R-SWTNet: A Context-Aware U-Net-Based Framework for Segmenting Rural Roads and Alleys in China with the SQVillages Dataset

Abstract

Rural road networks are vital for rural development, yet narrow alleys and occluded segments remain underrepresented in digital maps due to irregular morphology, spectral ambiguity, and limited model generalization. Traditional segmentation models struggle to balance local detail preservation and long-range dependency modeling, prioritizing either local features or global context alone. Hypothesizing that integrating hierarchical local features and global context will mitigate these limitations, this study aims to accurately segment such rural roads by proposing R-SWTNet, a context-aware U-Net-based framework, and constructing the SQVillages dataset. R-SWTNet integrates ResNet34 for hierarchical feature extraction, Swin Transformer for long-range dependency modeling, ASPP for multi-scale context fusion, and CAM-Residual blocks for channel-wise attention. The SQVillages dataset, built from multi-source remote sensing imagery, includes 18 diverse villages with adaptive augmentation to mitigate class imbalance. Experimental results show R-SWTNet achieves a validation IoU of 54.88% and F1-score of 70.87%, outperforming U-Net and Swin-UNet, and with less overfitting than R-Net and D-LinkNet. Its lightweight variant supports edge deployment, enabling on-site road management. This work provides a data-driven tool for infrastructure planning under China’s Rural Revitalization Strategy, with potential scalability to global unstructured rural road scenes.

1. Introduction

1.1. Background

Rural road networks serve as critical lifelines for rural development, underpinning economic growth, social equity, and sustainable livelihoods globally. Research has demonstrated that well-connected road infrastructure directly enhances rural household income by facilitating access to markets, healthcare, and education [1]. However, a significant portion of the global rural population remains disconnected from formal road networks [2], perpetuating poverty cycles by limiting mobility and resource access [3].

China, as the largest developing country in the world, is undergoing rapid urbanization [4]. By the end of 2024, the urbanization ratio of China surpassed 67% [5]. However, China is continuously investing in rural areas by prioritizing rural infrastructure development under its “Rural Revitalization Strategy” [6,7], aiming to construct safe, convenient, and efficient rural road systems. By the end of 2024, the total length of rural roads in China exceeded 4.64 million kilometers [8], representing a 19.5% increase since 2014 [9]—a milestone that has garnered international attention as a model for developing countries. This expansion, however, has intensified the demand for intelligent tools to support rural road identification, mapping, and management [10].

Rural roads present unique challenges for automated extraction from remote sensing imagery, distinct from their urban counterparts. Notably, narrow alleys within and between rural settlements—vital for daily mobility and social interaction—are consistently absent from mainstream commercial maps (e.g., Amap, OpenStreetMap) and public geographic databases. Their small width, irregular morphology, blurred spectral boundaries, and frequent occlusion by vegetation or buildings [11] make them particularly difficult to identify using traditional segmentation models like the original U-Net. Compounding this issue, most state-of-the-art road segmentation models are validated on well-structured, homogeneous road networks in Europe and North America (e.g., the Massachusetts dataset [12]), exhibiting limited generalizability to the highly heterogeneous, unplanned rural road networks of developing countries.

China’s dense and complex rural road networks, characterized by diverse landscapes (e.g., polders, hillsides, riverine villages) and mixed infrastructure types (paved/unpaved roads, alleys), thus present an ideal testbed for advancing road extraction technologies. Addressing the gaps in automated segmentation of such networks is critical to supporting data-driven rural planning and infrastructure management.

1.2. Related Applications

Road extraction is one of the major application scenarios of semantic segmentation, in which deep learning plays an important role. Starting from the AlexNet [13], the CNN (Convolutional Neural Networks) structure has built up a solid foundation for follow-up studies. In 2015, CNN was first implemented in pixel-wise labeling [14]. By introducing residual connection methods into CNN, the ResNet [15] structure significantly alleviates gradient vanishing of CNN, and provides a new prototype for complex feature extraction. The MobileNet [16], on the other hand, introduces the depthwise separable convolution (DSC) [17], making it possible to deploy deep learning models on mobile devices.

In the following studies, CNN has been improved to adapt to segmentation rather than classification tasks. Aoki and Saito [18] used CNN to recognize buildings and roads in aerial imagery by adding a new fully connected layer to resize the vector, which was later proven to be equivalent with another convolution layer in FCN (Fully Convolutional Network) [19], where there is no fully connected layer in the structure, making the segmentation of high-resolution aerial imagery possible [20].

Considering FCN’s drawbacks of rough upsampling, obscure details, and relatively low efficiency, DeconvNet [21], SegNet [22], and UNet [23] were introduced, forming an encoder–decoder structure, which has become a milestone in semantic segmentation tasks. In recent research, an encoder–decoder structure is nearly indispensable. Meanwhile, DeepLab series widened the receptive field of FCN by introducing the ASPP (Atrous Spatial Pyramid Pooling) [24] to fetch the multi-scale context of images.

From then on, improvements in backbones and structures above emerged. By combining ResNet with UNet structure, ResUNet was introduced to extract roads from aerial images [25]. By adding attention gates into the UNet structure, Oktay et al. established a model that learned to focus on target structures of varying shapes and sizes by suppressing irrelevant regions in an input image while highlighting salient features useful for a specific task [26]. By re-designing skip connections and using deep supervision, UNet++ achieved an average IoU gain of more than 3 points over UNet [27]. Inspired by the “Encoder-Decoder” structure, Wang et al. established the RD-Net to segment unstructured roads, improving overall accuracy [28].

The rise of Transformer architecture [29] has brought about a paradigm shift. The ViT (Vision Transformer) model has become one of the most important models in image processing [30], which later led to the Swin Transformer [31], where a SW-MSA (shifted window Multi-head Self-Attention) mechanism was brought to life, reducing overall resource cost. Researchers tried to combine the transformer architecture with other state-of-the-art models and achieved inspiring results. By adding transformer blocks into different sections of the UNet structure, TransUNet [32] and Swin-UNet were invented to capture long-range dependencies of biomedical scenarios [33] and aerial imagery scenarios [34] via shifted window self-attention.

The segmentation of aerial images, especially roads, with the demand of maintaining long-range dependencies calls for a more integrated and complex design of the models. Thus, it has become a trend to merge multiple structures to achieve well-optimized segmentation results. DPIF-Net [35] takes multiple measures like transformers and ASPP to merge global and local information, while Res_ASPP_UNet++ [36] uses depthwise separable convolution and ASPP to reduce model parameters. Sloan et al. [37] used multiple methods to map remote roads using satellite imagery. Yang et al. [38] designed R-MSFNet to extract roads of rural Xiong’an New Area using a self-built dataset and the DeepGlobe dataset.

1.3. Research Objectives

As most state-of-the-art models are trained and validated on well-structured road networks in Europe and North America, they exhibit significant limitations in generalizing to the unstructured, heterogeneous landscapes prevalent in rural China. To address the aforementioned challenges, this study aims to develop a robust deep learning framework for accurate segmentation of rural streets and alleys. The specific research objectives are as follows:

First, to propose a context-aware U-Net-based model optimized for rural road segmentation. This model will integrate advanced context-aware mechanisms (e.g., multi-scale feature fusion, long-range dependency modeling) to enhance segmentation accuracy in complex rural scenes. Key targets include mitigating blurred boundaries, severe occlusions, and varying road textures—issues that traditional U-Net and its variants struggle to resolve. By combining hierarchical feature extraction with adaptive attention mechanisms, the model will improve the representation of fine-grained rural road features.

