Research on Linpan Identification in Chengdu Plain Based on Object Detection Technology (2016–2023)—A Case Study of PiDu District

Abstract

Tens of thousands of ordinary traditional settlements remain clustered within specific geographic regions of China. Efficient and objective rapid identification of these settlements is crucial for preserving rural cultural heritage. This study takes the traditional settlement Linpan in the Chengdu Plain as a case study, focusing on Pidu District of Chengdu City in Sichuan Province, and proposes an innovative approach for rapid large scale surveys of common traditional settlements using object detection technology. Based on the technical requirements, the spatial characteristics of Linpan settlements in the Chengdu Plain were refined. High-resolution satellite images from 2016 and 2023 of Pidu were processed and cropped, and a diversified training dataset was constructed. After annotation, multiple rounds of training were conducted to develop a detection model based on YOLOv11. The model was then applied to identify thousands of rural settlements across the 438 km² area of Pidu, followed by an evaluation of various detection parameters. The results demonstrate that this method can complete the identification of Linpan settlements across the entire Pidu in just 6–7 min, achieving a precision of 96.59% and a recall rate of 94.39%. In terms of efficiency and accuracy, this approach significantly outperforms visual interpretation and remote sensing interpretation methods. Furthermore, based on the detection results, the spatiotemporal distribution characteristics of Linpan settlements during the study period were analyzed. This study aims to improve the surveying methods for traditional villages sand advance their conservation from “static observation” to “dynamic analysis”.

1. Introduction

“Linpan” is a distinctive rural settlement in western Sichuan, China, integrating production, living, ecology, culture, and landscape. Typically structured from outer to inner layers as “farmland-grove-dwelling-courtyard,” these settlements are interlaced with rivers and canals. Varying in size, they accommodate several to dozens of households, forming disk-shaped green islands scattered across the Chengdu Plain through its networked irrigation system (Figure 1).

Originating from the Qin and Han dynasties, Linpan has evolved over two millennia into a rural settlement unit that perfectly blends the Sichuan people’s livelihood, cultural beliefs, and social organization. It stands as an outstanding example of traditional human habitation; a unique testimony to agricultural civilization; the representative form of Han Chinese rice-farming culture harmonized with nature; and the residential expression of agricultural civilization. Meeting multiple UNESCO criteria for cultural heritage, Linpan possesses significant potential for heritage designation.

Like other contiguous traditional settlement clusters, the Linpan settlements face bottleneck issues in spatiotemporal preservation. The sheer number of Linpan is substantial—according to the 2014 government census, 121,100 such settlements remain. They are widely distributed across the Chengdu Plain, a 9100-square-kilometer alluvial basin formed by the Min and Tuo Rivers between the Longmen and Longquan Mountains in western Sichuan Basin [1]. Linpan exhibit diverse and complex forms: their sizes vary from clusters of a few households to dozens, and their layouts include circular, linear, fan-shaped, and irregular patterns shaped by topography. As living heritage sites, most rural residents of the Chengdu Plain inhabit these settlements. The 2014 census recorded 3.6556 million people living in Linpan across Chengdu, accounting for 22.38% of the city’s total population and 72.4% of its rural population [2].

Over the past decade, urbanization and rural revitalization have dramatically altered the quantity and distribution of Linpan. However, tracking these spatiotemporal changes remains challenging due to the settlements’ vast scale, extensive distribution, and morphological complexity, which make comprehensive censuses prohibitively costly. Since the 2014 survey, apart from the 6645 large and medium-sized Linpan officially designated for protection, no further comprehensive survey data or analysis exists for the remaining settlements.

Academic research on Linpan in the Chengdu Plain began in the 1980s. After 2004, the distinctive regional cultural value of Linpan gained widespread recognition, leading to a gradual expansion of studies in this field. Research on the spatial patterns of Linpan within the region has primarily relied on methods such as visual interpretation and field surveys to identify these settlements among various rural habitats. However, such census approaches are high-cost, time-intensive, and outdated, making data updates difficult [3,4,5,6]. Most studies cover limited areas, typically focusing on township or county-level units, with some county-level research forced to use data as old as seven years [7].

For multi-county studies, sampling methods have been adopted, making comprehensive coverage challenging [8]. Visual interpretation carries significant margins of error: one scholar visually identified 2961 Linpan in Pidu District in 2016 [9], while another using the same method counted 6239 Linpan in the same region during a similar period (2015) [10]. A scholar has also employed remote sensing interpretation techniques to identify forest patches across the entire Chengdu region. However, due to the inability to exclude non-forest patch elements within rural construction areas (such as agricultural plastic greenhouses, factory buildings, and resettlement communities), the identification precision remains relatively low, reaching merely 56.13% to 66.72% across different years [11].

In China’s specific geographical regions, numerous region-specific traditional settlements akin to Linpan are distributed across the landscape. Areas such as southeastern Guizhou, southern Fujian, and the Taihang Mountain region feature concentrated and contiguous traditional village settlements that form large-scale aggregations. Due to complex terrain and limited transportation access, traditional census methods struggle to promptly identify tens of thousands of such clustered traditional settlements. How to rapidly survey and identify traditional settlements across extensive regions while tracking and analyzing their spatiotemporal changes has remained a persistent challenge in this field.

A limited number of existing studies have moved beyond the limitations of official registries to examine the spatial distribution of numerous ordinary traditional settlements within specific regions. Defining such research subjects proves difficult: some studies omit comprehensive village counts [12], while others employ integrated approaches combining literature review, field surveys, and satellite image interpretation—yet these methods bring only a limited number of non-registered villages into research focus [13].

