黃成龍,柯宇曦,華向東,楊俊雅,孫夢(mèng)雨,楊萬能
邊緣計(jì)算在智慧農(nóng)業(yè)中的應(yīng)用現(xiàn)狀與展望
黃成龍1,柯宇曦1,華向東1,楊俊雅1,孫夢(mèng)雨1,楊萬能2
(1. 華中農(nóng)業(yè)大學(xué)工學(xué)院,武漢 430070;2. 華中農(nóng)業(yè)大學(xué)作物遺傳改良國家重點(diǎn)實(shí)驗(yàn)室,武漢 430070)
互聯(lián)網(wǎng)技術(shù)快速發(fā)展使得數(shù)據(jù)量劇增,云計(jì)算的數(shù)據(jù)集中處理模式存在實(shí)時(shí)性不足、能耗過高以及數(shù)據(jù)安全等一系列問題。邊緣計(jì)算是在靠近數(shù)據(jù)源端執(zhí)行計(jì)算的分散處理模式,與云計(jì)算相比具有低延遲、低成本、安全性高、個(gè)性化設(shè)計(jì)等優(yōu)勢(shì)。隨著智慧農(nóng)業(yè)迅速發(fā)展,結(jié)合深度學(xué)習(xí)的農(nóng)業(yè)應(yīng)用屢見不鮮,如作物病害檢測(cè)、生長環(huán)境監(jiān)測(cè)、作物自動(dòng)采摘、無人農(nóng)場(chǎng)管理等,邊緣計(jì)算可以為農(nóng)業(yè)多場(chǎng)景、復(fù)雜任務(wù)提供高效、可靠的新型數(shù)據(jù)處理方案。該研究概述了邊緣計(jì)算的發(fā)展,計(jì)算架構(gòu)及主要優(yōu)勢(shì);介紹了邊緣計(jì)算在農(nóng)業(yè)中的應(yīng)用背景,結(jié)合文獻(xiàn)量分析,歸納了邊緣計(jì)算在農(nóng)業(yè)上的主要應(yīng)用場(chǎng)景及相關(guān)智能農(nóng)業(yè)裝備,調(diào)研了現(xiàn)有常用邊緣計(jì)算設(shè)備及性能參數(shù),總結(jié)了適合邊緣計(jì)算的主流深度學(xué)習(xí)算法及模型壓縮方法。研究表明邊緣計(jì)算在智慧農(nóng)業(yè)中的應(yīng)用有效促進(jìn)了農(nóng)業(yè)的數(shù)字化、智能化,未來在多場(chǎng)景、多功能邊緣計(jì)算智能農(nóng)業(yè)裝備開發(fā)等領(lǐng)域?qū)⒚媾R重大挑戰(zhàn)和機(jī)遇。
物聯(lián)網(wǎng);邊緣計(jì)算;云計(jì)算;智慧農(nóng)業(yè);深度學(xué)習(xí);模型壓縮;模型部署
近年來,隨著互聯(lián)網(wǎng)技術(shù)的快速發(fā)展,遠(yuǎn)程高性能服務(wù)器集中解決計(jì)算與存儲(chǔ)問題的云計(jì)算模式推動(dòng)著萬物互聯(lián)和人工智能的飛速發(fā)展與廣泛應(yīng)用,極大改善了社會(huì)生活和工業(yè)生產(chǎn)方式[1]。在物聯(lián)網(wǎng)與云技術(shù)的蓬勃發(fā)展下,一系列農(nóng)業(yè)場(chǎng)景與云計(jì)算結(jié)合,實(shí)現(xiàn)數(shù)字化、自動(dòng)化,有力推動(dòng)了智慧農(nóng)業(yè)的發(fā)展[2]。思科全球云指數(shù)報(bào)告指出,2016年全球云數(shù)據(jù)中心數(shù)據(jù)量為6.0 ZB(1 ZB等于10億TB),到2021年,這一數(shù)字暴漲3倍,達(dá)到了19.5 ZB,云數(shù)據(jù)中心流量占總數(shù)據(jù)流量的95%[3]。在數(shù)據(jù)量急劇上升的萬物互聯(lián)時(shí)代,云計(jì)算的集中處理存在以下不足:1)實(shí)時(shí)性不足,隨著物聯(lián)網(wǎng)飛速發(fā)展,眾多終端設(shè)備產(chǎn)生的數(shù)據(jù)量劇增,使得網(wǎng)絡(luò)帶寬面臨巨大負(fù)擔(dān),導(dǎo)致數(shù)據(jù)傳輸延遲時(shí)間大大增加,難以滿足人們?nèi)粘9ぷ魃钚枨骩4]。2)能耗過高,云服務(wù)器數(shù)據(jù)激增,能耗大大增加,僅以中國數(shù)據(jù)中心來看,每年用電量以超過10%的速度增長,至2021年年耗電已超過1 000億kW·h[5]。3)數(shù)據(jù)安全問題,大數(shù)據(jù)時(shí)代下社會(huì)生活、工業(yè)生產(chǎn)等隱私數(shù)據(jù)直接上傳云數(shù)據(jù)中心會(huì)帶來一系列安全隱患,受隱私協(xié)議霸王條款、廠商技術(shù)漏洞和黑客攻擊等問題影響,隨時(shí)存在數(shù)據(jù)泄露與丟失的風(fēng)險(xiǎn)[6]。
為解決云計(jì)算實(shí)時(shí)性不足、能耗過高、及數(shù)據(jù)安全問題,邊緣計(jì)算采用在靠近數(shù)據(jù)源端執(zhí)行計(jì)算的分散處理模式,以此來降低云計(jì)算數(shù)據(jù)中心的計(jì)算負(fù)載,從而實(shí)現(xiàn)降低能耗以及減少網(wǎng)絡(luò)帶寬的壓力[7]。2016年11月30日,邊緣計(jì)算產(chǎn)業(yè)聯(lián)盟在北京成立,由華為、英特爾公司、中國信息通信研究院、軟通動(dòng)力等單位組成,在邊緣計(jì)算產(chǎn)業(yè)峰會(huì)上,正式發(fā)布《邊緣計(jì)算參考架構(gòu)2.0》[8]。其中邊緣計(jì)算定義為:在靠近數(shù)據(jù)源的一側(cè),采用網(wǎng)絡(luò)、計(jì)算、存儲(chǔ)的分布式平臺(tái),就近提供邊緣智能服務(wù)。