Second, to enhance the model’s capacity to identify narrow rural alleys. These critical pathways are systematically missing from existing maps due to their small scale and ambiguous spectral signatures. To achieve this, the study will optimize dataset construction by integrating multi-source data (high-resolution satellite and aerial imagery) and refining labeling protocols, ensuring that narrow alleys are accurately labeled and distinguishable from non-road features.

These objectives collectively address the technical and practical gaps in rural road segmentation, with the goal of supporting intelligent infrastructure planning under China’s Rural Revitalization Strategy.

2. Methodology

2.1. Dataset Preparation

2.1.1. Overview of Study Region

This study focuses on the Shuiyang River–Qingyi River basin and the paleo-Danyang Lake region, situated along the border of Jiangsu and Anhui Provinces in eastern China (Figure 1). This transprovincial area is hydrologically and ecologically distinct, characterized by a complex interplay of rivers, lakes, and anthropogenically modified landscapes, making it a representative testbed for investigating rural road networks in diverse and unstructured environments.

The region’s hydrological framework is dominated by the Shuiyang and Qingyi Rivers, which originate from the northern slopes of the Huangshan Mountains (Anhui Province). Flowing northward, these rivers traverse municipal administrative divisions, including Huangshan, Xuancheng, Wuhu (Anhui), and Nanjing (Jiangsu), collectively shaping the basin’s drainage network and influencing rural settlement patterns. Central to this system is the paleo-Danyang Lake, a historical lacustrine depression formed by upstream sedimentation and blockage of inflows to the Yangtze River [39]. Over centuries of natural infilling and anthropogenic land reclamation [40], this ancient lake has fragmented into a mosaic of smaller water bodies, including Shijiu Lake, Gucheng Lake, and Nanyi Lake. These are interconnected by an intricate network of natural rivers and artificial canals, giving rise to extensive polders—low-lying agricultural tracts enclosed by dykes to mitigate flooding—which define the region’s landscape.

The study area exhibits remarkable rural settlement diversity, with villages distributed across three primary geomorphic types, mostly polder villages, river-side villages, and plain villages. Polder villages cluster within dyke-enclosed areas, adapting to irregular reclaimed land boundaries, resulting in winding, narrow alleys and road networks that follow topographic gradients; river-side villages align along watercourses featuring linear road networks and frequent intersections with canals, posing challenges for segmentation due to spectral similarity between water and road surfaces; plain villages are located on flat alluvial plains with more regular layouts and mixed land uses (residential, agricultural, and public spaces) that complicate feature discrimination.

Within the study area, rural development exhibits diverse stage characteristics of “original-transitional-activated”: there are traditional villages preserving the original form of “built along ridges, lanes following polders”; transitional resettlement communities formed under the guidance of the “Three Zones and Three Lines” policy; and homestay villages revitalized through “culture+tourism” updates. This “multi-stage coexistence” pattern not only retains rural historical memory but also reflects rural adaptation attempts to modern needs in the urbanization process, further demonstrating the regenerative value of “cultural activation” for traditional spaces, thus providing a vivid case for exploring the spatial integration of tradition and modernity.

The study area is located in the triangular intersection area of Wu Culture (southern Jiangsu and northern Zhejiang), Huizhou Culture (Huangshan-Wuhu), and Lianghuai Culture (region between the Yangtze and the Huai river). The core area of the study area (e.g., Gaochun in Jiangsu, Xuanzhou in Anhui) belongs to the Xuanzhou subgroup of Wu Dialect, serving as the “Xin’an Cultural Sub-region” [41] on the western edge of Wu Culture; the study area connects to the Huangshan Mountains in the south and borders the Jianghuai Plain in the north. The intersection of the three cultures has ultimately formed a “pluralistic unity” cultural identity in rural settlements. This cultural pattern is not only reflected in dialects, customs, and architectural styles but also profoundly influences the spatial form and social structure of rural settlements.

This heterogeneity of spaces, construction, and culture mirrors the complexity of rural road networks across eastern China. Critically, the coexistence of narrow, unstructured alleys, occluded or shaded pathways, and mixed surface types provides a rigorous test for developing context-aware segmentation models. By capturing such diversity, the dataset derived from this region ensures robust model generalization to other rural contexts.

2.1.2. Data Source and Labeling

The dataset construction relies on multi-source data to ensure the accuracy and diversity of rural road features. Primary data sources include the following: (1) satellite imagery retrieved via the official API of Tianditu [42], a national geospatial platform developed by the National Geomatics Center of China, ensuring high-precision georeferencing1; (2) processed aerial imagery, obtained by exporting orthomosaics from oblique photography surveys of selected villages, providing sub-meter resolution details critical for distinguishing narrow alleys and occluded pathways.

After on-site investigations and landscape diversity assessments, more than 30 natural villages [43] were selected as dataset candidates, distributed across the Shuiyang River–Qingyi River basin and paleo-Danyang Lake region (Figure 1). This selection ensures coverage of diverse rural typologies (polders, plains, river-side villages) and infrastructure conditions (traditional villages, resettlement communities), mirroring the complexity of rural road networks in eastern China. For each village, a square satellite image centered on the settlement was fetched from Tianditu, with a resolution of 3072 × 3072 pixels (covering ~1.8 km × 1.8 km). This image is a mosaic of 144 tiles at zoom level 18, balancing spatial coverage and local detail (Figure 2).

When loading images from Tianditu API, a workflow originating from the name of the villages is carried out. In Tianditu, the villages are treated as POI (points of interest), and their longitude and latitude can be acquired through geocoding API. The longitude and latitude are then resolved as tile indices in the form of a triple tuple (x, y, z). The images’ centering coordinates (x, y) with a zoom level of z are then downloaded through Tianditu map API. The acquired images are 3072 × 3072 in resolution, merging 12 × 12 = 144 tiles, each with a resolution of 256 × 256.

Semantic segmentation requires pixel-wise alignment between images and masks. To address the ambiguity of rural road boundaries, a dual-source labeling strategy was adopted: (1) Manually labeling binary masks with a size of 3072 × 3072, where public spaces (roads, squares, parking lots) are labeled white; (2) cross-validating labels with high-resolution aerial imagery, which reveals fine-grained details indistinct in satellite imagery. This multi-source fusion reduces labeling errors caused by spectral confusion, particularly for unpaved or shaded rural pathways.

To adapt to model input size (512 × 512 pixels) and augment training diversity, large satellite images and their corresponding masks were cropped using a shifting window strategy (Figure 3). The window size was set to 512 × 512 pixels with a stride of 256 pixels (50% overlap), generating 121 non-overlapping sub-images per satellite image. This cropping strategy balances local feature integrity (e.g., preserving continuous alley segments) and dataset size, avoiding over-reliance on large homogeneous regions.

The cropped sub-images and masks form the backbone of the dataset, which is further split into training, validation, and test sets to ensure model generalization. This structured pipeline from multi-source data acquisition to precision labeling and adaptive cropping lays a robust foundation for training context-aware rural road segmentation models.

