robust and reproducible dataset for subsequent modeling and pattern analysis.

3.2. Remote Sensing Data and Preprocessing

The GF-2 satellite imagery used in this study was provided by the Qinghai Provincial Remote Sensing Center for Natural Resources (https://www.qhgfrs.cn/publiccms/, accessed on 4 June 2025). GF-2 data comprise four multispectral bands (red, green, blue, and near-infrared) with a spatial resolution of 3.2 m and a panchromatic band with 0.8 m resolution. Prior to analysis, all imagery underwent standard preprocessing in ENVI 5.6 (Harris Geospatial Solutions, Broomfield, CO, USA), including radiometric calibration, atmospheric correction, and orthorectification of both multispectral and panchromatic scenes. Subsequently, a nearest-neighbor diffusion-based pansharpening algorithm was applied to fuse the multispectral and panchromatic bands, yielding pan-sharpened products at 0.8 m spatial resolution [41,42,43,44]. This approach preserves spectral fidelity while enhancing spatial detail, which is critical for distinguishing fine geomorphic and archaeological features. Finally, the imagery was clipped and mosaicked to the Qinghai Lake Basin extent using a vector boundary mask derived from 1:100,000 topographic data. The resulting dataset provides high-resolution, radiometrically consistent coverage suitable for archaeological feature extraction and machine-learning-based detection [45].

3.3. Construction of the Ancient City Dataset (QHACD)

The QHACD was constructed using the GF-2 and Sentinel-2 imagery introduced in Section 3.1, combined with the data preprocessing steps described in Section 3.2. The pan-sharpening, geometric correction, and cloud masking operations in Section 3.2 ensured that all input data shared consistent spatial resolution and radiometric quality, which was essential for accurate sample annotation and subsequent model training. After preprocessing, ancient city sites were manually digitized based on GF-2 pan-sharpened images, while negative samples were extracted from non-archaeological areas with similar spectral or morphological characteristics. These procedures produced a standardized dataset suitable for training and evaluating AC-YOLOv11. For each site, image subsets were extracted within a 1.0–2.5 km window centered on the site, depending on city scale and topographic setting. Each subset was divided into uniform 450 m × 450 m (562 × 562 pixels) image tiles, corresponding to an approximate ground area of 57,372 km² in total. Manual interpretation and annotation were carried out to delineate “Ancient City” as a single object category. Bounding boxes were drawn to cover the complete extent of city walls and associated features. For sites with fragmentary boundaries or dual enclosures, multiple bounding boxes were applied. To minimize false positives, negative samples were also collected from areas with man-made geometric structures resembling city forms—such as modern settlements, agricultural grids, salt pans, industrial fences and quarry boundaries—maintaining a positive/negative ratio of 1:3 (Figure 2).

In total, 283,320 original tiles were generated across the study area, among which 37 contained verified ancient city instances. To mitigate overfitting and class imbalance inherent to the limited sample size, multiple data augmentation strategies were applied: Geometric augmentation: random rotations (0–360°), horizontal and vertical flips, scaling (0.5–1.5×), random cropping and translation; Radiometric augmentation: ±20% variation in brightness, contrast and saturation [46,47]; Multi-temporal expansion: incorporating GF-2 imagery acquired in different seasons and Sentinel-2 scenes from spring, summer and autumn to account for variations in vegetation cover, surface reflectance and snow; Synthetic augmentation: employing Mosaic and MixUp strategies to preserve semantic integrity while diversifying training samples (Figure 3) [48,49,50]. After augmentation, the number of positive samples increased from 37 to 296. A random 20% subset of the enhanced dataset was manually cross-validated by both archaeology and remote sensing specialists, and samples with ambiguous or indistinct boundaries were discarded. The final Qinghai Lake Ancient City Detection Dataset (QHACD) comprises 296 positive and 888 negative samples, totalling 1184 images, and serves as the foundation for subsequent model training and validation.

