An Open Dataset of Yangtze River Docks Based on OSM and Google Satellite Imagery (2024)

Abstract

As critical infrastructure at the sea-land interface, docks support regional economies and transportation networks. They include seaport docks and inland docks, the latter widely distributed along rivers, lakes, and canals. However, existing studies have mainly focused on seaport dock identification, with limited research and data products for inland docks. To address this gap, we developed an inland dock identification method for the Yangtze River by integrating crowdsourced OpenStreetMap (OSM) data with high-resolution Google satellite images. Based on OSM-derived prior information, we constructed a dedicated dock annotation dataset and applied YOLO-series models with a multi-scale detection strategy to identify and classify docks. The method achieved precision, recall, and F1-score above 0.9. A total of 3,562 docks were detected, including 2,738 floating and 824 vertical docks. To support diverse applications, the released dataset provides both bounding box representations and polygon vector delineations. As the first open inland dock dataset for the Yangtze River, it offers valuable support for studies on inland dock systems, waterway optimization, and regional economic analysis.

Background & Summary

As the core infrastructure at the sea-land interface in the port systems, docks serve as critical physical units for the realization of port functions. Based on their geographical locations and service functions, docks are generally classified into seaport docks and inland docks¹. Seaport docks are primarily distributed in areas along coastlines, facilitating international shipping and the import and export of bulk commodities², and thus constitute essential nodes in the global trade network. Inland docks, by contrast, are located along inland waterways such as rivers, lakes, and canals³. Compared with seaport docks, inland docks are more closely integrated with local industries, agriculture, and community life, and are characterized by decentralized layouts, diverse functions, and extensive service coverage. For example, Europe has more than 41,000 kilometers of navigable inland waterways connecting 25 EU member states and hundreds of European cities⁴. In Brazil, inland waterway transport reached 33.79 million tons in 2023, with the Amazon waterway carrying nearly half of the volume⁵. These examples highlight the crucial role of inland docks in spatial planning and regional economic development⁶.

Acquiring data on inland docks is essential not only for revealing the spatial structure and functional patterns of inland port systems, but also for supporting regional economic growth^7,8. Remote sensing technology refers to a non-contact method for acquiring information about the Earth’s surface using sensors, and it is considered one of the primary techniques for identifying or extracting port infrastructure^9–11. For example, Liu et al.⁹ identified ports in the Singapore region using polarimetric synthetic aperture radar (SAR) data. He et al.¹⁰ developed a port extraction algorithm based on high-resolution satellite images. Liu et al.¹¹ further proposed an unsupervised port detection method based on contextual and polarization-reflection-symmetry features to extract ports in Dalian (China), San Francisco (USA), and Singapore. Moreover, Zhang et al.¹² proposed a seaport segmentation algorithm integrating spatial and semantic alignment fusion. Despite these advancements, existing research still exhibit at least two major limitations.

Firstly, there is a lack of research on the identification of inland docks. Compared with seaports, inland docks are generally smaller in scale and more spatially constrained, making them easily confused with ships in inland waterways and therefore more difficult to identify accurately.

Secondly, there is a shortage of publicly available dataset related to inland docks. Research on the optimization of inland waterway transport networks and the socio-economic development of inland dock clusters still lacks sufficient data support.

The Yangtze River, the world’s third-longest river, stretches for 6,397 kilometers and drains a basin of 1.8 million square kilometers, accounting for 20% of China’s land area. It is the most developed, largest, and busiest inland waterway transport in China, contributing over 40% of the national Gross Domestic Product (GDP)¹³. Due to significant variations in topography and water level along its course, inland dock facilities distributed across the Yangtze River exhibit high structural diversity. Generally, these inland docks can be classified into two main types (Fig. 1): (1) floating docks, which are secured by floating platforms and anchoring systems, with connections between ship and shore provided by ramps or fixed trestle bridges; and (2) vertical docks, which are supported by reinforced concrete columns. This study selects the Yangtze River as the research area, as it not only reflects the typical characteristics of inland docks across the upper, middle, and lower reaches under diverse geographical and hydrological conditions, but also covers the main types of docks along the river, making it representative.

