A dataset of floating waste and river vegetation images collected from turbid rivers in Banjarmasin, Indonesia

Abstract

This dataset contains images of floating waste and river vegetation collected from several turbid river locations in Banjarmasin, South Kalimantan, Indonesia. The dataset combines still photographs and image frames extracted from video recordings, with video frames extracted at approximately one frame per second. The collected images were manually screened to remove images that did not contain relevant objects before annotation. The final dataset consists of 1,577 unique RGB images containing two object categories, namely floating waste and river vegetation. Each image was manually annotated using bounding boxes and stored in YOLO annotation format with object class identifiers and normalized bounding box coordinates. The dataset contains 5,792 annotated object instances, comprising 3,718 floating waste instances and 2,074 river vegetation instances. The dataset is partitioned into training, validation, and testing subsets containing 1,104, 315, and 158 images, respectively. The partitioning was performed using a group-aware strategy to ensure that images originating from the same video source were kept within the same subset. The dataset is intended to support computer vision, object detection, and environmental monitoring research under challenging turbid river conditions.

Specifications Table

Subject	Computer Sciences
Specific subject area	Computer vision dataset for floating waste and river vegetation detection in turbid rivers.
Type of data	JPEG images YOLO annotation files (.txt) YAML configuration file Dataset metadata
Data collection	Images were collected from several turbid river locations in Banjarmasin, South Kalimantan, Indonesia, under varying environmental conditions, including differences in lighting, water turbidity, viewing angles, object distribution, and background characteristics. The collected data consisted of still photographs and video recordings. Video recordings were converted into image frames at approximately one frame per second. The collected images were manually screened to remove images without relevant objects or unsuitable images before annotation. The final dataset contains 1,577 unique RGB images, comprising 1,281 selected video frames from 24 video sources and 296 still photographs. All images were manually annotated using bounding boxes for two object categories, namely floating waste and river vegetation, and the annotations were exported in YOLO format. The dataset was partitioned into training, validation, and testing subsets using a group-aware strategy, with images originating from the same video source retained within the same subset.
Data source location	Institution: Department of Information Technology, Faculty of Engineering, Universitas Lambung Mangkurat, Banjarbaru 70714, South Kalimantan, Indonesia. City/Town/Region: Banjarmasin, South Kalimantan. Country: Indonesia Latitude and longitude (approximate): 3.3186° S, 114.5944° E Description: The dataset was collected from several turbid river locations across Banjarmasin, South Kalimantan, Indonesia. Image acquisition was performed under various environmental conditions to capture variations in water turbidity, illumination, object distribution, and viewing angles. The reported coordinates represent the approximate center of the study area because the images were acquired from multiple sampling locations rather than a single fixed point
Data accessibility	Repository name: Mendeley Data Data identification number: doi: 10.17632/j26w4m645z.2 Direct URL to data: https://data.mendeley.com/datasets/j26w4m645z/2 Instructions for accessing these data: The revised dataset has been newly uploaded to Mendeley Data and is currently awaiting publication. Once published, the dataset will be publicly accessible and downloadable through the provided link.
Related research article	Not applicable.

1. Value of the Data

• •This dataset provides a documented collection of images of floating waste and river vegetation acquired from turbid river environments in Banjarmasin, South Kalimantan, Indonesia. The dataset reflects real-world environmental conditions, including variations in water turbidity, illumination, object density, and viewing angles.
• •The dataset provides manually annotated images suitable for object detection model development and evaluation, supporting researchers, students, and practitioners working in computer vision, environmental monitoring, remote sensing, and intelligent waste management [[1], [2], [3], [4]].
• •The dataset enables benchmarking and comparison of different object detection algorithms, data augmentation strategies, attention mechanisms, and deep learning architectures beyond YOLOv8, facilitating reproducible research using a common dataset [1,2,5].
• •The dataset may support future applications in automated river monitoring, floating waste detection, environmental conservation, and smart city initiatives, particularly for tropical urban river environments with challenging visual characteristics [[4], [5], [6]].

2. Background

River pollution caused by floating waste and aquatic vegetation has become an important environmental issue in many urban areas, particularly in regions with dense river networks such as Banjarmasin, South Kalimantan, Indonesia. Monitoring these objects manually is time-consuming, labor-intensive, and often affected by challenging environmental conditions, including high water turbidity, reflections, varying illumination, and complex backgrounds. Although deep learning-based object detection methods have demonstrated promising performance in environmental monitoring applications, the availability of publicly documented datasets representing turbid tropical river environments remains limited [1,2].

To address this limitation, a dataset containing images of floating waste and river vegetation was collected from several river locations in Banjarmasin. The images were manually annotated using bounding boxes in YOLO format to support object detection research. The dataset was designed to provide a representative benchmark for developing, evaluating, and comparing computer vision models under challenging real-world river conditions, while also supporting future studies in environmental monitoring, river pollution assessment, and intelligent waste management [[1], [2], [3], [4],7].

3. Data Description

The proposed dataset consists of 1,577 unique RGB images containing two object categories, namely floating waste and river vegetation, collected from several turbid river locations in Banjarmasin, South Kalimantan, Indonesia. The dataset comprises 1,281 selected image frames extracted from 24 video sources and 296 still photographs. The dataset was developed to support computer vision research for environmental monitoring in tropical river environments characterized by varying illumination conditions, water turbidity, object density, and background complexity [[1], [2], [3], [4], [5]].

Each image is accompanied by a corresponding annotation file in YOLO format containing normalized bounding box coordinates and object class labels. The dataset also includes a dataset.yaml configuration file for direct implementation using the YOLO object detection framework. The dataset is partitioned into training, validation, and testing subsets following a 70:20:10 ratio, consisting of 1,104 training images, 315 validation images, and 158 testing images. The partitioning was performed using a group-aware strategy to ensure that images originating from the same video source were retained within the same subset. Table 1 summarizes the overall characteristics of the proposed dataset [3,4,6].