邊緣計(jì)算可以為不同行業(yè)提供經(jīng)濟(jì)、可行、創(chuàng)新的解決方案:1)智慧水務(wù),基于邊緣計(jì)算的智慧供水系統(tǒng),實(shí)現(xiàn)故障自診斷、可預(yù)測(cè)性維護(hù),據(jù)華為云智能邊緣平臺(tái)報(bào)告指出結(jié)合邊緣計(jì)算的智慧水務(wù)系統(tǒng)故障時(shí)間和維護(hù)人力減少60%;2)智慧照明,基于邊緣計(jì)算的智慧照明系統(tǒng),實(shí)現(xiàn)路燈的遠(yuǎn)程、實(shí)時(shí)、自適應(yīng)控制,與傳統(tǒng)照明系統(tǒng)相比能耗降低80%,運(yùn)維成本降低90%;3)智能樓宇,基于邊緣計(jì)算的智慧樓宇,實(shí)現(xiàn)樓宇多系統(tǒng)協(xié)同控制和智能化運(yùn)營,比較供暖、通風(fēng)和空調(diào)系統(tǒng)耗能,相比傳統(tǒng)運(yùn)行方式節(jié)省了36.75%以上的能源[8-9]。綜上所述,邊緣計(jì)算滿足行業(yè)在敏捷聯(lián)接、實(shí)時(shí)業(yè)務(wù)、智能決策、數(shù)據(jù)安全等方面的關(guān)鍵需求,是行業(yè)數(shù)字化升級(jí)不可或缺的要素。
隨著中國老齡化加劇、城市化發(fā)展和氣候變化,傳統(tǒng)農(nóng)業(yè)發(fā)展面臨巨大挑戰(zhàn),智慧農(nóng)業(yè)作為農(nóng)業(yè)生產(chǎn)的高級(jí)階段[10],通過人工智能、物聯(lián)網(wǎng)、云計(jì)算等現(xiàn)代信息技術(shù)與傳統(tǒng)農(nóng)業(yè)相結(jié)合,實(shí)現(xiàn)農(nóng)業(yè)無人化、自動(dòng)化、智能化生產(chǎn)和管理。隨著智慧農(nóng)業(yè)的快速發(fā)展,越來越多智能農(nóng)業(yè)終端被應(yīng)用,通過實(shí)時(shí)現(xiàn)場(chǎng)數(shù)據(jù)收集、數(shù)據(jù)分析和執(zhí)行機(jī)構(gòu)控制,提高農(nóng)業(yè)生產(chǎn)的質(zhì)量和數(shù)量[11]。環(huán)境傳感器,可以獲取環(huán)境濕度、溫度、光照、二氧化碳含量,及土壤水分pH值,實(shí)現(xiàn)動(dòng)植物生長環(huán)境信息的動(dòng)態(tài)監(jiān)測(cè)[12]。動(dòng)植物生長監(jiān)測(cè)傳感器,可以獲取動(dòng)植物的光譜、圖像、聲音、電磁等信息,實(shí)現(xiàn)動(dòng)態(tài)生長、病害、產(chǎn)量等關(guān)鍵性狀的動(dòng)態(tài)解析[13]。智能裝備傳感器,可以獲取如拖拉機(jī)、收獲機(jī)、農(nóng)業(yè)機(jī)器人、無人機(jī)、和末端執(zhí)行器的作業(yè)狀態(tài)信息,實(shí)現(xiàn)農(nóng)業(yè)裝備的智能監(jiān)測(cè)和控制[14]。基于各種傳感器收集的多維度農(nóng)業(yè)信息,構(gòu)建大數(shù)據(jù)分析模型,可以為動(dòng)物養(yǎng)殖、植物生產(chǎn)裝備作業(yè)提供智能管理決策,如智能灌溉、變量施肥、精準(zhǔn)飼養(yǎng)、疾病診斷等,從而降低農(nóng)業(yè)生產(chǎn)、運(yùn)營成本[15]。智慧農(nóng)業(yè)按照“感知-決策-執(zhí)行”內(nèi)在邏輯,可以分為農(nóng)業(yè)智能感知,數(shù)據(jù)分析與決策,智能裝備執(zhí)行3個(gè)重要部分,其中數(shù)據(jù)分析與決策離不開云計(jì)算、邊緣計(jì)算平臺(tái)的支撐。
邊緣計(jì)算作為一種新型計(jì)算模式,將其應(yīng)用在智慧農(nóng)業(yè)上,實(shí)現(xiàn)在數(shù)據(jù)采集邊緣端完成數(shù)據(jù)處理和決策,可以有效克服云計(jì)算存在的瓶頸,顯著提高數(shù)據(jù)安全性、處理實(shí)時(shí)性,同時(shí)降低能耗、成本。本文介紹了邊緣計(jì)算的架構(gòu)、優(yōu)勢(shì),綜述了智慧農(nóng)業(yè)發(fā)展現(xiàn)狀,以及邊緣計(jì)算在農(nóng)業(yè)應(yīng)用上的文獻(xiàn)發(fā)表情況;分析了邊緣計(jì)算常用的核心設(shè)備,以及主流的邊緣計(jì)算人工智能算法;討論了邊緣計(jì)算主要智能農(nóng)業(yè)裝備以及農(nóng)業(yè)應(yīng)用場(chǎng)景;總結(jié)了現(xiàn)階段邊緣計(jì)算在智慧農(nóng)業(yè)應(yīng)用上存在的問題,并對(duì)未來發(fā)展進(jìn)行了展望。
邊緣計(jì)算的架構(gòu)如圖1所示,其在終端與云端之間引入邊緣計(jì)算端,代替云端處理部分?jǐn)?shù)據(jù)[16]。終端為用戶數(shù)據(jù)采集端,通過智能手機(jī)、工業(yè)相機(jī)等設(shè)備收集原始數(shù)據(jù)上傳至云端或邊緣計(jì)算端進(jìn)行計(jì)算與存儲(chǔ)。云端由多個(gè)高性能服務(wù)器與存儲(chǔ)設(shè)備構(gòu)成,可以從終端獲取訓(xùn)練數(shù)據(jù)完成復(fù)雜模型的訓(xùn)練和優(yōu)化,并將訓(xùn)練好的模型部署在邊緣計(jì)算端,實(shí)現(xiàn)對(duì)邊緣設(shè)備的有效調(diào)度以及針對(duì)特定任務(wù)的數(shù)據(jù)處理[17]。邊緣計(jì)算端,可以快速響應(yīng)終端請(qǐng)求并將處理結(jié)果反饋至終端,為用戶提供更好的實(shí)時(shí)服務(wù)。
