Google deploys five-camera airplane sweeps to reconstruct cities in 3D

Nat and Friends

Google combines multi-angle aerial photography with computer vision algorithms to convert 2D flyover images into textured 3D meshes, bypassing satellite limitations to monitor planetary changes.

AI-assisted summary of Google Earth’s Incredible 3D Imagery, Explained — original source. Check the original for full context.

Source published: 2017-04-18

Key ideas

Google Earth operates as a photorealistic digital mirror world covering all 196.9 million square miles of the planet, pairing satellite data with low-altitude aerial photography. (0:00)
To capture high-resolution 3D data, specialized aircraft fly systematic 5-hour lawnmower patterns equipped with five cameras recording downward and oblique angles simultaneously. (1:47)
The seamless global satellite mosaic, dubbed 'Pretty Earth,' merges 700,000 Landsat scenes into an 800-billion-pixel canvas that would take 60 years for a single computer to process. (5:12)
Google Earth Engine leverages the platform's multi-year historical archive as a planetary health monitor to analyze dynamic changes like Amazon deforestation and overfishing patterns. (6:18)

Detailed notes

## Background Published on YouTube by Nat and Friends on April 18, 2017, this behind-the-scenes video explores the technology and human effort behind Google Earth's photorealistic 3D imagery. The show notes explain how pixels, planes, and people collaborate to build and routinely update a digital mirror world of the planet.

## TL;DR The video breaks down how Google Earth constructs its detailed 3D environment, contrasting high-altitude 2D satellite views with low-flying camera planes that capture overlapping imagery in lawnmower-style flight paths. Through photogrammetry and computer vision, these multi-angle aerial photos are stripped of clouds and moving cars, converted into depth maps and 3D polygon meshes, and wrapped with texture mapping. Engineers also assemble the global cloud-free mosaic known as 'Pretty Earth' from Landsat data, spanning hundreds of billions of pixels across dozens of zoom tiers. Finally, the segment demonstrates how Google Earth Engine turns decades of historical imagery into an environmental health monitor, tracking dynamic changes like Amazonian deforestation and global fishing patterns over time.

## Atmosphere & style The piece carries an inquisitive, bright, and infectious curiosity driven by Nat's personable hosting and Google's engineers, Gopal and Kevin. It blends conversational tech explainer vibes with awe-inspiring reflections on space exploration, contrasting the historic, solitary snapshot of Apollo 8's 'Earthrise' with the sheer computational muscle of modern digital cartography. Visually and rhythmically, it feels like an energetic studio tour filled with playful metaphors—from street-view cars with wings to digital Russian nesting dolls—while remaining grounded in genuine reverence for Earth as a living, dynamic ecosystem.

## Bulletin - [00:00](https://www.youtube.com/watch?v=suo_aUTUpps&t=0s) Google Earth operates as a photorealistic digital mirror world covering all 196.9 million square miles of the planet, pairing satellite data with low-altitude aerial photography. - [01:47](https://www.youtube.com/watch?v=suo_aUTUpps&t=107s) To capture high-resolution 3D data, specialized aircraft fly systematic 5-hour lawnmower patterns equipped with five cameras recording downward and oblique angles simultaneously. - [05:12](https://www.youtube.com/watch?v=suo_aUTUpps&t=312s) The seamless global satellite mosaic, dubbed 'Pretty Earth,' merges 700,000 Landsat scenes into an 800-billion-pixel canvas that would take 60 years for a single computer to process. - [06:18](https://www.youtube.com/watch?v=suo_aUTUpps&t=378s) Google Earth Engine leverages the platform's multi-year historical archive as a planetary health monitor to analyze dynamic changes like Amazon deforestation and overfishing patterns.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=suo_aUTUpps&t=0s) Google Earth and Google Maps serve opposite purposes: Maps is designed for navigation and utility, whereas Earth is built for open-ended exploration and getting lost. - [01:47](https://www.youtube.com/watch?v=suo_aUTUpps&t=107s) Satellites provide the global 2D base layer, but detailed 3D imagery requires low-altitude aerial photography captured by airplanes under clear atmospheric conditions. - [01:47](https://www.youtube.com/watch?v=suo_aUTUpps&t=107s) Aerial capture missions involve flying back-and-forth 'lawnmower' paths over five-hour stretches to achieve the required depth and overlap. - [01:47](https://www.youtube.com/watch?v=suo_aUTUpps&t=107s) Planes carry a five-camera array recording straight down (nadir) and in four oblique directions to capture buildings and landscapes from multiple viewpoints. - [03:20](https://www.youtube.com/watch?v=suo_aUTUpps&t=200s) Before 3D reconstruction, raw imagery is preprocessed to remove atmospheric haze, balance color, and clean moving vehicles off streets. - [03:20](https://www.youtube.com/watch?v=suo_aUTUpps&t=200s) Photogrammetry algorithms combine overlapping aerial images with precise GPS coordinates to calculate depth maps and generate 3D polygon meshes. - [03:20](https://www.youtube.com/watch?v=suo_aUTUpps&t=200s) Texture mapping slices and drapes 2D aerial photographs over the generated polygon geometry to construct realistic facades and landscapes. - [03:20](https://www.youtube.com/watch?v=suo_aUTUpps&t=200s) Sharp architectural structures and solid rock faces reconstruct cleanly, whereas organic foliage remains difficult and often defaults to simplified shapes. - [05:12](https://www.youtube.com/watch?v=suo_aUTUpps&t=312s) The 'Pretty Earth' satellite layer algorithmically chooses the clearest, cloudless pixel for every location, producing a globally uniform springtime appearance. - [05:12](https://www.youtube.com/watch?v=suo_aUTUpps&t=312s) Navigating Google Earth involves over 20 nested zoom levels structured like modular puzzle pieces rather than a single expanding map. - [06:18](https://www.youtube.com/watch?v=suo_aUTUpps&t=378s) Global satellite maps refresh every two to three years, while rapidly developing metropolitan areas are re-flown and updated in under twelve months. - [06:18](https://www.youtube.com/watch?v=suo_aUTUpps&t=378s) Google Earth Engine turns historical satellite catalogs into an analytical platform to diagnose ecological trends, tracking deforestation and commercial fishing.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=suo_aUTUpps&t=0s) From "Earthrise" to a Digital Mirror World