Table 1.

Characteristics of the proposed dataset.

Characteristic	Description
Total images	1,577
Video sources	24
Video frames	1,281
Still photographs	296
Number of classes	2
Object classes	Floating waste, River vegetation
Annotation format	YOLO Bounding Box (.txt)
Image format	JPEG
Training images	1,104
Validation images	315
Testing images	158
Dataset split	70% training, 20% validation, 10% testing
Floating waste instances	2,054 training; 1,015 validation; 649 testing; 3,718 total
River vegetation instances	1,744 training; 204 validation; 126 testing; 2,074 total
Total bounding-box instances	3,798 training; 1,219 validation; 775 testing; 5,792 total

The dataset is organized into separate directories corresponding to the training, validation, and testing subsets. Each subset contains an images directory storing RGB images and a corresponding labels directory containing YOLO annotation files. A data.yaml configuration file is also included to define the dataset paths and class names required for implementation using the YOLO object detection framework. This standardized directory structure facilitates straightforward integration into YOLO-based object detection pipelines and supports reproducible experiments and future environmental monitoring research [6,7].

4. Experimental Design, Materials and Methods

The dataset consists of images collected directly from several turbid river locations in Banjarmasin, South Kalimantan, Indonesia. Image acquisition was conducted during daytime, approximately between 09:00 and 15:00 WITA, using smartphone cameras. During image acquisition, the camera was positioned at approximately 160 cm above the ground, approximately 3 m from the river, and at an angle of approximately 45° toward the water surface. The collected data consisted of both still photographs and video recordings. The images were acquired under diverse environmental conditions to represent real-world river environments, including variations in water turbidity, illumination, object density, camera viewpoints, and background complexity. The dataset focuses on two object categories, namely floating waste and river vegetation, which are frequently observed in polluted urban rivers and present considerable challenges for automatic object detection. Representative examples of the collected images are presented in Fig. 1 [2,3,5,6].

Fig. 1.

Representative examples of floating waste and river vegetation images collected from turbid river environments in Banjarmasin, South Kalimantan, Indonesia.

Following image acquisition, the video recordings were converted into image frames at approximately one frame per second. The resulting images from both photographs and extracted video frames were manually screened to remove images that did not contain relevant target objects or were unsuitable for annotation. The screening process resulted in 1,577 selected RGB images, comprising 1,281 image frames from 24 video sources and 296 still photographs. No image augmentation or additional cropping was applied during dataset preparation.

Following the screening process, the selected images were manually annotated using the bounding box technique through Roboflow to identify each object instance. The annotation process was performed manually to ensure consistency in object boundaries and class labels throughout the dataset. Annotation files were exported in YOLO format, with each image associated with a corresponding text file containing the object class identifier and normalized bounding box coordinates, as illustrated in Fig. 2.

Fig. 2.

Examples of manual object annotation using bounding boxes in YOLO annotation format.

Following the annotation process, the completed dataset was partitioned into training, validation, and testing subsets using a 70:20:10 ratio, consisting of 1,104 training images, 315 validation images, and 158 testing images, respectively. To reduce the risk of data leakage, a group-aware partitioning strategy was applied to the video-derived images, ensuring that all frames originating from the same video source were retained within the same subset and were not distributed across different subsets. This partitioning strategy separates the data used for model training, validation, and final evaluation, thereby supporting more reliable and reproducible model assessment. The proposed dataset was specifically developed to support computer vision research for environmental monitoring in challenging tropical river environments. Owing to its diverse environmental conditions, manually verified annotations, standardized directory organization, and compatibility with the YOLO framework, the dataset can serve as a benchmark for evaluating object detection models, attention mechanisms, data augmentation techniques, and other deep learning approaches designed for floating waste detection and river monitoring applications. Furthermore, the dataset is expected to facilitate reproducible research while supporting future developments in intelligent river monitoring systems and environmental conservation technologies [[1], [2], [3], [4], [5], [6]].

Limitations

The dataset has several limitations. First, the images were collected from several turbid river locations in Banjarmasin, South Kalimantan, so the dataset primarily represents the environmental and visual characteristics of this region and may not fully represent other river environments. Second, the dataset contains two main object categories, namely floating waste and river vegetation, which limits its applicability to other types of aquatic objects. Third, the images were manually annotated using bounding boxes, and the annotation process may still be affected by the visual complexity of turbid water, reflections, varying illumination, and complex backgrounds. Finally, the dataset remains dependent on the original visual characteristics of the collected images, which may limit its generalizability to river environments with substantially different visual conditions.

Credit author statement

Iphan F. Radam: Conceptualization, Supervision, Methodology, Project administration, Writing – review & editing. Yuslena Sari: Conceptualization, Supervision, Methodology, Project administration, Writing – review & editing. Muhammad Ryan Rizky Rahmadi: Data curation, Investigation, Software, Formal analysis, Visualization, Methodology, Writing – original draft, Writing – review & editing. Muhammad Alkaff: Supervision, Methodology, Writing – review & editing. Irham Maulani Abdul Gani: Data curation, Investigation. Mohammad Zaki Firmansah: Data curation, Validation. Natalie Grace Katiandagho: Validation, Visualization. Zainil Musthofa Radam: Data curation, Investigation, Validation. Fatih Alfayruz: Data curation, Validation, Visualization. Nashry Akbar: Investigation, Validation, Data curation.

Ethics Statement

The authors have read and followed the ethical requirements for publication in Data in Brief. This study does not involve human subjects, animal experiments, or data collected from social media platforms.