在網(wǎng)絡(luò)邊緣處理數(shù)據(jù)可以降低網(wǎng)絡(luò)負(fù)載和通信延遲,降低移動(dòng)節(jié)點(diǎn)的能源消耗,解決實(shí)時(shí)響應(yīng)和帶寬限制等問題。作為人工智能的重要分支,深度學(xué)習(xí)憑借大量神經(jīng)網(wǎng)絡(luò)層數(shù)可以解決眾多復(fù)雜問題,然而其巨大計(jì)算量導(dǎo)致硬件算力需求較高,而傳統(tǒng)的深度學(xué)習(xí)服務(wù)器存在體積大、移動(dòng)性差,成本高的不足,很難進(jìn)行大規(guī)模應(yīng)用[18]。邊緣計(jì)算設(shè)備包括現(xiàn)場(chǎng)可編程邏輯門陣列[19](Field-Programmable Gate Array,F(xiàn)PGA),數(shù)字信號(hào)處理器[20](Digital Signal Processor,DSP),片上系統(tǒng)[21](System-on-a-Chip,SOC),樹莓派[22](Raspberry Pi),Nvidia Jetson[23]和智能移動(dòng)終端等,均具有較強(qiáng)的本地運(yùn)算能力,可以部署深度學(xué)習(xí)人工智能模型,實(shí)現(xiàn)采集數(shù)據(jù)的快速準(zhǔn)確解析。施耐德電氣公司對(duì)邊緣計(jì)算部署的成本效益做出了分析,將邊緣計(jì)算處理器與存儲(chǔ)設(shè)備整合在1個(gè)機(jī)柜中,其工作處理能力相當(dāng)于13個(gè)機(jī)柜的云服務(wù)器的處理能力,尺寸縮減的同時(shí)提高了性能,邊緣計(jì)算數(shù)據(jù)中心相比于云計(jì)算數(shù)據(jù)中心的投資成本節(jié)省42%[24]。邊緣計(jì)算顯著降低了人工智能算法部署的硬件成本,提高了嵌入式開發(fā)的可行性,使得一系列人工智能應(yīng)用成為了可能。
圖1 邊緣計(jì)算架構(gòu)
邊緣計(jì)算具有低成本、低能耗、低延時(shí)、數(shù)據(jù)安全的優(yōu)勢(shì),已廣泛應(yīng)用工業(yè)生產(chǎn)、社會(huì)生活的智能嵌入式產(chǎn)品開發(fā)。王梓儒[25]分別在消費(fèi)級(jí)ARM平臺(tái)即樹莓派3B+,高性能嵌入式GPU Nvidia Jetson TX2以及Android智能手機(jī)上部署了深度目標(biāo)檢測(cè)網(wǎng)絡(luò),給出了3種不同平臺(tái)的邊緣計(jì)算部署方案。張釗[26]通過在Nvidia Jetson TX2上部署改進(jìn)后的YOLOV4-tiny算法,設(shè)計(jì)了基于邊緣計(jì)算的視頻監(jiān)控系統(tǒng),并對(duì)煤層氣井站視頻數(shù)據(jù)進(jìn)行實(shí)時(shí)監(jiān)測(cè),平均檢測(cè)精度達(dá)到92.15%,單張圖片檢測(cè)時(shí)長為0.102 s。Ma等[27]以華為atlas 200芯片作為智能處理芯片,采用參數(shù)量化的模型壓縮方法部署殘差網(wǎng)絡(luò)與特征金字塔融合網(wǎng)絡(luò),設(shè)計(jì)了電網(wǎng)結(jié)冰智能監(jiān)測(cè)裝置計(jì)算模塊,單幀檢測(cè)速度達(dá)170 ms。Kim等[28]通過在Nvidia Jetson NANO上部署YOLOV3-tiny目標(biāo)檢測(cè)模型,構(gòu)建了基于邊緣計(jì)算的對(duì)象運(yùn)動(dòng)與跟蹤系統(tǒng),通過分層次利用幀差計(jì)算、目標(biāo)檢測(cè)等輕任務(wù),自適應(yīng)地釋放不必要的待機(jī)對(duì)象運(yùn)動(dòng)和運(yùn)動(dòng)跟蹤模型,可以節(jié)省高達(dá)78.5%的GPU內(nèi)存占用。
綜上所述,相較于云計(jì)算,邊緣計(jì)算存在以下優(yōu)勢(shì):
1)低延遲,在靠近數(shù)據(jù)端進(jìn)行數(shù)據(jù)處理,避免了向云數(shù)據(jù)中心請(qǐng)求響應(yīng),可以降低網(wǎng)絡(luò)延遲,實(shí)現(xiàn)更快速、更高效的數(shù)據(jù)分析和處理,研究表明,與云端相比,基于邊緣計(jì)算的分支神經(jīng)網(wǎng)絡(luò)模型的推理延遲平均降低36%[29]。
2)低成本,網(wǎng)絡(luò)邊緣產(chǎn)生的大量數(shù)據(jù)無需全部上傳云端,減輕了網(wǎng)絡(luò)帶寬的壓力,同時(shí)降低了數(shù)據(jù)傳輸帶來的巨大能耗。施耐德電氣公司對(duì)邊緣計(jì)算數(shù)據(jù)中心的成本效益分析中提到,相同算力條件下,邊緣數(shù)據(jù)中心維護(hù)成本相較云數(shù)據(jù)中心每平方米節(jié)省1 600美元,節(jié)省成本42%[24],因此在本地設(shè)備上的數(shù)據(jù)管理成本大大低于云和數(shù)據(jù)中心網(wǎng)絡(luò)。
