Dataset and trained weights for detecting signatures of seafloor hydrothermal activity at Higashi-Aogashima Knoll Caldera

This repository supplements the paper titled "AI-Based Acoustic Observation Applied to Exploration of Seafloor Hydrothermal Activities" submitted to IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing.
It contains a dataset and trained model weights for detecting hydrothermal emission signatures from Multi-beam Echo Sounder (MBES) images.

* The formatted description can be found in the 00_ReadMe.md, by using markdown viewer.

## (1) 01_Dataset.zip

This zip file contains MBES images (heatmaps) and the annotations for signatures of seafloor hydrothermal activity. MBES data were obtained by two research cruises as follows. File names of images indicate the time in JST (Japan Standard Time) when images were captured.

### Data Collection

MBES data were collected from the following surveys conducted during two research cruises.

#### [KM-22-11C]
- Survey on the known (Central Cone) site at 70-100 kHz.
- Grid survey within whole area of the caldera at 70-100 kHz.

#### [KM23-01]
- Survey on the known (Central Cone) site at 70-100, 50 or 40 kHz.

### Capturing

During KM22-11C, MBES images were captured using a software Bandicam (Bandicam Company). During KM23-01, images were captured by slightly modifying YOLO-v7 (https://github.com/WongKinYiu/yolov7)'s detection code. After capturing, unique images were selected by comparing difference with previously captured images.

### Annotation

Annotations of hydrothermal activity signatures were provided by one of the authors (K.M.), using a python program labelImg (https://github.com/HumanSignal/labelImg).

### Subset splitting

The images and annotation were randomly split into two subsets, 90% for subset "train" and 10% for subset "val".

### Augmentation

By random cropping, images and annotations from KM22-11C cruise were augmented 5 times, while those from KM23-01 cruise were augmented 10 times. Note that data augmentation was performed after splitting into subsets, which means that two subsets do not share original MBES images.

### Final number of images, label files and objects

Train: 2575 images, 2085 label files, 2110 objects annotated.
Val: 305 images, 270 label files, 270 objects annotated.

## (2) 02_Detection_model.zip
This zip file contains the following files.

### init.pt
The initial weights for training, which had been pre-trained on larger dataset for MBES detection (Mimura et al., 2022). The architecture YOLO-v7 (https://github.com/WongKinYiu/yolov7) was used in this work.

### best.pt
The best weights trained on the dataset ("01_Dataset.zip").

### hyp.yaml, opt.yaml
The detailed training parameter files formatted to yolo-v7.

## Reference
Mimura, K. and Nakamura, K. (2022), “Datasets for hydrothermal plume detection”, Mendeley Data, V1, http://doi.org/10.17632/dg2595f68b.1.