科技版塊--如何統計全球太陽能裝置數目(下)
That is where the second technological trend comes in.Dr Kruitwagen and his colleagues trained a machine-learning system to spot the solar panels for them.Computer vision is a hot field.But the specifics of orbital reconnaissance meant that off-the-shelf software was not suitable for the task the researchers had in mind.Machine-learning systems are taught what to do by examining a "training set", which contains examples of what is being searched for.For common tasks such as facial recognition, pre-built training sets are often available.But Dr Kruitwagen's team had to build their own.For this, they turned to OpenStreetMap, an open-source rival to Google Maps in which volunteers had already tagged large numbers of solar plants.But there was little consistency.
這就體現了第二種技術趨勢的用武之地。克魯特瓦根博士和他的同事開發了一個機器學習系統來為他們識別太陽能電池板。計算機視覺是一個熱門領域。但軌道偵察的特殊性意味着現成的軟件不一定能完成研究人員心目中的任務。機器學習系統通過檢查「訓練集」來學習該做什麼，「訓練集」包含正在搜索的內容的示例。對於面部識別等常見任務，通常可以使用預先構建的訓練集。但是克魯特瓦根博士的團隊必須建立他們自己的「訓練集」。為此，他們求助於OpenStreetMap——谷歌地圖的開源競爭對手，志願者已經在這個平台上標記了大量的太陽能發電廠。但幾乎沒什麼一致性。
"Some people had just drawn rough outlines around an entire field,"Dr Kruitwagen says."Others had gone in and traced the outline of each row of panels separately."Fixing that involved a great deal of manual labour.Once the training data had been cleaned up, the learning algorithms had to be tweaked as well.From space, even big solar installations look small.Each pixel in the Sentinel images represented a ten-by-ten-metre square.Even for the higher-resolution SPOT satellites, the squares』 sides are one and a half metres long.Existing classifiers, trained for things like facial recognition or self-driving cars, are used to spotting objects that loom large in their field of vision.Hunting for smaller ones meant tinkering with the software to boost its ability to detect tiny features.
克魯特瓦根博士說：「有些人只是粗略地勾勒出了整個領域的輪廓。」「其他人走進去，分別畫出了每排嵌板的輪廓。」要解決這個問題，需要大量的體力勞動。一旦訓練數據被清理乾淨，學習算法也必須進行調整。從太空上看，即使是大型的太陽能裝置也看起來很小。「哨兵」號圖像中的每個像素代表一個10*10米的正方形。即使是高分辨率的SPOT衛星，正方形的邊長也有1.5米。現有的分類器受過面部識別或自動駕駛汽車等方面的訓練，可用於發現在他們的視野中逼近的物體。尋找較小物體意味着對軟件進行升級，以提高其檢測微小特徵的能力。
False positives—things like tennis courts and agricultural greenhouses that resemble solar panels from space—had to be removed.Though extraordinary, Dr Kruitwagen’s results are already out of date.The data-gathering phase of the project ended in 2018, meaning that the thousands of new plants built since then are not included.But the project, he says, proves that the method works.He intends to make his results, including the labour-intensive training set, available for others to use.One logical extension of his project, he says, would be to expand the analysis to include solar panels installed on domestic rooftops.
誤報信息——比如網球場和農業大棚，從太空看起來像太陽能電池板——必須被移除。儘管克魯特瓦根博士的研究成果非同尋常，但他已經過時了。該項目的數據收集階段於2018年結束，這意味着自那以後建造的數千座新裝置不包括在內。但他說，這個項目證明了這種方法是有效的。他打算將他的成果，包括勞動密集的「訓練集」，提供給其他人使用。他說，他的項目的一個合理延伸是擴大分析範圍，包括安裝在家庭屋頂上的太陽能電池板。
Such 「behind-the-meter」 installations are particularly tricky to track in other ways.More generally, Dr Kruitwagen hopes that his eye-in-the-sky approach—which, despite the planetary scale of the project, cost only around $15,000 in cloud-computing time—could presage more accurate estimates of other bits of climate-related infrastructure, such as fossil-fuel power stations, cement plants and terminals for ships carrying liquefied natural gas.The eventual result could be the assembly of a publicly available, computer-generated inventory of every significant bit of energy infrastructure on Earth.Quite apart from such a model's commercial and academic value, he says, an informed public would be one better able to hold politicians』 feet to the fire.
這樣的「表後」裝置在其他方面尤其難以追蹤。更廣泛地說，克魯特瓦根博士希望他的「天空之眼」方法——儘管該項目是規模龐大，但在雲計算時間上的花費僅為15000美元——能夠更準確地估計其他與氣候相關的基礎設施，如化石燃料發電站，水泥廠和運輸液化天然氣的船舶碼頭。最終的結果可能是列出一份公開可用的、由計算機生成的地球上每一個重要能源基礎設施的清單。他說，撇開這種模式的商業和學術價值不說，一個知情的公眾更能讓政客們按捺不住。
留言列表