3)安全性高,避免數(shù)據(jù)直接傳輸云端帶來的隱私泄露風(fēng)險(xiǎn),重要數(shù)據(jù)可以直接在邊緣計(jì)算端進(jìn)行加密處理或者保存。邊緣計(jì)算端更貼近數(shù)據(jù)采集設(shè)備,訪問攻擊的難度大幅提升,提高了數(shù)據(jù)安全性。
4)個(gè)性化設(shè)計(jì),通過將邊緣計(jì)算和人工智能結(jié)合,可以持續(xù)分析客戶數(shù)據(jù)及行為,提供實(shí)時(shí)交互,為智能設(shè)備提供自我修復(fù)、自我優(yōu)化的實(shí)時(shí)處理,實(shí)現(xiàn)即時(shí)個(gè)性化。
農(nóng)業(yè)是社會(huì)和國民經(jīng)濟(jì)的基礎(chǔ),及時(shí)獲取可靠的農(nóng)業(yè)信息,如作物生長和產(chǎn)量,對(duì)于制定糧食安全、減貧和可持續(xù)發(fā)展的相關(guān)政策和計(jì)劃至關(guān)重要[30]。隨著大數(shù)據(jù)、物聯(lián)網(wǎng)、云計(jì)算、人工智能等現(xiàn)代信息技術(shù)在農(nóng)業(yè)上的應(yīng)用,第三次農(nóng)業(yè)革命-農(nóng)業(yè)智能革命已經(jīng)到來[31]。智慧農(nóng)業(yè)是以信息和知識(shí)為核心要素,通過將物聯(lián)網(wǎng)、大數(shù)據(jù)、人工智能、云計(jì)算等先進(jìn)信息技術(shù)與農(nóng)業(yè)深度融合,實(shí)現(xiàn)農(nóng)業(yè)信息感知、智能控制、精準(zhǔn)決策、高效作業(yè)的全新的農(nóng)業(yè)生產(chǎn)管理方式,是農(nóng)業(yè)發(fā)展從信息化到智能化的高級(jí)階段[32]。農(nóng)業(yè)物聯(lián)網(wǎng)通過傳感器和軟件通過移動(dòng)平臺(tái)或者電腦平臺(tái)對(duì)農(nóng)業(yè)生產(chǎn)進(jìn)行控制,使得傳統(tǒng)農(nóng)業(yè)具有“智慧”,2020年7 500萬物聯(lián)網(wǎng)設(shè)備用于農(nóng)業(yè)領(lǐng)域?yàn)橹腔坜r(nóng)業(yè)提供了大量智能感知和控制終端[33]。農(nóng)業(yè)人工智能,通過研究圖像識(shí)別、智能控制、動(dòng)植物生長模型和專家系統(tǒng)等智能算法,實(shí)現(xiàn)對(duì)農(nóng)業(yè)大數(shù)據(jù)的智能分析處理、并作出有效決策,使農(nóng)業(yè)生產(chǎn)過程更加智能化、成本效益更高[34]。為實(shí)現(xiàn)農(nóng)業(yè)物聯(lián)網(wǎng)數(shù)據(jù)分析和處理,亟需能部署深度學(xué)習(xí)人工智能算法的高效、可靠、低成本計(jì)算平臺(tái)。
雖然云計(jì)算基礎(chǔ)架構(gòu)可以為分布式農(nóng)業(yè)物聯(lián)網(wǎng)傳感器、人工智能算法提供集中的強(qiáng)大算力基礎(chǔ),但是多個(gè)網(wǎng)絡(luò)層上傳感器異構(gòu)數(shù)據(jù)的傳輸、和集中的數(shù)據(jù)處理將帶來巨大網(wǎng)絡(luò)帶寬負(fù)擔(dān)、能源消耗、信息安全風(fēng)險(xiǎn)[35]。邊緣計(jì)算通過終端實(shí)時(shí)處理減少網(wǎng)絡(luò)負(fù)載和信息安全風(fēng)險(xiǎn),可以有效彌補(bǔ)云計(jì)算的不足,為智慧農(nóng)業(yè)提供了新的計(jì)算架構(gòu)[36]。Alharbi等[37]在智能農(nóng)業(yè)環(huán)境下,采用混合整數(shù)線性規(guī)劃進(jìn)行了數(shù)學(xué)建模,對(duì)結(jié)合邊緣計(jì)算的集成架構(gòu)模式與傳統(tǒng)的實(shí)現(xiàn)方法進(jìn)行了分析和比較,證明結(jié)合邊緣計(jì)算的新型架構(gòu)模式降低總能耗36%,碳排放量43%,可以將網(wǎng)絡(luò)流量減少86%,從而減少網(wǎng)絡(luò)擁塞,具有良好的應(yīng)用前景。此外,邊緣計(jì)算較云計(jì)算而言有著低延遲、低帶寬成本、移動(dòng)性支持和高可擴(kuò)展性等優(yōu)勢(shì),可以為農(nóng)業(yè)應(yīng)用提供成本低、實(shí)時(shí)性高、適用性強(qiáng)的解決方案[38],為智慧農(nóng)業(yè)發(fā)展提供新的技術(shù)支撐。
本文對(duì)2018—2021年國內(nèi)外關(guān)于邊緣計(jì)算在農(nóng)業(yè)應(yīng)用的相關(guān)文獻(xiàn)進(jìn)行統(tǒng)計(jì)分析,結(jié)果如圖2所示。其中國外文獻(xiàn)以Web of science為來源,國內(nèi)文獻(xiàn)以CNKI為來源,以邊緣計(jì)算,農(nóng)業(yè)為關(guān)鍵詞進(jìn)行篩選。邊緣計(jì)算概念是2016年底提出,2018年已經(jīng)有學(xué)者開始將邊緣計(jì)算應(yīng)用于農(nóng)業(yè)領(lǐng)域,2018—2019年為探索階段僅有少量相關(guān)文獻(xiàn)的。隨著數(shù)據(jù)量激增,云計(jì)算出現(xiàn)網(wǎng)絡(luò)延遲,能耗大,數(shù)據(jù)安全等一系列問題,大量國內(nèi)外研究者開始關(guān)注邊緣計(jì)算在農(nóng)業(yè)中的應(yīng)用,2020年相比前一年國內(nèi)文獻(xiàn)數(shù)量提高約3倍。2021年得益于系列邊緣計(jì)算產(chǎn)業(yè)聯(lián)盟成員的關(guān)注及投入,眾多高算力邊緣設(shè)備持續(xù)推出,邊緣計(jì)算在農(nóng)業(yè)中應(yīng)用的相關(guān)研究持續(xù)增長,且首次出現(xiàn)中文文獻(xiàn)發(fā)表量超過外文文獻(xiàn)[39]。綜上所述,目前邊緣計(jì)算在農(nóng)業(yè)上的應(yīng)用處于快速發(fā)展階段,可以預(yù)測(cè)未來將為越來越多的農(nóng)業(yè)場(chǎng)景提供新的解決方案。
