TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Show HN: Weather data created by phones for machine learning classifier training

7 点作者 jacobsheehy超过 6 年前

1 comment

jacobsheehy超过 6 年前
I&#x27;m building &#x27;virtual weather stations&#x27; out of disparate, unused data sources.<p>This effort is to create new data sources that describe our atmosphere. I don&#x27;t know if any of it will be useful but I cannot find much else out there that is working towards this kind of effort. So here we are.<p>The main experiments I&#x27;m working on are 1) sky photos, 2) sensor data like barometers in phones, and 3) user-submitted current weather.<p>1) Sky photos. I think that we can use a machine learning classifier that is built from user-submitted labelled photos of the sky. We should be able to them auto-tag weather data in the photos and get (numerically?) useful historical weather data from all outdoor sky photos. Additionally live photos could used to gain real-time information as well. Smartphone cameras are getting really good and can detect significant detail.<p>2) Most phones have a barometer now, other sensors as well like hygrometers etc. This data needs major QC and protection for user privacy and for preparation for integration and assimilation into models like WRF. This is a ways off but research has been done already here and there can be significant progress with good quality data. Here is what data from a single phone in Colorado looks like spanning a day or so: <a href="https:&#x2F;&#x2F;imgur.com&#x2F;a&#x2F;45lTZ6j" rel="nofollow">https:&#x2F;&#x2F;imgur.com&#x2F;a&#x2F;45lTZ6j</a>. The sensor code for Android is open source to encourage other apps and make the process of collecting atmosphere data from smartphones much easier. [2]<p>3) Current conditions reporting. I found that MADIS stations don&#x27;t update frequently enough to keep track of rapidly changing weather events. Users can submit current conditions and an algorithm picks which to show based on recency, trust, and other metrics. So far it can create a more accurate current conditions experience for the user.<p>I have taken significant effort to show the weather data from NOAA in a clear and readable way. The large graphics and animations have significant logic code behind them to attempt to most appropriately display the conditions described by MADIS stations. For example, even when most APIs will return a thunderstorm icon for a 30% chance (I find that misleading) All Clear takes an approach of showing a faded flashing bolt with a clear indication that it is &#x27;chance&#x27;.<p>Other efforts are made to improve weather data communication, such as automatically converting text forecasts to units set by the user. It&#x27;s not simple! But it is a useful thing for users. (See this forecast API result:<a href="https:&#x2F;&#x2F;api.weather.gov&#x2F;gridpoints&#x2F;OKX&#x2F;33,37&#x2F;forecast&#x2F;" rel="nofollow">https:&#x2F;&#x2F;api.weather.gov&#x2F;gridpoints&#x2F;OKX&#x2F;33,37&#x2F;forecast&#x2F;</a> for example). It has text forecast only in US units and I don&#x27;t see any options to change it otherwise.<p>Thanks!<p>[2] <a href="https:&#x2F;&#x2F;github.com&#x2F;JacobSheehy&#x2F;AllClearSensorLibrary" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;JacobSheehy&#x2F;AllClearSensorLibrary</a>