Therefore, for the investigation and identification of a large number of traditional villages over broad areas, a fast and accurate automated method is needed. The key task is to distinguish village patches from various built-up areas in rural regions, and further identify historical villages with traditional morphological characteristics from these village patches. Some studies have focused on village identification, but their methods can only detect built-up areas in large-scale satellite imagery, without being able to differentiate actual villages [14].

Deep learning-based computer vision techniques can directly learn the morphological features of objects from labeled images and understand their deep-level shape characteristics, making them well-suited to handling the complexity of human settlements. In recent years, such technologies have been increasingly applied in the study of rural settlement identification.

For example, in the field of image classification, the DeepLabV3+ algorithm has been used to identify ethnic architecture in aerial images of traditional villages [15]; convolutional neural network (CNN) techniques have been applied to create a traditional village feature database and classify a large number of current photos by province [16] or to identify heritage value elements of traditional villages from current photographs [17].

In the field of image segmentation, these methods have been used to identify and delineate key buildings in aerial imagery [18,19,20], segment objects such as roads, villages, farmlands, and mountains in remote sensing imagery [21,22], distinguish village from non-village areas [23,24], identify and map the boundaries of urban villages [25], map key buildings within urban villages [26], and propose an automated method for extracting traditional village buildings in northern China [27].

Image classification methods extract global statistical features and provide single-output, image-level semantic understanding, which is used to classify the types and quantities of objects in an image. Image segmentation methods focus on the pixel neighborhood features and assign semantic labels to each pixel in the image, producing dense spatial outputs and are often used for identifying the precise shape and boundaries of objects. However, to identify settlements with certain characteristics within a region, which requires instance-level localization and classification and outputs sparse object descriptions, object detection methods are more appropriate.

Object detection is a major research direction in computer vision, aiming to identify objects in images or videos and mark their locations and categories. It has been widely applied in fields such as lesion detection in medical imaging [28], real-time traffic environment monitoring [29,30], and plant disease detection [31,32]. Among the mainstream object detection algorithms, the deep learning-based YOLO (You Only Look Once) algorithm treats object detection as a regression problem, using a single neural network to directly predict bounding boxes and class probabilities from the entire image. Traditional settlements in specific areas have certain planar visual morphological features. By accurately labeling a number of satellite images of such areas and building a YOLO-based model, traditional settlements can be rapidly identified across larger regions and their locations marked. This enables accurate and comprehensive tracking of the spatiotemporal evolution of traditional settlement clusters and supports dynamic surveying and protection monitoring. This approach integrates morphological research from architecture, spatiotemporal analysis from geography, and deep learning techniques from computer vision. It has significant value for quickly identifying traditional villages among various rural settlements, clarifying their quantity and spatial distribution, analyzing the spatiotemporal evolution of traditional settlement spatial structures, and dynamically monitoring the status of traditional settlement morphology and protection.

Currently, such research is relatively limited. One study focused on identifying traditional villages in Conghua District, Guangdong Province, and compared image classification and object detection algorithms, finding that the Precision of the latter (95.61%) was significantly higher than that of the former (90.76%) [33]. Another study applied an early version of YOLO to recognize the morphology of Tulou in a small region of southwestern Fujian [34]. Object detection techniques can also be used to identify key buildings in a region, such as counting and dynamically monitoring the number of buildings in traditional villages [35], or identifying surviving vernacular houses within villages [36].

This efficient and objective automated analysis method will serve as a vital tool for detecting ordinary traditional settlements. Deep learning-based object detection technology has achieved remarkable progress in recent years and has been widely applied in fields such as lesion detection in medical imaging [28], real-time traffic environment monitoring [30], and plant disease identification [31,32].

The introduction of object detection technology in regions with concentrated traditional settlement clusters is of great significance for enabling rapid identification of traditional settlements among diverse rural habitats, mapping settlement quantities and spatial distribution patterns, analyzing spatiotemporal changes in their spatial structures, and facilitating dynamic monitoring of morphological conservation status.

This study selects a representative area of the Chengdu Plain (Pidu District) to conduct research on automated traditional settlement detection, aiming to address this challenge. The research holds not only theoretical and practical significance for Linpan identification and preservation in the Chengdu Plain but also provides reference value for rapidly detecting traditional settlements in similar contexts across China and internationally.

This study proposes an innovative approach for rapid large-scale surveying of ordinary traditional settlements using object detection technology. Specifically, we employed the YOLOv11 (You Only Look Once version 11) object detection model to address Linpan identification in Pidu District, evaluating a series of post-detection parameters including accuracy, F-measure, precision, recall, and mean Average Precision (mAP). Furthermore, the study assessed the model’s computational efficiency and influencing factors. YOLO’s ability for rapid response coupled with its relatively high accuracy may position it as the preferred choice for researchers conducting rapid surveys of diverse urban-rural settlement morphologies.

The primary contributions of this research are as follows: First, we adopted an interdisciplinary methodology integrating computer vision technology with architectural surveying. Second, based on the requirements of object detection technology, we distilled the spatial characteristics of Linpan settlements in the Chengdu Plain to facilitate rapid annotation for the identification system. The study extracted and cropped high-precision satellite imagery of Pidu District from 2016 and 2023, established a diversified training set according to the refined spatial features of Linpan, and conducted multiple rounds of annotation, training, and model optimization.

The results demonstrate that conducting object detection for Linpan across the entire Pidu District required only approximately 6 min, achieving precision rate is above 95%, the recall rate is above 94%, and the mean average precision (mAP) is 0.73.