3.4. Model Improvement Based on YOLOv11

To address the challenges posed by the blurred boundaries, weak linear textures, and fragmented or morphologically complex structures of ancient city remains in GF-2 remote sensing imagery of the Qinghai Lake Basin, this study proposes an enhanced deep residual architecture named AC-SENet (Ancient City Squeeze-and-Excitation Network) [51,52]. The AC-SENet replaces the original backbone of the YOLOv11 detector, aiming to strengthen the extraction of fine morphological features such as city walls, corner junctions, and moat edges. Built upon the ResNet-152 framework, AC-SENet integrates a Dual-Path Attention Block (DPAB) that jointly models channel-wise and spatial dependencies, alongside an adaptive multi-scale pooling strategy to enhance selective feature activation during the feature-fusion process [53,54,55]. These modifications enable the model to capture discriminative geometric information even under variable illumination, vegetation cover, and terrain complexity—conditions typical of high-altitude archaeological landscapes [56,57]. Compared with the baseline YOLOv11 configuration, the proposed AC-SENet introduces dual-path excitation and deformable calibration to expand the receptive field adaptively and emphasize structural continuity across broken walls and irregular urban perimeters [58,59]. The resulting AC-YOLOv11 architecture achieves improved robustness in detecting low-contrast, partially eroded, or geomorphically disturbed ancient sites—laying the foundation for high-precision archaeological mapping over the Qinghai–Tibet Plateau.

3.4.1. Compression Operation of AC-SENet

In the feature-extraction process of ancient city remote sensing imagery, the compression operation aims to achieve efficient feature dimensionality reduction and channel re-weighting, thereby enhancing the model’s sensitivity to key structural components such as city walls, corner junctions and moat edges, as well as the linear geometry of rectangular enclosures. Unlike the conventional Squeeze-and-Excitation Network (SENet), which relies solely on Global Average Pooling (GAP) for feature compression, the proposed AC-SENet introduces a dual-modal pooling strategy that combines Soft Pooling and Global Max Pooling (GMP) to mitigate the loss of locally salient information inherent to GAP [60].

Ancient city features in high-resolution satellite imagery often manifest as localized high-response regions—e.g., compacted-earth wall edges or angular corner nodes—whose responses may be suppressed by GAP’s averaging operation. To address this issue, Soft Pooling is adopted to dynamically weight each pixel’s contribution through an exponential attention mechanism. Given an input feature map $R^{H\times W\times C}$, the importance weight $w_i$ of each feature value $x_i$ is computed using a Softmax function:

$$w_i={\displaystyle \frac{e^{x_i}}{{\sum }_{j\in U}{e}^{x_j}}}$$

(1)

Based on this weight distribution, a channel-wise descriptor vector is produced as:

$$y_c=\sum _{i\in U}{w}_i\times x_i$$

(2)

This formulation amplifies highly responsive local features via the exponential operator, enabling the network to concentrate on prominent archaeological morphologies such as ramparts and angular boundaries.

To further compensate for the reduced sensitivity of soft pooling to sparse or weakly preserved traces—typical of partially eroded sites—the AC-SENet parallelly incorporates GMP, which retains the maximum activation within each channel $max\left(U_c\right)$, generating a descriptor $Z_{max}\in R^{1\times 1\times C}$. This addition enhances spatial invariance and improves robustness against illumination variation and discontinuous boundaries [61]. Finally, the outputs of the two pooling branches are fused via channel-wise concatenation followed by adaptive weighting:

$$Z_{\mathit{fused}}=\alpha \cdot y_c+\left(1-\alpha \right)\cdot z_{max},\alpha \in \left[0, 1\right]$$

(3)

where $\alpha \in \left[\mathrm{0,1}\right]$ is a learnable parameter optimized during back-propagation.

The fused representation simultaneously preserves the fine-grained sensitivity of soft pooling and the noise-resilient global response of max pooling, achieving superior feature compression under low-contrast or topographically complex conditions—typical of archaeological landscapes in the Qinghai–Tibet Plateau.

3.4.2. Excitation Operation of AC-SENet

Ancient city remains in remote sensing imagery typically appear as low-contrast, locally high-reflectance, and structurally heterogeneous objects, where subtle variations in wall edges, corner junctions, and moat shadows play decisive roles in visual recognition. To enhance sensitivity to these morphological cues while suppressing background noise, the Dual-Path Excitation Block (DPEB) is embedded within the backbone of AC-SENet to achieve dynamic channel recalibration and spatial modulation in a coordinated manner. The detailed structure of the Dual-Path Excitation Block (DPEB) is illustrated in Figure 4.