Almost fifty years after the Apollo 8 mission headed to the Moon, Nat reflects on the famous moment when astronaut Bill Anders spotted our planet rising in lunar orbit and took an unscheduled photograph. Anders famously marveled over the headset, “Oh, my god, look at that picture over there. There's the earth coming up. Wow, is that pretty,” despite a crewmate warning that the shot was not on the checklist. That photograph, immortalized as "Earthrise," allowed humanity to gaze back at its home planet in a way never before possible. Traversing the 240,000-mile distance took three days in a spaceship, but today anyone can pull up an interactive, photorealistic representation of all 196.9 million square miles of the planet in roughly a single second.

Exploring how that digital globe exists, Nat and Lo visit Google to understand the mechanics and human effort powering Google Earth over its ten-plus years of development. Gopal explains the fundamental philosophy that differentiates the platform from Google Maps: “Google Maps is for finding your way. Google Earth is about getting lost.” While Maps prioritizes navigation and utility, Google Earth functions as the world's largest repository of geo-imagery, aiming to build a complete mirror world where users can explore everything from mountain peaks to cityscapes and ocean floors.

- **Apollo 8 vantage point:** A 240,000-mile journey requiring three days of spaceflight, captured impromptu in the 1968 "Earthrise" photo. - **Surface area captured:** The digital globe maps all 196.9 million square miles of Earth. - **Platform goal:** Creating a photorealistic "mirror world" repository of geo-imagery for free exploration.

---

[01:47](https://www.youtube.com/watch?v=suo_aUTUpps&t=107s) Planes, Cameras, and "Mowing the Lawn"

Satellites and airplanes perform distinct, complementary roles in assembling Google Earth. While satellites provide the global 2D imagery wrapped around the planetary sphere, high-resolution 3D models require aerial photography gathered by low-flying aircraft. Nat admits she originally assumed all overhead photography came strictly from orbit, jokingly reframing Google’s camera planes as "street view cars with wings." Capturing aerial photography for 3D generation is heavily constrained by environmental conditions, particularly cloud cover and haze. Kevin points out that clear skies are mandatory, while Gopal reveals that capturing a completely cloud-free model of London required flying repeat passes over the city across a prolonged period.

To acquire the required depth and overlap, planes fly systematic, five-hour missions over target areas. Instead of traveling point-to-point, pilots fly back-and-forth zigzag patterns across the target landscape. Kevin describes this flight geometry: “So it's north, and turn south. It's sort of like mowing the lawn.” Each aircraft is outfitted with an array of five specialized cameras—one pointing straight down, and four angled forward, backward, to the left, and to the right—so every surface feature is recorded simultaneously from multiple vantage points.

- **2D vs. 3D division:** Satellites provide 2D global wrap coverage, while airplane flights supply data for close-to-the-ground 3D reconstructions. - **Flight duration and path:** Standard photo runs last around five hours, flying lawnmower-style overlapping north-south passes. - **Camera setup:** Five simultaneous cameras per airplane capturing nadir (downward) and oblique (forward, backward, left, right) perspectives.

---

[03:20](https://www.youtube.com/watch?v=suo_aUTUpps&t=200s) Photogrammetry: From 2D Pixels to 3D Geometry