The subsequent sections of this paper are structured as follows: Part II describes the case study area and research methodology; Part III presents the results and analysis; Part IV provides the discussion and conclusions.

2. Case Study Area and Research Methodology

2.1. Study Area

Located in the northwest of Chengdu, Sichuan Province (Figure 2), Pidu District occupies the heartland of the Chengdu Plain between 103°42′–104°02′ E and 30°43′–30°52′ N, covering a total area of 437 km² and administering 3 towns and 9 sub-districts, with a permanent population of 1.697 million as of late 2023 (Chengdu Municipal Statistics Bureau, 2024), characterized by a subtropical humid monsoon climate.

As the case study area, Pidu District exhibits two distinctive structural characteristics: it functions as both a traditional agricultural heartland and an emerging industrial zone. Firstly, situated in the core of the Chengdu Plain with fertile soil and extensive water networks, Pidu possesses abundant agricultural resources supporting a long-established and advanced farming sector. Dominated by rice, vegetables, and ornamental horticulture, its specialty product Pixian doubanjiang (broad bean paste) enjoys nationwide renown. Additionally, the district hosts 8700 diverse Linpan settlements according to 2012 census data (data source: Chengdu Urban and Rural Planning and Design Research Institute), which serve not only as agricultural foundations but also as vital cultural and ecological components. Thus Pidu represents a quintessential model of agricultural settlement evolution in the western Sichuan Plain.

Secondly, the elongated district borders Chengdu metropolis to its east. Amid accelerating urbanization, industrial zones and urban expansion have dramatically transformed traditional settlements—creating extensive urban built-up areas in the east while preserving substantial traditional Linpan and rural clusters in the northwest. This spatial dichotomy positions Pidu as a prime example of traditional settlement transformation under urban encroachment.

Furthermore, Pidu offers rich data for comparative Linpan studies, with multiple scholars having documented its settlements across different years using varied methodologies. This extensive research legacy enables robust comparison between our object detection results and prior findings.

2.2. Research Methods

For the linpan detection task, YOLOv11 demonstrates superior applicability compared with earlier YOLO models. Linpan targets often exhibit complex shapes, scale variations, irregular boundaries, and significant vegetation interference. By incorporating a more efficient multi-scale feature fusion structure and lightweight design, YOLOv11 markedly enhances the perception and localization of small and irregular objects. It maintains high detection accuracy while substantially reducing parameter size, thereby improving inference efficiency on typical satellite imagery. Moreover, the optimization strategies adopted during training strengthen the model’s generalization in complex natural scenes, enabling more accurate and stable recognition of linpan structures. Consequently, YOLOv11 achieves a better balance of precision and efficiency for detection tasks.

This study employs the latest YOLOv11 object detection model. Compared to its predecessors (e.g., YOLOv8), this model has been comprehensively optimized in terms of architectural design and training strategies. YOLOv11 enhances the Backbone and Neck structures by adopting more efficient feature extraction modules, significantly improving detection precision and generalization capability in complex scenes. Simultaneously, its optimized training pipeline further accelerates model convergence speed and achieves a better balance between computational efficiency and detection performance.

Experiments demonstrate that YOLOv11 achieves superior mean Average Precision (mAP) over YOLOv8 on the COCO dataset. Furthermore, YOLOv11m (the medium-sized variant) reduces the parameter count by 22% compared to YOLOv8m, substantially boosting computational efficiency while maintaining high precision. These improvements make it particularly suitable for object detection tasks in forest plot aerial imagery, enabling faster and more accurate identification and analysis of complex ground features (Figure 3).

2.3. Research Workflow

The object detection process for Linpan using YOLOv11 consists of three phases: data preprocessing, model processing, and result presentation with visualization (Figure 4).

2.3.1. Data Preprocessing Phase

This stage begins with literature review and data collection to ensure comprehensive and diverse datasets. As the original images were excessively large, a dedicated program was developed to tile them into training-appropriate dimensions. These processed images, combined with Linpan images sourced from online repositories, formed the training dataset.

2.3.2. Model Training Phase

This phase encompasses model architecture design, training, evaluation, and optimization. The model was constructed based on features from the training set, with continuous hyperparameter tuning to enhance performance.

2.3.3. Result Presentation and Visualization

Final detection outcomes and performance metrics were visually demonstrated to evaluate the model’s effectiveness.

2.3.4. Result Verification

The recognition results were validated to ensure their reliability. Key evaluation metrics included precision, recall, and F1 score.

Precision reflects the proportion of correctly predicted samples among the total samples. Its formula is:

$$\mathit{Precision} = {\displaystyle \frac{\mathit{TP}}{\mathit{TP}+\mathit{FP}}}$$

(1)

TP (True Positive): The number of samples predicted as positive and actually positive.

FP (False Positive): The number of samples predicted as positive but actually negative.

Recall reflects the proportion of actual positive samples that are correctly predicted by the model. Its formula is:

$$\mathit{Recall} = {\displaystyle \frac{\mathit{TP}}{\mathit{TP}+\mathit{FN}}}$$

(2)

FN (False Negative): the number of samples predicted as negative but actually positive.

F1Score, harmonic mean of precision and recall.

$$\mathit{F}1\mathit{Score}={\displaystyle \frac{2 \times \mathit{Precision} \times \mathit{Recall}}{\mathit{Precision}+ \mathit{Recall}}}$$

(3)

2.3.5. Spatial Analysis of Recognition Results

To investigate the spatial distribution and evolution trends of linpan in Pidu District, the recognition results were analyzed using kernel density estimation and average nearest neighbor analysis.