Given the compressed feature descriptor $Z\in R^{1\times 1\times C}$ obtained from the preceding compression stage, the DPEB first performs channel re-weighting. The input vector is passed through a fully connected layer with parameters $W_1\in R^{{\displaystyle \frac{C}{r}}\times C}$ (r = 16), producing an intermediate representation: $Z_{mid}\in R^{1\times 1\times C/r}$, followed by a LeakyReLU activation (slope = 0.2) to alleviate gradient vanishing and preserve informative negative activations. A subsequent expansion layer: $W_2\in R^{C\times C/r}$ restores dimensionality: $A_{channel}\in R^{1\times 1\times C}$. The channel activation vector is normalized through a Hard-Sigmoid gating function to constrain values to the range $\left[0, 1\right]$, balancing computational efficiency with non-linear expressiveness [62]. The resulting weight vector $S\in R^{1\times 1\times C}$ is then applied to the feature map $U\in R^{H\times W\times C}$ via elementwise multiplication, selectively amplifying channels representing ramparts, corners and other structural edges, while suppressing those dominated by agricultural grids, modern infrastructure or shadow artifacts.

To further capture the fine-grained spatial patterns of moats and eroded foundations, the DPEB integrates a spatial attention branch. The recalibrated feature map $U_{recalibrated}$ is convolved with a depthwise separable convolution (3 × 3) to generate a spatial attention map $M_{spatial}\in R^{H\times w}$:$M_{\mathrm{spatial}}=\sigma \left(f_{\mathrm{DWConv}}\left(U_{\mathrm{recalibrated}}\right)\right)$, where $f_{\mathrm{DWConv}}$ denotes depthwise convolution and $\sigma$ is the sigmoid activation function. The attention map is multiplied elementwise with $U_{recalibrated}$ to form the spatially modulated output $U_{DPEB}$. This dual-path excitation mechanism allows AC-SENet to dynamically adapt to heterogeneous spectral–spatial conditions and degraded archaeological morphologies, substantially improving robustness against erosion, agricultural disturbance and surface sediment coverage [63].

3.4.3. Adjustment Operation of AC-SENet

Based on the multi-scale feature maps generated from the excitation module, the adjustment operation dynamically redistributes weights and fuses features across scales to optimize the model’s adaptability to eroded, incomplete, and morphologically irregular city structures, while enhancing the detection specificity of corners and linear fortification edges. The output feature maps $\{{X^{\prime }}_1,{X^{\prime }}_2,{X^{\prime }}_3\}\in R^{1\times 1\times C}$ correspond to different spatial scales of city-wall structures. Each feature set is processed through a lightweight $1\times 1$ convolution to generate scale-wise weight vectors $W\in R^{3\times C}$. A channel-wise Softmax normalization is then applied to dynamically evaluate the contribution of each scale to the current channel response. The high-resolution feature ${X^{\prime }}_1$ primarily governs corner and short-edge localization, whereas the low-resolution feature ${X^{\prime }}_3$ emphasizes global rectangular contours. The weighted fusion process can be formulated as:

$$X_{\mathrm{fused}}=\sum _{k=1}^3{w}_{k,c}\cdot X_k^{\prime }$$

(4)

where $w_{k,c}$ denotes the normalized weight of the $k$-th scale for channel $c$. Given that most ancient city walls in the Qinghai Lake Basin have undergone severe morphological degradation due to erosion, cultivation, and modern construction, fixed-kernel convolutions are often incapable of accurately capturing their non-rigid and discontinuous boundaries. To address this limitation, a Deformable Feature Calibration (DFC) mechanism (Figure 5) is introduced, which learns sampling offsets $\Delta p$ to dynamically adjust the receptive field of the convolution [64,65]. This allows the model to adaptively represent irregular geometric structures such as collapsed or partially eroded city perimeters while maintaining overall spatial consistency. The calibrated feature map $X_{\mathrm{calibrated}}$ is subsequently merged with the original input $X_{\mathrm{input}}$ through a residual connection, ensuring gradient stability and accelerating convergence:

$$X_{\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}}=X_{\mathrm{calibrated}}+X_{\mathrm{i}\mathrm{n}\mathrm{p}\mathrm{u}\mathrm{t}}$$

(5)

This residual fusion preserves shallow geometric cues while enhancing high-level semantic abstraction. Consequently, the AC-SENet adjustment module effectively balances detail sensitivity and structural robustness, yielding improved detection accuracy for ancient city sites under complex topographic and illumination conditions in the Qinghai–Tibet Plateau [66].

3.4.4. Integration of AC-SENet with the YOLOv11 Framework

After the optimization of AC-SENet, this network was integrated as the backbone of the YOLOv11 (https://github.com/ultralytics/ultralytics) detection framework to construct the AC-YOLOv11 (Ancient-City YOLOv11) model. In this configuration, the original YOLOv11 backbone was replaced by AC-SENet, while the Neck (PANet) and Head (Detection Head) structures were retained to ensure the stability of the detection pipeline. The overall architecture is illustrated in Figure 6. The AC-SENet extracts multi-scale hierarchical features (C3, C4, C5) from the pan-sharpened GF-2 imagery, which are then fused by the Path Aggregation Network (PANet) to enhance the flow of semantic and localization information across different feature levels. The resulting fused features are fed into the YOLOv11 detection head, where anchor-free detection layers predict bounding boxes and confidence scores through decoupled classification and regression branches.