Turning overlapping aerial snapshots into interactive 3D spaces relies on photogrammetry, which Gopal defines as “just a fancy word for taking all of this imagery that we collect from the plane and constructing a 3D model.” Before building the geometry, the raw photography undergoes meticulous cleanup: teams color-correct the images, eliminate atmospheric haze and clouds, and remove moving vehicles from street grids so the cities appear undisturbed. Once prepped, computer vision algorithms take over, cross-referencing features visible across overlapping frames while incorporating precision data from onboard GPS antennas to determine exactly where each camera was positioned and angled in space.

```mermaid flowchart TD accTitle: Photogrammetry Pipeline accDescr: Diagram illustrating how Google Earth turns aerial photography into textured 3D meshes.

subgraph Acquisition["Data Capture & Preprocessing"] A["5-Camera Aerial Capture\n(Lawnmower flight paths)"] --> B["Image Cleanup\n(Haze/cloud removal, car extraction)"] end

subgraph Reconstruction["3D Geometry Generation"] B --> C["Computer Vision Feature Matching\n+ GPS Camera Positioning"] C --> D["Depth Map Calculation"] D --> E["Polygon Mesh Generation"] end

subgraph Texturing["Surface Application"] E --> F["Texture Mapping\n(Draping aerial photos over geometry)"] F --> G["Final 3D Mirror World"] end ```

By calculating the visual displacement of matching features from known positions, the software creates depth maps that establish precise distances between objects and the lens. These depth maps are stitched together into a continuous 3D polygon mesh, representing the physical shape of structures and terrain. The aerial photographs are then sliced and projected onto the mesh through texture mapping, wrapping the photography over building facades like cut paper. Gopal notes that geometric, hard-edged structures resolve cleanly, but organic shapes remain difficult: trees often turn into simplistic "lollipop" shapes due to intricate leaf-and-branch layouts, though algorithms are improving, as shown in high-fidelity scans of Yosemite National Park capturing subtle rock contours.

- **Computer vision automation:** Matches identical features across overlapping photos using exact camera GPS coordinates and angles. - **Depth maps and meshes:** Translates optical feature parallax into depth maps, aggregating them into a unified structural mesh. - **Texture mapping:** Slices and projects aerial images onto the sides and roofs of 3D polygonal models. - **Organic rendering challenges:** Complex vegetation often renders as "lollipops," whereas sheer rock faces (like Yosemite) achieve high structural accuracy.

---

[05:12](https://www.youtube.com/watch?v=suo_aUTUpps&t=312s) Scale, Zoom Layers, and "Pretty Earth"

The volume of imagery underpinning Google Earth is staggering, totaling tens of millions of distinct images. For the seamless global satellite view alone—internally referred to as "Pretty Earth"—engineers combined data from approximately 700,000 Landsat scenes. The algorithm reviews overlapping passes to pick the single clearest, cloudless pixel for every location, standardizing vegetation so that across the entire digital globe, it appears to be springtime everywhere.

The resulting global mosaic contains roughly 800 billion pixels. If an individual attempted to print this image at native resolution using a standard household printer, the paper required would match the footprint of an entire city block. Gopal underscores the processing load behind that feat: “If you took a single computer, it would take 60 years to process that.” Because Google Earth offers more than twenty discrete zoom levels, navigating through the platform is not a single flat map expanding, but rather like opening a cascading series of Russian nesting dolls made entirely out of puzzle pieces.

- **Pretty Earth composite:** Built from around 700,000 Landsat scenes to create an uninterrupted, cloudless planetary surface. - **Springtime effect:** Algorithms select optimal pixels across seasons, creating an eternal springtime appearance worldwide. - **800-billion-pixel canvas:** Printing the global mosaic would require a sheet the size of a city block; computing it on one machine would take 60 years. - **Zoom architecture:** Features over 20 distinct zoom levels layered together seamlessly.