Fig. 1.

Yangtze River inland docks and classification: (a) floating dock; (b) vertical dock (Maps Data ©2024 https://www.google.com/maps/).

This study aims to develop an openly accessible inland dock dataset for the Yangtze River. The methodology primarily involves using vector data of docks (from OpenStreetMap, OSM) provided by volunteers as prior knowledge to construct an annotated dock dataset for the Yangtze River. Subsequently, YOLO-series object detection models are employed to identify inland docks along the Yangtze River. Our primary contributions include:

1. Proposed a method for inland dock identification that integrates crowdsourced geographic information (OSM) and high-resolution satellite images, effectively reduces misclassification between docks and ships.

2. Based on the proposed method, we developed an open dock dataset for the Yangtze River, detecting a total of 3,562 docks, including 2,738 floating docks and 824 vertical docks.

Methods

Overview

This study proposes a method for identifying inland docks (hereinafter referred to as “dock”) by integrating open-source geographic data from OpenStreetMap (OSM) and high-resolution Google satellite images (Fig. 2). First, an M × M regular grid is generated around the geometric center of OSM dock labelling data. The actual dock boundaries within each grid cell are manually annotated using the corresponding Google images to create the annotation set. Subsequently, various object detection models are trained using the annotation set and corresponding Google images to determine the optimal model. Finally, the optimal model is applied to all Google images partitioned into specific sizes across the study area, detecting dock locations and delineating their boundaries for the construction of the Yangtze River Dock Dataset.

Fig. 2.

The workflow of our approach.

Data acquisition

The OSM data (©OpenStreetMap https://www.openstreetmap.org/copyright) refers to the map database collaboratively edited online by global volunteers. This data is not only openly accessible but also contains diverse geographic features such as roads, buildings, water systems, land use, and land cover. In this study, the third-party platform Geofabrik (https://download.geofabrik.de/) was first utilized to obtain water data covering the Yangtze River Basin, which served as the study area (represented as polygon vectors). Subsequently, using the QuickOSM plugin (https://docs.3liz.org/QuickOSM/) in QGIS and applying the tags “Man_made = pier” and “Waterway = dock” (Table 1), we extracted a total of 3,614 polygon features related to docks within the study area. All the OSM data were acquired in 2024.

Table 1.

Tags related to docks in OSM data (https://wiki.openstreetmap.org/wiki/Tags).

Tags*	Description
Waterway = dock	A place for constructing, repairing and storing ships out of the water.
Man_made = pier	A pier is a raised walkway over water, supported by widely spread piles or pillars.

*OSM tag consists of a key-value pair, with the “ = ” indicating the key on the left and the value on the right.

We acquired 0.1 m resolution Google Satellite images for the study area using the QuickMapServices plugin in QGIS (https://nextgis.com/blog/quickmapservices/). The images used in this study were all acquired in 2024. At present, this represents the only freely accessible sub-meter resolution satellite images. While medium- and low-resolution satellite data (e.g., Sentinel-2 and Landsat) are also freely accessible, their spatial resolution (e.g., 10 meters or lower) is insufficient for identifying small docks along inland rivers.