The purpose of integrating AC-SENet into the YOLOv11 framework is to enhance multi-scale feature extraction, strengthen the representation of weak linear and polygonal structures, and reduce false detections in complex geomorphological backgrounds. The dual-path attention mechanism is designed to improve the response to rectangular fortification outlines, corner transitions, and trench-like edges, while the deformable convolution module provides adaptive feature calibration for irregular wall remnants and partially degraded enclosures [67]. In addition, the deep residual structure inherited from ResNet-152 offers strong representational capacity under small-sample conditions typical of archaeological remote sensing tasks. The comparative performance of AC-YOLOv11 and the baseline YOLOv11 is presented in Section 4.

4. Comparative Analysis and Results

4.1. Training Environment and Parameter Configuration

All model training and testing were performed on a local workstation environment running 64-bit Windows 11 (Microsoft Corporation, Redmond, WA, USA). The experiments were implemented using PyTorch 2.1.0 (https://pytorch.org) with the Python 3.9 programming language (https://www.python.org) [68,69]. The hardware configuration consisted of an NVIDIA GeForce RTX 4070 Ti GPU (NVIDIA Corporation, Santa Clara, CA, USA), with CUDA 11.2 support (NVIDIA Corporation, Santa Clara, CA, USA), and the entire workflow was managed within PyCharm 2023.3.7 (JetBrains s.r.o., Prague, Czech Republic).

To ensure the comparability and reproducibility of the results, no pre-trained weights were used, and all hyperparameters were kept fixed across runs. The dataset was divided into training and validation subsets in an 8:2 ratio, and model training was conducted using the parameter settings listed in

Table 1. Parameter settings for AC-YOLOv11 training.

Parameters	Setup
Epochs	200
Batch size	16
Workers	4
Learning rate	0.001
Optimizer	Adam
Imgsz	640
Ratio of training set to validation set	8:2

.

4.2. Evaluation Metrics

The model performance was evaluated using five standard metrics: loss function, precision, recall, mean average precision (mAP), and the F1 score (

Table 2. Evaluation metrics used for model performance assessment.

Metric	Formula	Definition	Purpose
Loss function (L)	$L=-{\displaystyle \sum _{i=1}^C}{y}_i·\mathrm{l}\mathrm{o}\mathrm{g}\left(p_i\right)$	Measures the error between predicted and ground-truth values across $C$ classes. In this single-class detection task, YOLO combines bounding box regression loss and distribution focal loss.	Evaluates training stability and effectiveness of learning rate scheduling.
Precision	$\mathrm{Precision}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$	Proportion of correctly predicted positives among all predicted positives.	Reflects the accuracy of positive predictions.
Recall	$\mathrm{Recall}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$	Proportion of actual positives correctly identified by the model.	Reflects the completeness of positive detection.
Mean Average Precision (mAP)	$\mathrm{mAP}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{\mathrm{A}\mathrm{P}}_i$	The mean of AP values across all N classes. Since N = 1, $mAP$ equals AP for the “Ancient City” class. Includes ${mAP}_{50}$ (IoU = 0.5) and ${mAP}_{50-95}$ (averaged across IoU thresholds from 0.5 to 0.95).	Provides a comprehensive measure of detection precision and localization accuracy.
F1 Score	$F_1=2·{\displaystyle \frac{\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}$	Harmonic mean of precision and recall.	Provides a balanced evaluation when precision and recall are uneven.

) [70,71,72].

4.3. Ablation Experiments

To verify the effectiveness of the proposed modules for ancient city detection, we conducted a series of ablation experiments on the Qinghai Lake ancient city remote sensing dataset (QHACD). The original YOLOv11 model served as the baseline. Using a controlled-variable approach, we incrementally introduced: (1) a ResNet-152 backbone, (2) a standard SE attention module, (3) the proposed Dual-Path Excitation Block (DPEB), and (4) the Deformable Feature Calibration mechanism. All experiments employed identical hyperparameters and data splits to ensure fair comparison [73]. Training settings were kept consistent: input image size = 450 × 450, initial learning rate = 0.001, optimizer = Adam, batch size = 16, and total epochs = 200. Evaluation metrics included Precision, Recall, F1-score, mAP@0.5 and mAP@0.5:0.95.