---

[06:18](https://www.youtube.com/watch?v=suo_aUTUpps&t=378s) Keeping the Planet Alive with Earth Engine

Because landscapes and human settlements evolve constantly, Google continuously refreshes its imagery archive. While the high-altitude global satellite view is updated roughly once every two to three years, high-density metropolitan areas with rapid development are re-flown and updated in under a year. This rolling temporal record allows researchers and engineers to track changes across the surface of the planet using Google Earth Engine, a planetary-scale analytical platform that pairs historical satellite catalogs with computer vision algorithms.

By identifying changes in tree coverage over time, Earth Engine tracks deforestation trends across the Amazon rainforest and generates predictive heat maps highlighting regions experiencing heavy logging. The same analytical framework tracks global commercial fishing patterns to identify overfished waters. Gopal describes this capability as a vital diagnostic tool: “Think about it as a health monitor for the planet.” He emphasizes that maintaining Google Earth is an ongoing commitment to mirroring dynamic ecosystems rather than creating a fixed museum piece: “It's not a static planet. It is fully alive at a macro scale.”

Nat wraps up by urging viewers to explore recent updates—including the "I'm Feeling Lucky" dice feature that teleports users to unexpected places across the globe—and teases upcoming episodes exploring virtual reality through Google Earth VR.

- **Update frequency:** Whole-globe satellite views refresh every two to three years; major urban centers update within twelve months. - **Google Earth Engine capabilities:** Uses computer vision over time-series data to map deforestation in the Amazon and identify overfished oceanic zones. - **Macro health monitor:** Serves as an environmental dashboard tracking dynamic natural cycles on a living planet. - **Exploration features:** Includes an "I'm Feeling Lucky" dice roll that teleports users to randomized locations.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=suo_aUTUpps&t=0s) Google Maps is for finding your way. Google Earth is about getting lost. — Gopal

> [06:18](https://www.youtube.com/watch?v=suo_aUTUpps&t=378s) It's not a static planet. It is fully alive at a macro scale. — Gopal

## Mentioned Google Earth, Google Maps, Google Earth Engine, Google Earth VR, Google, Apollo 8, Bill Anders, Earthrise, Landsat, Amazon rainforest, Yosemite National Park, London, Nat and Friends, Gopal, Kevin

## Video description Google Earth is the most photorealistic, digital version of our planet. But how?! Where do the images come from? How are they they put together? And how often are they updated? In this video, learn about the pixels, planes, and people that create Google Earth’s 3D imagery.

Play with Google Earth → https://www.google.com/earth/ Share this video on Facebook → https://goo.gl/byTisB Follow us → https://twitter.com/NatAndFriends Subscribe → https://goo.gl/CEsJyN

Thanks to friends Vibe Mountain for the music in this video, and Freddy Arenas (http://www.freddyarenas.com) for the animations.

-------

Watch ALL the videos → https://goo.gl/8UyyHP

:) Nat & Friends http://youtube.com/natandfriends -- We make videos where we go behind the scenes at Google. What are you curious about? Let us know!

[Source](https://www.youtube.com/watch?v=suo_aUTUpps)

中文简报

谷歌出动五相机航拍机群进行3D城市重建谷歌将多角度航空摄影与计算机视觉算法相结合，将2D俯视图转化为带纹理的3D网格，突破卫星限制以监测地球变化。该视频解析了Google Earth如何构建其精细的3D环境，对比了高空2D卫星视角与以割草机式航线采集重叠图像的低空航拍飞机。通过摄影测量学和计算机视觉技术，这些多角度航拍照被剔除云层和移动车辆，转化为深度图与3D多边形网格，并完成纹理贴图。工程师还利用Landsat数据合成了名为“Pretty Earth”的全球无云镶嵌图，覆盖数十个缩放级别、包含数千亿像素。最后，视频展示了Google Earth Engine如何将数十年历史影像转化为环境健康监测工具，以追踪亚马逊森林砍伐和全球捕捞模式等动态变化。 Google Earth作为涵盖地球全部1.969亿平方英里的逼真数字镜像世界运行，结合了卫星数据与低空航空摄影。 (0:00) 为采集高分辨率3D数据，特种飞机按系统化的5小时割草机式航线飞行，机载五台相机可同时记录垂直和倾斜角度。 (1:47) 被称为“Pretty Earth”的无缝全球卫星镶嵌图融合了70万景Landsat影像，构成一张8000亿像素的画布，单台电脑需耗时60年才能处理完成。 (5:12) Google Earth Engine利用该平台多年的历史档案作为地球健康监测器，用以分析亚马逊森林砍伐和过度捕捞模式等动态变化。 (6:18)

Detailed notes — zh

