The instinct is to look for a question no human being has ever asked. That instinct is wrong twice over: genuinely untouched questions are vanishingly rare, and the ones you find are usually untouched because they are impossible to answer with free data and a laptop.

What actually wins competitions and gets published is a novel combination:

(a dataset few people use) × (a place or scale few people study) × (a question someone has a stake in) The three‑term test

Each term does specific work. The dataset term means nobody has the drop on you. The place term means the result is yours and not a re‑run. The stakes term is the one students skip, and it is the one reviewers respond to hardest.

The strongest version of the third term is auditing

Someone publicly claims an environmental outcome — a company says it protected a forest, an agency says it paid farmers to plant cover crops, a government says it plugged leaking wells — and you check the claim from orbit.

This framing is unusually powerful for a student, for two reasons that have nothing to do with how clever the analysis is:

• Novelty is not arguable. Nobody has audited that specific claim before. You do not have to win a debate about whether your topic is new.
• The result has consequences. A finding that money did or did not buy the outcome it promised is interesting to someone outside your school.

The two things that decide success

1. Do the novelty search before you write any code

Budget a full week. Search Google Scholar for your exact dataset × region × question triple, then read the reference lists of the three closest hits — that is where the paper that scoops you is hiding. Also search the local watershed association and any nearby university's environmental science department, because a great deal of this work lives in gray literature and unpublished master's theses that Scholar indexes badly.

2. Have a control group, and quantify your uncertainty

The characteristic failure of a student remote‑sensing project is a convincing‑looking correlation with nothing to compare it against. Buried streams run hotter than open streams — but buried streams are also under parking lots. Unless you match them against open streams with similar pavement, slope, and upstream drainage area, you have not discovered anything about burial. You have rediscovered that cities are paved.

Matched controls, error bars, and validation against ground measurements are the entire difference between a paper and a poster.

The moment to build on

The richest environmental detail in Spirited Away is not the Stink God. It is Haku.

Haku cannot remember his own name, and so he cannot leave the bathhouse. Chihiro remembers it for him: he is the spirit of the Kohaku River — 琥珀川 — the river she fell into as a small child, which carried her safely to the bank instead of drowning her. The reason he lost his name is stated plainly in the film. The river was filled in, and apartments were built on top of it.

A river was paved over. Its spirit forgot what it was. Nobody remembers it was ever there.

Both halves of that map onto named science

Stream burial

Cities worldwide have driven their small streams into culverts, pipes, and concrete channels and built over the top. Headwater streams are hit hardest, because they make up the largest share of total stream length and are the cheapest to bury — that is the finding of Elmore & Kaushal (2008), the canonical paper. The documented consequences are specific: destroyed channels, downstream habitat degradation, fragmented aquatic habitat, faster transport of water and toxic contaminants, and the loss of nutrient and sediment retention.

Shifting baseline syndrome

And here is why nobody notices. Soga & Gaston (2018) define it as a gradual change in the accepted norms for the condition of the natural environment, caused by a lack of experience, memory, or knowledge of its past condition. It is also called environmental generational amnesia, and it was coined by the fisheries scientist Daniel Pauly in 1995.

The literature's own favourite example is rivers. Children once swam in clear local streams and caught minnows in jars. Today a murky, straightened concrete channel is simply accepted as what a river is.

Haku forgetting his name is shifting baseline syndrome. The mapping the whole project rests on

That is not a metaphor forced onto the film. It is a precise correspondence between a fictional device and a citable scientific concept — and it is what makes the framing read as intellectual rather than decorative.

Research questions

Method

1. Historical layer. Georeference historical USGS topographic quadrangles — free from USGS topoView — plus Sanborn fire insurance maps. Digitize the blue lines: streams that used to be mapped.
2. Topographic layer. Take USGS 3DEP LiDAR elevation data and compute flow accumulation, using richdem or whitebox in Python, or QGIS. This tells you where a stream should be from the shape of the land alone.
3. Define a ghost reach. Topography says stream + historical map says stream + modern NHD says nothing → buried. This follows the established Elmore & Kaushal approach, which means your method is defensible rather than invented.
4. Overlay. Census income and race, HOLC redlining grades, NLCD canopy and imperviousness, city 311 open data, and Landsat land surface temperature via Earth Engine.
5. Statistics. Matched comparison of buried against still‑open reaches, matched on imperviousness, slope, and upstream contributing area. Then test Moran's I, and move to a spatial error or lag model if autocorrelation is present.

Where the gap is

Skip Baltimore and Detroit. Your own city is both more novel and a better story. Note also that there is a public art installation called “Ghost Rivers” in Baltimore about the buried Sumwalt Run — excellent precedent to cite, and it means that exact title is taken there.

Feasibility

High. QGIS plus Python (geopandas, richdem or whitebox), all data free, and no access wall of the kind that complicates the ECOSTRESS project on sheet H‑2.

The historical map georeferencing is genuinely tedious — and that is an asset. It is visible, honest labour that reviewers and judges respect, and it is the part that cannot be automated away.

Alternative — The Half‑Life of a River Cleanup

If hands in the water beats hands in a GIS, build on the other scene, which has the strongest autobiographical link to Miyazaki himself. The River Spirit arrives at the bathhouse as the reeking Stink God; Chihiro finds a thorn in its side, and when it is pulled a torrent of garbage pours out — and a bicycle. Miyazaki has confirmed this came from cleaning a polluted river in his hometown, where the crew hauled a bicycle out of the muck. “I cleaned a river once.”