As shown in

Table 3. Ablation experiment results of AC-YOLOv11 on the QHACD.

ID	Model Variant	Precision (%)	Recall (%)	F1-Score (%)	mAP@0.5 (%)	mAP@0.5:0.95 (%)	Inference Speed (ms)
1	YOLOv11 (Original Backbone)	88.6	83.1	85.7	78.5	70.4	5.1–5.3
2	+ ResNet-152 Backbone	90.3	86.7	88.4	80.9	72.6	5.4–5.6
3	+ ResNet-152 + SE Attention	94.5	89.2	91.7	81.8	73.4	5.7–5.9
4	+ ResNet-152 + DPEB (Ours)	97.5	91.2	94.2	82.3	74.9	6.2

, the baseline YOLOv11 demonstrates moderate performance (Precision = 88.6%, Recall = 83.1%, mAP@0.5 = 78.5%). Replacing the backbone with ResNet-152 yields overall improvements, particularly in Recall and mAP. Incorporation of a standard SE attention module further increases Precision to 94.5%, indicating that channel attention enhances feature representation. However, SE alone offers limited robustness under complex terrain and noisy backgrounds. By contrast, the proposed DPEB achieves a more balanced enhancement across channel and spatial dimensions, yielding a Precision of 97.5%, Recall of 91.2%, F1 = 94.2%, and mAP@0.5/mAP@0.5:0.95 of 82.3% and 74.9%, respectively. These results demonstrate that the DPEB significantly improves the model’s ability to detect linear edges and corner features of ancient city remains in complex GF-2 imagery, validating the overall effectiveness of the AC-YOLOv11 architecture.

Compared with the baseline YOLOv11, integrating the DPEB module leads to a 9.9% relative increase in F1-score (from 85.7% to 94.2%) and a 3.8 percentage-point gain in mAP@0.5 (from 78.5% to 82.3%), indicating substantial improvements in both detection accuracy and robustness.

4.4. Model Comparison Experiments

To further validate the effectiveness of the proposed AC-YOLOv11 model for ancient city detection, we conducted a systematic comparison with several state-of-the-art object detection frameworks, including YOLOv3, YOLOv4, YOLOv7, and the lightweight EfficientDet-D3 [74]. All models were trained and evaluated under identical hyperparameter and data-split configurations to ensure fairness and comparability [75,76]. Evaluation metrics included Precision, Recall, F1-score, mAP@0.5, and mAP@0.5:0.95. As shown in

Table 4. Comparison of detection performance among different models on the QHACD.

Model	Precision (%)	Recall (%)	F1-Score (%)	mAP@0.5 (%)	mAP@0.5:0.95 (%)	Inference Speed (ms)
YOLOv3	89.1	84.3	86.6	78.9	70.5	7.8
YOLOv4	91.5	86.9	89.1	81.2	72.4	8.5
YOLOv7	92.3	88.1	90.1	82	73.5	5
EfficientDet-D3	90.2	85.4	87.7	80.1	72	12
AC-YOLOv11 (Ours)	97.5	91.2	94.2	82.3	74.9	6.2

, conventional YOLO variants exhibit stable performance in ancient city detection, with YOLOv7 outperforming earlier versions in terms of mean average precision [10]. EfficientDet-D3 demonstrates superior computational efficiency and inference speed but suffers from reduced accuracy under complex terrain, vegetation cover, and shadow interference [77]. In contrast, the proposed AC-YOLOv11 achieves the best results across all metrics, with Precision = 97.5%, Recall = 91.2%, F1 = 94.2%, mAP@0.5 = 82.3%, and mAP@0.5:0.95 = 74.9%. These outcomes clearly indicate that AC-YOLOv11 provides more accurate detection of city-wall segments, corners, and moat features in high-resolution GF-2 imagery, maintaining robustness even in low-contrast or topographically complex environments.

Compared with the best baseline (YOLOv7), AC-YOLOv11 yields a +7.4 percentage-point gain in Precision, +3.1 points in Recall, and a +2.9-point improvement in mAP@0.5, highlighting its superior balance between accuracy and generalization.

4.5. Full-Scale Evaluation and Model Performance

4.5.1. Training and Validation Losses

To comprehensively assess the performance of the proposed AC-YOLOv11 model in ancient city detection, we conducted full-scale training and evaluation on the QHACD. The model was trained using 450 × 450 input patches, the Adam optimizer, and 200 epochs. Both training and validation losses exhibit stable, monotonic convergence without signs of overfitting.