2.4. Data Collection

This study utilizes data sourced from Google Maps satellite imagery and aerial photographs from UAV surveys.

The research focuses on linpan settlements in Pidu District, Chengdu as the study area. This region not only represents a typical exemplar of West Sichuan’s agricultural heritage but also constitutes the core concentration zone of linpan clusters, exhibiting distinctive regional characteristics and significant research value.

To comprehensively capture the spatiotemporal evolution patterns of linpan landscapes, the research team acquired high-resolution imagery of Pidu District for two temporal periods (2016 and 2023) through the Google Maps platform (Figure 5). This establishes a temporally comparative research framework for longitudinal analysis.

3. Construction of Joint Rules for Linpan

3.1. Feature Extraction of Linpan

Pidu District’s rural areas feature diverse building patches, including single-household farms, multi-household settlements, modern resettlement communities, factories, agricultural plastic greenhouses, small market towns, and agricultural facilities. These structures exhibit significant differences from traditional linpan in spatial organization, functional layout, and architectural form. They lack the distinctive settlement pattern, farmland-surrounded characteristics, and ecological embeddedness inherent to linpan.

Prior to training set annotation, a “Conjunctive Rule” must be established to distinguish linpan from other rural building patches. Linpan patches must satisfy all predefined criteria of this rule (e.g., area, aspect ratio), with any unmet condition leading to exclusion.

To demonstrate the urban planning applications of object detection, this study engaged laypersons in the annotation process. Consequently, the labeling rules were designed to be intuitive, straightforward, and non-technical (

Table 1. Joint Rules for Linpan Annotation Oriented to Object Detection.

Rule Category	Morphological Element	Judgment Rule	Remarks
Linpan Scale	Number of Buildings	≥3
External Morphology	Aspect Ratio [9] (λ, form compactness index)	2 > λ >0.5	λ is the aspect ratio of the settlement cluster
	Vegetation Coverage Rate	>20%	Vegetation area within the settlement cluster form
	Tree Enclosure Degree	>25%	Degree of tree enclosure around the settlement cluster boundary
Internal Morphology	Internal Axes	None	Excluding modern residential communities
	Geometric Road Network	None	Such as arc-shaped or fishbone-shaped road networks (characteristic of modern residential communities)
	Large-Scale Internal Buildings	None	Excluding industrial buildings
	Building Roof Color	Light color ≤ 1/5	Excluding modern buildings

Notes: Among the 16 sampled Linpan settlements in the Western Sichuan Plain, only two mountainous Linpan settlements exhibit a length-to-width ratio (λ) greater than 2 [9].

and

Table 2. The diagram of Joint Rules for Linpan Annotation Oriented to Object Detection.

Element	Introduction	Linpan	Non-Linpan
Buildings	Among them, Linpan are classified into single-household courtyards (less than 3 households), Linpan (3–40 households),and large-scale settlements (more than 40 households).
Aspect Ratio	Residential houses in Western Sichuan can be classified into linear settlements (λ < 0.5 or 2 < λ) and Linpan(2 > λ > 0.5) according to the aspect ratio λ of their external morphology.
Vegetation Coverage Rate	Residential houses in Western Sichuan can be classified into Linpan (vegetation coverage rate > 20%) and ordinary residential houses (vegetation coverage rate < 20%) according to the vegetation coverage rate.
Tree Enclosure Degree	Residential houses in Western Sichuan can be classified into Linpan (tree enclosure degree > 25%) and ordinary residential houses (tree enclosure degree < 25%) according to the tree enclosure degree.
Internal Axes	Residential houses in Western Sichuan can be classified into modern residential communities and Linpan settlements according to the presence or absence of internal axes and geometric road networks.
Geometric Road Network	Residential houses in Western Sichuan can be classified into Linpan (no geometric road network inside) and ordinary residential houses (have geometric roads network inside) according to the tree enclosure degree.
Large-Scale Internal Buildings	Residential houses in Western Sichuan can be divided into Linpan and modern settlements according to the presence of large buildings inside.
Building Roof Color	Residential houses in Western Sichuan can be divided into Linpan and modern settlements according to the presence of colorful building roofs inside.

).

3.2. Data Tiling

Given the extensive scale of the study area, direct annotation and analysis of the entire region would incur low data processing efficiency and excessive computational resource consumption. To enhance research precision and workflow efficiency, this study developed an automated tiling program using Python 3.70 and OpenCV-based image processing libraries.

The program divides raw map imagery into standardized 1000 × 1000-pixel JPEG units according to predefined parameters. This process generated an image array comprising 22 rows, with each row containing 31 sub-images, yielding 682 standardized tiles (Figure 4).

This gridded processing approach achieves three critical objectives: Reduces per-image data volume for streamlined deep learning model input. Ensures complete spatial coverage with seamless segmentation. Establishes a standardized data foundation for fine-grained annotation and quantitative analysis. Through this standardized preprocessing, spatial detail characteristics of original imagery are preserved while significantly improving large-scale image processing efficiency.

3.3. Data Filtering

During dataset construction, an initial image collection was generated using a grid-based cropping method with strict quality control. The dataset covers high-resolution Google Earth satellite imagery (zoom level 17) from 2016 and 2023, yielding 1364 preliminary JPG images. Using automated Python scripts combined with manual verification, invalid images containing blank margins or unclear linpan contours were removed. A total of 482 valid images were finally selected as training samples, accounting for approximately 35.35% (347 from 2016 and 135 from 2023). These samples maintain clear landform boundaries and consistent annotation quality, standardized to 1000 × 1000 pixels in JPG format. The resulting high-quality dataset ensures both representativeness and reliability for linpan detection model training.