The question nobody has properly answered: how long does a cleanup last?

Volunteer cleanups are enormously popular, well funded, and emotionally satisfying — and the rate at which the trash comes back is barely measured. That is a striking gap for such a widespread practice.

1. Pick 6–10 reaches of a local urban stream.
2. Standardized baseline litter survey: fixed transects, item counts by material category, following an established protocol.
3. Clean each reach completely.
4. Re‑survey at fixed intervals — 2, 4, 8, and 16 weeks.
5. Fit a re‑accumulation curve and estimate a half‑life.
6. Test what predicts fast re‑accumulation: upstream impervious area, storm events between surveys, road proximity, adjacent land use, outfall locations.

It produces a real number with error bars, it is cheap, and it yields a direct policy recommendation: cleanups here last three months and here three weeks, so put the interception device in the second place.

The college application arc

1. You noticed something specific — not “the movie is about nature,” but a spirit who lost his name because his river was paved over.
2. You went looking, and found the real thing has a name: stream burial. And the forgetting has a name too: shifting baseline syndrome.
3. You went hunting for your own city's ghost rivers — georeferencing century‑old maps, running flow accumulation over LiDAR.
4. You found them. And you found who lost them.
5. Haku gets his name back and is freed. You gave your city's rivers their names back.

The original idea

Nearly all urban heat island research uses Landsat, which passes over at about 10:30 in the morning. But heat does not kill people at 10:30 in the morning. It kills them across nights that never cool down, when the body gets no recovery period at all. So: which built‑environment features predict slow overnight cooling, and does that track income or historical redlining?

The instrument that makes this possible is ECOSTRESS, a thermal sensor mounted on the International Space Station. Because the station's orbit precesses, its overpass time drifts — which is normally an inconvenience, but here is the entire point. It samples the night.

What we found on the fourth search

What survives

Two things, and they are both narrow but real.

1. Everyone measures levels. Nobody measures the rate. This is the pattern across every single paper in the urban heat equity literature: temperature is sampled at a snapshot time and compared between neighborhoods. But the physically interesting quantity is the nocturnal decay parameter — how fast a place sheds heat after sunset. You get it by fitting a diurnal temperature cycle model, a three‑ to six‑parameter curve with separate day and night functions, to the irregular ECOSTRESS overpasses.

This matters because it is a different physics question. Daytime peak temperature is set by radiative load: albedo, tree shade, pavement. Nocturnal cooling rate is set by thermal inertia and geometry — the heat capacity of the built mass, and the sky view factor, meaning how much open sky a surface can actually radiate to. A narrow street canyon cannot cool no matter how pale you paint it.

2. It was one city. The authors say so themselves, in print: “It is important to note that LA is a rather uniquely situated city… may not be extendable to other U.S. cities without further investigation.” A multi‑city version with city as a random effect in a multilevel model, and variance partitioning to see how much of the disparity is a within‑city versus between‑city phenomenon, is a genuinely different contribution.

The feasibility surprise

Week‑one go / no‑go

Before committing, check exactly one thing: how many usable nighttime ECOSTRESS scenes exist over your candidate city? If you cannot get enough night observations spread across enough hours to constrain a decay curve, the entire project is impossible, and you want to know that in week one rather than month three. Note that coverage runs roughly 52°N to 52°S.

Auditing a claim · someone said they did an environmental good

Blind spots · nobody looked because old sensors could not see it

Wrong variable · everyone measured the thing that is easy to measure

Field work · boots, not code

Order of operations

1. 1

   Choose the study area

   Everything depends on this — the novelty check, the data audit, and the story. What you want: a real stream network, city 311 open data, USGS 3DEP LiDAR coverage, historical topographic quadrangles back to about 1900, and a personal connection. Avoid Baltimore and Detroit.

2. 2

   Novelty search — budget one week

   Search Scholar for "buried stream" OR "stream burial" OR daylighting + [city]. Then search the city's watershed association and the nearest university's environmental science department, because much of this work is gray literature and unpublished theses. Read the reference lists of the three closest hits.

3. 3

   Data availability audit — one afternoon

   • USGS topoView — confirm historical quad coverage and the earliest available date
   • USGS 3DEP — confirm LiDAR coverage and resolution
   • City open data portal — confirm 311 flooding categories exist and are geocoded
   • NHD — download modern flowlines
   • Mapping Inequality — check whether your city has a HOLC redlining map
4. 4

   Write the formal proposal

   Hypotheses stated so they could be falsified, methods, the statistical plan written before you see results, and a timeline. This document is also most of your competition paperwork.

5. 5

   Pilot on one small watershed

   Do not scale to the whole city first. Georeference one quad, run flow accumulation on one sub‑basin, find one ghost reach end to end. You will discover the three things that are harder than expected while they are still cheap to fix.

6. 6

   Scale, then write

   Post the preprint to EarthArXiv as soon as the analysis is stable — it is free, citable, and timestamps your work. Then check the AGU abstract deadline and the competition calendar below.

Competitions

Publication

The four standing rules

1. 1

   Install Git for Windows — optional, but do it

   Git for Windows provides Git Bash, which enables Claude Code's Bash tool. Without it, Claude Code falls back to using PowerShell as its shell. You will not need to learn Git yourself.

   1. Go to git-scm.com/downloads/win and download the installer.
   2. Run it and click Next on every screen to accept the defaults. There are many screens; you do not need to change anything on any of them.
   3. If it asks you to choose an editor, keep the default.
   4. When you reach “Adjusting your PATH environment,” keep the recommended option selected.