圖2 邊緣計(jì)算農(nóng)業(yè)應(yīng)用國內(nèi)外文獻(xiàn)量
現(xiàn)階段,邊緣計(jì)算的農(nóng)業(yè)應(yīng)用通常與人工智能算法結(jié)合,旨在實(shí)現(xiàn)動(dòng)植物生長動(dòng)態(tài)監(jiān)測(cè)、環(huán)境實(shí)時(shí)檢測(cè)、和農(nóng)業(yè)裝備作業(yè)智能決策。根據(jù)現(xiàn)有文獻(xiàn)報(bào)道,邊緣計(jì)算在農(nóng)業(yè)中應(yīng)用的場(chǎng)景如表1所示,主要分為環(huán)境監(jiān)測(cè)與病蟲害識(shí)別、作物生長及產(chǎn)量預(yù)測(cè)、農(nóng)業(yè)偵察與路徑規(guī)劃等方面。此外,表中對(duì)不同邊緣計(jì)算農(nóng)業(yè)應(yīng)用場(chǎng)景下,測(cè)量目標(biāo)、采用的邊緣計(jì)算設(shè)備、網(wǎng)絡(luò)模型、檢測(cè)速度與精度指標(biāo),進(jìn)行了歸納總結(jié)。
1)環(huán)境監(jiān)測(cè)與病蟲害識(shí)別
病蟲害識(shí)別與環(huán)境檢測(cè)是目前邊緣計(jì)算最常見農(nóng)業(yè)應(yīng)用場(chǎng)景。劉蘇偉[40]基于邊緣計(jì)算與深度學(xué)習(xí)構(gòu)建了玉米葉片病害識(shí)別系統(tǒng),通過采集玉米葉片圖像,對(duì)葉斑病、葉枯病、銹病以及健康葉片進(jìn)行識(shí)別,選用ResNet18模型測(cè)試精確率達(dá)85.4%,當(dāng)終端和邊緣設(shè)備連接并傳輸數(shù)據(jù)時(shí),最大速度達(dá)5.58 MB/s。牛愷銳等[41]基于深度學(xué)習(xí)框架構(gòu)建了一個(gè)特征提取網(wǎng)絡(luò),并部署在海思Hi3559A芯片上,實(shí)現(xiàn)小麥、水稻病蟲害識(shí)別,模型準(zhǔn)確率分別為92%、97%,識(shí)別速度達(dá)20.0幀/s且功耗小于5 W,該邊緣計(jì)算嵌入式終端相較于傳統(tǒng)服務(wù)器具有低成本、低功耗、輕量化等優(yōu)勢(shì)。李鳳迪[42]構(gòu)建了基于深度學(xué)習(xí)的松材線蟲病樹檢測(cè)方法,選用樹莓派 4B作為邊緣計(jì)算平臺(tái)部署訓(xùn)練好的MobileNetv2-SSDLite模型并集成在大疆M600無人機(jī)上,實(shí)現(xiàn)松材線蟲病樹的在線監(jiān)測(cè),識(shí)別速度達(dá)到5幀/s。孫志朋[43]通過在樹莓派部署卷積神經(jīng)網(wǎng)絡(luò)對(duì)水稻害蟲圖像進(jìn)行識(shí)別,準(zhǔn)確率可達(dá)到89%,利用邊緣設(shè)備完成了害蟲在線識(shí)別計(jì)數(shù)、水稻生長環(huán)境監(jiān)測(cè),減少了云端計(jì)算壓力。Guillén等[44]基于深度學(xué)習(xí)搭建了農(nóng)業(yè)低溫預(yù)測(cè)邊緣計(jì)算平臺(tái),以Nvidia Jetson AGX Xavier為邊緣設(shè)備部署LSTM (Long Short-Term Memory)模型實(shí)現(xiàn)溫度預(yù)測(cè),推理時(shí)間為0.3 s,預(yù)測(cè)值的平均誤差小于0.8 ℃,設(shè)備耗電量小于0.08 kW·h。綜上所述,通過邊緣設(shè)備部署人工智能算法,可以實(shí)現(xiàn)高精度、實(shí)時(shí)性的環(huán)境監(jiān)測(cè)和病蟲害識(shí)別,為農(nóng)業(yè)人工智能應(yīng)用提供了新的技術(shù)途徑。
2)作物生長及產(chǎn)量預(yù)測(cè)
作物生長及產(chǎn)量預(yù)測(cè)是邊緣計(jì)算在農(nóng)業(yè)上的應(yīng)用領(lǐng)域之一,通過邊緣設(shè)備上部署機(jī)器學(xué)習(xí)預(yù)測(cè)模型,可以大大減少預(yù)測(cè)環(huán)節(jié)所用時(shí)間。Park等[45]將邊緣計(jì)算技術(shù)融入智能農(nóng)場(chǎng)中,分析環(huán)境和生長數(shù)據(jù)獲取關(guān)鍵參數(shù),以此來預(yù)測(cè)作物生長及最終產(chǎn)量,通過在樹莓派上部署LSTM模型對(duì)櫻桃番茄的產(chǎn)量進(jìn)行預(yù)測(cè),得到預(yù)測(cè)值均方誤差為0.045,預(yù)測(cè)精度較高。