内容来源：https://blog.google/company-news/outreach-and-initiatives/sustainability/speciesnet-open-source-ai-wildlife/

内容总结：

开源AI模型SpeciesNet助力全球野生动物保护，多国项目成效显著

在人类视线之外，野生动物如何活动？从黎明时分哥伦比亚丛林中的美洲狮，到澳大利亚漫步的食火鸡，触发式相机为我们捕捉了无数珍贵瞬间。然而，海量的图像数据却成为野生动物管理者、生物学家和保护工作者的沉重负担——人工识别分类耗时耗力。

为此，谷歌于一年前免费开源了人工智能模型SpeciesNet。该模型自2019年起已通过“野生动物洞察”平台应用，能够自动识别近2500种哺乳动物、鸟类和爬行动物，显著提升了相机陷阱数据的处理效率。目前，全球多个保护项目正借助该技术加速科研与监测进程。

在非洲，坦桑尼亚塞伦盖蒂国家公园的“Snapshot Serengeti”项目自2010年起与坦桑尼亚野生动物研究所合作，布设了大量相机。项目初期依赖在线志愿者分析图像，但数据量远超人力负荷。北卡罗来纳州维克森林大学的项目负责人利用SpeciesNet，在数日内完成了积压的1100万张照片分析，处理了相当于数十年的数据，从而助力对该生物多样性热点区域动物行为与种群动态的长期研究。

在南美洲，哥伦比亚洪堡研究所通过“野生动物洞察”平台使用SpeciesNet，监测正经历快速变化的亚马逊雨林生物。该机构近期拓展了国家级的“Red Otus”监测网络，在全国公私土地布设相机。通过对数万张图像的分析，项目发现了鸟类迁徙时间的变化以及野生动物日活动模式的转变：部分哺乳动物可能为躲避威胁而更趋夜行性，开发区域的鸟类清晨出现时间也似乎更晚。

在北美洲，美国爱达荷州渔猎局等多家机构利用SpeciesNet识别相机照片中的动物。该州在森林茂密的北部地区广泛部署相机陷阱，作为空中调查的补充。模型预先按物种分类图像，专家再进行最终复核，使每年数百万张图像的分析效率大幅提升。

在澳大利亚，野生动物观测站基于开源模型，针对本地特有物种（如鹤鸵、红腿小袋鼠等）进行了专项训练，使其能够有效监测这些标志性、受威胁或濒危物种的种群状况，助力区域性保护。

SpeciesNet能够在多角度、不同光照及动物局部可见的条件下进行识别。目前，该模型已在全球多地保护实践中发挥作用，帮助研究者更高效地理解并保护与我们共享地球的野生动物。

中文翻译：

我们的开源AI模型SpeciesNet如何助力野生动物保护事业

从黎明时分潜行于哥伦比亚丛林的美洲狮，到漫步于澳大利亚的食火鸡，触发式相机让我们得以窥见人类不在场时动物的真实状态。但对于野生动物管理者、生物学家和保护工作者而言，将海量抓拍图像转化为可行动数据曾是极其耗时的工作。

这正是SpeciesNet的价值所在。这款AI模型经过训练，能自动识别近2500种哺乳动物、鸟类和爬行动物。自2019年起，该模型已通过Wildlife Insights平台投入使用。一年前我们将其作为免费开源工具正式发布，如今全球研究团队正运用它以前所未有的速度解析相机陷阱数据。

非洲合作伙伴：塞伦盖蒂快照计划

2024年拍摄的这组图像展示了夜间活动的象群、雄狮、斑马侧影以及一只似乎正注视镜头的疣猪。图片来源：塞伦盖蒂快照计划；T.M.安德森

在非洲，塞伦盖蒂快照计划自2010年起与坦桑尼亚野生动物研究所合作，于坦桑尼亚塞伦盖蒂国家公园部署相机陷阱。项目初期曾招募线上志愿者分析图像，但海量数据远超人工处理能力。北卡罗来纳州维克森林大学的项目负责人托德·迈克尔·安德森运用SpeciesNet分析了积压的1100万张照片，仅用数日便完成了原本需数十年的数据处理工作。该项目正通过图像分析，长期追踪非洲生物多样性最丰富区域之一的动物行为与种群动态。

南美合作伙伴：哥伦比亚洪堡研究所

2025年3月至5月拍摄的图像中，展示了在美国南部和墨西哥濒危、但在南美仍较常见的野生小猫科动物虎猫，以及一只美洲狮。图片来源：卢西塔尼亚项目/安第斯大学/奥图斯网络

在哥伦比亚，我们的长期合作方洪堡研究所将SpeciesNet应用于Wildlife Insights平台。该机构监测的许多物种栖息于正经历剧变的哥伦比亚亚马逊雨林——全球生物多样性热点区域。近期他们拓展建立了覆盖全国公私土地的奥图斯国家监测网络，通过分析数万张图像，揭示了哥伦比亚鸟类迁徙时序变化及野生动物昼夜活动规律。研究表明，部分哺乳动物可能为躲避威胁而趋于夜行性，发达地区的鸟类清晨活动时间也出现延迟。

北美合作伙伴：爱达荷州渔猎部

2025年7月至9月拍摄的图像记录了黑熊家族、郊狼、骡鹿和美洲赤鹿，这些正是爱达荷州渔猎部为监测种群健康状态而追踪的物种。图片来源：爱达荷州渔猎部

作为美加地区众多使用SpeciesNet的野生动物及交通管理机构之一，爱达荷州渔猎部在南部地区常进行航空调查的同时，于全州（尤其是森林茂密的北部）部署了数百个相机陷阱。虽然最终仍需专家审核，但SpeciesNet的事先物种分类使每年数百万张图像的处理效率大幅提升。