Subsequently, in accordance with the Conjunctive Rule (Table 1), all samples were annotated using the Label Img tool with standardized labeling protocols. Each identified linpan instance was uniformly tagged with the category label “linpan” (Figure 6), with annotations converted into YOLO-compliant format (each image corresponding to a .txt file containing normalized target center coordinates, width, and height). To ensure data quality, a dual verification mechanism was established: Precise annotation of linpan targets. Cross-validation to guarantee labeling consistency.

Additionally, statistical analysis was performed on the dataset’s class distribution. Sample augmentation techniques were applied to optimize data balance, thereby providing standardized, high-quality data support for subsequent YOLO-series object detection model training. This methodical framework ensures both annotation accuracy and alignment with practical model training requirements.

To further enhance the accuracy and generalization ability of the image recognition model, the team systematically collected and organized publicly available auxiliary materials such as Linpan floor plans and structural diagrams from online sources. These multi-source data not only enrich the diversity of training samples but also provide a more comprehensive basis for feature learning in deep learning models. The integration of multi-temporal remote sensing images with professional floor plans has laid a solid data foundation for subsequent research on forest settlement morphology identification, pattern analysis, and evolution studies.

4. Linpan Detection Model

4.1. Parameter Configuration

The initial training epochs were set to 70, a value validated to ensure model convergence while preventing overfitting. The batch size was set to 16 to maximize memory utilization on the NVIDIA RTX A2000 22 G GPU (22 GB memory; NVIDIA Corporation; Santa Clara, CA, USA). The input image size (img-size) was adjusted to 1000 × 1000 pixels—this super-resolution setting specifically addresses multi-target detection requirements in the scenario, enhancing small-object recall rates by preserving finer feature details. Considering data loading efficiency and stability, the number of data loading workers was temporarily set to 0 to avoid potential CUDA memory conflicts.

4.2. Model Training

4.2.1. Model Training Procedure

In the YOLOv11-based Linpan morphology object detection task, the Label Img tool was first employed to annotate multi-scale and multi-category Linpan targets within images, generating standardized YOLO-format annotation files. These data were imported into the model training pipeline via the yolov11.yaml configuration file.

The model’s backbone innovatively adopts C3K2 modules to replace traditional C2f blocks. Through dynamic convolution and cross-stage connection techniques, this module significantly enhances the model’s feature extraction capability for multi-scale targets against complex backgrounds in forest settlement environments.

Features are then fed into the neck for multi-level feature fusion and enhancement, utilizing an improved Bidirectional Feature Pyramid Network (BiFPN) architecture. Its weighted feature fusion mechanism effectively integrates semantic information and detailed features across different hierarchical levels.

The final head outputs detection results across three dimensions: bounding box coordinates, object categories, and confidence scores. The bounding box prediction is optimized using CIoU Loss, while Focal Loss addresses class imbalance in classification tasks.

Upon training completion, the model outputs two weight files: best.pt (optimal validation performance) and last.pt (final training state), enabling efficient and accurate intelligent monitoring and analysis of forest settlements (Figure 7).

To enhance the accuracy of the forest settlement target detection model, we implemented a dual-track optimization strategy, systematically refining the YOLOv11 model through coordinated increases in both epoch count and training sample size.

During deep learning model training, epoch configuration requires scientific calibration rather than simplistic “more-is-better” approaches. Experimental findings reveal that when training iterations exceed the optimal convergence point, models tend to over-memorize intricate details of training data. This triggers regularization effects where high accuracy on training sets contrasts with performance degradation on validation/test sets. To systematically evaluate epoch count’s impact on forest settlement detection performance, we designed a controlled variable experiment: Under identical hardware environments and datasets, three configurations were tested for 70 epochs (baseline group), 200 epochs (standard group), and 300 epochs (extreme group).

Sample size expansion does not guarantee better outcomes. We observed that when sample volume exceeds a certain threshold, maintaining annotation quality becomes significantly more challenging, potentially introducing detrimental noise that compromises model performance. To systematically evaluate the impact of temporal relevance and data diversity on forest settlement detection, we designed two comparative experiments:

Group 1: Temporally homogeneous data (176 samples collected exclusively in 2016).

Group 2: Spatiotemporally mixed data (404 samples from 2016 and 2023, supplemented with 30 web-sourced Linpan floor plans).

Experimental results revealed that despite larger volume in Group 2, the validation set mAP decreased by 2.5 percentage points compared to Group 1 (

Table 3. Training mAP Results Summary Table.

	70 Epoch	200 Epoch	300 Epoch
Images from 2016 only	0.500	0.609	0.765
2016, 2023, network images	0.600	0.665	0.730

). This indicates that in sample selection, we should focus more on the consistency of data quality and the matching degree of spatiotemporal features, rather than blindly pursuing sheer volume. Moving forward, we will construct a sample quality evaluation system. Through automatic detection of annotation consistency (using cross-validation) and spatiotemporal feature matching algorithms (based on geospatial information verification), we aim to enhance the training dataset.