2.1.3. Dataset Outline

To ensure the dataset captures the heterogeneous characteristics of rural road networks, 18 natural villages were selected from over 30 candidates to form the SQVillages dataset, with their geographic distribution visualized in Figure 4. This selection prioritizes diversity in geomorphic types and infrastructure conditions across the study region, comprising 2178 cropped image–mask pairs (512 × 512 pixels) derived from the shifting window strategy detailed in Section 2.2.2, with 121 image–mask pairs for each village.

The dataset is split into training and validation sets using a village-level stratified strategy. In Figure 4, the training set is labeled as dark red bold rings, the validation set as black bold rings, the test set as red rings, and the rest, which stands for dataset candidates, as gray thin rings. Of the 18 villages selected in the SQVillages dataset, the validation set exclusively uses 242 samples from Gaogang Village (GG) and Xiaohu Village (XH), while the remaining 1936 samples form the training set. This separation ensures the validation set represents an independent rural scene, avoiding data leakage. The proportion of the training set in SQVillages is 88.9%, which falls in the commonly used domain of 85% to 90% for small samples.

Training samples undergo adaptive augmentation: random rotations (−90 to 90), random horizontal flips, and color jitter (brightness and contrast enhancement level 0.3, hue enhancement level 0.1) to simulate varying lighting and spectral conditions. Each image is normalized using channel-wise mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225) from the ImageNet dataset [44], aligning with pre-trained model initialization. Augmentation is applied three times per training sample, expanding the training set to 5808 samples—sufficient to train deep models while preserving rural road feature diversity. The basic information of the villages in the SQVillages dataset is listed in

Table 1. The basic information of the villages in the SQVillages dataset.

Name	Abbr.	Location	Natural Landscape	Honors and History	Construction and Development	Foreground Ratio of the 3K image (%)
Shenghong Village	SH	Xinfeng Country, Huangshan District, Huangshan, Anhui Province	Hill-side	Third Batch of Chinese Traditional Village	-	3.00
Yaduan Village	YD	Jishan Town, Nanling County, Wuhu, Anhui Province	Plain, River-side	-	-	1.75
Matou Village	MT	Qinxi Town, Jing County, Xuancheng, Anhui Province	Hill-side, River-side	Fourth Batch of Chinese Traditional Village	-	2.53
Huangtian Village	HT	Langqiao Town, Jing County, Xuancheng, Anhui Province	Hill-side	First Batch of Chinese Traditional Village	-	2.62
Badushu Village	BDS	Yangliu Town, Wanzhi District, Wuhu, Anhui Province	Hill-side	-	-	0.91
Suda Village	SD	Wanzhi Town, Wanzhi District, Wuhu, Anhui Province	Hill-side	Third Batch of Anhui Traditional Village	-	3.51
Qianjia Village	QJ	Yangjiang Town, Gaochun District, Nanjing, Jiangsu Province	Polder	-	-	3.26
Shuibiqiao Village	SBQ	Zhuanqiang Town, Gaochun District, Nanjing, Jiangsu Province	Polder Dyke, River-side	-	River Channel Widening; Pumping Station Construction; Two Resettlement Housing Complexes; Demolition of the Western Sector	4.47
Mojiazui Village	MJZ	Zhuqiao Country, Xuanzhou District, Xuancheng, Anhui Province	Polder Dyke	-	-	4.51
Qiqiao Village	QQ	Qiqiao Sub-district, Gaochun District, Nanjing, Jiangsu Province	Plain	Second Batch of Chinese Traditional Village	-	12.62
Xiaohu Village	XH	Shuidong Town, Xuanzhou District, Xuancheng, Anhui Province	Plain, Hill-side	Fourth Batch of Chinese Traditional Village	-	5.32
Gaogang Village	GG	Qiqiao Sub-district, Gaochun District, Nanjing, Jiangsu Province	Plain	-	Collaboration with Tsinghua University, Utilizing Vacant Buildings [45] and Holistic Village Operation [46]	6.14
Hujiaba Village	HJB	Qiqiao Sub-district, Gaochun District, Nanjing, Jiangsu Province	Plain, Reservoir-side	Reservoir Relocation Village	Construction of Homestays	5.48
Wangjia Village	WJ	Dongba Sub-district, Gaochun District, Nanjing, Jiangsu Province	River-side, Close to Xu River	-	Resettlement	8.21
Zhangdu Village	ZD	Yunling Town, Jing County, Xuancheng, Anhui Province	River-side	Third Batch of Chinese Traditional Village	-	4.39
Baduhe Village	BDH	Gongshan Town, Nanling County, Wuhu, Anhui Province	Hill-side	Anhui Traditional Village	-	1.85
Yangtian Village	YT	Liqiao Town, Xuanzhou District, Xuancheng, Anhui Province	Hill-side	-	-	2.73
Pei Village	PC	Feili Town, Langxi County, Xuancheng, Anhui Province	Lake-side	Second Batch of Chinese Traditional Village	-	1.29

, where the villages used as the validation set are in italic format.

To rigorously evaluate model generalization to unseen rural scenes, eight additional villages from the same region (excluded from SQVillages) were selected as the independent test set, ensuring a comprehensive assessment of model robustness, as shown in

Table 2. Information about the villages used for testing.

Name	Abbr.	Location	Natural Landscape	Honors and History	Construction and Development
Lingxiasu Village	LXS	Yongfeng Country, Huangshan District, Huangshan, Anhui Province	Hill-side, River-side	First Batch of Chinese Traditional Village	-
Chitan Village	CT	Qinxi Town, Jing County, Xuancheng, Anhui Province	River-side	Fifth Batch of Chinese Traditional Village	-
Xuantan Village	XT	Taoxin Town, Wanzhi District, Wuhu, Anhui Province	Polder	-	-
Youzhagou Village	YZG	Zhuqiao Country, Xuanzhou District, Xuancheng, Anhui Province	Polder Dyke	-	-
Longtan Village	LT	Yangjiang Town, Gaochun District, Nanjing, Jiangsu Province	Polder Dyke	-	-
Shanmen Village	SM	Gangkou Town, Ningguo City, Xuancheng, Anhui Province	Plain	Fourth Batch of Chinese Traditional Village	-
Hejia Village	HJ	Gucheng Sub-district, Gaochun District, Nanjing, Jiangsu Province	Plain, Lake-side	Fifth Batch of Jiangsu Traditional Village	Renewal of Vacant Buildings
Lixi Village	LX	Qiqiao Sub-district, Gaochun District, Nanjing, Jiangsu Province	Plain	-	-

. Test set preprocessing follows the same pipeline as the training data (512 × 512 sliding windows, georeferencing) but with no data augmentation applied to reflect real-world inference conditions. This consistency in preprocessing ensures fair comparison, while the exclusion of test villages from training guarantees unbiased evaluation of the model’s ability to segment narrow alleys, occluded roads, and complex rural topologies in novel environments.

2.2. Network Architecture

2.2.1. General Structure

The R-SWTNet network structure (Figure 5) adopts an encoder–decoder architecture optimized for rural road segmentation, building on the classic UNet framework with three core components, encoders, bottleneck block, and decoders, each modified to address the complexity of rural road features. The detailed parameters of the network are listed in

Table 3. The parameters of R-SWTNet.