Development of the dock label set

We first generated a regular M × M grid centered on the centroid of 3,614 OSM dock polygon features (Fig. 3a). After evaluating multiple grid sizes, we set M to 500 meters, as this specific dimension can adequately encompass the vast majority of docks. Subsequently, Google images within each grid cell were acquired (Fig. 3b). The dock boundaries and types, visually interpreted from the imagery, were manually delineated and annotated using LabelImg. This involved adding an attribute field to each bounding box to record the dock type. Specifically, for floating docks, only the pontoon (i.e., a rectangular, flat-bottomed vessel without propulsion) was annotated, while auxiliary facilities such as movable gangways or access bridges were excluded (e.g., the instance shown in Fig. 3c was labeled as a “floating dock”). For vertical docks, annotations generally only included the upper structural platform used for cargo handling and storage, whereas fixed access bridges and other auxiliary structures were not labeled. Theoretically, annotations for 3,614 docks could be obtained. However, due to data quality issues in OSM (such as volunteer editing errors or outdated information), annotations for only 2,717 docks were ultimately acquired. The remaining 897 docks were excluded as no corresponding dock structures could be identified in the Google images, and the final dock annotation dataset consists of 2,717 validated samples.

Fig. 3.

Schematic diagram of dock annotation dataset development.

Model training

To develop the Yangtze River Dock Dataset and identify the optimal object detection model, this study compares multiple classic models from the YOLO series, including YOLOv5, YOLOv8, YOLOv11 and YOLOv13, along with their enhanced variants. These improved models are based on the YOLOv5 baseline architecture and are individually integrated with the bi-directional feature pyramid network (BiFPN)¹⁴, the Coordinate Attention (CA)¹⁵, the Global Attention Mechanism (GAM)¹⁶, and the Simple, Parameter-Free Attention Module (SimAM)¹⁷. In total, eight object detection models are compared in this study. The complete dock annotation dataset, consisting of 2,717 validated samples and their corresponding Google satellite images, was divided into training, validation, and testing sets using a 7:1:2 ratio. The model was implemented in the PyTorch framework and trained on an NVIDIA RTX 2060 Ti GPU with 12 GB of memory. After several rounds of optimization, the key hyperparameters were finalized as follows: 200 training epochs, an initial learning rate of 0.01, and a batch size of 16.

Model application

Theoretically, the study area could be divided into grids, and an object detection model could be used to identify docks within the high-resolution remote sensing image of each grid (referred to in this study as single-scale detection). However, in inland river contexts, floating docks in remote sensing images exhibit features similar to those of ships (Fig. 4), which can lead to frequent misclassification between docks and ships.

Fig. 4.

Schematic diagram of a floating dock and a ship in Google images.

To address this issue, this study proposes a multi-scale detection strategy. The core concept of this strategy involves partitioning the study area using grids of two different sizes and performing dock detection on each grid. The final results are derived by comparing the two detection outcomes to eliminate ships that may have been misidentified as docks. The main procedural steps are as follows:

Step 1: A regular grid of N × N size is established over the entire study area, and the corresponding sub-meter high-resolution Google satellite images are acquired. The optimal object detection model is applied to detect docks within each grid, recording the results as Set I (Fig. 5a).

Fig. 5.

Schematic diagram of multi-scale detection strategy.

Step 2: Centered on the grids where docks were identified in Set I, a larger regular grid of 3 N × 3 N size is created, and the corresponding sub-meter high-resolution Google satellite images are obtained. The same model is then applied again to detect docks within these larger grids, recording the results as Set II (Fig. 5b).

Step 3: Calculate the similarity between each dock bounding box in Set I and each dock bounding box in Set II sequentially. The specific calculation formula is as follows:

$$\mathit{Similarity}=\frac{{\mathit{Area}}_{I\cap \mathit{II}}}{{\mathit{Area}}_I}$$ 1

${\mathit{Area}}_I$ denotes the area of a single dock bounding box in Set I, ${\mathit{Area}}_{I\cap \mathit{II}}$ represents the intersection area between a single dock bounding box in Set I and corresponding bounding box in Set II. When the Similarity exceeds a certain threshold (termed the Similarity Threshold), it indicates consistent detection between the two sets, and the corresponding dock in Set I is retained; otherwise, it is discarded.

Therefore, after performing single-scale detection using the eight object detection models, this study implemented multi-scale detection by establishing 3 N × 3 N regular grids centered on the grids where docks were identified in the single-scale results. Similarity thresholds of 0.3, 0.5, and 0.7 were selected for experimentation, because an excessively low threshold may introduce false positives by misclassifying ships as docks, while an excessively high threshold may lead to false negatives by incorrectly discarding actual docks.