Specifically, the training losses—train/box_loss, train/cls_loss, and train/dfl_loss—decreased from 3.65/7.24/3.79 at epoch 1 to 0.98/0.61/1.89 at epoch 200 (Figure 7). Correspondingly, validation losses—val/box_loss, val/cls_loss, and val/dfl_loss—declined from 3.62/6.46/4.37 to 1.21/0.70/2.04. The most rapid loss reduction occurred within the first 40 epochs, followed by a gradual convergence phase, stabilizing after approximately 100 epochs. The learning rate decayed monotonically from 6.75 × 10⁻² to 1.50 × 10⁻⁵, consistent with the smooth convergence of loss curves, confirming the model’s stable optimization process.

4.5.2. Model Performance Metrics

Corresponding to the loss convergence trends, all major performance indicators of AC-YOLOv11 show continuous improvement throughout the training process (Figure 8) [78]. The values of Precision, Recall, mAP@0.5, and mAP@0.5:0.95 increased from 0.567/0.047/0.166/0.077 at epoch 1 to 0.956/0.932/0.967/0.843 at epoch 200, with a corresponding F1-score of 0.944. During the initial 20 epochs, the metrics improved rapidly, followed by a gradual enhancement phase between epochs 60 and 120, and stabilization after epoch 120. Several transient peaks were observed: Precision reached 0.983 at epoch 144, Recall peaked at 0.937 at epoch 196, while mAP@0.5 and mAP@0.5:0.95 achieved 0.972 and 0.849 at epoch 188, respectively. Considering both performance stability and generalization consistency, the final converged weights were adopted as the default model for subsequent evaluations.

During model training, the F1-score increased significantly with epoch progression (Figure 9). In the early stage (epochs 1–10), the model remained in the feature-learning phase, with F1 values fluctuating between 0.09 and 0.45. Between epochs 10 and 40, rapid convergence occurred, with F1 rising from 0.45 to 0.77. After epoch 60, the curve entered a steady growth phase, surpassing 0.90 after epoch 100 and reaching its maximum of 0.948 at epoch 188.

The evolution trend of F1 closely mirrors that of Precision and Recall, indicating that the model maintains a good balance between accuracy and completeness in detecting ancient city targets. Minor oscillations in the later epochs reflect ongoing optimization on high-confidence samples. Overall, the continuous rise and eventual stabilization of the F1-score confirm that AC-YOLOv11 achieves excellent convergence and generalization on the QHACD.

4.5.3. Validation Set Detection Results

In the QHACD validation set, the AC-YOLOv11 model achieved high-confidence detection of ancient city sites across most regions. The model not only accurately localized key structural elements such as rammed-earth walls, moats, and corner bastions, but also successfully identified multiple cities and their associated architectural units—such as gate enclosures (wengcheng), citadels, and inner–outer city layouts—within a single GF-2 scene (Figure 10).

Across representative sites in the Qinghai Lake Basin—Shinaihai, Jiangxigou, Ganzihekou, and the Quanji River valley—the model demonstrated high confidence and precise spatial localization results that closely matched archaeological ground-truth positions. Moreover, the AC-YOLOv11 exhibited robust discrimination between archaeological remains and environmental or anthropogenic features such as natural landforms, farmlands, and modern settlements. Even in imagery characterized by strong topographic relief, dense vegetation cover, or uneven illumination, the model maintained consistent detection performance, confirming its robustness to blurred boundaries and weak-texture targets.

The average inference time was approximately 6.2 ms per image (excluding I/O), indicating the model’s computational efficiency and suitability for large-scale archaeological site recognition and rapid regional mapping [79].

5. Discussion

5.1. Application and Detection Results of the Model in the Qinghai Lake Basin

After completing model training and validation, the AC-YOLOv11 model was applied to pan-sharpened GF-2 satellite imagery covering the Qinghai Lake Basin to evaluate its applicability and stability for large-scale archaeological site recognition. The imagery has a spatial resolution of 0.8 m and covers an area of approximately 7000 km². With a confidence threshold of 0.75 and an input size of 450 × 450 pixels, the model automatically performed inference on all image tiles across the study area.

A total of 309 potential ancient city targets were detected, including complete rectangular enclosures, degraded rammed-earth walls, moat-like linear features, and corner bastion structures (Figure 11). Following manual screening and archaeological background verification, 74 sites were identified as highly probable ancient city remains. These are primarily distributed along the western shore of Qinghai Lake, the Buha River Basin, and the Qiabuqia–Daotang River corridor (Figure 12).