Blocks	Shape of Input (C, H, W)	Shape of Output (C, H, W)	Type	Kernel	Stride	Padding	Window_Size
Encoder 1	3, 512, 512	64, 256, 256	Conv	7 × 7	2	3	-
Encoder 2	64, 256, 256	64, 256, 256	ResNet Conv × 3	3 × 3	1	1	-
Encoder 3	64, 256, 256	128, 128, 128	ResNet Downsampling	3 × 3	2	1	-
	128, 128, 128	128, 128, 128	ResNet Conv × 3	3 × 3	1	1	-
Encoder 4	128, 128, 128	256, 64, 64	ResNet Downsampling	3 × 3	2	1	-
	256, 64, 64	256, 64, 64	ResNet Conv × 5	3 × 3	1	1	-
	256, 64, 64	256, 64, 64	Transformer (256) × 2	-	-	-	8
Bottleneck	256, 64, 64	512, 32, 32	Patch Merging	3 × 3	2	1	-
	512, 32, 32	512, 32, 32	ASPP	Multiple	-	-	-
	512, 32, 32	512, 32, 32	Transformer (512) × 2	-	-	-	16
Decoder 1	512, 32, 32	256, 64, 64	Pixel Shuffle	2 × 2	2	0	-
	256, 64, 64	256, 64, 64	Gated Skip Connection (Encoder 4)	-	-	-	-
	256, 64, 64	256, 64, 64	CAM-Residual	1 × 1	1	0	-
	256, 64, 64	256, 64, 64	Transformer (256) × 2	-	-	-	8
Decoder 2	256, 64, 64	128, 128, 128	Pixel Shuffle	2 × 2	2	0	-
	128, 128, 128	128, 128, 128	Gated Skip Connection (Encoder 3)	-	-	-	-
	128, 128, 128	128, 128, 128	CAM-Residual	1 × 1	1	0	-
Decoder 3	128, 128, 128	64, 256, 256	Pixel Shuffle	2 × 2	2	0	-
	64, 256, 256	64, 256, 256	Gated Skip Connection (Encoder 2)	-	-	-	-
	64, 256, 256	64, 256, 256	CAM-Residual	1 × 1	1	0	-
Decoder 4	64, 256, 256	64, 512, 512	ConvTranspose	2 × 2	2	0	-
	64, 512, 512	64, 512, 512	Conv, ReLU	3 × 3	1	1	-
	64, 512, 512	1, 512, 512	Conv, Sigmoid	1 × 1	1	0	-

.

The encoders perform hierarchical feature extraction via four downsampling blocks, progressively reducing spatial dimensions while enriching semantic information. It uses ResNet34 as the backbone, with early blocks extracting low-level features (textures, edges) through consecutive convolution layers, and Swin Transformer modules integrated gradually to model long-range spatial dependencies—critical for connecting occluded or fragmented road segments. By combining local feature refinement (via ResNet) and contextual correlation (via Transformer), the encoder captures multi-scale road characteristics, from narrow alley edges to large-scale network topology.

The bottleneck block deepens semantic integration by merging high-level encoder features. It first downsamples features via patch merging, then employs a positional encoding block and an ASPP module to aggregate multi-scale context, enabling the model to adapt to varying road widths and occlusions. Two subsequent Swin Transformer blocks further strengthen global feature correlation, ensuring coherent representation of rural road networks despite their unstructured layout.

The decoder module reconstructs high-resolution segmentation maps by reversing downsampling, leveraging skip connections to fuse encoder-derived multi-scale features. Each decoder block upsamples features via pixel shuffle and gated skip connections and refines the fused map using CAM-Residual blocks—these enhance channel-wise attention to critical road features while suppressing noise. Selective integration of Transformer modules in later decoder stages further optimizes boundary precision, ensuring accurate recovery of narrow alleys and blurred edges. The final output is a binary segmentation mask generated via a 1 × 1 convolution with sigmoid activation, classifying road pixels from background.

This architecture integrates ResNet34’s hierarchical feature extraction, Swin Transformer’s long-range dependency modeling, ASPP’s multi-scale context fusion, and CAM-Residual’s channel-wise attention, collectively enhancing the model’s ability to capture blurred boundaries, occluded segments, and narrow alleys in rural scenes.

2.2.2. The Structure of Swin Transformer Module

The Swin Transformer module in R-SWTNet is a computationally optimized adaptation of the Swin Transformer architecture (Figure 6), designed to enhance the network’s capacity for capturing long-range dependencies and global context, which is critical for segmenting rural roads with fragmented or occluded segments. The module comprises two consecutive transformer blocks, which work synergistically to balance local detail preservation and cross-region information interaction.

The first block employs a Window Multi-head Self-Attention (W-MSA) mechanism, partitioning the feature map into non-overlapping local windows and confining self-attention computations within these windows to reduce computational complexity. By anchoring attention to local regions, W-MSA prioritizes fine-grained details such as road edges, texture variations, and small-scale occlusions, ensuring precise modeling of subtle rural road features without the cost of global attention.

The second block introduces a Shifted Window Multi-head Self-Attention (SW-MSA) mechanism, which shifts the window partitions of the previous layer to create overlapping regions between adjacent windows. This design enables cross-window information communication: features from neighboring windows (e.g., a partially occluded road segment in one window and its continuation in an adjacent window) can interact, effectively expanding the model’s receptive field. Importantly, the sequential arrangement of W-MSA and SW-MSA ensures local features are first stabilized before integrating contextual cues from neighboring regions, preventing noise from diluting critical details (e.g., distinguishing a narrow alley from surrounding field ridges).

The self-attention operation within each block is computed as:

$$\mathrm{A}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}(Q,K,V)=\mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}}+B\right)V$$

(1)

where Q, K, and V represent query, key, and value matrices, respectively, and $d_k$ denotes the dimension of the query/key vectors.

The shifted window strategy enriches inter-window information interaction, enabling the model to aggregate context from adjacent regions and adaptively adjust feature weights based on semantic relevance (e.g., emphasizing distant but structurally continuous road segments). When deployed in R-SWTNet’s encoders, this module outperforms conventional transformer blocks by balancing three key advantages: (1) capturing long-range dependencies to connect fragmented road segments (e.g., roads interrupted by buildings or trees); (2) preserving multi-scale features, from fine-grained alley edges (low-level) to large-scale network topology (high-level); and (3) maintaining computational efficiency, as window-based attention reduces resource consumption compared to global self-attention.

By deploying these transformer blocks at different depths of the encoder, R-SWTNet adaptively models long-range dependencies across varying feature scales: shallow layers focus on short-range context (e.g., local road texture coherence), while deeper layers capture long-range relationships (e.g., connectivity between distant road segments). This hierarchical adaptation enhances the model’s ability to interpret complex rural road topologies, particularly in scenarios with severe occlusions or narrow, winding alleys—ultimately improving segmentation accuracy and robustness for unstructured rural road networks.

2.2.3. The Structure of Atrous Spatial Pyramid Pooling

The Atrous Spatial Pyramid Pooling (ASPP) block (Figure 7) in R-SWTNet is a critical component in the bottleneck layer, designed to enrich high-level features with multi-scale contextual information—essential for capturing rural roads of varying widths (from narrow alleys to main roads) and mitigating the impact of occlusions.