Coviello等[46]通過智能手機(jī)對(duì)葡萄產(chǎn)量進(jìn)行測(cè)算,使用設(shè)計(jì)的計(jì)數(shù)網(wǎng)絡(luò)GBCNet在兩個(gè)原始數(shù)據(jù)集 CR1和CR2 上進(jìn)行測(cè)試,檢測(cè)的平均百分比誤差在0.85%~11.73%,手機(jī)拍攝和處理單張圖片時(shí)間小于1 s,具有較好的便攜性和較高的預(yù)測(cè)效率。綜上所述,與服務(wù)器端數(shù)據(jù)采集、上傳、分析及模型預(yù)測(cè)的步驟相比,邊緣計(jì)算設(shè)備可以直接實(shí)現(xiàn)現(xiàn)場(chǎng)數(shù)據(jù)采集與模型預(yù)測(cè),且具有較高的預(yù)測(cè)精度和效率,可為精準(zhǔn)農(nóng)業(yè)發(fā)展助力。
3)農(nóng)業(yè)偵察與無人機(jī)路徑規(guī)劃
農(nóng)業(yè)偵察與無人機(jī)路徑規(guī)劃,是農(nóng)業(yè)裝備智能作業(yè)的重要內(nèi)容。與所有數(shù)據(jù)傳輸?shù)皆撇煌ㄟ^邊緣節(jié)點(diǎn)與無人機(jī)等傳感器連接提供了近數(shù)據(jù)端、低延時(shí)、低成本的智能數(shù)據(jù)處理與決策方案。Yang等[47]結(jié)合邊緣計(jì)算提出了一種無人機(jī)自適應(yīng)作物偵察機(jī)制,將EDANet模型部署在Nvidia Jetson TX2上,結(jié)合無人機(jī)在多個(gè)角度對(duì)水稻進(jìn)行偵察,可以將稻田偵察速度提高36%,準(zhǔn)確率達(dá)99.25%。Chen等[48]結(jié)合邊緣計(jì)算建立了無人機(jī)害蟲智能識(shí)別系統(tǒng),在Nvidia Jetson TX2上部署基于YOLOv3-tiny的無人機(jī)果園乳頭狀錐蟲智能識(shí)別模型,實(shí)現(xiàn)害蟲快速準(zhǔn)確定位,并規(guī)劃出最優(yōu)無人機(jī)農(nóng)藥噴灑路徑,與傳統(tǒng)路徑相比縮短19%,且減少了87%的水消耗量,節(jié)省了53%的工作時(shí)間;此外還可以將害蟲位置和產(chǎn)生情況傳輸?shù)皆贫艘员阌涗浐头治鲎魑锷L情況。由此可知,通過嵌入式邊緣計(jì)算設(shè)備和無人機(jī)結(jié)合,可以部署復(fù)雜的人工智能模型,實(shí)現(xiàn)高精度農(nóng)業(yè)偵察和最優(yōu)路徑規(guī)劃。
表1 邊緣計(jì)算在農(nóng)業(yè)中的應(yīng)用場(chǎng)景
根據(jù)文獻(xiàn)報(bào)道目前基于邊緣計(jì)算的智能農(nóng)業(yè)裝備如圖3所示,主要分為智能農(nóng)業(yè)無人機(jī)[66]、智能農(nóng)業(yè)機(jī)器人[67]以及農(nóng)業(yè)智能移動(dòng)終端[68]。邊緣設(shè)備與無人機(jī)結(jié)合常用于執(zhí)行雜草、蟲害檢測(cè)、路徑規(guī)劃和農(nóng)藥自動(dòng)噴灑等任務(wù);與地面農(nóng)業(yè)機(jī)器人結(jié)合實(shí)現(xiàn)農(nóng)作物實(shí)時(shí)檢測(cè),可完成作物采摘、除草、實(shí)時(shí)環(huán)境監(jiān)測(cè)等任務(wù);結(jié)合智能移動(dòng)端開發(fā)的應(yīng)用程序?yàn)橛脩籼峁┝烁涌旖莘奖愕霓r(nóng)業(yè)圖像采集和數(shù)據(jù)處理方案。
1)智能農(nóng)業(yè)無人機(jī),作為一種新型的信息獲取載體,無人機(jī)因其操作靈活、適應(yīng)性高,廣泛應(yīng)用于各種農(nóng)業(yè)場(chǎng)景,尤其是在農(nóng)藥噴灑、作物蟲害監(jiān)測(cè)、地形勘測(cè)等方面[69]。通過在無人機(jī)上部署邊緣計(jì)算核心設(shè)備,在空中作業(yè)的過程中,對(duì)采集到的圖像進(jìn)行實(shí)時(shí)處理,自動(dòng)進(jìn)行路徑規(guī)劃、作物病害識(shí)別,完成除草、農(nóng)藥噴灑、地圖繪制等作業(yè),減少后續(xù)數(shù)據(jù)傳輸、遠(yuǎn)程處理等步驟,提高工作效率。如Ukaegbu等[70]基于無人機(jī)和樹莓派3B,開展飛行作業(yè)過程中雜草檢測(cè)與除草劑自動(dòng)噴灑研究,實(shí)現(xiàn)0.5 m的飛行高度下雜草檢測(cè)時(shí)間小于1 s,精度大于98%;Camargo等[55]在邊緣設(shè)備Nvidia Jetson AGX Xavier上部署ResNet-18 DCNN(Dynamic Convolution Neural Network,動(dòng)態(tài)卷積神經(jīng)網(wǎng)絡(luò))模型實(shí)現(xiàn)雜草與作物智能檢測(cè),總體準(zhǔn)確率為94%,檢測(cè)速度達(dá)到2.2 幀/s,實(shí)現(xiàn)雜草地圖的在線繪制;Partel等[62]開發(fā)了一種智能除草噴霧器,以Nvidia Jetson TX2作為邊緣計(jì)算端部署YOLOV3-tiny模型完成目標(biāo)雜草識(shí)別,平均檢測(cè)精度達(dá)90%,速度達(dá)到22幀/s。綜上所述,結(jié)合邊緣計(jì)算端與無人機(jī)設(shè)備可以在飛行過程中執(zhí)行數(shù)據(jù)分析任務(wù),減少了數(shù)據(jù)交互帶來的時(shí)間成本,使得自主路徑規(guī)劃作業(yè)成為可能,提高了工作效率。