Overall, the AC-YOLOv11 model demonstrated stable performance in complex surface environments. It exhibited strong capability in extracting high-response morphological features such as walls and moats, while maintaining high detection confidence even in areas with surface fragmentation, dense vegetation cover, or modern construction disturbance. Some detection results further revealed the presence of multiple adjacent settlement units—including main fortresses, subsidiary gate enclosures, and satellite settlements—within a single detection window, suggesting the model’s potential for hierarchical settlement system identification in archaeological landscape analysis [80].

5.2. Field Survey and Verification

To validate the accuracy and archaeological relevance of the AC-YOLOv11 detection results, eight high-confidence suspected ancient city sites were selected across the Qinghai Lake Basin for on-site field surveys and artifact sampling. Field investigations included documentation of surface features, collection of pottery and porcelain fragments, geomorphological observations, and comparative analysis with known archaeological sites and historical records. The purpose of the field verification was to evaluate the model’s practical reliability under different landscape conditions and to clarify how well its detections correspond to genuine archaeological features.

Based on the fieldwork and cross-validation with archaeological data, three of the eight locations were confirmed as ancient city remains, one was identified as a peripheral settlement associated with Fusi City, and four sites lacked sufficient surface evidence to be classified as ancient cities (

Table 5. Field verification results of suspected ancient city sites detected by the AC-YOLOv11 model.

ID	Coordinates (Lat, Long)	Distance to Known Site	Surface Remains	Surface Artifacts	Geomorphic Setting	Final Interpretation
1	37°02′57.23″N, 99°31′47.39″E	~5.2 km from Fusi City	Rammed-earth wall segments and low stone foundations; irregular plan	Painted porcelain sherds	Piedmont alluvial-fan terrace; gentle slope	Non-ancient city (modern disturbance)
2	37°02′44.10″N, 99°32′15.70″E	~4.4 km from Fusi City	Fragmented rammed-earth wall bases with fissures and voids	Red sandy pottery and glazed ware	Piedmont alluvial-fan terrace	Non-ancient city
3	37°01′25.49″N, 99°34′26.08″E	~0.7 km SW of Fusi City	Low mounds and exposed rammed layers	Dark-red sandy pottery, brown-green glaze	Second terrace on north bank of Qieji River	Peripheral settlement of Fusi City
4	37°01′30.9″N, 99°36′14.5″E	~1.2 km SE of Fusi City	Rammed-earth platforms and low wall remains	Decorated gray pottery with cord and grid patterns	Terrace of the Qieji River; flat terrain	Outer city of Fusi (east wall, southern section)
5	36°52′15.00″N, 100°46′12.87″E	~20 km from Xihai County City	Rammed-earth wall (height ~1 m)	Red sandy and brown-glazed pottery	Terminal fan terrace	Non-ancient city
6	37°10′40.86″N, 99°44′10.32″E	~16 km from Beixiangyang City	Elongated mound with residual rammed base	Dark-red sandy pottery, brown-green glaze	Junction of lake terrace and piedmont fan	Non-ancient city
7	37°22′17.45″N, 98°48′55.93″E	~4.5 km from Jinquan City	Earthen mounds with traces of compaction	Gray and glazed pottery, red sandy ware	Terrace near Buha River	Non-ancient city
8	37°28′16.42″N, 98°36′8.20″E; 37°28′10.41″N, 98°35′56.38″E	~200 m and 800 m from Dalai-Mane No. 2 City	No visible rammed-earth remains; covered by grass	No artifacts found; no cultural layer	Second terrace on north bank of Buha River	Ancient city remains (Dalai-Mane No. 3 and No. 4)

and

Table 6. Comparison between model-detected sites and field verification results in the Qinghai Lake Basin.

ID	Model Output	Field (Overview)	Field (Detail)	Artifacts
1
2
3
4
5
6
7
8

). Among them, Site 3, southwest of Fusi City, contained rammed-earth sections and sediment accumulations but no decorated gray pottery; it was therefore interpreted as a peripheral settlement of Fusi City. Site 4 preserved low rammed-earth platforms and wall foundations, yielding cord-marked and net-patterned gray pottery identical to finds from the inner Fusi City, confirming its identity as the southeastern segment of Fusi City’s outer-east wall. Site 8, located north of Dalai Mani II City, showed no exposed cultural layers or surface artifacts; however, satellite imagery revealed a clear square-enclosure pattern, and combined with epigraphic clues, the site is tentatively attributed to the Northern and Southern Dynasties period and provisionally named Dalai Mani III and IV Cities [81].