The ASPP block consists of multiple parallel convolutional branches that process the input feature map at distinct scales, then concatenate their outputs to aggregate multi-scale context. By leveraging atrous convolution with varying dilation rates, ASPP effectively captures context at multiple scales: small dilation rates focus on local details like narrow alley edges, while larger rates capture broader contextual cues such as the spatial relationship between a road and surrounding buildings or vegetation. The global average pooling branch further enhances the model’s awareness of “road-like” regions in complex rural scenes, where roads may be fragmented by trees or occluded by houses. Together, these mechanisms enable the bottleneck layer to synthesize rich, multi-scale context, which is critical for accurately segmenting rural roads with blurred boundaries, irregular morphologies, and diverse widths.

In the R-SWTNet architecture, the ASPP block processes high-level features from the encoder’s deepest layer (after patch merging), enriching them with multi-scale context before passing to subsequent Swin Transformer blocks. This ensures that the decoder receives feature maps encoding both fine-grained details and global structure, ultimately improving the model’s ability to reconstruct continuous, accurate road networks in unstructured rural environments.

2.2.4. The Structure of CAM-Residual Block

The CAM-Residual (Channel Attention Mechanism-Residual) block is a key component in R-SWTNet’s decoders, designed to enhance feature refinement while maintaining computational efficiency (Figure 8). It embeds a channel attention mechanism within residual connections, enabling adaptive emphasis on critical road features while suppressing irrelevant background noise.

Structurally, the block processes concatenated features from encoders and skip connections: first, residual paths preserve low-level details to mitigate gradient vanishing during training; second, the channel attention module, which replaces traditional fully connected layers with lightweight convolutions, dynamically weights feature channels based on their contribution to road segmentation, reducing parameter complexity. This dual design strengthens the model’s ability to discern intricate topological details.

In rural road segmentation, CAM-Residual addresses two core challenges: (1) enhancing discriminability between spectrally similar features via channel-wise weighting; (2) stabilizing training for imbalanced datasets (fewer road pixels) by preserving gradient flow through residual paths. By refining decoder features, it ensures accurate reconstruction of rural road planar morphology, particularly for narrow, occluded alleys critical to village mobility networks.

2.3. Loss Function

2.3.1. Focal Loss

Binary cross-entropy loss (BCE loss) is one of the most popular loss functions in binary classification scenarios. The CE of a pixel with a predicted value of p and a true label of y is as follows:

$$\mathrm{C}\mathrm{E}(p,y)=\left\{\begin{array}{cc}-\log (p)& \mathrm{i}\mathrm{f} y=1\\ -\log (1-p)& \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\end{array}\right.$$

(2)

If we assign a parameter $p_t$ as

$$p_t=\left\{\begin{array}{cc}p& \mathrm{i}\mathrm{f} y=1\\ 1-p& \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\end{array}\right.$$

(3)

then the CE value can be simplified to

$$\mathrm{C}\mathrm{E}(p,y)=\mathrm{C}\mathrm{E}\left(p_t\right)=-\log (p_t)$$

In tasks where the foreground and background are imbalanced, focal loss [47] is more suitable than cross-entropy. In Table 1, the foreground ratio of each village stays below 10%, indicating the applicability of focal loss. Compared with cross-entropy, focal loss introduces factors $\alpha$ and $\gamma$ in the form of

$$\mathrm{F}\mathrm{L}\left(p_t\right)=-{\alpha }_t(1-p_t{)}^{\gamma }\log (p_t),$$

(4)

where

$${\alpha }_t=\left\{\begin{array}{cc}\alpha & \mathrm{f}\mathrm{o}\mathrm{r} y>1\\ 1-\alpha & \mathrm{o}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{w}\mathrm{i}\mathrm{s}\mathrm{e}\end{array}.\right.$$

(5)

Generally, the value of $\alpha$ stays around 0.75, and $\gamma$ is often set to 2. By introducing these two factors, the easy-to-classify samples are less important, and the hard-to-classify ones are amplified, forcing models to focus on underrepresented classes. This adaptability ensures better recognition of hard-to-classify cases, balancing performance across all categories and enhancing model robustness in imbalanced datasets. The focal loss of a whole image is the arithmetic mean of $\mathrm{F}\mathrm{L}\left(p_t\right)$.

2.3.2. Tversky Loss

Tversky loss [48], an extension of the Tversky index, is optimized for imbalanced segmentation tasks by introducing adjustable weights for false positives (FP) and false negatives (FN), making it particularly effective for rural road segmentation—where narrow alleys (small targets) and class imbalance (fewer road pixels) demand precise control over error penalties. Unlike Dice loss (which treats FP and FN equally), Tversky loss dynamically balances these errors via hyperparameters α and β, enabling targeted optimization for rural road challenges like under-segmentation of narrow paths or over-segmentation of background clutter.

The Tversky loss for a binary segmentation task is defined as:

$$\mathrm{T}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{s}\mathrm{k}\mathrm{y} \mathrm{L}\mathrm{o}\mathrm{s}\mathrm{s}=1-{\displaystyle \frac{TP+ϵ}{TP+\alpha FP+\beta FN+ϵ}}$$

(6)

where TP (true positives), FP (false positives), and FN (false negatives) form the confusion matrix; $ϵ$ (often 1 × 10⁻⁸) avoids division by zero; α and β weight FP and FN, respectively. For rural roads, we set α = 0.3 and β = 0.7 to prioritize reducing FN, which is critical for capturing narrow alleys and maintaining road continuity.

This design enhances sensitivity to small, sparse structures: by amplifying FN penalties (β = 0.7), the model focuses on recovering fragmented road segments, while controlled FP weighting (α = 0.3) balances precision. Compared with Dice loss [49] (which averages TP/FP/FN), Tversky loss adaptively aligns with rural road characteristics, ensuring both continuity of major roads and preservation of narrow alley details, critical for reconstructing complete village mobility networks.

2.3.3. Dynamic Hybrid Loss

To address the dual challenges of class imbalance (fewer road pixels) and fine-grained structure preservation (narrow alleys) in rural road segmentation, we propose a Dynamic Hybrid Loss (DHLoss) that adaptively combines Focal loss and Tversky loss via a task-aware weight scheduler. This design leverages the strengths of both losses: Tversky loss optimizes spatial overlap for small targets, while Focal loss prioritizes hard-to-classify pixels, ensuring robust segmentation of complex rural scenes.

DHLoss is formulated as:

$$\mathrm{D}\mathrm{H}\mathrm{L}\mathrm{o}\mathrm{s}\mathrm{s}=\lambda \cdot \mathrm{F}\mathrm{o}\mathrm{c}\mathrm{a}\mathrm{l} \mathrm{L}\mathrm{o}\mathrm{s}\mathrm{s}+(1-\lambda )\cdot \mathrm{T}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{s}\mathrm{k}\mathrm{y} \mathrm{L}\mathrm{o}\mathrm{s}\mathrm{s}$$

(7)

In this expression, $\lambda$ is the dynamic ratio increasing epoch-wise in the first several warming-up epochs from 0 to ${\lambda }_{\mathrm{m}\mathrm{a}\mathrm{x}}$ (generally ranging from 0.5 to 0.7), so that in the early stages of training, the DHLoss consists mainly of Tversky loss, which focuses on maximizing spatial overlap between predicted and ground-truth road regions. This stabilizes initial convergence by prioritizing the recovery of large-scale road topology and mitigating false negatives (FN) for narrow alleys. As training progresses, Focal loss gradually gains weight, emphasizing hard-to-classify pixels (e.g., blurred alley boundaries, occluded segments by vegetation). By down-weighting easy samples and up-weighting hard samples (e.g., road-edge pixels confused with field ridges), Focal loss refines boundary precision and reduces false positives (FP) from background clutter, complementing Tversky loss’s focus on spatial overlap.