   Already have Git, or not sure? Install it anyway. Reinstalling causes no problems.

2. 2

   Open PowerShell

   PowerShell is Windows' built‑in terminal. It is pre‑installed on every Windows computer.

   Press Win + X and choose Windows PowerShell or Terminal from the menu. A window with a blinking cursor appears.

   PowerShell or CMD? This matters

   Windows has two command‑line programs that look nearly identical but take different commands. Read the start of your prompt line:

   • PowerShell shows PS C:\Users\YourName> — note the PS.
   • CMD shows C:\Users\YourName> — no PS.

   Not the x86 one

   If the Start menu offers Windows PowerShell (x86), do not use it. That entry runs as a 32‑bit process, and the installer will refuse with “Claude Code does not support 32‑bit Windows.” Pick the entry without (x86) in its name.

3. 3

   Run the installer

   Copy the line below, paste it into PowerShell with Ctrl + V or a right‑click, and press Enter. irm fetches the installer and iex runs it.

   PowerShellCopy

   irm https://claude.ai/install.ps1 | iex

   You will see text scrolling. When it finishes you should see Claude Code successfully installed!

   If you are in CMD rather than PowerShell, use this instead:

   CMDCopy

   curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd

   Two errors that mean you're in the wrong shell

   • The token '&&' is not a valid statement separator → you ran the CMD command in PowerShell. Use the PowerShell one.
   • 'irm' is not recognized as an internal or external command → you ran the PowerShell command in CMD. Use the CMD one, or open PowerShell.

   Native installs update themselves in the background, so this is the only time you run an install command.

4. 4

   Verify it worked

   PowerShellCopy

   claude --version

   A working install prints a version number followed by (Claude Code) — something like 2.1.211 (Claude Code).

   For a fuller check, claude doctor prints read‑only diagnostics — install health, settings file validation errors, and any warnings with suggested fixes — without starting a session.

   PowerShellCopy

   claude doctor

   If you get 'claude' is not recognized, the install directory is not on your PATH yet. That fix is on sheet W‑4.

5. 5

   Log in

   Navigate to a folder you want to work in, then start Claude Code. In PowerShell, cd changes directory:

   PowerShellCopy

   cd C:\Users\YourName\Documents\ghost-rivers
   claude

   On first run you are prompted to log in, and a browser window opens for you to sign in. Once you have logged in, your credentials are stored and you will not need to do it again. To switch accounts later, type /login inside a running session.

   If the ANTHROPIC_API_KEY environment variable is set, Claude Code skips the browser and asks you once to approve that key instead.

Other ways to install

Native Windows or WSL?

You can run Claude Code directly on Windows or inside the Windows Subsystem for Linux. Choose based on where your project files live and whether you need sandboxing.

For a student research project on a Windows laptop, native Windows is the right default. Your QGIS install, your Python environment, and your downloaded map files are all Windows‑side, and crossing the WSL filesystem boundary is slow and confusing. If you go the WSL route instead, you install and launch claude inside the WSL terminal, not from PowerShell — and WSL setups do not need Git for Windows.

Pointing Claude Code at Git Bash

With Git for Windows installed, Claude Code uses Git Bash for its Bash tool. If it cannot find it, set the path in your settings file at %USERPROFILE%\.claude\settings.json:

{
  "env": {
    "CLAUDE_CODE_GIT_BASH_PATH": "C:\\Program Files\\Git\\bin\\bash.exe"
  }
}

Note the doubled backslashes — that is required in JSON. To find where your Git actually lives:

Get-Command git | Select-Object Source

Uninstalling

If you installed natively:

Remove-Item -Path "$env:USERPROFILE\.local\bin\claude.exe" -Force
Remove-Item -Path "$env:USERPROFILE\.local\share\claude" -Recurse -Force

WinGet: winget uninstall Anthropic.ClaudeCode. npm: npm uninstall -g @anthropic-ai/claude-code.

Six things about living in a terminal

• You cannot click on things. Use the arrow keys to move around.
• Esc interrupts Claude mid‑task. This is the single most useful key. If it starts down the wrong path, stop it rather than letting it finish.
• Enter sends your message. Type in plain English.
• ↑ recalls previous commands, and Tab completes them.
• Type / to list every command available to you.
• exit, or Ctrl + D twice on an empty prompt, leaves.

Starting up

PS C:\Users\You> cd Documents\ghost-rivers
PS C:\Users\You\Documents\ghost-rivers> claude

  ╭──────────────────────────────────────────╮
  │  Claude Code v2.1.211                    │
  │  model: claude-opus-5                    │
  │  cwd:   C:\Users\You\Documents\...       │
  ╰──────────────────────────────────────────╯

  /help for commands · /resume to continue a previous conversation
> 

Above the prompt you see the version, the current model, and the working directory. That last one matters: Claude Code can only see files in and below the directory you started it in. Start it in the wrong place and it will not find your project.

Ask about the project before asking for changes

Claude Code reads your files as needed — you do not have to attach or paste anything. Let it look around first:

It will also answer questions about itself, which is the fastest way to learn what it can do:

Making your first change

Describe the outcome you want in plain language.

Claude Code will find the right file, show you the proposed change, ask for approval, and then make the edit. Whether it asks each time depends on your permission mode.

Permission modes — Shift + Tab

Press Shift + Tab to cycle. This is the control that decides how much rope Claude has.