Results evaluation

To validate the effectiveness of the developed Yangtze River Dock Dataset, we randomly sampled 10% of the N × N grids (356 grids in total) where docks were detected, and extracted an equal number of grids from areas where no docks were detected. The reference dock boundaries within these sampled grids were then annotated through visual interpretation of Google satellite images. The dataset validity was quantitatively assessed using three metrics: precision, recall and F1-score. Furthermore, we analyzed the effects of different grid sizes (300 m, 500 m, 700 m), different similarity thresholds (0.3, 0.5, 0.7), and the multi-scale detection strategy on dock identification performance.

Finally, utilizing the optimal parameter combination, all docks and their bounding boxes within the study area were identified. The dock boundaries were manually vectorized based on visual interpretation of Google satellite imagery, focusing on the pontoon structures of floating docks and the upper structural platforms of vertical docks, resulting in the Yangtze River Dock Dataset. To validate the accuracy of the vectorized dock boundaries, an Inter-Annotator Agreement approach was adopted¹⁸. Specifically, the centroid of each vectorized dock polygon was extracted, and approximately 10% of these centroids were randomly selected as samples. Two professional surveyors from the water resources sector were then invited to independently delineate the dock boundaries corresponding to the sampled locations based on visual interpretation of Google images. Finally, the vectorization results produced by each surveyor were compared with our dock dataset, and Cohen’s kappa coefficient was calculated to quantify the level of agreement¹⁹. The Cohen’s kappa value greater than 0.81 indicates a high level of consistency among different vectorization results²⁰, thereby validating the accuracy of the dock boundary delineation.

Data Records

The 2024 Yangtze River Dock Dataset developed in this study comprises two types of vector data in Esri Shapefile format: (1) Dock Bounding Box Dataset: Represented by polyline vectors, containing six fields such as the id of each dock bounding box, dock type, the coordinates of the top-right and bottom-left corners of the dock bounding box (as shown in Table 2). (2) Dock Dataset: Represented by polygon vectors, containing four fields including the id of each dock and the ID of its corresponding bounding box, dock perimeter and area (as shown in Table 3). This dataset product contains vector data for a total of 3,562 docks, consisting of 2,738 floating docks and 824 vertical docks, as illustrated in Fig. 6. The dataset has been made public under Figshare²¹. The data is projected using the World Geodetic System WGS84 and the Pseudo-Mercator projection coordinate system (EPSG: 3857).

Table 2.

The Yangtze River dock bounding box dataset.

Field	Description	Type
BID	The ID of each dock bounding box	Integer
Class	Dock type (floating or vertical)	String
Left_X	The x-coordinate of the bottom-left corner of the dock bounding box	String
Left_Y	The y-coordinate of the bottom-left corner of the dock bounding box	String
Right_X	The x-coordinate of the top-right corner of the dock bounding box	String
Right_Y	The y-coordinate of the top-right corner of the dock bounding box	String

Table 3.

The Yangtze River dock dataset.

Field	Description	Type
DID	The ID of each dock	Integer
BID	The ID of each dock bounding box	Integer
Perimeter (m)	Perimeter of the dock	Double
Area (m2)	Area of the dock	Double

Fig. 6.

Schematic diagram of the Yangtze River dock dataset.

Technical Validation

Comparative evaluation of different object detection models

Table 4 presents the dock detection results obtained using different object detection models, where “Floating” and “Vertical” denote the evaluation results for floating docks and vertical docks, respectively, and “All-type” represents the evaluation results for all dock types. In most cases, the precision, recall, and F1-score of the different models approach or even exceed 0.8, indicating that the YOLO series models achieve overall strong performance in inland dock detection tasks. In terms of precision, the YOLOv13 model performs best for “Floating” and “All-type” docks (0.866 and 0.871), respectively. Regarding recall, the YOLOv5 model achieves the highest values across “Floating”, “Vertical” and “All-type” docks (0.922, 0.942 and 0.928). In terms of F1-score, YOLOv5 achieves the highest F1-score for “Vertical” and “All-type” docks (0.927 and 0.879), while the YOLOv5_BiFPN model yields the best F1-score for “Floating” docks (0.870).