## Background 该视频由 YouTube 频道 Nat and Friends 于 2017 年 4 月 18 日发布。视频简介指出，Google Earth 是地球上最逼真的数字化镜像，本期内容带观众深入幕后，揭秘构成其 3D 影像背后的像素、航拍飞机与研发团队。 ## TL;DR 视频深入剖析了 Google Earth 打造逼真 3D 数字星球的技术原理。与注重导航的 Google Maps 不同，Google Earth 旨在让用户沉浸式探索；其高精度的三维地貌并非单纯依靠卫星，而是由低空航拍飞机携带 5 向相机阵列以“割草式”往复拍摄，再通过摄影测量法与计算机视觉算法重建 3D 多边形网格并贴图。此外，宏观图层通过精选 8000 亿像素的晴空卫星图像构建无缝春季景观，并结合 Earth Engine 记录森林砍伐、过度捕捞等生态变迁，使数字地球成为反映活体行星动态演变的重要档案。 ## Atmosphere & style 全片节奏明快且充满探索的好奇心，巧妙地将宏大的航天历史与硬核的计算机视觉技术交织在一起。主持人 Nat 与工程师们的对谈轻松风趣，用“割草机式的航拍”和“俄罗斯套娃般的缩放层级”等生动比喻化解了算法与工程的枯燥感。从阿波罗任务初见地球的诗意震撼，到深入自然保护与生态监测的宏观使命，整期视频给人一种兼具科技温度与人文敬畏的沉浸感。 ## Bulletin - [05:12](https://www.youtube.com/watch?v=suo_aUTUpps&t=312s) Google Earth 的宏观全球图层由 Landsat 卫星的约 70 万个独立场景拼接而成，总像素量达 8000 亿，若用单机处理需要耗时整整 60 年。 - [02:08](https://www.youtube.com/watch?v=suo_aUTUpps&t=128s) 逼真的城市 3D 景观并非来自卫星，而是依靠装载 5 向航拍相机的专用飞机进行拉网式低空飞行采集，结合摄影测量法重建多边形网格与纹理贴图。 - [06:18](https://www.youtube.com/watch?v=suo_aUTUpps&t=378s) 图像更新频率因分辨率与地理位置而异：外太空宏观视图每隔几年刷新一次，而人口密集的大城市核心区通常在一年内就会完成更新。 - [06:18](https://www.youtube.com/watch?v=suo_aUTUpps&t=378s) 借助 Earth Engine 平台与计算机视觉算法，Google Earth 已成为监测地球健康的宏观仪器，可自动分析亚马逊森林砍伐及全球过度捕捞等环境演变。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=suo_aUTUpps&t=0s) 产品定位差异：Google Maps 核心在于导航与指路（Finding your way），而 Google Earth 旨在提供沉浸式的数字镜像世界（Mirror world），让人神游并迷失其中。 - [02:08](https://www.youtube.com/watch?v=suo_aUTUpps&t=128s) 高精度 3D 城市建模主要依赖低空飞机而非卫星，飞机被称为“长了翅膀的街景车”，通过微小折返间距进行“割草”式扫描拍摄。 - [02:08](https://www.youtube.com/watch?v=suo_aUTUpps&t=128s) 航拍飞机机腹装备了 5 台专业高清相机（垂直向下以及前、后、左、右四向倾斜），单次任务飞行通常持续约 5 小时。 - [02:08](https://www.youtube.com/watch?v=suo_aUTUpps&t=128s) 多变的天气与云层遮挡是航拍采样的最大挑战，例如团队曾反复飞越伦敦上空无数次，才拼凑出无云遮挡的完整城市全貌。 - [03:52](https://www.youtube.com/watch?v=suo_aUTUpps&t=232s) 摄影测量法结合高精度机载 GPS 天线位姿数据与特征点匹配算法，推算物体距相机的距离并生成深度图（Depth Maps）。 - [03:52](https://www.youtube.com/watch?v=suo_aUTUpps&t=232s) 通过将离散的深度图缝合为 3D 多边形网格（Mesh），再精准切片贴附真实照片完成纹理贴图（Texturing），实现逼真的三维还原。 - [03:52](https://www.youtube.com/watch?v=suo_aUTUpps&t=232s) 计算机视觉算法处理规则几何体相对容易，而树木、岩石等自然有机形状（Organic shapes）的重建难度极高，目前建模保真度已大幅提升。 - [05:12](https://www.youtube.com/watch?v=suo_aUTUpps&t=312s) 全局“美丽地球”（Pretty Earth）图层通过算法筛选出各地最晴朗无霾的时段像素，使整个宏观地球永久呈现繁花似锦的“春季景观”。 - [05:12](https://www.youtube.com/watch?v=suo_aUTUpps&t=312s) 平滑的缩放体验背后隐藏着超过 20 层的瓦片金字塔缩放层级（Zoom levels），架构如同多层嵌套的俄罗斯套娃。 - [06:18](https://www.youtube.com/watch?v=suo_aUTUpps&t=378s) 地球图像具有动态生命周期，Earth Engine 能将时间序列数据转化为环境分析热力图，使 Google Earth 成为观察地球生态规律的活体档案。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=suo_aUTUpps&t=0s) 从“地出”到数字镜像地球主持人 Nat 回顾了将近 50 年前的阿波罗 8 号任务。1968 年，宇航员比尔·安德斯（Bill Anders）在环月轨道上拍下了一张意料之外的惊艳照片，当时他惊叹道：“Oh, my god, look at that picture over there. There's the earth coming up. Wow, is that pretty.”（天啊，看那边的画面！地球正升起来，真是太美了。）尽管同行机组成员开玩笑提醒这不在计划拍摄任务中，这张被命名为“地出”（Earthrise）的照片依然成为了人类历史上首次以全景视角俯瞰地球的划时代影像。在过去，想要获得这样的视角需要航行 24 万英里、花费数天时间；而今天，任何人在屏幕前点击一个链接，仅需大约一秒钟就能俯瞰我们整个 1.969 亿平方英里大小的星球。 Google Earth 团队的 Gopal 阐述了 Google Maps 与 Google Earth 的根本区别。他指出，Google Maps 专注于导航与指路，而 Google Earth 则是让人沉浸与迷失其中。团队的目标是构建全网最庞大的地理影像库，打造出地球最逼真的数字版本，正如 Gopal 所说：“We're trying to create a mirror world so people can go anywhere.”（我们正努力创造一个镜像世界，让人人都能神游任何地方。）从高耸的山脉、繁华的城市到神秘的海底，Google Earth 在过去十年间不断迭代，画面解析度越来越高。为了揭开这颗数字星球背后的数据来源、图像规模以及拼接算法，Nat 与搭档 Lo 一同采访了团队核心成员 Gopal 和 Kevin。 - **地球总表面积**：约 1.969 亿平方英里（约 5.1 亿平方公里）。 - **产品核心定位**：Google Maps 用来寻找路线（Finding your way），Google Earth 用来探索迷失（Getting lost）。 - **数据层级划分**：高空宏观全球视图依赖卫星（Satellites）采集并平铺包裹为 2D 贴图；接近地面的局部高精度 3D 视图则主要依赖飞机（Planes）采集。 --- ### [02:08](https://www.youtube.com/watch?v=suo_aUTUpps&t=128s) 航拍飞行动图与摄影测量法多数人往往误以为所有俯瞰图都是由高空卫星拍摄的，但要生成逼真的 3D 影像，必须依靠低空飞行器——Nat 形象地将其称作“长了翅膀的街景车”。Kevin 指出，航拍过程中面临的最大挑战就是多变的天气，团队始终力求在万里无云的晴空下进行拍摄。Gopal 补充说，为了拼凑出一整套完全没有云层遮挡的伦敦全貌，飞机不得不反复飞越伦敦上空无数次。一架执行采集任务的航拍飞机单次飞行时长通常在 5 小时左右。飞机并不是长途跨国巡航，而是在特定空域以微小的折返间距进行“之”字形来回拉网式扫描，Kevin 将其比喻为“割草”（mowing the lawn）。机腹装备了 5 台专业高清相机：一台垂直朝下，另外四台分别朝向前、后、左、右。连续重叠拍摄的画面随后被送入后台，通过名为“摄影测量法”（Photogrammetry）的算法体系构建出高精度的 3D 模型。 ```mermaid flowchart TD accTitle: 航拍 3D 建模流程图 accDescr: 展示从飞机航拍采集到计算机视觉算法重建 3D 网格的完整工艺步骤 subgraph Prep["图像采集与前期准备"] A["5 向航拍阵列<br/>(下/前/后/左/右)"] --> B["预处理<br/>(去云雾/色彩校正/移除车辆)"] end subgraph CV["计算机视觉处理"] B --> C["特征点匹配<br/>(结合高精度 GPS 天线位姿)"] C --> D["生成深度图 (Depth Maps)"] end subgraph Mesh["模型与贴图输出"] D --> E["缝合 3D 多边形网格 (Mesh)"] E --> F["纹理贴图烘焙 (Texture Mapping)"] end ``` - **航拍作业模式**：南北交替折返，重叠覆盖同一区域。 - **前期图像预处理**：去云除雾（Removing clouds & haze）、统一色彩校正（Color correcting），甚至手动或自动剥离街道上的运动车辆，以保证场景静止纯净。 - **摄影测量学定义**：Gopal 将其概括为“taking all of this imagery we collect from the plane and constructing a 3D model”（采集飞机航拍的所有图像并据此构建 3D 模型的统称）。 --- ### [03:52](https://www.youtube.com/watch?v=suo_aUTUpps&t=232s) 计算机视觉与 3D 网格重建近年来 3D 呈现效果的飞跃，关键在于计算机视觉（Computer Vision）技术的突破。机载的高精度专用 GPS 天线能够精确记录快门按下瞬间相机的空间坐标与拍摄倾角。算法系统自动在海量重叠照片中寻找相同的视觉特征点，计算出物体与相机之间的距离，从而生成细致入微的“深度图”（Depth Map）。各个视角相机的深度图会被缝合为一个完整的 3D 三角多边形网格（Mesh），然后再将航拍拍摄的真实照片切片并精准烘焙贴附到网格表面，这就是“纹理贴图”（Texturing）。Gopal 解释，规则几何体的边缘很容易提取重建，但自然界中的有机形状（Organic shapes）极为棘手。比如树木拥有复杂的枝杈和叶片，算法往往只能粗糙地将其还原成“棒棒糖”形状。不过团队的模型解析力正在大幅进化，在近期优胜美地国家公园（Yosemite National Park）的采集中，系统已经能够以极高保真度重建岩壁那细腻曲折的天然纹理。 - **多边形网格（Mesh）**：将分散的深度图数据缝合成整体空间立体骨架。 - **纹理贴图（Texturing）**：如同裁剪照片纸片贴在立体建筑边缘，赋予 3D 模型真实光影与细节。 - **有机物建模突破**：从早期的“棒棒糖”式树木到高精细度的优胜美地花岗岩峭壁。 --- ### [05:12](https://www.youtube.com/watch?v=suo_aUTUpps&t=312s) 8000 亿像素的“美丽地球”与套娃缩放体系当被问及 Google Earth 究竟是由多少张照片组成时，Gopal 表示其数量级在“数千万张”。以全局尺度的“美丽地球”（Pretty Earth）图层为例，该无缝拼接的全球视图精选自陆地卫星（Landsat）捕获的约 70 万张独立场景。团队利用算法在同一地点的不同时段照片中挑选出最为晴朗通透的像素点进行组合，因此在 Google Earth 的宏观图景中，全球各地都永远呈现出繁花似锦的“春季景观”。整个无缝春季地球图层由高达 8000 亿（800 billion）个像素拼接而成。如果将这张图用家用打印机以原始比例打印出来，需要一张如同一个城市街区（City block）那么大的纸张；若仅依靠一台单机电脑来渲染处理这批数据，整整需要 60 年的时间。更令人震撼的是，用户在缩放时所体验到的无缝平滑浏览，背后实际上隐藏着超过 20 层的缩放层级（Zoom levels）。Nat 将其比喻为一套用无数拼图碎片构成的“俄罗斯套娃”（Russian dolls），每次放大缩小都在穿梭于不同精度的庞大图层之间。 | 维度 / 项目 | 数据与技术指标 | 备注说明 | |:--- |:--- |:--- | | 全局图层数据源 | 约 700,000 个 Landsat 场景 | 算法挑选最优像素剔除阴霾 | | 全局像素总量 | 8000 亿像素 (800 Billion Pixels) | 打印需要街区大小纸张 | | 单机处理时间 | 约 60 年 | Google 借助超大规模分布式计算处理 | | 缩放层级深度 | 20+ 个层级 | 类似套娃结构的瓦片金字塔体系 | --- ### [06:18](https://www.youtube.com/watch?v=suo_aUTUpps&t=378s) 动态更新与地球引擎（Earth Engine）公众最关心的核心问题莫过于图像的更新频率。Gopal 透露，外太空视角的全球宏观卫星底图大约每隔几年全面更新一次；而越是深入地面、人口稠密的核心大都市圈，更新频率就越高，通常在一年之内就会刷新。这种高密度的时序图像积累，让 Google Earth 不再只是一张静态地图，而是成为记录地球演变的动态档案。借助 Earth Engine 平台与计算机视觉算法，科研人员能够自动分析环境变化规律。例如，通过追踪亚马逊雨林树冠的萎缩与扩张来监测滥伐情况，生成全球重度采伐热力图；亦或是监控海洋捕捞船队的轨迹以识别过度捕鱼区域。Gopal 感慨道：“Watching these cycles happen on the planet, you realize that this is a living thing. And the product we're building has to be a living thing to reflect that. It's not a static planet.”（观察地球上的这些宏观周期，你会意识到地球本身是一个活生生的生命体。我们打造的产品也必须是一个活生生的实体才能反映出这一点。它绝不是一颗静态的星球。）最后，Nat 推荐观众体验最新上线的“手气不错”（I'm Feeling Lucky）掷骰子随机传送功能，并预告团队后续将推出关于虚拟现实深度探索的内容，包括 Google Earth VR 的技术揭秘。 - **更新周期规律**：全球太空视图数年一更；重点大城市更新周期压缩至 1 年以内。 - **Earth Engine 核心应用**： - 追踪亚马逊森林砍伐，绘制全球采伐热力图。 - 监控全球水域船只分布，识别过度捕捞热点。 - 充当全人类观测地球健康的宏观体检监视器（Health monitor for the planet）。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=suo_aUTUpps&t=0s) We're trying to create a mirror world so people can go anywhere. — Gopal ## Mentioned Google Earth, Google Maps, Nat, Lo, Gopal, Kevin, 比尔·安德斯 (Bill Anders), 阿波罗 8 号 (Apollo 8), Earth Engine, Google Earth VR, 陆地卫星 (Landsat), 优胜美地国家公园 (Yosemite National Park), 伦敦 (London), 亚马逊雨林 (Amazon Rainforest), Vibe Mountain, Freddy Arenas ## Video description Google Earth is the most photorealistic, digital version of our planet. But how?! Where do the images come from? How are they they put together? And how often are they updated? In this video, learn about the pixels, planes, and people that create Google Earth’s 3D imagery. Play with Google Earth → https://www.google.com/earth/ Share this video on Facebook → https://goo.gl/byTisB Follow us → https://twitter.com/NatAndFriends Subscribe → https://goo.gl/CEsJyN Thanks to friends Vibe Mountain for the music in this video, and Freddy Arenas (http://www.freddyarenas.com) for the animations. ------- Watch ALL the videos → https://goo.gl/8UyyHP :) Nat & Friends http://youtube.com/natandfriends -- We make videos where we go behind the scenes at Google. What are you curious about? Let us know! [Source](https://www.youtube.com/watch?v=suo_aUTUpps)

Source quotes

[0:00] Gopal: We're trying to create a mirror world so people can go anywhere.

Source quotes

[0:00] Gopal: Google Maps is for finding your way. Google Earth is about getting lost. [6:18] Gopal: It's not a static planet. It is fully alive at a macro scale.

Original transcript