Git without learning Git

Version control is worth having on a research project — it is how you undo a bad afternoon. You can run it conversationally:

Four habits worth forming immediately

Shell commands · run from PowerShell

Keyboard

The dozen slash commands you will actually use

Worth knowing about

'claude' is not recognized

The installer worked, but the folder it installed into is not on your PATH, so Windows cannot find the program. Run these two lines in PowerShell:

$currentPath = [Environment]::GetEnvironmentVariable('PATH', 'User')
[Environment]::SetEnvironmentVariable('PATH', "$currentPath;$env:USERPROFILE\.local\bin", 'User')

Then close PowerShell entirely and open a new window — PATH changes do not apply to already‑open terminals. Try claude again.

'irm' is not recognized as an internal or external command

You are in CMD, not PowerShell. Close the window and open PowerShell instead (Win + X → Windows PowerShell), or use the CMD form of the installer:

curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd

The token '&&' is not a valid statement separator

The mirror image of the previous error: you ran the CMD command in PowerShell. Use irm https://claude.ai/install.ps1 | iex instead.

Could not create SSL/TLS secure channel

Usually an older Windows 10 system defaulting to an outdated TLS version. Run this first, in the same window, then retry:

[Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12
irm https://claude.ai/install.ps1 | iex

Claude Code does not support 32-bit Windows

On a 64‑bit machine this almost always means you opened Windows PowerShell (x86), which runs as a 32‑bit process. Close it, open the Start menu entry without (x86) in the name, and run the installer again.

Claude Code on Windows requires either Git for Windows (for bash) or PowerShell

Neither shell was found, and Claude Code needs at least one. Work through these in order:

1. Make sure powershell.exe is on your PATH. Its default location is C:\Windows\System32\WindowsPowerShell\v1.0\. Alternatively install PowerShell 7, which provides pwsh.
2. Or install Git for Windows to get Git Bash.
3. If Git is installed but Claude Code cannot find it, point at it explicitly:

$env:CLAUDE_CODE_GIT_BASH_PATH="C:\Program Files\Git\bin\bash.exe"

That lasts only for the current window. To make it permanent, put it in settings.json as shown on sheet W‑1. If your Git lives elsewhere, find it with Get-Command git | Select-Object Source and use that Git\bin folder.

syntax error near unexpected token '<', or HTML in your terminal

The install URL returned a web page instead of the installer script. If that page says App unavailable in region, Claude Code is not available in your country — check Anthropic's supported countries list. Otherwise just run the command again; transient failures happen.

Login fails, or there is no Claude Code option

Search or file discovery fails

Claude Code normally bundles ripgrep. If searching your project silently returns nothing, that is the usual culprit — check the official search troubleshooting page.

Everything worked, then a command hangs forever

You have most likely triggered something waiting for input that Claude cannot see — an interactive prompt, a pager, or a confirmation dialog. Press Esc to interrupt. When asking Claude to run a tool, prefer non‑interactive flags.

Set the project up once

Start in your project folder and run /init. Then tell it the things it cannot guess, and they will persist across every future session:

Stage by stage

Setting up the Python environment

This is the part that historically eats a weekend on Windows, because the geospatial stack has awkward binary dependencies. Ask for it explicitly:

Georeferencing the historical quads

Note the last sentence. You need the error figure for your limitations section, so ask for it while you are doing the work rather than reconstructing it later.

Flow accumulation

Defining ghost reaches

The matched comparison

Spatial statistics

Figures

Where you must not trust it

上面这张图是电脑合成的，但方法是真的。

不管上面盖了什么，水还是往低处流。所以你可以这样做：拿激光测出来的地面高度数据，让电脑算「水会往哪里流」。它就会画出这座城市一定有过的那些河——包括今天任何地图上都找不到的那些。因为一百年前有人把它们塞进了地下管道，然后在上面盖了楼。

这里面有什么

现在做到哪一步了

大部分人第一个想法是：我要找一个全世界都没人问过的问题。

这个想法错了两次。第一，完全没人碰过的问题非常少。第二，你能找到的那些，之所以没人碰，通常就是因为用免费数据加一台笔记本根本做不出来。

真正能获奖、能发表的，是一个新的组合：

（很少人用的数据）×（很少人研究的地方）×（有人真的想知道答案的问题） 三个条件

这三个条件各有用处：

• 数据这一项，保证别人抢不了先。
• 地方这一项，保证结果是你的，不是重复别人。
• 有人在乎这一项，学生最容易忘掉。但评委最看重的就是这个。

第三项最好用的玩法：查账

有人公开说自己做了一件环保好事。比如：

• 某公司说它花钱保护了一片森林
• 某部门说它发钱让农民种了保护土壤的作物
• 某政府说它把漏气的废井都封好了

而你，用卫星去查：他说的是真的吗？

这个思路对学生特别有威力。原因跟你的分析有多聪明没关系：

• 没人能说你的题目不新。在你之前，没有人查过这一件具体的事。这一点没法争。
• 结论真的有用。「这笔钱到底买到了它承诺的东西吗」——这是你们学校以外的人也会关心的问题。

决定成败的两件事

一、动手写代码之前，先查有没有人做过

给这件事整整一个星期。

去 Google Scholar 搜「你的数据 + 你的地区 + 你的问题」。然后把最接近的三篇论文的参考文献全部看一遍。抢先做了你这个题目的那篇论文，往往就藏在那里面。

