農業真正的農業掌握在AI手中

05-03

The Future of Humanitys Food Supply Is in the Hands of AI

Author: Matt SimonMatt Simon

人類給自己出了一個難題——在人口大爆炸持續下去的同時，地球卻是不會變的更大的。當地球在2050年迎來第100億個人時，她依舊得用同樣數量的土地來養活人們，再加上全球變暖造成水源的持續減少，人類將來到底能不能吃飽將會是個大問題。

也許冥冥之中自有天意，人工智慧在這個時候及時粉墨登場。真正的智能機器人和機器學習演算法將會帶來一個綠色革命，一個將用這個星球有限的土地來餵飽人類的革命。你可以想像一下，自動探測乾旱痕迹的衛星，可以自動識別並殺死生病植物的拖拉機，以及能自動通知農場主農作物病害的AI手機。

忘掉稻草人吧，農業的未來在智能機器的手中。

數碼農業專家

A Digital Green Thumb

Deep learning is a powerful method of computing in which programmers don』t explicitly tell a computer what to do, but instead train it to recognize certain patterns. You could feed a computer photos of diseased and healthy plant leaves, labeled as such. From these it will learn what diseased and healthy leaves look like, and determine the health of new leaves on its own.

深度學習是一種威力強大的演算法，編程員不再明確地告訴電腦目標，而是訓練它去識別幾種模式。你可以給電腦提供標明健康和病變的植物葉子照片。從這些照片中，它就可以學會如何分辨健康和病變的葉子，進而判斷其他葉子是否健康。

That』s exactly what biologist David Hughes and epidemiologist Marcel Salathé did with 14 crops infected by 26 diseases. They fed a computer more than 50,000 images, and by learning on its own, the program can correctly identify 99.35 percent of the new images they throw at it。

這正是生物學家大衛·休斯（David Hughes）和流行病學家馬塞爾·賽拉瑟（Marcel Salathé）對感染了26種不同疾病的14種農作物所做的。他們給一個電腦提供了超過5萬張相片，通過自我學習，電腦從相片判斷植物是否健康的準確率達到了99.35%。

Still, those are manipulated images, with uniform lighting and backgrounds so it』s easier for the computer to make sense of the leaves. Pluck an image of a diseased plant from the Internet and feed it to the computer and the accuracy is around 30 to 40 percent.

但是，這些相片都是經過人為處理的，它們都有著統一的亮度和背景，這樣電腦的判斷會變得輕鬆得多。在辨認一張從網路上找到的生病植物的相片，電腦的準確率只有30%~40%。

Not terrible, but Hughes and Salathé hope to see this AI power their app, PlantVillage, which currently allows farmers around the world to upload a photo of their ailing plants to a forum for experts to diagnose. To smarten up the AI, theyll continue feeding it photos of diseased plants. 「More and more images from various sources, in terms of how the pictures were taken, time of year, location, and so on,」 says Salathé. 「And the algorithm can just pick up on that and learn."

這結果雖然不好，但是也不差。休斯和塞拉瑟希望把這個AI運用到他們的手機應用Plant Village上。目前，全世界的農民都可以把他們生病的作物拍照上傳到Plant Village的論壇里來判斷感染的疾病。只不過，目前幫助判斷的是人類專家。為了提高應用的智能，它將通過持續提供的生病植物的相片來學習。「我們會給它更多各種各樣的相片，例如不同拍照的手段、季節、地點或者其他的屬性」，塞拉瑟說，「這個演算法將自己學會分辨。」

This isn』t simply a matter of ferreting out infections: Plenty of other things beat plants up. 「Most diseases that hamper growers are physiological stresses, so not enough calcium or magnesium or too much salt or too much heat,」 says Hughes. 「People often think its a bacterial or fungal disease.」 Misdiagnoses can lead to farmers wasting money and time on pesticides or herbicides. In the future, AI could help farmers quickly and accurately pinpoint the problem.