Table 4.

Comparison of dock detection results across different object detection models.

Object Detection Model	Class	Precision	Recall	F1-score
YOLOv5	Floating	0.805	0.922	0.859
	Vertical	0.913	0.942	0.927
	All-type	0.835	0.928	0.879
YOLOv5_BiFPN	Floating	0.836	0.907	0.870
	Vertical	0.924	0.864	0.893
	All-type	0.860	0.894	0.876
YOLOv5_CA	Floating	0.811	0.907	0.856
	Vertical	0.905	0.870	0.887
	All-type	0.836	0.896	0.865
YOLOv5_GAM	Floating	0.813	0.910	0.859
	Vertical	0.886	0.857	0.871
	All-type	0.833	0.894	0.862
YOLOv5_SimAM	Floating	0.810	0.912	0.858
	Vertical	0.924	0.864	0.893
	All-type	0.840	0.898	0.868
YOLOv8	Floating	0.865	0.799	0.831
	Vertical	0.838	0.740	0.786
	All-type	0.858	0.782	0.818
YOLOv11	Floating	0.843	0.760	0.799
	Vertical	0.863	0.695	0.770
	All-type	0.849	0.740	0.791
YOLOv13	Floating	0.866	0.780	0.821
	Vertical	0.875	0.767	0.817
	All-type	0.871	0.774	0.819

“Floating” and “Vertical” correspond to floating and vertical dock categories, respectively, and “All-type” refers to all dock types.

The recall of YOLOv5 is notably higher than that of YOLOv8, YOLOv11, and YOLOv13. This phenomenon may be explained by the fact that, under conditions of relatively limited training data, models with simpler architectures and fewer parameters, such as YOLOv5, are less prone to overfitting and can more effectively capture the general characteristics of docks. Consequently, they tend to detect a larger number of potential targets, which is reflected in relatively higher recall^22,23. Later iterations of the YOLO series (e.g., YOLOv8, YOLOv11, and YOLOv13) are generally more complex, and their performance advantages depend on abundant and diverse training data²⁴. Consequently, in scenarios with limited sample sizes or highly heterogeneous target geometries, these potential advantages may not be fully realized^25,26. Although the improved models based on YOLOv5 (YOLOv5_BiFPN, YOLOv5_CA, and YOLOv5_SimAM) show slight gains in precision, they underperform the baseline YOLOv5 in recall and F1-score. Potential reasons for this include: the BiFPN module may be more prone to losing weak feature targets in complex backgrounds, while the attention modules may excessively suppress certain target features due to strong attention constraints, thereby reducing recall¹⁵.

Furthermore, Fig. 7 illustrates the dock detection performance of different object detection models. As shown in Fig. 7, The YOLOv5 model accurately identified both the locations and boundaries of docks across all sample areas. However, in certain sample regions, all other models exhibit unavoidable missed detections (Figs. 7b-3, 7c-4, 7d-3, 7e-1, 7f-2, 7g-1, and 7h-2), in which contributes relatively lower recall rates. These observations are generally consistent with the quantitative results presented in Table 4.

Fig. 7.

Comparative analysis of dock detection performance across different object detection models.

Evaluation based on multi-scale detection strategy

Table 5 presents the evaluation results of the multi-scale detection strategy using the best-performing model YOLOv5, under different grid sizes and similarity thresholds. For comparative purposes, the single-scale detection results based on the YOLOv5 model are also included. For brevity, Table 5 presents only the evaluation results for all docks (“All-type”). As shown in Table 5, the multi-scale detection strategy significantly improves dock identification compared to the single-scale approach. Specifically, when employing grid sizes of 300 m, 500 m, and 700 m, the multi-scale method achieves maximum precision improvements of 14.34%, 8.5%, and 65.65%, respectively, relative to the single-scale method. Corresponding maximum improvements in F1-score are 5.55%, 3.41%, and 28.26%, indicating that the multi-scale strategy effectively reduces false positives.

