Side-scan sonar imaging data of underwater vehicles for mine detection

Unmanned vehicles have become increasingly popular in the underwater domain in the last decade, as they provide better operation reliability by minimizing human involvement in most tasks. Perception of the environment is crucial for safety and other tasks, such as guidance and trajectory control, mainly when operating underwater. Mine detection is one of the riskiest operations since it involves systems that can easily damage vehicles and endanger human lives if manned. Automating mine detection from side-scan sonar images enhances safety while reducing false negatives. The collected dataset contains 1170 real sonar images taken between 2010 and 2021 using a Teledyne Marine Gavia Autonomous Underwater Vehicle (AUV), which includes enough information to classify its content objects as NOn-Mine-like BOttom Objects (NOMBO) and MIne-Like COntacts (MILCO). The dataset is annotated and can be quickly deployed for object detection, classification, or image segmentation tasks. Collecting a dataset of this type requires a significant amount of time and cost, which increases its rarity and relevance to research and industrial development.

a b s t r a c t Unmanned vehicles have become increasingly popular in the underwater domain in the last decade, as they provide better operation reliability by minimizing human involvement in most tasks.Perception of the environment is crucial for safety and other tasks, such as guidance and trajectory control, mainly when operating underwater.Mine detection is one of the riskiest operations since it involves systems that can easily damage vehicles and endanger human lives if manned.Automating mine detection from side-scan sonar images enhances safety while reducing false negatives.The collected dataset contains 1170 real sonar images taken between 2010 and 2021 using a Teledyne Marine Gavia Autonomous Underwater Vehicle (AUV), which includes enough information to classify its content objects as NOn-Mine-like BOttom Objects (NOMBO) and MIne-Like COntacts (MILCO).The dataset is annotated and can be quickly deployed for object detection, classification, or image segmentation tasks.Collecting a dataset of this type requires a significant amount of time and cost, which increases its rarity and relevance to research and industrial development.

Data source location
The data was collected over the years along the Portuguese coast during missions performed by the Portuguese Navy's sappers' divers group number three ( Destacamento de Mergulhadores Sapadores -DMS 3).This group is responsible for all activities related to mine warfare at sea.

Value of the Data
• Using this data can unlock the full potential of sonar image object detection and classification, facilitating the development of new algorithms and applications.• These data were gathered in real-time using a state-of-the-art Autonomous Underwater Vehicle (AUV) and hold immense potential for validating developed models.Utilizing this data makes it possible to unlock new research possibilities, enabling the achievement of breakthroughs that would otherwise be impossible.• Utilizing this dataset allows the researcher to eliminate the costly and time-intensive process of collecting field data.This will save resources and enable the researchers to focus on the critical aspects of their projects.• Investing in a real dataset is a crucial step towards improving the safety and efficacy of mine detection systems.By incorporating real-world data, we can significantly reduce the probability of accidents and enhance the existing architectures to serve their purpose better.
• Civilian and military scientists can benefit from accurate data gathered by AUVs, which have already been annotated, facilitating algorithms and machine learning implementations.

Background
The data was collected over the years along the Portuguese coast during missions performed by the Portuguese Navy's sappers' divers group number three ( Destacamento de Mergulhadores Sapadores -DMS 3).Due to the research and development effort s developed by the Portuguese Navy (PoN), every opportunity is used to gather data that is usually used when participating in national and international projects in the maritime domain.The last project where the PoN Research Center ( Centro de Investigação Naval -CINAV) was responsible for the gathering and use of underwater sonar images for mine detection was the Open Cooperation for European Maritime Awareness (OCEAN 2020) project, funded under the European Union's Preparatory Action Plan on Defence Research (PADR) that occurred between 2018 and 2021 [1] .This project supported maritime surveillance and interdiction missions at sea using Unmanned Vehicles (UVs) integrated into fleet operations [2] .