圖3 基于邊緣計(jì)算的智能農(nóng)業(yè)裝備[68]
2)智能農(nóng)業(yè)機(jī)器人,隨著人工智能技術(shù)飛速發(fā)展,智能農(nóng)業(yè)機(jī)器人在內(nèi)部嵌入邊緣計(jì)算平臺(tái),可以直接在邊緣側(cè)實(shí)現(xiàn)對(duì)圖像信息的分析和決策,可以完成智能播種、種植、耕作、采摘、收割、分選等一系列工作[71]。部署人工智能算法的農(nóng)業(yè)機(jī)器人,相較于傳統(tǒng)控制作業(yè)方式更加高效智能,可以應(yīng)用于復(fù)雜的農(nóng)業(yè)作業(yè)場(chǎng)景,如棉花打頂、智能除草、精準(zhǔn)灌溉等。Nilay等[49]結(jié)合FPGA設(shè)備設(shè)計(jì)的水果采摘機(jī)器人,對(duì)采集到的圖像信息進(jìn)行處理,目標(biāo)水果識(shí)別精度為95.8%,識(shí)別速度達(dá)30 幀 /s,實(shí)現(xiàn)了目標(biāo)水果的自動(dòng)采集;Wang等[54]結(jié)合邊緣計(jì)算設(shè)計(jì)育苗機(jī)器人,通過在邊緣設(shè)備Nvidia Jetson TX2上部署YOLOV4-tiny模型實(shí)現(xiàn)了盆花的實(shí)時(shí)檢測(cè)與定位,平均檢測(cè)準(zhǔn)確率89.72%,檢測(cè)速度達(dá)到16 幀/s,完成了盆栽的自動(dòng)化管理;Chechliński等[50]設(shè)計(jì)的自主除草機(jī)器人,采用樹莓派3B作為邊緣計(jì)算設(shè)備以超過10 幀/s的檢測(cè)速度實(shí)現(xiàn)了雜草實(shí)時(shí)檢測(cè)。因此,將智能農(nóng)業(yè)機(jī)器人與邊緣計(jì)算技術(shù)結(jié)合,突破了傳統(tǒng)農(nóng)業(yè)機(jī)器人在復(fù)雜任務(wù)、復(fù)雜環(huán)境下作業(yè)的瓶頸,是智慧農(nóng)業(yè)的重要發(fā)展方向。
3)農(nóng)業(yè)智能移動(dòng)終端,隨著智能移動(dòng)終端的快速發(fā)展,其算力和存儲(chǔ)性能不斷提高,使得復(fù)雜深度學(xué)習(xí)模型部署成為了可能[72]?;谥悄芤苿?dòng)終端設(shè)計(jì)人工智能應(yīng)用程序,可實(shí)現(xiàn)便攜式、高精度的農(nóng)業(yè)信息采集與分析。如Liu等[56]在移動(dòng)智能手機(jī)上部署GoogLeNet模型,實(shí)現(xiàn)21種葡萄分類識(shí)別,準(zhǔn)確率達(dá)99.91%;Buzzy等[60]將YOLOV3-tiny部署在智能手機(jī)端,實(shí)現(xiàn)了植物葉片的檢測(cè)與計(jì)數(shù),檢測(cè)時(shí)間小于0.1 s;Ai等[59]將邊緣計(jì)算與深度學(xué)習(xí)結(jié)合,以卷積神經(jīng)網(wǎng)絡(luò)為基礎(chǔ)構(gòu)建了Inception-ResNet-v2模型,并部署在手機(jī)端,應(yīng)用于植物病蟲害的識(shí)別和檢測(cè),總體識(shí)別準(zhǔn)確率為86.1%。綜上所述,智能移動(dòng)終端,具有拍照、聲音采集等通用傳感器,基于通用的Android開發(fā)平臺(tái)設(shè)計(jì)移動(dòng)端人工智能應(yīng)用程序,可以為智慧農(nóng)業(yè)提供便攜式、低成本智能檢測(cè)方案。
隨著邊緣計(jì)算的快速發(fā)展,越來越多的計(jì)算設(shè)備為邊緣AI (Artificial Intelligence)應(yīng)用程序和嵌入式設(shè)備而設(shè)計(jì),農(nóng)業(yè)中常用的邊緣計(jì)算核心設(shè)備如圖4所示,主要包括樹莓派、英偉達(dá)小型計(jì)算平臺(tái)、FPGA、和手機(jī)處理器等,該類設(shè)備具有體積小、結(jié)構(gòu)緊湊、功耗低、算力高等優(yōu)勢(shì)[73]。常用的邊緣計(jì)算設(shè)備算力、功耗等性能參數(shù)如表2所示。樹莓派3B自2016年發(fā)布以來,因其高便攜性、低功耗受到了科研工作者的廣泛關(guān)注,到2019年樹莓派4B發(fā)布,計(jì)算能力相較于樹莓派3B有顯著提升,較高性價(jià)比以及較小的體積使其常作為邊緣計(jì)算核心設(shè)備集成于各類智慧農(nóng)業(yè)平臺(tái)中[74]。此外,Xilinx PYNQ-Z2、海思Hi3559等FPGA、SOC芯片的計(jì)算能力相較于樹莓派提升了近百倍,可以加載更加復(fù)雜的模型,并提高模型推理速度[75]。近年來,Nvidia Jetson推出的一系列邊緣計(jì)算設(shè)備如NANO、TX2、AGX等,其算力為0.5~10 T不等,可為不同農(nóng)業(yè)應(yīng)用場(chǎng)景提供最佳性價(jià)比的檢測(cè)方案[76]。
圖4 邊緣計(jì)算核心設(shè)備圖