他們的目標不僅僅是為了排除受到感染的植物，因為植物可以因為多種原因而得病。「大部分影響植物生長的疾病都是生理上的壓力，例如缺鈣或者缺鎂、土壤含鹽量太高，或者天氣太熱了，」休斯表示，「農民卻經常會認為這是細菌性或者真菌性的疾病。」誤診將會導致農民們在農藥上浪費金錢與時間。在未來，他們的AI將幫助農民迅速並精準地找出問題的根源。

After that, the humans will wrest back control—because while an app might be able to find the problem, only an extension expert can tailor a solution to a specific climate or soil or time of year. The UN』s Food and Agriculture Organization considers such technology a 「useful tool」 for crop management, but the expert』s word is doctrine. Thus, says Fazil Dusunceli, a plant pathologist with the FAO, such electronic results are welcome, but 「final pest management decisions should be taken in collaboration with experts on the ground.」

不過，找出問題之後的後續工作就需要人來做了。雖然一個應用可以幫助找到問題，只有人類專家才可以根據具體的氣候、土壤以及季節來找出解決問題的答案。聯合國農糧組織認為這項技術對於農作物管理是一個「有效的工具」，但是專家的意見才能作為定論。糧農組織植物病理學家法齊爾·杜桑切利（Fazil Dusunceli）表示，這類的電子判斷結果很好，但是「最終的農害管理」需要和當地的專家一起決定。

農機訓練者

Tractor Trainer

While the developing world is hungry for agricultural knowledge, the developed world is drowning in pesticides and herbicides. In the US each year, farmers use 310 million pounds of herbicide—on just corn, soy, and cotton fields. It』s the spray-and-pray approach, not so much sniping as carpet bombing.

當發展中國家急需農業知識時，發達國家正在被農藥淹沒。在美國，每年僅僅被傾撒在玉米，大豆和棉花上的農藥就有3.1億磅。這是一種「噴洒後祈禱」的方式，和狙擊式去蟲害相比，更像是地毯轟炸。

A company called Blue River Technology may have hit upon solution, at least as far as lettuce is concerned. Its LettuceBot looks like your typical tractor, but in fact it』s a machine-learning-powered … machine.

一個名為Blue River Technology的公司很有可能找到了解決這個問題的方式，起碼解決了生菜的問題。它的Lettuce Bot看起來和一般的拖拉機一樣，但是事實上它是一個基於深度學習的…機器。

Blue River claims the LettuceBot can roll through a field photographing 5,000 young plants a minute, using algorithms and machine vision to identify each sprout as lettuce or a weed. If that seems too impossibly fast to you, "its well within the computing of machine learning and computer vision," says Jeremy Howard, founder of deep-learning outfit Enlitic. A graphics chip can identify an image in just .02 seconds, he adds.

Blue River 宣稱，Lettuce Bot可以在開過一片田地時，以每分鐘5千株的速度對菜苗拍照，並通過演算法以及機器視覺來判斷每株植物是生菜還是雜草。如果你覺得它的速度快的難以想像，「這離機器學習和機器視覺的極限還遠得很」，深度學習演算法公司Enlitic的創辦人傑瑞米·霍華德（Jeremy Howard）對此說道。他還表示，圖像處理晶元只需要0.02秒就可以識別一張相片。

With an accuracy within a quarter inch, the bot pinpoints and sprays each weed on the fly. If it eyeballs a lettuce plant and determines it isn』t growing optimally, it』ll spray that too (farmers overplant lettuce by a factor of five, so they can sacrifice plenty of extras). If two sprouts ended up too close to one another during planting (not ideal), the machine can discern them from, say, one particularly large plant, and zap them as well.

Lettuce Bot的精準度可以達到1/4英寸，這代表它可以在運行過程中準確的找出每株雜草並向它們噴洒除草藥。如果它判定一株生菜苗沒有健康生長，它也會噴一下（通常情況下，農民種下的菜苗會比預期收穫的高5倍，所以他們可以犧牲許多菜苗）。如果兩株菜苗被發現長的太近（不健康），這個機器不會把它們誤認為一株大植物，而是會分別噴洒一下。

Now, consider the alternative: spraying a field with herbicides willy-nilly. 「Its akin to saying if a few people in the city of San Francisco had an infection, your only solution would be to give every man woman, and child in the city an antibiotic,」 says Ben Chostner of Blue River Technology. 「People would be cured, but its expensive, its not using the antibiotics to the best of their potential.」

現在考慮替代方法：在田間噴洒除草劑。「這就像說，如果舊金山市有幾個人感染了病毒，你唯一的解決方式就是給每人發一片抗生素」， Blue River 的本·科斯特納（Ben Chostner）對此說道，「人們會被治癒，但是治癒的成本實在是太貴了，抗生素的潛能也沒有被開發出來。」

With the LettuceBot, on the other hand, Chostner says farmers can reduce their use of chemicals by 90 percent. And the machine is already hard at work—Blue River treats fields that supply 10 percent of the lettuce in the US annually.