4.2.2. Model Training Results

The model trained exclusively on the 2016 dataset achieved the highest mAP value. However, due to insufficient sample diversity in its training data and the research objective of analyzing Linpan changes in Pidu District, Chengdu between 2016 and 2023, the model with mAP = 0.730 was ultimately selected as the detection model. Training results are presented below:

As shown in Figure 8, the first row of images focuses on internal metrics, demonstrating the convergence of training errors and loss functions. The second row visualizes external performance indicators on the validation set.

Box loss, a measure of the discrepancy between predicted and true bounding boxes in object detection, shows a steady decline with increasing iterations, indicating the model is progressively learning to localize targets. Classification loss, which gauges the accuracy of object category prediction, also exhibits a continuous downward trend, signaling enhanced classification capability. Distribution Focal Loss, is another metric that declines, reflecting more stable regression.

Metrics/precision (B), or precision, quantifies the proportion of correct positive predictions. The curve’s upward trajectory, stabilizing around 0.65, suggests a reduction in the false—positive rate. Metrics/recall (B), or recall, denotes the ratio of true positives accurately identified. Its gradual ascent to approximately 0.65 implies a lower false-negative rate.

Figure 9 shows how the model’s mAP changes during training. It illustrates the precision-recall trade-off at different confidence thresholds. The closer the curve is to the top-right corner, the better the model can maintain high recall while ensuring high precision. The figure shows mAP@0.5 is 0.73, meaning the model’s average precision across all categories is 0.73 when the IoU threshold is 0.5. This shows the model performs well in detection and classification tasks.

Figure 10a,b present images from the validation set with manual annotations. In contrast, Figure 10b,d show the corresponding machine-generated annotations. The model shows a preliminary ability to identify Linpan and distinguish it from other construction land types with some accuracy. However, the model has low confidence in some areas, indicating uncertainty and a risk of misclassification. This may stem from poor recognition of regions with blurry boundaries or indistinct visual features. Despite initial success, there is room for model optimization, particularly in enhancing robustness and accuracy in complex scenarios.

5. Results and Analysis

5.1. Result Visualization

Using the optimal model, detection was performed, taking 6 and 7 min, respectively. The team’s Python scripts were then used to reconstruct the 1000 × 1000-pixel images back into the original Pidu District imagery, producing linpan detection results for 2016 and 2023, as shown in Figure 11a,c. Subsequent spatial analysis indicates, as illustrated in Figure 11b,d, that Hezuo, Xipu, and Anjing streets consistently represent low-density areas for linpan. By 2023, linpan density across the district had declined markedly, with the most pronounced reductions observed in Deyuan and Hongguang streets, reflecting intense land-use changes and landscape pattern evolution during the urbanization process in this region (Figure 11).

5.2. Verification

The calculation in

Table 4. Main Indicators of Linpan Object Detection Results in 2016 and 2023.

	Precision	Recall	F1 Score
Detection results in 2016	97.85%	98.49%	98.17%
Detection results in 2023	96.59%	94.39%	95.48%

shows that for the 2016 satellite image detection, the model has a precision of 97.85%, a recall of 98.49%, and an F1 Score of 98.17%. For the 2023 satellite images, the precision is 96.59%, the recall is 94.39%, and the F1 Score is 95.48%.

For geographic remote sensing object detection, the typical ranges for precision and recall are 0.7–0.85 and 0.6–0.85, respectively. In this detection task, both precision and recall exceed 0.9, indicating a low false positive rate and demonstrating that the model successfully detects the majority of genuine Linpan. An F1-score > 0.7 signifies a viable detection model. The current F1-scores for both years exceed 0.9, confirming the model’s strong performance.

Comparing the 2023 results to the 2016 detection sample, both precision and recall show a slight decline. Possible reasons include:

(1)

Degraded feature quality of Linpan, where characteristic differences may have rendered some instances unrecognizable to the model;

(2)

Insufficient model adaptability for 2023 imagery, as the training data predominantly originated from 2016 images, resulting in inconsistencies in resolution and color profiles with the 2023 data.

5.3. Spatio-Temporal Distribution Analysis of Linpan

Based on GIS spatial analysis, Linpan settlements in Pidu District exhibit a pronounced clustering pattern, predominantly concentrated in the northwest. Their overall distribution extends along a northeast–southwest axis, forming a distinct belt-shaped corridor. High-density clusters are notably observed in Tangchang, Xinminchang, and Tangyuan towns, where a polycentric and group-based pattern emerges. This spatial configuration indicates a strong correlation between Linpan distribution and regional economic development, land-use layout, and transportation networks.

5.3.1. Kernel Density Analysis

The analysis results indicate that in 2016, Linpan in Pidu District exhibited a distinct northwest-southeast band-like distribution structure. High-density clusters were primarily concentrated in the northwestern part of the district, especially in Tangchang Town and its bordering areas with Tangyuan Town and Xinminchang Town. This region featured dense Linpan distribution with relatively well-preserved morphological integrity.

In contrast, extending towards the southeast, particularly in areas approaching downtown Chengdu, the number of Linpan significantly decreased. This revealed a spatial trend of gradual Linpan sparsification from the urban-rural fringe towards the urban core.

The 2023 kernel density analysis results show that the northwest-southeast band-like distribution pattern of Linpan persisted. High-density areas remained concentrated in Tangchang Town and its adjacent zones with Xinminchang Town, indicating continuity in Linpan preservation within this region.

However, compared to 2016, Linpan density in some peripheral areas decreased. This suggests that urbanization development has exerted an influence on the spatial pattern of Linpan. Notably, the concentration of Linpan around Tangchang Town exhibited a tendency of contraction toward the core area (Figure 12).