Table 5.

Evaluation results of the multi-scale detection strategy under different grid sizes and similarity thresholds.

Grid size (N)	Detection Method	Similarity Threshold (St)	Precision	Recall	F1-score
300 m	Single-Scale Detection	—	0.725	0.876	0.793
	Multi-Scale Detection	0.3	0.815	0.846	0.830
		0.5	0.826	0.848	0.837
		0.7	0.829	0.841	0.835
500 m	Single-Scale Detection	—	0.835	0.928	0.879
	Multi-Scale Detection	0.3	0.905	0.904	0.904
		0.5	0.906	0.911	0.909
		0.7	0.908	0.896	0.902
700 m	Single-Scale Detection	—	0.559	0.927	0.690
	Multi-Scale Detection	0.3	0.919	0.854	0.885
		0.5	0.909	0.852	0.879
		0.7	0.926	0.835	0.878

Considering the combinations of different grid sizes (300 m, 500 m, and 700 m) and similarity thresholds (0.3, 0.5, and 0.7), the multi-scale detection strategy achieves optimal performance with a grid size of 500 m and a similarity threshold of 0.5. Under these settings, the precision, recall, and F1-score reach 0.906, 0.911, and 0.909, respectively. This combination of parameters produces the best overall results.

Figure 8 illustrates the dock detection performance of the multi-scale detection strategy under different grid sizes and similarity thresholds, with comparisons against the single-scale detection method. As shown in Fig. 8, the multi-scale strategy more effectively mitigates the misclassification of ships as docks compared to the single-scale approach (Fig. 8a-2). Furthermore, an overly small grid size (e.g., 300 m) may result in incomplete dock detection (Fig. 8b-3), while an excessively large grid size (e.g., 700 m) could lead to missed detections due to the small size of docks and complex surrounding features (Fig. 8d-3). Therefore, the analysis indicates that optimal dock detection performance is achieved with a grid size of 500 meters and a similarity threshold of 0.5.

Fig. 8.

Comparison of dock detection performance under different grid sizes and similarity thresholds.

Vectorization accuracy assessment

A random sample of 10% (356) of the total 3,562 docks was selected for evaluation. The sampled docks were independently vectorized by two professional surveyors and then compared with the corresponding dock boundary data provided in this study. The resulting Cohen’s kappa coefficients were 0.937 and 0.965, respectively, both well above the commonly accepted threshold of 0.81, indicating a high level of accuracy in the dock boundary vector data.

Usage Notes

Applications

(1) The dataset enables the quantification of the spatial density and distributional imbalance of dock facilities along different sections of the Yangtze River. For instance, inland waterway planning, port system research, and related personnel can calculate indicators such as the number and density of docks at the scale of river sections or administrative units. These can then be integrated with economic statistics (e.g., GDP, industrial output) at the county or prefecture level to conduct spatial correlation or regression analyses. This helps identify the spatial relationship between dock infrastructure allocation and regional economic development levels⁷, providing a quantitative basis for optimizing inland waterway network layouts and evaluating infrastructure investment^27,28.

(2) Dock location data can also serve as a spatial proxy for the intensity of human activities, supporting environmental impact assessments at the basin scale. For instance, water environment studies and watershed ecological protection and relevant management departments can overlay this data with multi-source geospatial data such as vessel Automatic Identification System (AIS) trajectory data, water quality monitoring data, and shoreline land-use data. By statistically analyzing vessel activity intensity or changes in water quality indicators within specified buffer distances around docks, researchers can analyze the spatial associations between dock clusters, shoreline hardening levels, and water pollution risks, thereby providing spatially explicit references for watershed ecological zoning and environmental management^29–31.