This dynamic combination ensures both initial topology capture and late-stage detail optimization: Tversky loss establishes a baseline road network structure, avoiding under-segmentation of narrow alleys, while focal loss sharpens boundary definition and suppresses background interference, enhancing overall segmentation fidelity. For rural road scenes—characterized by imbalanced data, small targets, and ambiguous boundaries—DHLoss balances global structural integrity and local detail accuracy, outperforming single-loss strategies (e.g., Focal or Tversky alone) in both quantitative metrics (IoU, F1-score) and visual continuity of segmented roads.

3. Experiment and Results

3.1. Optimization and Learning Rate Management

In deep learning-based segmentation tasks, the choice of optimizer and learning rate scheduling directly impacts model convergence, generalization, and training stability, which is particularly critical for complex rural road scenes with imbalanced data and fine-grained features. This study employs a robust optimization pipeline combining the AdamW optimizer [50,51], with a sequential learning rate scheduler complemented by memory-efficient training strategies, to enable stable training of the R-SWTNet architecture.

The AdamW optimizer is selected for its adaptive gradient descent capabilities and effective weight decay regularization, configured with a base learning rate of 1 × 10⁻⁴, weight decay of 1 × 10⁻⁴, and batch size of 4. To balance training stability and convergence speed, the learning rate is dynamically adjusted via a two-stage sequential scheduler: (1) A linear warm-up phase [52] over the first 10 epochs, starting from 0.1× the base rate (1 × 10⁻⁵) and gradually increasing to the full base rate. This mitigates unstable early training caused by large initial gradients. (2) A cosine annealing scheduler with warm restarts [53], which oscillates the learning rate between the base value and a minimum of 1 × 10⁻⁶, with an initial cycle length of 15 epochs that doubles after each restart ($T_{multi}$ = 2). This strategy encourages exploration of the loss landscape, facilitating convergence to flat minima [54]—critical for robust generalization [55] to unseen rural road scenes.

To accommodate larger batch sizes under computational constraints, the learning rate is scaled proportionally to the square root of the batch size (e.g., doubling the batch size increases the learning rate by $\sqrt{2}$), a conservative scaling strategy that avoids overfitting compared with linear scaling. Additionally, memory efficiency is enhanced via gradient checkpointing and mixed-precision training (using autocast for 16-bit floating-point operations), enabling training with larger batch sizes without compromising model capacity.

This integrated optimization framework balances three key objectives: (1) adaptive gradient updates via AdamW to handle imbalanced rural road data; (2) stable convergence via warm-up and cosine annealing to avoid local minima; and (3) memory-efficient training via checkpointing and mixed precision to support larger batch sizes, ultimately accelerating convergence and improving model robustness.

3.2. Evaluation Metrics

To comprehensively assess the performance of rural road segmentation models, we select four core metrics: Precision, Recall, Intersection over Union (IoU), and F1-score. These metrics are tailored to address the unique challenges of rural road scenes, including class imbalance (fewer road pixels), narrow target preservation (alleys < 3 m), and boundary ambiguity (blurred edges from occlusions).

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{TP}{TP+FP}}$$

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{TP}{TP+FN}}$$

$$\mathrm{I}\mathrm{o}\mathrm{U}={\displaystyle \frac{TP}{TP+FP+FN}}={\displaystyle \frac{|A\cap B|}{|A\cup B|}}$$

$$\mathrm{F}1 \mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\displaystyle \frac{2\times TP}{2\times TP+FP+FN}}={\displaystyle \frac{2|A\cap B|}{\left|A\right|+\left|B\right|}}={\displaystyle \frac{2\times \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}}$$

Precision focuses on suppressing background clutter (e.g., field ridges or dry riverbeds) by evaluating the proportion of correctly predicted road pixels, directly mitigating false positives that arise from spectral similarity between roads and non-road features. Recall prioritizes capturing narrow alleys (<3 m) and occluded segments by measuring the model’s ability to recover all true road pixels, reducing false negatives critical for preserving rural road networks’ continuity. IoU, the primary spatial metric, quantifies overlap between predicted and ground-truth road regions, explicitly reflecting the model’s capacity to maintain spatial integrity (e.g., avoiding fragmented segmentation of winding paths). F1-score, the harmonic mean of Precision and Recall, balances these two metrics to counteract class imbalance (road pixels < 5% of total) and prevents bias toward over-segmentation (high Precision but missed narrow lanes) or under-segmentation (high Recall but excessive background misclassifications). Together, these metrics collectively validate the model’s robustness across rural-specific scenarios, from fine-grained edge preservation to large-scale topological consistency.

3.3. Comparative Experiment

3.3.1. Experiment Design

To systematically evaluate the performance of the proposed R-SWTNet architecture, comparative experiments were conducted against five benchmark models, spanning classic, transformer-based, residual-enhanced, and lightweight variants (

Table 4. The basic information of the models.

No.	Model Name	Key Features	Params (M)	FLOPs (G)
1	U-Net	Foundational encoder–decoder structure	31.04	219.08
2	Swin-UNet	As defined in related literature	23.77	105.38
3	R-Net	U-Net using ResNet34 as encoder backbone with classic decoder	24.35	117.42
4	R-Net-DSC	R-Net with depthwise separable convolution	4.30	25.19
5	D-LinkNet [56]	Similar to R-Net, using dilated convolution in upsampling	23.69	110.96
6	R-SWTNet	The proposed model prototype	34.24	126.28
7	R-SWTNet-DSC	R-SWTNet with depthwise separable convolution	20.93	58.13

).

All models were trained under identical experimental conditions to ensure fair comparison: 100 epochs with a batch size of 16, early stopping to prevent overfitting on the validation set, and the same optimization pipeline (as detailed in Section 3.1). Input images were standardized to 512 × 512 pixels, and data augmentation was applied uniformly to all models.

To investigate the impact of lightweight design on rural road segmentation, two DSC-based variants were constructed by replacing all standard convolutional layers in the ResNet34 backbone blocks with depthwise separable convolutions. This modification decouples spatial and channel-wise operations, drastically reducing parameter counts. These variants enabled analysis of the gains and trade-offs between computational efficiency and segmentation accuracy, critical for deploying models in resource-constrained rural environments.

By spanning architectural spectrums—from vanilla U-Net to hybrid transformer convolutional models, and from full-capacity ResNet to lightweight DSC variants—this experimental design isolates the contributions of key components in R-SWTNet: residual feature extraction (ResNet34), long-range dependency modeling (Swin Transformer), multi-scale context fusion (ASPP), and channel-wise attention (CAM-Residual). Collectively, these comparisons validate whether the proposed architecture achieves superior performance by synergistically integrating these mechanisms, rather than relying on isolated innovations.

3.3.2. Experiment Results

The quantitative performance of all models is summarized in

Table 5. The quantitative metrics of the models.

