ReefCloud

ReefCloud is a digital tool that supports underwater coral reef benthic data collection and analysis, allowing the world's coral reef monitoring community to work together in real time to improve the storage, sharing and interpretation of in-water ecological observations that help us track global reef health. Co-funded by the Australian Department of Foreign Affairs and Trade (DFAT) and the Australian Institute of Marine Science (AIMS), ReefCloud provides an end-to-end solution for integrating, synthesising, reporting and communicating coral reef monitoring data using technology developed at AIMS.The online platform, ReefCloud.ai, leverages digital technologies (including cloud-based computing, Artificial Intelligence and Bayesian Statistics) to make coral reef data more integratable and sharable, with the ultimate goal of strengthening the linkage between underwater ecological scientific insights and data-driven actions to manage coral reefs. Key innovations - what does ReefCloud do?Collect and manage benthic monitoring imagesAIMS staff and researchers outside of AIMS are able to upload underwater digital photos from monitoring surveys along with associated data - location, date, habitat, depth - to the ReefCloud Data Portal to rapidly assess the ecological condition of coral reefs. An automated pipeline secures and organises this data, making ReefCloud a global cloud-based repository for benthic imagery, and there are options to share data at multiple levels to foster collaborations between scientists and organisations for improved insights into regional and global reef health status.Analyse and automatically annotate images using AIReefCloud employs machine-learning from trained experts to automate image analysis, extracting relevant information on benthic composition quickly and efficiently. This helps researchers to draw out advanced taxonomic detail from their submitted images, increase data analysis efficency and validate automated methods.How does this work?The AIMS Long Term Monitoring Program (LTMP) performs annual, image-based surveys of 80 reefs on Australia’s Great Barrier Reef. This dataset consists of millions of quality-controlled point annotations made by expert coral reef ecologists. AIMS use this large, high-quality dataset to train a machine learning model. As new images are uploaded to ReefCloud by external users, the trained model converts them into feature vectors which are classified into coral reef categories using smaller, faster classification machine learning models.Users need only label a part of their dataset to allow the model to pick out the specifics of the new images and provide accurate labels. Users can then output a spreadsheet that contains the proportion of different reef species detected in each submitted photo - vital data for understanding reef condition.ReefCloud crops newly imported images into "patches" centred on sampling points (number and distribution per image defined by user). The algorithm views all 65,000 pixels within each 256 by 256 pixel patch, extracting information on colours and contrasts and neighbours of each pixel, and compressing all that information into a 128-digit “feature vector”. ReefCloud assigns a label to a point, that feature vector is associated with a certain class, defined by users during training. The machine that runs over the AI performs a classification on remaining non-human annotated points by inferencing using the numbers. This is faster than using traditional machine learning where imagery is directly compared – but has the same accuracy. The difference in time for “inferencing” between these two methodologies is significant – minutes for a number vs hours to days for pixels. Even though it takes a little additional time to create the feature vector for each point annotated, the benefit is ReefCloud can re-inference entire datasets easily whenever required.Synthesise insights and view automatically updated reportsInteractive dashboards, report cards and data access tools enable rapid interpretation, reporting and communication on coral reef conditions across geographies. Data are modelled for presentation on the ReefCloud Public Reporting Dashboard.What data are available on the Public Dashboard and how is it generated?All publicly available data from ReefCloud from a specific region is compiled into a single data set that includes information about the sampling source, depth, date, location and number of photo points identified as belonging to each of 12 broad benthic groups (Hard Coral, Soft Coral, Macroalgae, Turf Algae, Crustose Coralline Algae, Cyanobacteria, Seagrass, Hard Substrate, Rubble, Soft Sediment, Other Invertebrates and Other). These data inform broad spatial-scale statistical models, which predict annual hard coral cover and macroalgae cover (median and upper/lower credibility intervals) across all known coral reefs (both monitored and unmonitored) in the region by incorporating a spatial grid and information about the environment (currently cyclone activity and degree heating weeks).ReefCloud employs two different models to provide summary statistics and trends for an indivudal monitoring site vs a region with a broader spatial scale. These two models serve two different purposes. The site-specific models provide an estimate of the cover at a single site and are informated only by data submitted from that local site. The broader spatial scale models provide estimates of cover at all sites regardless of whether photos exist inside ReefCloud or not. Importantly, these models are informed by all available monitoring and environmental data for the entire region and thus, the predictions for any given site and the result of both the observed data extracted from photos for that site as well as the more general patterns in its neighbourhood. Consequently, it is possible that the two different models may yield slightly different esrunares if thet are both compared for a single site or an area that only had a single monitored site.