5.3.2. Spatial Correlation Analysis

To investigate the spatial distribution characteristics and evolution trends of Linpan in Pidu District across different time points, this study conducted Average Nearest Neighbor (ANN) analysis on the 2016 and 2023 Linpan point location datasets. This method effectively reveals spatial clustering or dispersion patterns of point features, providing critical insights into the spatial organization of Linpan.

The analysis results show that in 2016, the observed average distance between Linpan in Pidu District was 202.17 m, significantly smaller than the theoretical expected mean distance of 242.43 m. The corresponding ANN was 0.8339. This indicates that the 2016 Linpan distribution significantly deviated from randomness, exhibiting distinct spatial clustering characteristics.

The 2023 analysis results further demonstrate a marked intensification of Linpan clustering. The observed average distance for this year was 258.30 m, below the theoretical expected mean distance of 340.82 m, yielding a nearest neighbor ratio of 0.7579. The increased clustering intensity resulted from a substantial decline in Linpan numbers within peripheral areas between 2016 and 2023. Consequently, the remaining Linpan distribution became spatially more compact compared to the core clustering zones (Figure 13).

This intensification of clustering may be driven by multiple factors, such as: The removal or degradation of peripheral Linpan under urbanization pressures; the preservation and concentrated conservation of Linpan in key areas; and policy inclination toward protecting traditional features in core zones. Overall, Linpan in Pidu District demonstrates a spatial evolution pattern of “contraction at the periphery and intensified clustering in the core” between 2016 and 2023.

6. Discussion

6.1. Methodological Advancement

By comparing the number of linpan in Pidu District recorded by different scholars using various methods (

Table 5. Data Comparison of Linpan Identification in Pi Du District in Previous Studies.

Date Time	Number of Linpan	Researcher	Data Source	Research Method
2005	8587	Guo Yingman (2017)	Google Imagery	Visual Interpretation Visual Interpretation
2015	6239		Google Imagery
2006	10,623	Wang Yao (2019)	CAD Topographic Map	Visual Interpretation
2016	2961		Google Imagery, GDEMV2 30 M resolution digital elevation data
2018	4980	Zhong Bo (2022)	Landsat Remote Sensing Imagery	Remote Sensing Interpretation
2016	3465	This Study (2025)	Google Imagery	Object Detection (YOLOv11)
2023	1616		Google Imagery

), this study further validates the effectiveness and reliability of its approach. In 2016, this study identified 3465 linpan, a figure highly consistent with the 2961 linpan reported by researcher Wang Yao through manual visual interpretation. Although manual interpretation is time-consuming, labor-intensive, and difficult to scale, its accuracy has long been recognized in academic research. The close alignment between the two results indicates that the YOLO deep learning model employed here achieved a level of automated extraction accuracy comparable to human interpretation.

In contrast, this number is substantially lower than the 4980 linpan reported by researcher Zhong Bo in 2018 using traditional remote sensing classification methods, likely pixel-based classifiers. The discrepancy may stem from methodological differences: traditional remote sensing classification relies primarily on spectral features, which cannot effectively distinguish linpan—organic complexes of buildings and vegetation—from other spectrally similar objects such as rural settlements, sparse forests, or orchards. This limitation often results in severe overestimation.

By contrast, the YOLO model adopted in this study, as an object detection framework based on deep convolutional neural networks, can more effectively capture the overall morphology, texture, and contextual relationships of linpan within their surrounding environments, thereby achieving higher discriminative capacity in complex landscapes. Furthermore, the automated and efficient linpan recognition method explored here reduces the time required for a county-level survey to within approximately five minutes. It is therefore feasible to complete a survey of an entire metropolitan area (~12,000 km²) within only a few hours. Compared with traditional methods such as field surveys and manual visual interpretation, this approach delivers significant improvements in efficiency. Relative to conventional remote sensing interpretation, it provides higher accuracy. Moreover, by utilizing openly accessible satellite imagery, the method reduces survey costs to a negligible level, ensuring substantial cost savings.

In summary, compared with manual interpretation, the proposed approach achieves similar accuracy while dramatically enhancing efficiency and reducing labor demands. Compared with traditional remote sensing classification methods, it markedly improves recognition accuracy and effectively avoids misclassification and omission caused by spectral confusion. This methodological advancement provides an innovative and practical technical pathway for rapid, accurate, and large-scale investigation and dynamic monitoring of traditional rural settlement landscapes.

6.2. Socioeconomic and Policy Analysis

Comparative results show that the number of linpan in Pidu District declined sharply from 2016 to 2023, with a reduction of 53.36%. Spatially, the decrease was less pronounced in the core area but more significant in peripheral zones, revealing a spatial pattern in which linpan loss intensified closer to Chengdu’s central urban area.

This sharp decline is primarily driven by the district’s rapid development as part of a national central city. First, urban expansion and large-scale infrastructure projects—such as the Chengdu Ring Expressway, Chengdu–Dujiangyan High-Speed Railway corridor, and the Electronic Information Industry Functional Zone—consumed extensive agricultural and ecological land, resulting in large-scale requisition and demolition of traditional linpan. Second, under the policies of rural revitalization and comprehensive land consolidation, some scattered and small-scale linpan were merged or transformed through approaches such as chaiyuan bingyuan (courtyard demolition and consolidation) to improve settlement concentration and land-use efficiency. In addition, changes in agricultural practices and evolving residential preferences among farmers accelerated the natural disappearance of traditional linpan. Overall, the drastic reduction in linpan in Pidu District reflects the combined effects of rapid urbanization, spatial restructuring strategies, and ecological protection processes.