[0:00] [MUSIC PLAYING] [0:05] NAT: Almost 50 years ago, Apollo 8 [0:07] left Earth for the moon. [0:09] While in orbit, astronaut Bill Anders [0:12] decided to take an unexpected photo. [0:14] BILL ANDERS: Oh, my god, look at that picture over there. [0:16] There's the earth coming up. [0:18] Wow, is that pretty. [0:19] SPEAKER: Hey, don't take that. [0:20] It's not scheduled. [0:23] Later named "Earthrise," the photo [0:26] captured our planet in a way that it had never [0:29] been seen before. [0:31] And while the 240,000-mile trip to get to that vantage point [0:35] was, I'm sure, amazing-- [0:37] I mean, if you ever get that offer, take it-- [0:40] if you don't have a spaceship and three days to spare, [0:44] lucky for you, all you need to do [0:46] is click a link to get almost the same view [0:48] in about a second. [0:50] [MUSIC PLAYING] [0:54] This is the Earth, all 196.9 million square miles of it. [1:01] GOPAL: Google Maps is for finding your way. [1:03] Google Earth is about getting lost. [1:05] Google Earth is the biggest repository [1:07] of geo-imagery, the most photorealistic digital version [1:12] of our planet. [1:13] We're trying to create a mirror world [1:14] so people can go anywhere. [1:16] NAT: From mountains, to cities, [1:18] to the bottom of the ocean. [1:21] Google Earth has been around for about the last 10 years. [1:24] And just like our Earth, it's been evolving over this time. [1:29] The imagery has been getting better and better. [1:32] I was really curious to know how is Google Earth created? [1:36] How many images actually make it up? [1:39] And where do they come from? [1:41] Last year, together with Lo, I met up with Gopal and Kevin [1:45] to find out. [1:47] GOPAL: So how do we build Google Earth? [1:49] The way it starts is we look at places [1:53] that we want to collect in imagery, [1:54] and then we collect it through a variety of different ways. [1:57] One is satellites. [1:59] And satellites give you the global views. [2:01] And that's all 2D imagery that's wrapped around the globe. [2:03] When you get closer to the ground, [2:05] we have 3D data that we collect from planes. [2:08] NAT: Yes, you heard right-- planes. [2:11] I'd always assumed that every overhead photo [2:13] of the Earth I'd ever seen was taken by a satellite. [2:17] But I learned creating 3D imagery requires [2:19] special conditions. [2:21] So Google flies planes or, as I now like to think of them, [2:24] street view cars with wings. [2:27] LO: What are some of the challenges that you guys have? [2:30] KEVIN: The biggest challenge in doing something [2:32] like this is weather. [2:33] Our preference is always to have clear skies. [2:37] GOPAL: It took us a really long time to get London, [2:39] because we had to fly over London a lot, [2:41] before we got a fully cloud-free image of London. [2:44] KEVIN: Come on in. [2:45] LO: Thanks. [2:47] NAT: Kevin told us that a typical flight to take photos [2:50] is around five hours, except the planes aren't going [2:53] across the country, they're making little zigzags [2:55] over the same area. [2:57] KEVIN: So it's north, and turn south. [3:01] It's sort of like mowing the lawn. [3:04] NAT: This pattern helps the photos overlap. [3:06] And multiple cameras help capture a place [3:09] from different angles. [3:10] KEVIN: The planes have five different cameras-- [3:12] one looking down, and forward, back, left and right. [3:15] NAT: In my mind, I'm picturing, [3:17] like, photograph, photograph, photograph, photograph. [3:19] KEVIN: Yep. [3:20] NAT: And then, something puts them all together. [3:23] Something called an algorithm? [3:25] KEVIN: Photogrammetry, yeah. [3:26] GOPAL: Which is just a fancy word [3:27] for taking all of this imagery that we collect from the plane [3:31] and constructing a 3D model. [3:33] NAT: The first step to creating a place in 3D [3:35] is a little bit of photo editing. [3:37] GOPAL: So all the imagery is prepped. [3:39] That would be removing clouds, removing haze, [3:41] color correcting. [3:43] You'll actually see that a lot of the cities [3:44] don't even have cars in them. [3:46] We actually take the time to extract those. [3:49] NAT: Then the 3D science begins. [3:52] KEVIN: The big breakthrough that's [3:53] happened in the last few years has been the introduction [3:56] of computer vision. [3:59] The computer looks for features within the overlapping [4:02] images that are the same. [4:04] We use a special GPS antenna that [4:07] allows us to know where that camera was, [4:08] so we know roughly where things were taken and at what angle. [4:12] NAT: And this allows them to create a depth map. [4:15] GOPAL: And that's just our understanding of how far [4:18] the things are from the camera. [4:19] And we take all of