EcoRRAP Benthic Data 2021-2023

This dataset contains benthic data from generated from imagery collected by divers conducting photogrammetric monitoring at EcoRRAP sites (see SOP #14: Reef monitoring sampling methods | AIMS). These images used are the underlying images of the EcoRRAP 3D models and are used here to assess benthic community composition within EcoRRAP plots.

A subset of images were extracted for this dataset by using a python script that filtered for image position within 3D models, to enable the best spatial stratification, as well as for aspect to maximise extraction of images facing the benthos. A 1000-pixel bounding box was applied to the edges of the image to remove areas of distortion (original image size was 8256x5504 pixels). Images were extracted for 352 plots in 2021, and 336 plots in 2022, and 327 plots in 2023 due to field imaging constraints.

Twelve images per plot with 50 points per image were analysed in ReefCloud. Images from a subset of plots were manually annotated in ReefCloud to train an automated image analysis model and the remaining plots were analysed using the automated model. Images were annotated using the EcoRRAP Community Composition Label Set.

The ReefCloud ML model was trained with a goal of achieving 80% accuracy for identification of benthos. Training proceeded as an iterative process with model diagnostics assessed after each round of training. Once performance targets were reached across course taxonomic classes, targeted annotation was undertaken to improve the model performance for specific classes.

本数据集包含由潜水员在EcoRRAP监测点开展摄影测量监测（photogrammetric monitoring）所采集的影像生成的底栖生物数据（benthic data）（详见标准作业程序（Standard Operating Procedure, SOP）#14：珊瑚礁监测采样方法 | 澳大利亚海洋科学研究所（Australian Institute of Marine Science, AIMS））。本次使用的影像为EcoRRAP三维模型的原始基底影像，用于评估EcoRRAP样地内的底栖生物群落组成（benthic community composition）。

本数据集通过Python脚本（Python）提取部分影像：脚本首先依据三维模型内的影像位置进行筛选，以实现最优空间分层；同时根据影像朝向，最大化提取面向底栖环境的影像。随后对影像边缘施加1000像素的边界框（bounding box），以去除畸变区域（原始影像尺寸为8256×5504像素）。受野外成像条件限制，2021年、2022年、2023年分别提取了352个、336个、327个样地的影像。

每个样地选取12张影像，每张影像标注50个点位，依托ReefCloud平台开展分析。首先对部分样地的影像进行人工标注，以训练自动化影像分析模型；剩余样地则采用训练完成的自动化模型完成分析。所有影像均采用EcoRRAP群落组成标注集（"EcoRRAP Community Composition Label Set"）进行标注。

ReefCloud机器学习模型（Machine Learning Model）的训练目标为实现80%的底栖生物识别准确率。模型训练采用迭代式流程，每轮训练完成后均对模型诊断结果进行评估。当大类分类单元（course taxonomic classes）的性能指标达到预设目标后，再开展针对性标注以优化模型在特定分类单元上的识别性能。