(3) Furthermore, the docks provided in this dataset can be incorporated as nodes into inland waterway network models. By integrating waterway data and vessel AIS trajectory data³², a topological network consisting of waterways, dock nodes, and vessel movement paths can be constructed. This enables analysis of the connectivity and criticality of different docks within the network, as well as simulation-based evaluation of the impacts of dock restrictions or closures on route length, transport efficiency, and overall network connectivity. This spatial data is essential for optimizing inland logistics planning and emergency response strategies³³.

Limitations

(1) This study proposes a dock identification method that integrates open-source geographic data from OSM and Google high-resolution satellite images, as both sources are freely accessible. Although OSM data may contain quality issues^34,35, previous studies have demonstrated that most OSM data exhibits high positional accuracy^36,37. Therefore, it is feasible to obtain high-quality training labels based on OSM data. While Google high-resolution satellite images may suffer from untimely updates, it remains the only freely available sub-meter resolution remote sensing data, providing essential support for accurately identifying small inland docks. Nevertheless, previous studies have found that insufficient data completeness remains a major limitation when using OSM data^38,39, and Google satellite imagery may also suffer from data quality issues⁴⁰. In addition, when vessels are berthed at docks, the similarity in their visual characteristics may lead to the two being identified as a single object (Fig. 8a-1) or to missed dock detections (Fig. 8d-3), thereby affecting precision and recall. Therefore, future studies could consider overlaying our dock dataset with high-resolution remote sensing imagery to develop more reliable dock annotation datasets, as well as incorporating long-term time-series high-resolution remote sensing data (such as the Gaofen series^41,42 or the WorldView series⁴³) to provide data support for the development of a reliable, multi-temporal dataset of inland docks along the Yangtze River.

(2) This study compares eight different object detection models from the YOLO series, which have been widely applied to detect various features in remote sensing images, including vehicles^44,45, ships⁴⁶, and buildings²⁶. We found that the baseline YOLOv5 model performed best in identifying inland docks, even outperforming the latest YOLOv13. This may be attributed to YOLOv5 more lightweight and stable architectural design, its robustness in training on small samples and irregular shaped objects, and the fact that its network structure and hyperparameter configurations have been extensively validated and optimized in previous studies^45,47,48. Conversely, more complex YOLO series models (e.g., YOLOv11) may have advantages when handling large-scale datasets²⁴. Nonetheless, future work could explore other object detection models (e.g., Faster R-CNN⁴⁹ or SSD series⁵⁰) or additional improved YOLO variants to further enhance the accuracy of inland dock identification.

(3) The Yangtze River was selected as the study area because it is not only one of the world’s most important inland shipping routes but also hosts a large number and diverse types of inland docks along its banks, making it a representative region. The developed Yangtze River Inland Dock Dataset may support applications such as infrastructure management along the river⁵¹, sustainable economic development of the basin⁵² and optimization of shipping and logistics scheduling³. Although the dataset does not cover other regions globally, the proposed method of integrating OSM and Google high-resolution satellite images for dock identification could be applied to other study areas. This approach would provide methodological support for acquiring inland dock data in relevant regions and promote research and applications in global inland dock monitoring and management.

Author contributions

Q.Z. developed the idea, researched and reviewed. F.R. processed the data, wrote and revised the manuscript. H.Z. conducted experiments. A.Z. and H.L. reviewed and revised.

Funding

National Natural Science Foundation of China (Grant No. 42471492); Beijing Jinshui Information Technology Development Co., Ltd. (Grant 24-HB21KFF20240434).

Data availability

The data that support the findings of this study are openly available in Figshare at 10.6084/m9.figshare.30555779.

Code availability

The QGIS python scripts for data processing and the code for model training are available online through the GitHub repository: https://github.com/haoming222/Open_dock.

Competing interests

The authors declare no competing interests.