6.3. Method Generalizability

Beyond linpan settlements, various other types of traditional settlements exist across China. Fujian Tulou, as the most representative clustered dwellings in southeastern China, exhibit unique historical, cultural, and architectural values. Settlements in the southern Taihang Mountains represent typical mountainous habitation in central China, reflecting residents’ adaptive strategies and survival wisdom in complex terrains. Minority settlements in Yunnan embody the country’s ethnic and cultural diversity, carrying distinctive cultural traits and spatial forms. Although these three settlement types differ in cultural background and regional environment, their spatial forms share similarities with linpan, exhibiting a degree of regularity, aggregation, and recurring structural patterns that facilitate efficient capture and analysis through image-based machine recognition methods.

However, existing studies on these settlements still largely rely on manual visual interpretation and field surveys [37,38,39], which are traditional, time-consuming, and limited in efficiency, making large-scale and high-precision investigations challenging. Applying the YOLO-based automatic detection and classification method proposed in this study to these three settlement types could enable rapid extraction of spatial distributions, intelligent identification of morphological types, and preliminary assessment of structural conditions. This approach would substantially reduce labor and time costs, significantly improve the efficiency and coverage of cultural heritage surveys and protection, and provide a reliable technical pathway for systematic, multi-scale conservation and monitoring of traditional settlements.

7. Conclusions

7.1. Key Findings and Practical Implications

The key findings indicate that deep learning–based object detection models (e.g., YOLO) can efficiently and accurately identify traditional rural settlements with complex morphologies and mixed backgrounds. This approach enabled automated monitoring of the spatial distribution and dynamic changes in Linpan in the Chengdu Plain, and quantitatively revealed urbanization and economic drivers as the primary causes of their sharp decline. The method carries strong practical significance and broad transferability.

7.2. Academic Contributions

(1)

Methodological Innovation: Development of an efficient automated recognition model for complex rural settlements.

This study represents the first application of the advanced YOLOv11 object detection model to spatial identification of linpan, a typical rural settlement in western Sichuan. To address challenges such as irregular shapes, complex textures, and high background interference, the model was specifically optimized and trained, enabling rapid and accurate automated extraction of linpan boundaries. Compared with traditional remote sensing interpretation methods, this approach maintains high precision while substantially improving processing efficiency, providing a novel technical solution for large-scale, high-frequency dynamic monitoring of traditional rural settlements.

(2)

Practical Applicability: Construction of the first high-precision, multi-temporal dataset of linpan spatial distribution in Pidu District, Chengdu.

The study generated spatial distribution and quantity datasets of linpan in Pidu District for two key time points, 2016 and 2023. These datasets are not only highly accurate but also possess explicit spatiotemporal dimensions, quantitatively revealing that linpan numbers in Pidu District declined by over 53.36% over seven years. The datasets provide valuable baseline data for subsequent studies and are of fundamental significance for investigating rural landscape changes under rapid urbanization.

(3)

Theoretical Significance: Integration of artificial intelligence with urban and rural planning research.

The study demonstrates the substantial potential of multidisciplinary methods to address traditional geographic challenges, offering valuable insights for theoretical innovation and methodological advancement in urban and rural planning.

7.3. Limitations of This Study

7.3.1. Need for Expanded Training Samples

The current training primarily utilized multi-temporal satellite imagery from 2016 and 2023. While integrating multi-year remote sensing data enhances the model’s recognition capability in dynamic surface environments, significant differences exist between the two datasets in resolution, spectral characteristics, and land cover—attributable to factors like sensor upgrades, landscape transformations, and seasonal variations.

This cross-year data combination effectively improves model generalization:

It strengthens recognition of time-invariant features (e.g., roads, buildings), reducing interference from lighting/seasonal changes;

It enhances adaptability to time-varying features (e.g., vegetation growth, urban expansion), improving detection accuracy of landscape change patterns.

Future work could incorporate intermediate-year imagery (e.g., 2010, 2020) to construct continuous time series, coupled with multi-source data (e.g., nighttime lights, elevation data). This would further optimize the model’s spatiotemporal feature extraction capability for complex scenario object detection.

7.3.2. Need for Training Parameter Optimization

During model training, parameter optimization critically impacts final performance. The current configuration can be enhanced through adjustments to key hyperparameters:

First, appropriately increase the number of training epochs to allow more sufficient learning time, terminating training only after the validation loss convergence stabilizes.

Second, experiment with adjusting the batch size—employing larger batches within GPU memory constraints improves gradient update stability. This should be coordinated with corresponding learning rate adjustments (e.g., linear scaling rule) to maintain training efficiency.

Additionally, introduce progressive adjustment strategies, such as using higher learning rates with smaller batches during initial training stages for rapid convergence, followed by gradual refinement in later phases.

These parameter optimizations require systematic hyperparameter tuning (grid search or Bayesian optimization) based on validation set metrics to identify optimal combinations. Furthermore, implement early stopping and model checkpoints to prevent overfitting and ensure optimal model performance.

7.3.3. Need for Deeper Geospatial Analysis

This study underutilizes GIS’s capabilities for multidimensional spatiotemporal analysis. Current work only achieves basic spatial visualization, lacking systematic deep geospatial examination of YOLO-detected Linpan data.

Future research should conduct: Quantitative spatial pattern analysis: measuring clustering patterns, density gradients, and spatial statistics. Spatial coupling relationship exploration between Linpan distribution and natural/socioeconomic drivers. Such advanced spatial analytics would significantly elevate the scientific value and practical utility of findings.