No.	Model Name	Train Precision (%)	Val Precision (%)	Train Recall (%)	Val Recall (%)	Train IoU (%)	Val IoU (%)	Train F1 Score (%)	Val F1 Score (%)	Overfitting (%) 1	Relative Overfitting (%) 2
1	U-Net	31.69	43.97	59.43	32.07	26.05	22.76	41.34	37.08	4.26	11.49
2	Swin-UNet	55.90	46.11	74.78	70.77	47.03	38.73	63.98	55.84	8.14	14.58
3	R-Net	81.04	69.73	92.60	69.25	76.01	53.15	86.43	69.41	17.02	24.52
4	R-Net-DSC	75.24	65.63	88.50	64.20	68.55	48.05	81.34	64.91	16.43	25.31
5	D-LinkNet	86.40 3	73.81	95.72	67.79	83.18	54.64	90.82	70.67	20.15	28.51
6	R-SWTNet	79.55	69.99	91.41	71.11	74.01	54.88	85.07	70.87	14.20	20.04
7	R-SWTNet-DSC	77.98	66.77	90.40	65.01	72.02	49.12	83.73	65.88	17.85	27.09

¹ The overfitting metric is calculated as training f1 minus validation f1. ² Relative overfitting metric is the overfitting over F1-score on the validation set. ³ The overall best metric of each column is emphasized with bold font.

, which reveals distinct disparities in segmentation accuracy and parameter efficiency across architectural paradigms. Of all the trained models, R-Net, D-LinkNet, and R-SWTNet converge around epoch 35–40, after which the metrics on the validation set struggle to improve, while the others take more time to converge.

Overall, the baseline U-Net exhibits the lowest performance across all metrics, even with the second-highest parameter counts, with validation IoU and F1-score values significantly below those of other models, confirming its inadequacy for complex rural road segmentation. In contrast, transformer-based and residual-enhanced models demonstrate marked improvements, though their performance varies with architectural design.

Swin-UNet, representing pure transformer architectures, achieves higher training precision and recall than U-Net but shows a substantial drop in validation metrics, particularly in IoU and F1-score, indicating limited generalization to unseen rural scenes. Residual-based models, by comparison, perform consistently better: R-Net (ResNet34 backbone) outperforms Swin-UNet in both training and validation metrics, with validation IoU and F1-score exceeding those of Swin-UNet by notable margins. R-Net, D-LinkNet, and the proposed R-SWTNet share similar validation metrics and FLOPs, yet the proposed R-SWTNet has the least overfitting with the most parameters. Generally speaking, the more the parameters, the more possible to overfit, the seemingly contradictory phenomenon indicates the superiority of the R-SWTNet.

Lightweight variants with depthwise separable convolutions (DSC) exhibit clear parameter efficiency trade-offs. R-Net-DSC and R-SWTNet-DSC reduce parameter counts compared with their non-DSC counterparts. However, this parameter reduction is accompanied by consistent drops in both training and validation metrics, with validation IoU decreasing by 7–8 percentage points for both variants and a slight increase in overfitting.

The actual segmentation results vary across models, with visual patterns in Figure 9 corroborating the quantitative trends in Table 5. Visually, the U-Net model struggles to distinguish rural roads from background features, often misclassifying buildings as roads, which is a tendency rooted in its architectural bias toward large, high-contrast regions. U-Net’s encoder–decoder structure prioritizes local, block-like features and loses fine-grained details during downsampling, failing to capture narrow alleys or occluded segments. In low-occlusion scenarios (e.g., scenarios A, C, and H), the U-Net model performs better, but is still far from optimal. This results in blurred boundaries and incomplete road networks, aligning with its low IoU scores (22.76% validation).

The pure transformer-based Swin-UNet also suffers from misclassifications. Unlike U-Net, the Swin-UNet can recognize semantics mostly properly, but due to the lack of capabilities in local feature extraction, the boundaries of the foreground remain blurred. Experiments were taken to replace the original four times downsampling with two times downsampling in the first layer, only to find a massive increase in training cost. In contrast, residual-based models exhibit clearer segmentation. R-Net (ResNet34 backbone) outperforms Swin-UNet visually, with sharper road boundaries and fewer false positives, confirming residual connections’ role in stabilizing local feature extraction.

The proposed R-SWTNet demonstrates incremental but critical visual improvements over benchmarks. It reaches the best balance of local and global features. The pure CNN-based models, like the R-Net, D-LinkNet and R-Net-DSC, if looked closely, have lots of spurs and small white dots in their segmentation results, while R-SWTNet, thanks to its transformer-based encoder3 and bottleneck, reduces this greatly. Moreover, in scenario F, R-SWTNet performs the best in segmenting the playground in the middle, though far from optimal. In addition, the comparison of D-LinkNet and R-SWTNet correlates with the metrics in Table 5. In scenarios B, D, E, and F, the alleys R-SWTNet recognizes are a little bit wider than the alleys D-LinkNet recognizes, which is cross-validated by the difference in Precision and Recall on the validation set.

Lightweight DSC variants (R-Net-DSC, R-SWTNet-DSC) exhibit noticeable visual degradation, particularly in scenario B. In scenario B, R-Net-DSC fails to identify the Chitan Old Town parking lot, a distinct white square in high-resolution imagery. Their non-DSC counterparts (R-Net, R-SWTNet), on the contrary, both segment it correctly. The degradation reflects reduced feature extraction capacity from parameter compression, aligning with the quantitative IoU drops (5–6 percentage points), confirming DSC’s trade-off between efficiency and representational power for complex rural scenes.

4. Discussion

4.1. Necessity of Transformer-CNN Conjunction

For rural road segmentation, addressing occlusion and discontinuous road segments hinges on effectively modeling long-range dependencies—a capability inherently limited in vanilla U-Net architectures. Two primary strategies exist to mitigate this limitation: data augmentation and architectural enhancement. Among these, architectural innovations have proven most impactful for capturing the complex topological relationships of rural road networks.

Transformers are indispensable for segmenting occluded rural roads, where long-range contextual reasoning is critical. Unlike CNNs, which rely on local receptive fields and implicitly construct connections between distant pixels with extended receptive fields by stacking multiple layers, transformers dynamically and explicitly weight semantic relevance across an entire image. For instance, the hierarchical SW-MSA mechanism in the Swin Transformer progressively builds global context by correlating disconnected segments across non-overlapping windows, enabling coherent reconstruction of occluded roads (Figure 9). This explains the drastic performance gap between transformer-equipped models (Swin-UNet, R-SWTNet) and the baseline U-Net, whose convolutional inductive bias prioritizes local patterns, leading to fragmented segmentation of rural alley networks.

Despite their superior ability to model long-range dependencies, transformer mechanisms exhibit inherent limitations in local feature granularity and training convergence for pixel-wise segmentation tasks. These drawbacks necessitate hybrid architectures that merge the strengths of both paradigms: CNNs for local feature precision, and transformers for long-range contextual reasoning. The proposed R-SWTNet exemplifies this synergy by integrating ResNet34 and Swin Transformer modules. ResNet34’s hierarchical convolution layers efficiently extract low-level features and preserve spatial resolution, mitigating the “local blur” issue of pure transformers. Meanwhile, Swin Transformer blocks model long-range dependencies, correlating fragmented road segments across distant image regions. This hybrid design outperforms pure transformer architectures and pure CNN architectures, achieving higher validation IoUs or lower overfitting. For rural road segmentation—where both narrow alleys (local) and network continuity (global) are critical—such hybrid integration transitions from architectural choice to necessity.