Through analysis of the results, the false-positive cases fall primarily into three categories: (1) natural terraces and fan surfaces with sharp slope breaks or rectilinear erosion edges, which resemble degraded rammed-earth walls in GF-2 imagery; (2) agricultural grids, whose regular rectangular plots and high-contrast boundaries mimic the planimetric patterns of ancient enclosures; (3) abandoned construction sites or industrial platforms, which exhibit artificial right-angled edges and compacted surfaces similar to archaeological ramparts [82].

These misidentifications reveal how AC-YOLOv11 internally prioritizes morphological cues. Visual inspection shows that the model responds strongly to continuous linear edges, corner angles, and banded micro-relief—diagnostic traits of ancient fortified structures. When similar geometric features occur in natural or modern contexts, the network may generalize its learned representations and assign high confidence to non-archaeological targets. This not only explains the specific cases of misclassification but also demonstrates that the model is extracting meaningful archaeological morphology rather than relying on spurious texture patterns.

Overall, the field verification confirms that AC-YOLOv11 is capable of detecting archaeologically significant remains while maintaining robustness in complex geomorphic environments. At the same time, the error analysis highlights the need to further enrich the negative-sample set—especially with terraces, agricultural grids, and modern infrastructures—to enhance the model’s discriminative ability in future iterations.

5.3. Spatial Pattern of Ancient City Distribution and Its Relationship with the Physical Environment

By integrating 37 confirmed ancient cities with 74 highly probable model-identified sites, the spatial analysis reveals a clear coupling between settlement distribution and hydro-geomorphic conditions in the Qinghai Lake Basin [83]. Approximately three-quarters of all sites are located within 10 km of major rivers or the lakeshore, and more than 90% fall within the 3500–4000 m elevation belt with gentle slopes below 2° (Figure 13, Figure 14 and Figure 15). This concentration along low-slope lacustrine plains, river terraces and piedmont fans indicates that ancient polities deliberately selected locations that balanced access to water resources, arable land and pastures with the need for defensive visibility and transport connectivity [84,85].

These patterns echo previous archaeological and palaeo-environmental studies that emphasize the role of the Qinghai Lake Basin as a key agro-pastoral transition zone and a strategic hub along the “Qinghai Road” of the Silk Road [86,87]. The clustering of cities along river corridors such as the Buha River and the Qiabuqia–Daotang River corridor suggests the existence of a corridor-like settlement system linking the interior plateau with the Hexi Corridor and the Central Plains [88]. Against the background of fluctuating lake levels and Holocene climate variability, the observed distribution thus reflects long-term human adaptation to water availability, terrain constraints and inter-regional exchange demands in a high-altitude arid environment [89,90].

6. Conclusions

This study proposes AC-YOLOv11, a deep learning framework specifically designed for the automatic detection of ancient city sites in high-resolution satellite imagery on the northeastern Tibetan Plateau. By integrating a ResNet-152 backbone with squeeze-and-excitation attention, the Dual-Path Excitation Block, and deformable feature calibration, the model effectively enhances the recognition of eroded rectangular enclosures, linear wall segments, and weak geomorphic traces characteristic of archaeological remains in this region. Experimental results demonstrate that AC-YOLOv11 consistently outperforms YOLOv3, YOLOv4, YOLOv7, YOLOv11 and EfficientDet-D3 in both accuracy and robustness, while maintaining an efficient inference time suitable for large-area archaeological prospection.

Application of the model to the Qinghai Lake Basin generated 309 candidate detections, among which 74 were identified as highly probable ancient city remains. Field validation at selected locations confirmed several previously undocumented sites, highlighting the method’s potential to substantially expand archaeological inventories in areas where traditional survey coverage is limited. The spatial analysis further reveals that ancient cities cluster along low-slope lacustrine plains and river corridors within the 3500–4000 m elevation belt, offering new insights into high-altitude settlement strategies and human–environment interactions on the northeastern Tibetan Plateau.

Despite its advantages, the framework remains constrained by the limited number of positive samples and the reliance on high-resolution commercial data. Future work will focus on expanding training datasets through multi-regional collaboration, incorporating multi-temporal or multi-sensor imagery, and exploring semi-supervised learning strategies to enhance adaptability and reduce annotation requirements. Nevertheless, the present results demonstrate that AC-YOLOv11 provides a practical, interpretable and scalable approach for archaeological remote sensing, offering meaningful support for cultural heritage assessment, site monitoring and early-stage prospection in complex plateau environments.