AI in Space in High Detail Speeding up the Processing of Space Images

太空AI高处见微加速太空图像处理

Much of the information that is beamed back from space is useless. Pictures taken by satellites orbiting the Earth might take days to download, only to show lots of cloud obscuring the area of interest. The subject matter may also be surrounded by irrelevant information. All this uses up a lot of valuable bandwidth.

从太空传回的信息大多是无用的。在地球轨道上运行的卫星所拍摄的照片可能要花好几天才能下载回地面，结果却只看到目标区域被重重云层遮盖。主要内容还可能伴随着大量无关紧要的信息。这一切耗费了大量宝贵的带宽。

Processing data in space, before transmission, would reduce clutter, but this can be tricky. Cosmic rays randomly flip the ones and zeroes that computers operate on, introducing unpredictable errors. High levels of radiation can also damage electronic circuits. KP Labs, based in Gliwice, Poland, is building a satellite to overcome some of these problems. Their device, called Intuition-1, is controlled by a neural network, a form of artificial intelligence modelled on the human brain. The satellite is what is known in the trade as a 6U CubeSat, which means it is composed of six standard-sized 10x10x11.5cm modules.

传输前在太空先行处理数据可以减少无用信息，但操作起来有难度。宇宙射线会导致计算机二进制信息的1和0发生随机翻转，导致不可预测的错误。强辐射也可能损坏电路。位于波兰格利维采（Gliwice）的KP实验室（KP Labs）正在打造一颗卫星来克服其中一些问题。这颗卫星名为“直觉一号”（Intuition-1），由一个神经网络（一种模拟人脑的人工智能形态）控制。按业内的说法，这是一种6U立方卫星（6U CubeSat）——由六个10x10x11.5cm的标准尺寸模块组成。

Intuition-1 will be equipped with a hyperspectral imager, which takes 150 pictures of every scene it looks at. Each picture is at a different spectral frequency, so contains different information. The neural network stitches these together using powerful graphics chips hardened against radiation. The developers have also built error correction into their software.

“直觉一号”将配备一台高光谱成像仪，对每一目标场景拍摄150张照片，每张照片的光谱频率不同，因而包含不同信息。卫星内的神经网络运用经抗辐射加固的图形芯片将照片拼合起来。开发人员还在软件中加入了纠错功能。

Intuition-1 will view a 15km-wide swathe of Earth at a resolution of 25 metres per pixel. This will be able to reveal details such as how well crops are growing or allow the number of trees in a forest to be counted.

“直觉一号”将以每像素25米的分辨率拍摄15公里范围内的地球图像。这能够显示诸如作物长势或森林树木数量等细节。

But instead of transmitting back every last bit of image data, the satellite will summarise what the user requests as useful information. This might, for instance, be a heat-map showing areas of weeds in a field or the location of a forest fire. Reducing the data load means that some of this information can be transmitted live.

但该卫星并不会把图像数据巨细靡遗地传回地球，而是提炼出用户所请求的有用信息，例如显示农田内的杂草区域或森林火灾位置的热点图。数据量减少意味着一些信息可被实时传输。

The satellite will be used to prove that a hardened neural network can survive in space. This could pave the way for other space applications. For example, the Curiosity rover on Mars was successfully upgraded in 2016 with a set of algorithms to detect “interesting” rocks for investigation, instead of picking them randomly. A neural network could provide future rovers and deep-space probes with a better ability to make decisions.

该卫星将用于证明一个经加固的神经网络能在太空持续使用。这可以为其他太空应用铺平道路。例如，火星上的好奇号探测车在2016年成功升级，运用一套算法来挑选“有趣”的岩石进行研究，而不再是随机选取。神经网络可为未来的星球探测车和深空探测器提供更好的决断力。

The neural network and hyperspectral imager have already been built and tested by KP labs. The kit will go into a satellite body being constructed by Clyde Space, a satellite producer based in Scotland, and launched in 2022. After that there will be more intelligence in space.

KP实验室已制造并测试了上述神经网络及高光谱成像仪。该套件将安装到由苏格兰卫星制造商Clyde Space制造的一颗卫星上，于2022年发射。往后，太空将变得更智能。

英文、中文版本下载：http://www.yingyushijie.com/shop/source/detail/id/1683.html