ReefCloud是一款支持水下珊瑚礁底栖（benthic）数据采集与分析的数字化工具，可让全球珊瑚礁监测社群实现实时协同，优化水下生态观测数据的存储、共享与解读工作，助力我们追踪全球珊瑚礁健康状况。该平台由澳大利亚外交贸易部（Department of Foreign Affairs and Trade, DFAT）与澳大利亚海洋科学研究所（Australian Institute of Marine Science, AIMS）联合资助，依托AIMS开发的技术，为珊瑚礁监测数据的整合、综合分析、报告编制与传播提供端到端解决方案。

在线平台ReefCloud.ai依托云计算、人工智能（Artificial Intelligence）、贝叶斯统计（Bayesian Statistics）等数字技术，提升珊瑚礁数据的可整合性与共享性，最终目标是强化水下生态科学认知与数据驱动的珊瑚礁管理行动之间的关联。

### 核心创新——ReefCloud的功能是什么？
1. 采集与管理底栖监测图像
AIMS内部及外部的科研人员可将监测调查获取的水下数码照片，连同位置、日期、生境、水深等关联数据上传至ReefCloud数据门户，以快速评估珊瑚礁的生态状况。自动化流程可对该类数据进行安全存储与组织管理，使ReefCloud成为全球首个基于云端的底栖图像存储库，同时支持多层级数据共享，以促进科研机构与组织间的协作，深化对区域及全球珊瑚礁健康状况的认知。

2. 利用人工智能开展图像分析与自动标注
ReefCloud依托经过专家训练的机器学习（machine learning）模型实现图像分析自动化，快速高效地提取底栖群落组成的相关信息。这可帮助科研人员从提交的图像中提取精细的分类学细节，提升数据分析效率，并验证自动化分析方法的可靠性。

#### 工作原理
澳大利亚海洋科学研究所长期监测项目（Long Term Monitoring Program, LTMP）每年会对澳大利亚大堡礁的80处珊瑚礁开展基于图像的调查，该数据集包含数百万经质量管控的、由资深珊瑚礁生态学家完成的点标注（point annotations）。AIMS利用这一大型高质量数据集训练机器学习模型。当外部用户向ReefCloud上传新图像时，预训练模型会将图像转换为特征向量（feature vector），再通过小型轻量化分类机器学习模型将其归类至珊瑚礁类别。

用户仅需对部分数据集进行标注，即可让模型识别新图像的特定细节并生成准确的标注。随后用户可导出包含每张提交照片中检测到的不同珊瑚礁物种占比的电子表格——这是理解珊瑚礁健康状况的关键数据。

ReefCloud会将新导入的图像裁剪为以采样点为中心的“图像块（patches）”，图像块的数量与分布可由用户自定义。算法会读取每个256×256像素图像块内的全部65000个像素，提取每个像素的色彩、对比度及邻域像素信息，并将所有信息压缩为128维的“特征向量”。ReefCloud会为每个采样点分配标签，该特征向量将与用户在训练阶段定义的类别相关联。AI模型会通过基于这些数值的推理（inferencing），对其余非人工标注的采样点完成分类。

该方法比传统的直接图像对比法更快，且精度相当。两种推理方法的耗时差异显著：前者仅需数分钟，后者则需数小时至数天。尽管为每个标注采样点生成特征向量需要额外花费少量时间，但ReefCloud可在后续需要时轻松地对整个数据集重新进行推理，这一优势尤为突出。

3. 整合分析结果并查看自动更新的报告
交互式仪表板、报告卡片与数据访问工具可实现跨区域珊瑚礁状况的快速解读、报告编制与传播。相关数据会经过建模处理，呈现在ReefCloud公共报告仪表板中。

#### 公共仪表板上的公开数据与生成逻辑
来自特定区域的所有ReefCloud公开数据会被整合为单一数据集，其中包含采样来源、水深、日期、位置，以及被归类为12大类底栖生物类别的照片点数量：硬质珊瑚、软质珊瑚、大型藻类、藻席藻类（Turf Algae）、壳状珊瑚藻、蓝细菌、海草、硬质基底、碎石、软沉积物、其他无脊椎动物及其他类别。

这些数据将用于构建大范围空间统计模型，通过整合空间网格与环境信息（当前为气旋活动与热度日数），预测该区域内所有已知珊瑚礁（包括已监测与未监测区域）的年度硬质珊瑚盖度与大型藻类盖度（中位数及上下可信区间）。

ReefCloud采用两种不同的模型，分别为单个监测站点与更大空间尺度的区域提供汇总统计数据与趋势分析。这两种模型各司其职：站点专属模型仅利用该本地站点提交的数据估算单一场点的生物盖度；大范围空间尺度模型则可估算所有场点的盖度，无论该场点是否有图像上传至ReefCloud。重要的是，这类模型会利用整个区域的所有可用监测与环境数据进行训练，因此，任意给定场点的预测结果，会同时结合该场点从照片中提取的观测数据，以及其周边区域的整体分布模式。

因此，若对仅拥有单个监测站点的单一场点或区域同时使用两种模型进行分析，二者的结果可能会存在细微差异。