另外还要搜两个地方：

• 当地的流域保护协会（这种组织常有自己的报告）
• 附近大学环境科学系的网页

因为这类研究有很多是没正式发表的报告和硕士论文。Google Scholar 收不全它们。

二、一定要有对照组，一定要给误差范围

学生做遥感项目最常见的失败是：画出一条看起来很有说服力的曲线，但没有任何东西可以拿来对比。

举个例子。被埋掉的河道确实比露天的河道热。但被埋的河道上面往往是停车场——水泥地本来就热。

所以你必须去找一些露天的河道，它们的水泥地比例差不多、坡度差不多、上游集水面积也差不多，然后拿这些来比。这叫配对。

不做配对的话，你什么都没发现。你只是重新发现了「城市铺了水泥」。

有配对的对照组、有误差范围、跟地面实测数据核对过——这就是「论文」和「海报」的全部区别。

要抓住的那一幕

《千与千寻》里最重的环境细节，不是腐烂神，是白龙。

白龙记不起自己的名字，所以他走不出汤屋。最后是千寻帮他想起来的：他是琥珀川的河神。就是千寻小时候掉进去的那条河。那条河没有淹死她，反而把她推到了岸上。

他为什么会丢掉名字？电影里说得很直白：那条河被填平了，上面盖起了公寓楼。

一条河被铺平了。它的神忘了自己是谁。没有人记得它曾经在那里。

这两件事，现实里都有名字

第一件：河道填埋

全世界的城市都干过同一件事：把小河塞进地下管道和水泥渠，然后在上面盖房子。

被埋得最多的是源头小溪。原因有两个：它们加起来占了河流总长度里最大的一份；而且它们最细，埋起来最便宜。这是 Elmore 和 Kaushal 2008 年那篇论文的核心发现，也是这个领域最基础的一篇。

埋掉之后会发生什么，已经有研究记录下来了：

• 自然河道被毁掉
• 下游的生物栖息地变差
• 水里的生物被切断，走不通了
• 雨水和有毒污染物冲得更快
• 河道原本能留住的养分和泥沙，留不住了

第二件：基线漂移

这一件解释了「为什么没人发现」。

Soga 和 Gaston 在 2018 年给它下了定义。简单说就是：因为你没见过环境以前的样子，你就以为它现在这样是正常的。一代一代往下滑，每一代都觉得自己看到的是「本来的样子」。

它还有个更好懂的名字叫「环境代际失忆」。最早是渔业科学家 Daniel Pauly 在 1995 年提出来的。

这个领域自己最爱用的例子就是河：

上一代人小时候在清澈的小河里游泳、用玻璃瓶捞小鱼。今天的小孩看到的是一条浑浊的水泥渠，然后觉得——河本来就是这样的啊。

白龙忘掉自己的名字，就是基线漂移。 整个项目立在这一句上

这不是硬套在电影上的比喻。它是一个虚构设定和一个真实科学概念的精确对应。也正因为这样，这个框架读起来才像想法，不像装饰。

要回答的四个问题

具体怎么做

1. 先拿老地图。美国地质调查局有个免费网站叫 topoView，上面有一百年前的地形图。把老地图跟今天的坐标对上（这一步叫「配准」），然后把上面的蓝线一条条描出来——那些就是以前画在图上的河。
2. 再算水往哪流。拿 3DEP 的激光测高数据，让电脑算「水会汇到哪里」。用 Python 的 richdem 或 whitebox，也可以直接用 QGIS 这个免费软件。这一步只看地面的形状，就能告诉你河应该在哪。
3. 然后判断哪一段被埋了。三个条件同时成立就算：地形说这里有河 + 老地图说这里有河 + 现在的官方水系图（叫 NHD）什么都没有。
   这个做法是跟着 Elmore 和 Kaushal 走的，所以是站得住脚的，不是你自己发明的。
4. 再把别的信息叠上去。人口普查的收入和族裔、以前的红线区等级、水泥地和树冠比例、市政投诉数据、还有卫星测的地表温度。
5. 最后跑统计。把被埋的河段和露天的河段配对来比（按水泥地比例、坡度、上游面积配对）。然后检查一件事：挨着的地方会不会因为挨着而长得像？这个检查叫 Moran's I。如果有，就换成专门处理这种情况的模型。

空白在哪里

避开巴尔的摩和底特律。你自己的城市既更新，也是更好的故事。

另外提醒一句：巴尔的摩有个公共艺术项目就叫「Ghost Rivers」，讲的是被埋掉的 Sumwalt Run。这是很好的引用例子，但也说明那个名字在当地被占了。

做得出来吗？做得出来

要用的东西：QGIS（免费）加 Python 的几个库（geopandas、richdem 或 whitebox）。数据全免费，而且不像 H‑2 那个项目那样有取数据的麻烦。

把老地图一张张对上坐标，确实很枯燥。但这恰好是一个优点。这是看得见的、老实的劳动，评审和评委都认这个。而且这一部分没法让机器替你做。

备选方案：《清理河流的半衰期》

如果你更想把手伸进水里，而不是伸进电脑，那就抓另一幕。这一幕跟宫崎骏本人的关系最直接。

河神变成臭气熏天的「腐烂神」来到汤屋。千寻发现它身上插着一根刺，刺一拔出来，垃圾像洪水一样喷出来——里面还有一辆自行车。

宫崎骏本人确认过：这来自他在家乡清理一条脏河的经历，当时大家真的从泥里拖出了一辆自行车。他说：「我清理过一条河。」

真正没人好好回答的问题：清一次能干净多久？

志愿者清河这件事到处都有，钱也不少，做完还很有成就感。但是——垃圾多快会回来？几乎没有人测过。

对一件这么普遍的事来说，这是个惊人的空白。

怎么做：

1. 在本地一条城市小河上选 6 到 10 段。
2. 先数一遍垃圾。要有固定的样带，按材料分类计数，照着一个现成的规范做。
3. 把每一段彻底清干净。
4. 之后按固定时间回去重新数：第 2、4、8、16 周。
5. 画出「垃圾回来」的曲线，算出一个半衰期（也就是垃圾回到一半要多久）。
6. 再看什么因素决定它回得快：上游水泥地多不多、这期间下了几场雨、离马路多近、旁边是什么用途的地、附近有没有排水口。