these various depth maps [4:21] from the different cameras, stitch those together [4:24] in what's called a mesh, which is basically [4:25] a big 3D reconstruction of the place, and we texture it. [4:30] And texture is applying the photography [4:32] that we took to the sides of these 3D buildings. [4:35] It's almost like taking pieces of paper and cutting them up. [4:38] You can actually extract the edges of something, [4:41] and then stitch those together, and then understand [4:44] what that shape might be. [4:47] And organic shapes are what's hard to render. [4:49] It gets even more complicated when [4:51] you're talking about trees, because trees [4:52] have branches and leaves. [4:54] And often, you might see them as a lollipop, [4:56] because that's as good as we can get. [4:59] But we're getting better at modeling those organic shapes. [5:03] We did a collect of Yosemite National Park. [5:05] And we were able to actually capture [5:07] that in really high fidelity, with all the different bends [5:09] that a rock face might have. [5:12] NAT: Do you know how many different images make up [5:15] what I see as Google Earth? [5:18] GOPAL: Yeah. [5:18] It's probably on the orders of tens of millions. [5:22] One interesting stat is to look at what we call Pretty Earth. [5:26] And this is the global view. [5:28] So we have a full, seamless image. [5:31] And that comes from about 700,000 scenes from Landsat. [5:34] And what we're doing is we're finding [5:36] the best pixel from each. [5:38] So if you look at Google Earth, it's springtime everywhere. [5:41] NA: To be precise, a 800-billion-pixel spring globe, [5:45] which is so big, if you wanted to print that out on your home [5:48] printer, you would need to find a piece of paper [5:50] the size of a city block. [5:51] GOPAL: If you took a single computer, [5:53] it would take 60 years to process that. [5:56] NAT: And you can just keep multiplying this times [5:58] all the different levels of zoom that [6:00] exist in Google Earth, which are over 20. [6:03] So even though using Earth feels like one seamless world [6:07] that you're just zipping around, it's [6:10] really more like you're traveling [6:11] through a series of Russian dolls, [6:14] all made up of puzzle pieces. [6:18] LO: I think everybody else [6:19] wants to know, how often do the images get updated? [6:22] It's like the number one question. [6:24] GOPAL: We try to update it as often as possible. [6:26] The image all the way from space, [6:28] when you're looking at the whole globe, [6:29] we try to update that maybe once every couple of years. [6:33] As you start to dive in, we update that imagery [6:35] more and more frequently. [6:36] So for big city populations, it'll be under a year. [6:40] What that allows you to do is look [6:41] at how the planet has changed over time. [6:43] And we use this product, called Earth Engine, [6:45] that allows you to look at all of this data [6:48] and, using computer vision, draw out insights from the things [6:51] that are changing. [6:52] So we can track deforestation in the Amazon, [6:55] because we can see how the trees are shrinking and growing. [6:57] And then, from that, we can generate a heat map of the most [7:00] logged places on the planet. [7:01] We can also do that with fishing and see the most overfished [7:05] areas. [7:06] Think about it as a health monitor for the planet. [7:09] Watching these cycles happen on the planet, [7:11] you realize that this is a living thing. [7:13] And the product we're building has to be a living thing [7:15] to reflect that. [7:16] It's not a static planet. [7:20] It is fully alive at a macro scale. [7:23] And that is very eye-opening, when you're actually [7:25] watching the things change right in front of your eyes, [7:27] when you have that perspective to see that. [7:32] NAT: Thanks for watching. [7:34] And if you haven't played with Google Earth recently, [7:36] I highly recommend this new feature they just launched [7:39] called, I'm Feeling Lucky. [7:41] You roll the dice, and then it teleports you [7:43] to a random spot in the world. [7:48] Also, to give you a heads-up, in the coming weeks [7:51] we'll be diving deep into VR, including Google Earth VR. [7:54] So there's that to look forward to. [7:57] OK, that's all from me. [7:58] Bye. [7:59] [MUSIC PLAYING]