Data Description
The dataset contains 1170 side-scan sonar images [3] collected using a 90 0-180 0 kHz Marine Sonic dual frequency side-scan sonar of a Teledyne Marine Gavia Autonomous Underwater Vehicle (AUV) [4] , as illustrated in Fig. 1 .All the images were carefully analyzed and annotated, including the image coordinates of the Bounding Box (BB) of the detected objects divided into NOn-Mine-like BOttom Objects (NOMBO) and MIne-Like COntacts (MILCO) classes.The number of dataset images and the respective number of MILCO and NOMBO occurrences per year are described in Table 1 .the absolute object BB coordinates (x, y, w, h ) can be obtained using the following relation: where x, y are the absolute coordinates of the BB in the image and w, h are its absolute width and height.This representation follows the standard You Only Look Once (YOLO) object detection deep neural network format [5] , and the annotation can be easily verified using LabelImg, a graphical image annotation tool commonly used during the image annotation process [6] .It is also possible to use LabelImg to convert the performed annotations to Pattern Analysis, Statistical Modelling, and Computational Learning (PASCAL) Visual Object Classes (VOC) [7] , the format used by ImageNet [8] or Create Machine Learning (CreateML) [9] formats.Fig. 4 illustrates annotated images from each year, showing the complexity of the annotation process.
Object detection is a common task in Computer Vision (CV), which involves estimating the coordinates of the bounding box of objects in an image.Regardless of the object detection algo-   rithm used, the annotation format must be adapted or converted to suit the chosen architecture.The dataset provides annotations in the YOLO format, but using LabelImg or another equivalent tool, it can be easily converted as required.Fig. 5 shows a standard schematic used for object detection, where most algorithms, after being trained (in supervised training), provide confidence values for the classes and the respective estimated BB coordinates.
To perform an initial object detection using the provided dataset, we have implemented the YOLO v4 algorithm [10] .The necessary files were customized to meet the specific requirements of our object detection task, changing the original "yolov4.cfg" file to configure the batch size (number of samples used in a single training iteration) to 64, the subdivisions (assumed division in the training batches) to 16, the maximum number of batches (after reaching this value the training stops) to 60 0 0, and configuring steps (batches were the learning rate is adjusted multiplying it by a 0.1 factor) at 4800 and 5400.We also set the network image input dimensions to 512 × 512 pixels (width x height) to optimize detection accuracy.As initial implementation, we have initialized the network weights with pre-trained values in the Common Objects in Context (COCO) dataset [11] using the file "'yolo4.conv.137" to perform transfer learning.In this preliminary test, using our dataset of 1170 images, over the first 50 0 0 training iterations, we achieved an average Intersection over Union (IoU) of 60%, a mean Average Precision (AP) of 75%, a Precision of 82%, and a Recall of 64% with a confidence threshold of 0.25.All the considered files, including the final training weights "yolov4-custom_50 0 0.weights " and a Jupyter Notebook example "Real_time_object_classifier.ipynb" with some implementation notes, are available in the dataset repository in a separate file named "Training.zip."It is worth noting that this implementation provides a basic algorithm for initial object detection training.With detailed tuning and optimization of the model parameters and training process, significantly better results can be achieved.

Experimental Design, Materials and Methods
As described, the dataset was acquired using a 90 0-180 0 kHz Marine Sonic dual frequency side-scan sonar of a Teledyne Marine Gavia Autonomous Underwater Vehicle (AUV) [4] , as illustrated in Fig. 6 .The vehicle is composed of several modules that are 200 mm in diameter.These modules can be adapted to fit a variety of AUV missions.The vehicle's weight can range from 48 to 100 kg, and its size can vary from 1.7 to 3 m.The maximum depth that the vehicle can reach is 200 m, and its maximum speed is 5 knots.In addition to the side-scan sonar described earlier, the vehicle is also equipped with a Global Positioning System (GPS), an inertial navigation system, and a high-resolution camera.

Fig. 4 .
Fig. 4. Examples of annotated images per year, with the image number in the top left corner.

Table 1
Summary of the dataset.

Table 2
Annotation format description with an example.