这个研究能给出一个带误差范围的真数字，成本很低，而且能直接给出一条建议：

这里清一次能维持三个月，那里只能维持三个星期。所以拦垃圾的装置应该装在第二个地方。

大学申请文书可以怎么写

1. 你注意到的是一个很具体的东西。不是「这部电影讲自然」，而是一个因为自己的河被铺平、所以丢了名字的神。
2. 你去查了，发现现实里这件事有名字：河道填埋。而「忘记」也有名字：基线漂移。
3. 你开始找你自己城市的幽灵河流。对一百年前的地图，跑水流计算。
4. 你找到了。而且你还查出了是谁失去了它们。
5. 白龙拿回名字，得到自由。你把你的城市的河的名字还给了它们。

本来的点子

几乎所有研究城市热岛的人，用的都是 Landsat 卫星。而这颗卫星大约在上午十点半经过。

但热不是在上午十点半害死人的。

真正危险的是那种一整晚都降不下温的夜晚。人的身体一夜都没得到休息，第二天接着热。

所以问题是：什么样的街区晚上降温慢？这跟收入或者以前的红线区有关系吗？

能做这件事的仪器叫 ECOSTRESS，装在国际空间站上，专门测温度。

它有个特别之处：因为空间站的轨道会慢慢转，它每次经过同一个地方的时间都不一样。平时这算个麻烦，但在这里正好是关键——它能拍到夜里。

我们查到第四条时发现的东西

那还剩下什么？

剩两样，都很窄，但都是真的。

第一：所有人都在测温度有多高，没人测降温有多快。

这是这个领域一贯的做法——在某个时刻测一下温度，然后比较不同社区。

但真正有意思的是降温速度：太阳下山之后，一个地方多快能把热量放掉？

要算这个，你得把卫星那些零散的观测时刻拟合成一条曲线（白天一段、夜里一段），然后看夜里那一段掉得多快。

为什么这是个不同的问题？因为背后的物理不一样：

• 白天有多热，取决于晒进来多少：地面颜色深浅、有没有树荫、水泥地多不多。
• 晚上降得多快，取决于两件事。一是建筑材料攒了多少热（水泥和砖攒热特别厉害）。二是这块地面能看到多少天空——热量是往天上散的，如果四周都是高楼，热就散不出去。

所以一条又窄又高的街道，不管你把墙刷得多白，晚上都凉不下来。

第二：那篇论文只做了一座城市。

作者自己在论文里就写了：洛杉矶的位置比较特殊，结论不一定能推到美国其他城市。

所以做一个多城市的版本，看看这个规律在不同城市之间是不是一样——这是一个真的不一样的贡献。

拿数据时的一个意外

第一周就该做的判断：这个题目到底能不能做

投入之前只查一件事：你想研究的那座城市上方，到底有多少张能用的夜间 ECOSTRESS 照片？

如果夜间的观测太少，或者时间点太集中，你就画不出那条降温曲线，整个项目就不成立。

这件事你要在第一周就知道，不要等到第三个月。

另外注意：这颗仪器只覆盖大约南北纬 52 度之间。

查账 · 有人说自己做了好事，你去查

盲区 · 没人看，因为老卫星看不见

测错了东西 · 大家都去测那个容易测的

要下野外 · 靠雨靴，不靠代码

按这个顺序做

1. 1

   先定研究哪座城市

   后面每一步都得等这个——查重、查数据、还有故事本身。

   你要找的城市：有真的河、有公开的市政投诉数据、有激光测高数据、有 1900 年左右的老地图，而且跟你本人有关系。

   避开巴尔的摩和底特律。

2. 2

   查有没有人做过 —— 留一个星期

   在 Google Scholar 上搜 "buried stream" OR "stream burial" OR daylighting + [城市名]。

   然后再搜这座城市的流域保护协会，还有最近那所大学的环境科学系。因为这类研究很多是没正式发表的报告和硕士论文。

   最后把最接近的三篇的参考文献看完。

3. 3

   查数据能不能拿到 —— 一个下午

   • USGS topoView —— 看有没有你那个城市的老地图，最早到哪一年
   • USGS 3DEP —— 看有没有激光测高数据，精度多少
   • 市政数据网站 —— 看有没有「积水」这类投诉，而且带地址坐标
   • NHD —— 下载现在的官方水系图
   • Mapping Inequality —— 看这座城市有没有老的红线区地图
4. 4