澳大利亚合作伙伴：澳大利亚野生动物观测站

2025年8月至11月澳大利亚春季期间，WildObs合作伙伴拍摄到这组图像：一对红腿小袋鼠、午间漫步的食火鸡，以及一只凝视镜头的食火鸡。图片来源：澳大利亚野生动物观测站

在澳大利亚，我们的合作方野生动物观测站基于开源SpeciesNet模型进行本地化训练，使其能识别对当地至关重要但未包含在初始模型中的特有物种。针对澳大利亚众多全球独有物种，定制版SpeciesNet帮助各地机构重点监测其区域内的标志性、受威胁或濒危物种，以维系野生种群存续。

SpeciesNet具备多角度识别能力，可适应不同光照条件，即使动物仅部分入镜也能准确辨识。有时动物们会好奇地直视镜头，留下生动的肖像瞬间。

上述项目仅是我们全球合作网络的部分范例。衷心感谢所有实地运用这项工具来认知与保护地球生灵的合作伙伴。若想深入了解SpeciesNet的研发历程、模型训练与性能表现，欢迎阅读我们在谷歌研究博客发布的专题文章。

英文来源：

How our open-source AI model SpeciesNet is helping to promote wildlife conservation
From a puma prowling through the Colombian forest at dawn to a cassowary wandering across Australia, motion-triggered cameras give us an unprecedented view of what animals do when humans aren't around. But for wildlife managers, biologists and conservationists, turning millions of these candid snapshots into actionable data is incredibly time-consuming.
That's where SpeciesNet comes in. SpeciesNet is an AI model trained to automatically identify nearly 2,500 categories of mammals, birds and reptiles. The model has been used since 2019 via Wildlife Insights. We launched it as a free, open-source tool a year ago, and today, research groups are using it to make sense of their camera trap data faster than ever.
Africa partner: Snapshot Serengeti
These images from 2024 show a group of elephants at night, a male lion, a zebra in profile, and a warthog that appears to be looking at the camera. Image credit: Snapshot Serengeti; T.M. Anderson
In Africa, the Snapshot Serengeti project has operated camera traps in Tanzania’s Serengeti National Park, in collaboration with the Tanzanian Wildlife Research Institute, since 2010. At first the project recruited online volunteers, but it had too many images for the volunteers to analyze. Project leader Todd Michael Anderson at North Carolina’s Wake Forest University used SpeciesNet to analyze a backlog of 11 million photos, processing decades’ worth of data in just days. The project is analyzing these images to get a long-term view of fauna behavior and abundance in one of Africa’s most biodiverse regions.
South American partner: Colombia’s Humboldt Institute
These images were captured between March and May 2025. They show an ocelot, a small wild cat that’s endangered in the southern U.S. and Mexico but is still common in South America, and a puma (also known as a cougar or mountain lion). Image credit: Project Lucitania/Universidad de los Andes/Red Otus
In Colombia, our longtime collaborators at the Humboldt Institute use SpeciesNet as part of the Wildlife Insights platform. Many of the species the institute monitors live in Colombia’s Amazon Rainforest, an extremely biodiverse region that is undergoing rapid changes. The group recently expanded in launching Red Otus, a national-scale network that captures camera trap images on public and private land across the country. The Red Otus project has analyzed tens of thousands of images it has collected to discover changes in the timing of bird migrations and the daily patterns of wildlife across Colombia. Analysis suggests that some mammals are becoming more nocturnal, perhaps to avoid threats, and birds appear later in the morning in developed areas, perhaps to avoid predators.
North American partner: Idaho Department of Fish and Game
These images, captured from July through September 2025, show some of the species IDFG monitors to ensure the population is healthy and stable. The photos show a family of black bears, a coyote, a mule deer and an elk. Photo credit: Idaho Department of Fish and Game
The Idaho Department of Fish and Game (IDFG) is among many state wildlife and transportation agencies in the U.S. and Canada that are using the SpeciesNet AI model to identify animals in their camera trap photos. While aerial surveys are frequently flown in southern Idaho, the agency deploys hundreds of camera traps across the state, particularly in the more forested, northern areas. Human experts conduct a final review, but having SpeciesNet sort the images by species beforehand greatly speeds up reviewing the millions of images collected each year.
Australian partner: Wildlife Observatory of Australia
The images above were captured by WildObs partners in Australian springtime, from August to November 2025. They show a pair of red-legged pademelons, cassowaries out for a midday stroll, and a cassowary peering into the camera. Photo credit: Wildlife Observatory of Australia
In Australia, our collaborators at the Wildlife Observatory of Australia (WildObs) took the open-source SpeciesNet model and trained it to identify species that weren’t part of the initial model, but that are important locally. Australia is home to many species not found anywhere else in the world, and those species are a priority for monitoring and conservation. A version of SpeciesNet trained on local wildlife lets groups keep an eye on iconic, threatened or endangered species specific to their region in order to sustain wild populations.
SpeciesNet can identify species from multiple angles, in different types of light, and when only a portion of the animal is visible. But sometimes animals get curious and look straight at the camera, producing a true portrait.
The projects above represent just a sample of the groups we’ve worked with to help run SpeciesNet to interpret camera trap photos. We’re grateful to all of our partners who are applying this tool on the ground to better understand and protect the wildlife that also call our planet home. To learn more about the history of SpeciesNet, its model training and performance, read our post on the Google Research Blog.