4.2. The Gains and Trade-Offs of DSC Module

Depthwise Separable Convolution (DSC) decouples spatial and channel-wise operations, drastically reducing parameters by factorizing standard convolution into depthwise (spatial) and pointwise (channel) convolutions, drastically reducing parameters and FLOPs.

However, this parameter compression introduces critical trade-offs. Quantitatively, DSC variants exhibit consistent performance degradation: R-SWTNet-DSC’s validation IoU drops by 5.76% (from 54.88% to 49.12%), and R-Net-DSC’s validation IoU decreases by 5.10% (from 53.15% to 48.05%). Visually, DSC models fail to segment fine-grained rural features due to reduced model capacity, creating a “shallow representation bottleneck”.

Training dynamics further highlight DSC limitations. DSC-equipped models (e.g., R-Net-DSC) converge more slowly than their standard convolution counterparts, partly because DSC replaces ResNet34’s pre-trained kernels with randomly initialized depthwise layers, disrupting transfer learning benefits.

For rural road segmentation, DSC’s trade-offs are context-dependent: it offers a feasible compromise for resource-constrained edge devices (e.g., on-site mapping with limited memory), but its accuracy loss makes it unsuitable for precision-critical tasks (e.g., narrow alley preservation). R-SWTNet’s non-DSC variant thus remains preferable for most rural applications, balancing performance and efficiency without sacrificing fine-grained feature extraction.

4.3. Labeling Noise and Training Ambiguities

Labeling noise and training ambiguities collectively undermine the robustness of rural road segmentation models, stemming from the intrinsic complexity of rural road semantics and limitations in remote sensing data. Rural roads lack standardized definitions, unlike their urban counterparts, which have clear functional boundaries and hardened surfaces. This semantic fuzziness is exacerbated by morphological diversity: pathways may be unpaved, grass-covered, or barely distinguishable from adjacent features like field ridges or footpaths. When narrowed, their spectral signatures in satellite imagery converge with the surrounding terrain. For example, a sunlit earth alley and a dry field ridge often exhibit near-identical reflectance under midday light, making spectral cues alone insufficient for reliable classification.

Compounding this, manual labeling introduces subjective errors due to inconsistent criteria for defining “roads”. Annotators may label a 1.5 m wide compacted path as “foreground” (due to frequent pedestrian use) or “background” (due to lack of vehicular traffic), creating conflicting labels for identical features. Such noise propagates through training: models learn unstable patterns when locally similar receptive fields (e.g., a 2 m wide alley vs. a 2 m wide ridge) have contradictory labels, leading to erratic segmentation. For instance, scenario D (LT) in Figure 9 shows misclassification of a canal dyke as a road (Figure 10), likely because of the spectral similarity of concrete dyke and road with similar curvatures, and the opposite labels.

Training ambiguities further amplify these issues, driven by satellite imagery with relatively low resolution that homogenizes inner-village textures, blurring subtle topological details (e.g., alley edges vs. courtyard walls). This forces models to rely on global context over local features, increasing vulnerability to labeling errors. For example, squares and parking lots are labeled “foreground” (road), while courtyards are “background”, but their spectral similarity and blurred boundaries (e.g., walls indistinct in satellite imagery) make this distinction unreliable. As shown in scenario H of Figure 9, none of the models clearly segments the “open space” at the road junction—a failure rooted in training data, where such spaces were inconsistently labeled, leaving the model to “speculate” based on incomplete cues.

These intertwined challenges—labeling noise (semantic fuzziness and subjective annotation) and training ambiguity (feature confusion from low resolution and inconsistent labels)—create a bottleneck for rural road segmentation. Models trained under such conditions struggle to balance recall for narrow alleys and precision for background clutter, highlighting the need for multi-source validation (e.g., cross-referencing satellite and aerial imagery) and expanded datasets with diverse “open space” semantics to reduce reliance on unstable spectral cues.

5. Conclusions

This study addresses the critical challenge of accurately segmenting unstructured rural roads and narrow alleys using deep learning, proposing a context-aware framework and dataset to advance rural road extraction technology. By integrating hierarchical feature extraction, long-range dependency modeling, and multi-scale context fusion, the research provides a robust solution for data-driven rural planning under China’s Rural Revitalization Strategy.

5.1. Context-Aware Hybrid Architecture: Advancing Rural Road Segmentation

The proposed R-SWTNet architecture achieves superior segmentation performance by synergistically integrating ResNet34, Swin Transformer, ASPP, and CAM-Residual modules. ResNet34’s hierarchical convolutions extract low-level features (textures, edges) with high spatial resolution, while the Swin Transformer models long-range dependencies via shifted window attention, correlating fragmented road segments across distant image regions. The ASPP module enriches multi-scale contexts to adapt to varying road widths, and CAM-Residual blocks refine decoder features by suppressing background noise and emphasizing critical road boundaries.

Quantitatively, R-SWTNet outperforms benchmark models on the SQVillages dataset: it achieves a validation IoU of 54.88% and F1-score of 70.87% (Table 5), exceeding U-Net (22.76% IoU) and Swin-UNet (38.73% IoU), and is slightly superior to R-Net (53.15% IoU) and D-LinkNet (54.64% IoU), with remarkably lower overfitting. Visually, it demonstrates superior capability in segmenting narrow alleys and maintaining road continuity, confirming its robustness (Figure 9). This hybrid design resolves the limitations of pure CNNs (poor long-range reasoning) and pure transformers (local detail loss), establishing a new paradigm for unstructured rural road segmentation.

5.2. SQVillages Dataset: Enabling Robust Model Generalization

To address the scarcity of rural-specific road datasets, this study constructs the SQVillages dataset, integrating multi-source remote sensing data (satellite and aerial imagery) and adaptive augmentation. The dataset includes 18 villages across diverse geomorphic types (polders, plains, river-side villages) and infrastructure conditions (traditional villages, resettlement communities).

Key innovations of SQVillages include:

(1)

Multi-source labeling (cross-validating satellite imagery with sub-meter aerial orthomosaics to reduce spectral confusion errors);

(2)

Village-level stratified splitting (isolating validation/test sets by village to avoid data leakage). This dataset ensures model generalization to unseen rural scenes, as evidenced by R-SWTNet’s consistent performance across training and validation sets.

5.3. Limitations and Future Work

Despite the promising results, this study has several limitations. First, the SQVillages dataset, while diverse, is geographically focused on eastern China, and model performance on villages in vastly different landscapes (e.g., high-altitude plateaus or deserts) remains unexplored. Second, the computational cost of the full R-SWTNet model may hinder its deployment on extremely resource-constrained edge devices.

These limitations point to clear directions for future research. Our immediate next steps include:

(1)

Expanding the SQVillages dataset to incorporate villages from western China’s mountainous regions to test model generalization further;

(2)

Exploring more advanced model compression and quantization techniques to enhance deployment feasibility without significant accuracy loss;

(3)

Investigating semi-supervised learning paradigms to reduce the heavy reliance on costly pixel-wise annotations.

In conclusion, this study achieved its initial aim of developing a robust deep learning framework for segmenting rural streets and alleys in China. By addressing both architectural innovation and dataset construction, we have advanced the state-of-the-art in rural road extraction and laid a foundation for intelligent rural infrastructure management.