   写正式的研究方案

   要包括：你的假设（写成「如果错了，什么证据能推翻它」的样子）、方法、统计怎么做、时间表。

   统计方案要在看到结果之前就定下来。不然你很容易不知不觉地挑一个「结果好看」的算法。

   另外，这份文件本身就是你竞赛报名材料的大部分内容。

5. 5

   先在一个小流域上试一遍

   不要一上来就铺开整座城市。

   先对一张地图、在一个小流域上跑一次水流计算、把一条幽灵河段从头到尾走通。

   你会发现有三件事比你想的难——而这时候改还很便宜。

6. 6

   铺开，然后写

   分析一稳定就把稿子传到 EarthArXiv。这是个免费网站，能让别人引用你，还能给你的工作打上时间戳（证明你先做的）。

   然后去查 AGU 的摘要截止日期和各个竞赛的时间。

可以投的竞赛

可以投的期刊

四条一直要记着的规矩

1. 1

   装 Git for Windows —— 可选，但建议装

   装了它，Claude Code 就能用一个叫 Bash 的工具。不装的话，它会改用 PowerShell，也能工作。

   你自己不需要学 Git。装它只是为了让 Claude Code 多一个工具。

   1. 打开 git-scm.com/downloads/win 下载。
   2. 运行安装程序，每一屏都点「下一步」。屏数很多，但没有一屏需要你改东西。
   3. 如果问你选编辑器，保持默认。
   4. 看到「Adjusting your PATH environment」那一屏，保持它推荐的选项。

   已经装过、或者不确定装没装？再装一次也没关系，不会出问题。

2. 2

   打开 PowerShell

   PowerShell 是 Windows 自带的命令行，每台电脑上都有。

   按 Win + X，在菜单里选 Windows PowerShell 或者 终端。会出现一个黑窗口，里面有个闪烁的光标。

   PowerShell 还是 CMD？这个一定要分清

   Windows 有两个命令行程序，长得很像，但吃的命令不一样。看每行开头就能分：

   • PowerShell：显示 PS C:\Users\你的名字>，前面有 PS。
   • CMD：显示 C:\Users\你的名字>，没有 PS。

   别选带 (x86) 的那个

   如果开始菜单里有 Windows PowerShell (x86)，不要用。那个是 32 位版本，安装会失败，报错说不支持 32 位 Windows。

   选名字里没有 (x86) 的那一个。

3. 3

   运行安装命令

   复制下面这一行，用 Ctrl + V 或者右键粘贴到 PowerShell 里，然后按回车。

   PowerShell复制

   irm https://claude.ai/install.ps1 | iex

   这行命令的意思是：irm 把安装脚本下载下来，iex 把它跑起来。

   你会看到一堆文字往上滚。装完会显示 Claude Code successfully installed!

   如果你开的是 CMD 不是 PowerShell，用这一行：

   CMD复制

   curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd

   两个报错，意思都是「你开错窗口了」

   • 看到 The token '&&' is not a valid statement separator → 你在 PowerShell 里跑了 CMD 的命令。换成上面 PowerShell 那条。
   • 看到 'irm' is not recognized as an internal or external command → 你在 CMD 里跑了 PowerShell 的命令。换成 CMD 那条，或者去开 PowerShell。

   这样装的版本会自己在后台更新，所以安装命令你只跑这一次就够了。

4. 4

   检查装好了没有

   PowerShell复制

   claude --version

   装好的话，它会打印一个版本号，后面跟着 (Claude Code)，比如 2.1.211 (Claude Code)。

   想查得更全，用下面这条。它会告诉你安装有没有问题、配置文件有没有写错，还会给修复建议。而且它不会启动对话。

   PowerShell复制

   claude doctor

   如果出现 'claude' is not recognized，说明装好了但 Windows 不知道去哪找它。解决办法在 第 W‑4 页。

5. 5

   登录

   先切到你要工作的文件夹，再启动。在命令行里，cd 就是「进入某个文件夹」的意思：

   PowerShell复制

   cd C:\Users\YourName\Documents\ghost-rivers
   claude

   第一次运行会让你登录，并且自动打开浏览器。登录过一次之后就记住了，以后不用再登。

   以后想换账号，在对话里输入 /login。

其他安装方式

直接在 Windows 上跑，还是在 WSL 里跑？

WSL 是 Windows 里的一个 Linux 环境。你两种都可以选。

对一个跑在 Windows 笔记本上的学生项目，直接在 Windows 上跑就对了。

因为你的 QGIS、你的 Python、你下载的地图文件，全都在 Windows 这一边。跨到 WSL 那一边去读文件既慢又容易搞混。

如果你还是要用 WSL：安装和启动 claude 都要在 WSL 的窗口里面做，不是在 PowerShell 里。而且用 WSL 就不需要装 Git for Windows 了。

告诉它 Git Bash 在哪

装了 Git for Windows 之后，Claude Code 会自己去找 Git Bash。如果找不到，就在这个文件里告诉它：%USERPROFILE%\.claude\settings.json

{
  "env": {
    "CLAUDE_CODE_GIT_BASH_PATH": "C:\\Program Files\\Git\\bin\\bash.exe"
  }
}

注意那两道斜杠 \\。这种文件里必须写两道，写一道会出错。

想知道你的 Git 装在哪，跑这个：

Get-Command git | Select-Object Source

怎么卸载

如果你是用官方脚本装的：

Remove-Item -Path "$env:USERPROFILE\.local\bin\claude.exe" -Force
Remove-Item -Path "$env:USERPROFILE\.local\share\claude" -Recurse -Force