而通過Lettuce Bot，科斯特納表示農民可以減少90%的農藥用量。目前，此機器已經被用於市場上：供應美國10%生菜的田正在使用 Blue River 的產品。

LettuceBot is so powerful because it uses machine learning to make one of the few things robots are already great at even better: precision. Robots can』t run like us or manipulate objects quite like we do, but they』re consistent and meticulous—the perfect agricultural snipers.

Lettuce Bot之所以威力強大，是因為它通過機器學習來放大了機器自帶的優勢：精準。機器人不能像人類一樣奔跑或者操縱物體，但是它們一絲不苟，從而能成為完美的農業狙擊手。

天上的AI

Life From Above

Orbiting over 400 miles above your head, NASA』s Landsat satellites provide a downright magical survey of Earth』s surface in a slew of bandwidths far beyond the visible spectrum. All of these layers of information are hard to digest for a human, to be sure, but for machine learning algorithms, they ain』t no thing.

在我們頭上400英里的軌道上，NASA的Landsat衛星利用可見光譜之外的電磁波給人們提供了可謂是神奇的地球表面探勘數據。這些數據對於人類來說形同天書，但是對於機器學習演算法來說，它們一點難度也沒有。

And that could be extremely valuable for monitoring agriculture, particularly in developing countries, where governments and banks face a dearth of data when making decisions about which farmers they give loans or emergency assistance to. During a drought in India, for instance, not only will regions suffer to different degrees, but within those regions some farmers might have better means to procure water than others.

而這，對農業監測來說簡直就是無價之寶。尤其是在發展中國家，當政府和銀行嚴重缺乏數據來幫助他們決定哪些農民需要貸款或者救助時，這些數據將擁有者極大的影響力。就拿印度的一場旱災來說，每個地區之間缺水的程度都會不同，就算在同個地區之內，有些農民相對而言都會有更多的手段取得水。

So a startup called Harvesting is analyzing satellite data on a vast scale with machine learning, with the idea to help institutions distribute money more efficiently. 「Our hope is that in using this technology we would be able to segregate such farmers and villages and have banks or governments move dollars to the right set of people,」 says Harvesting CEO Ruchit Garg. While a human analyst can handle 10, maybe 15 variables at a time, Garg says, machine learning algorithms can handle 2,000 or more. That』s some serious context.

所以，一個名為Harvesting的創業公司正在使用機器學習分析龐大的衛星數據，意圖幫助政府部門和銀行更加有效的分配資源。Harvesting的首席執行官魯什特·加格（Ruchit Garg）對此表示：「我們的目標是通過這個技術把農民和村莊歸類，好讓銀行或者政府把錢發放在正確的人群手中。」一名人類的分析員可以處理10到15個不同的變數，而機器學習演算法可以處理超過2000個變數，這是2個量級的差距。

Choosing where to allocate resources is a particularly pressing problem for governments as a warming Earth sends the climate into chaos. Traditionally, farming in India has been a relatively predictable affair, at least as far as humans holding dominion over their environment goes. 「So what I learned from my father, my grandfather, thats how I grow, these are the seasons I know,」 Garg says. 「However because of drastic climate change, things are no longer what my father or my grandfather used to do.」

全球變暖所帶來的氣候改變意味著政府正在面臨如何合理分配資源的難題。一直以來，印度的傳統農業都非常規律。「我從我父親和祖父那學會了如何種地，如何分辨季節，」加格說，「但是由於氣候的變化，我父親和祖父的經驗已經無用了。」

It』s the new world order, folks. Farmers can take the punches, or they can farm smarter. More data, more AI, and more chemical-spraying robots.

現在我們面臨的是新世界秩序。農民可以選擇被淘汰，或者選擇學習更加智能的農業方式——更多數據，更多AI，更多的農藥噴洒機器人。

As for those tomato plants you keep neglecting—that one』s on you, I』m afraid.

微博：數據智農

微信：數據智農

知乎：數據智農

郵箱：dataintellagr@126.com

美編|潘金超

責編|劉國輝