用 WinGet 装的：winget uninstall Anthropic.ClaudeCode

用 npm 装的：npm uninstall -g @anthropic-ai/claude-code

关于命令行的六件事

• 什么都点不了。用方向键移动。
• Esc 可以把它打断。这是最有用的一个键。它一开始走错方向，就按 Esc 停掉，别等它做完。
• Enter 发送。用大白话打字就行。
• ↑ 调出上一条，Tab 自动补全。
• 打一个 / 就会列出所有能用的命令。
• 打 exit，或者在空行上按两次 Ctrl + D，就退出。

启动之后长这样

PS C:\Users\You> cd Documents\ghost-rivers
PS C:\Users\You\Documents\ghost-rivers> claude

  ╭──────────────────────────────────────────╮
  │  Claude Code v2.1.211                    │
  │  model: claude-opus-5                    │
  │  cwd:   C:\Users\You\Documents\...       │
  ╰──────────────────────────────────────────╯

  /help for commands · /resume to continue a previous conversation
> 

上面那个框里显示：版本、现在用的模型、还有当前文件夹（cwd）。

最后那一项很关键：它只能看到你启动它时所在的那个文件夹，以及里面的子文件夹。在错的地方启动，它就找不到你的文件。

先让它了解项目，再让它改东西

它会自己去读你的文件，你不用复制粘贴给它。先让它转一圈：

它也会回答关于它自己的问题。这是了解它能干什么最快的办法：

第一次让它改东西

用大白话说你想要什么结果。

它会：找到该改的文件 → 把改动给你看 → 问你同不同意 → 然后才动手。

会不会每次都问你，取决于下面这个设置。

权限模式：按 Shift + Tab 切换

这个设置决定「给它多大自由」。按 Shift + Tab 循环切换。

不用学 Git 也能用 Git

Git 是记录「你什么时候改了什么」的工具。研究项目值得用它，因为它就是你的「撤销一整个糟糕下午」按钮。

你可以直接用说话的方式让它做：

值得马上养成的四个习惯

在 PowerShell 里打的命令

快捷键

你真的会用到的那十几个

其他值得知道的

'claude' is not recognized

意思是：装是装好了，但 Windows 不知道去哪找它。

在 PowerShell 里跑这两行：

$currentPath = [Environment]::GetEnvironmentVariable('PATH', 'User')
[Environment]::SetEnvironmentVariable('PATH', "$currentPath;$env:USERPROFILE\.local\bin", 'User')

然后把 PowerShell 整个关掉，重新开一个新窗口。这一步不能省——已经开着的窗口不会知道路径变了。

再试一次 claude。

'irm' is not recognized as an internal or external command

意思是：你在 CMD 里，不是 PowerShell。

关掉这个窗口，按 Win + X 开 PowerShell。或者用 CMD 版本的命令：

curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd

The token '&&' is not a valid statement separator

意思是：正好反过来——你在 PowerShell 里跑了 CMD 的命令。

改用 irm https://claude.ai/install.ps1 | iex。

Could not create SSL/TLS secure channel

意思是：你的 Windows 比较老，用的加密方式过时了，连不上服务器。

先在同一个窗口里跑第一行，然后再跑安装：

[Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12
irm https://claude.ai/install.ps1 | iex

Claude Code does not support 32-bit Windows

意思是：如果你的电脑是 64 位的（现在基本都是），那这个报错几乎肯定是因为你开的是 Windows PowerShell (x86)。

关掉它，从开始菜单里开那个名字不带 (x86) 的，重新跑一次。

Claude Code on Windows requires either Git for Windows (for bash) or PowerShell

意思是：它一个能用的命令行工具都没找到，但它至少需要一个。

按顺序试：

1. 确认 powershell.exe 在系统路径里。它一般在 C:\Windows\System32\WindowsPowerShell\v1.0\。或者去装 PowerShell 7。
2. 或者装 Git for Windows，这样就有 Git Bash 了。
3. 如果 Git 已经装了但它找不到，就直接告诉它位置：

$env:CLAUDE_CODE_GIT_BASH_PATH="C:\Program Files\Git\bin\bash.exe"

但这一行只对当前窗口有效。想让它永久生效，就按 第 W‑1 页的办法写进 settings.json。

如果你的 Git 装在别的地方，用 Get-Command git | Select-Object Source 找出来。

屏幕上出现一堆 HTML 代码

意思是：下载地址返回的是一个网页，不是安装脚本。

如果那个网页写着 App unavailable in region，说明 Claude Code 在你所在的国家还不能用。

如果不是，就再跑一次命令。偶尔会失败一次。

登录失败，或者根本没有 Claude Code 这个选项

它搜不到你的文件

Claude Code 自带一个搜索工具叫 ripgrep。如果在你的项目里搜索总是什么都搜不到，一般就是这个东西出问题了。去查官方的搜索排错页面。

本来都好，突然卡住不动了

意思是：很可能有个程序在等你输入什么，但那个提示 Claude 看不见。

按 Esc 打断它。

以后让它跑工具的时候，尽量让它加上「不需要交互」的参数。

项目开头设置一次就好

在项目文件夹里启动，跑一次 /init。然后把它猜不到的事告诉它。这些以后每次都生效：

一步一步来

第一步：装 Python 环境

这一步历来最能吃时间，因为处理地图的那几个库特别难装。所以要说清楚你要什么：

第二步：把老地图对上坐标

注意最后一句。你的论文里要写「局限」这一节，那时候需要这个误差数字。所以做的时候就问出来，不要等到最后再回头补。

第三步：算水往哪流

第四步：判断哪些河段被埋了

第五步：配对比较

第六步：统计

第七步：画图

哪些地方不能直接信它