樹莓派以較低成本與高便攜性受到了許多研究者們的青睞。Kundu等[77]提出了Custom-Net模型用于檢測(cè)珍珠粟疾病,并將模型部署在樹莓派3B上實(shí)現(xiàn)了98.78%的分類準(zhǔn)確率。Mishra等[61]采用樹莓派3B并結(jié)合由專用CNN(Convolutional Neural Network)硬件塊組成的Intel Movidius神經(jīng)計(jì)算棒作為邊緣設(shè)備,部署訓(xùn)練好的深度CNN模型,實(shí)現(xiàn)玉米葉片的病害識(shí)別,準(zhǔn)確率達(dá)88.46%;Tarek等[63]將MobileNetV3部署在樹莓派4B上,實(shí)現(xiàn)番茄疾病的快速準(zhǔn)確診斷,檢測(cè)精度達(dá)98.99%,檢測(cè)效率為每張圖250~350ms;Emebo等[64]構(gòu)建了一個(gè)番茄葉片病害分類模型,部署在手持式設(shè)備的樹莓派上,模型平均精度達(dá)99.01%;Tufail等[65]提出了一種基于紋理、形狀和顏色特征組合的支持向量機(jī)分類器,并將該算法部署在樹莓派4B上進(jìn)行實(shí)時(shí)監(jiān)測(cè),分類準(zhǔn)確率達(dá)96%,檢測(cè)效率為6幀/s;Meng等[51]開發(fā)了一種水下無人機(jī),配備360°全景攝像頭作為圖像采集端,并在樹莓派3B上部署深度學(xué)習(xí)魚類識(shí)別模型,模型準(zhǔn)確率達(dá)87%。
Nvidia Jetson系列開發(fā)板以寬泛、出色的算力在眾多邊緣設(shè)備中脫穎而出,且廠商提供了豐富的軟硬件支持服務(wù),因此以其作為邊緣計(jì)算設(shè)備的文獻(xiàn)報(bào)道最多。Seo等[52]以Nvidia Jetson NANO作為邊緣計(jì)算端,基于YOLOV4-tiny設(shè)計(jì)了復(fù)雜場(chǎng)景下生豬識(shí)別和定位算法,檢測(cè)精度達(dá)97.66%,檢測(cè)速度為34.38 幀/s,實(shí)現(xiàn)養(yǎng)豬場(chǎng)生豬智能監(jiān)測(cè);Deng等[53]針對(duì)無人機(jī)對(duì)雜草識(shí)別及精準(zhǔn)噴藥問題,構(gòu)建雜草識(shí)別輕量級(jí)的網(wǎng)絡(luò)架構(gòu),并將其部署在Nvidia Jetson TX2上,實(shí)現(xiàn)4.5幀/s的檢測(cè)速度和80.9%的檢測(cè)準(zhǔn)確度。
表2 常用邊緣計(jì)算設(shè)備及參數(shù)
注:每秒浮點(diǎn)運(yùn)算次數(shù)(Floating-point Operations Per Second, FLOPS),1GFLOPS等于每秒十億(109)次的浮點(diǎn)運(yùn)算,1TFLOPS等于每秒一萬億(1012)次的浮點(diǎn)運(yùn)算。
Note: FLOPS is the floating-point operations per second, 1GFLOPS means one billion (= 109) floating-point operations per second, and 1TFLOPS means one trillion (= 1012) floating-point operations per second.
除此之外,F(xiàn)PGA、DSP以及手機(jī)處理器也具有極強(qiáng)的算力,可用于邊緣端數(shù)據(jù)處理[78]。He等[57]提出了一種基于深度學(xué)習(xí)的油菜害蟲檢測(cè)方法,在移動(dòng)智能手機(jī)上部署SSD w/Inception模型,實(shí)現(xiàn)油菜害蟲實(shí)時(shí)診斷,平均檢測(cè)精度達(dá)77.14%;Ahmed等[58]基于深度學(xué)習(xí)開發(fā)了一種植物葉片疾病自動(dòng)診斷移動(dòng)式平臺(tái),在Android移動(dòng)端對(duì)14種作物常見的38種疾病進(jìn)行分類,總體分類準(zhǔn)確率達(dá)到94%。Liu等[79]開發(fā)了一款基于Android的便攜式植物表型分析應(yīng)用程序,實(shí)現(xiàn)15個(gè)整株性狀、25個(gè)葉片性狀和5個(gè)莖稈性狀的便攜式、實(shí)時(shí)檢測(cè)。綜上所述,面對(duì)不同的農(nóng)業(yè)應(yīng)用場(chǎng)景,用戶可以選擇合適算力的邊緣計(jì)算設(shè)備,為農(nóng)業(yè)生產(chǎn)提供具有成本效益的解決方案。
深度學(xué)習(xí)作為一種智能數(shù)據(jù)處理方法,廣泛應(yīng)用于智慧農(nóng)業(yè)研究與生產(chǎn)實(shí)踐,然而大多數(shù)深度學(xué)習(xí)方法對(duì)計(jì)算設(shè)備的算力和內(nèi)存需求較高[80]。雖然云計(jì)算可以提供較高算力和內(nèi)存支撐,但會(huì)導(dǎo)致高延遲和巨大的網(wǎng)絡(luò)帶寬壓力[81]。而基于邊緣計(jì)算的深度學(xué)習(xí)模型部署,為人工智能應(yīng)用提供了一種近數(shù)據(jù)端、低延時(shí)、低成本的檢測(cè)方案[82]。與云服務(wù)器不同,受邊緣計(jì)算設(shè)備算力限制,部署于邊緣端的模型運(yùn)算速度與模型大小密切相關(guān)[83]。根據(jù)現(xiàn)有文獻(xiàn),目前邊緣計(jì)算在農(nóng)業(yè)中應(yīng)用的深度學(xué)習(xí)算法主要采用輕量化深度學(xué)習(xí)網(wǎng)絡(luò),包括SSD(Single Shot Multibox Detector),YOLO(You Only Look Once)等算法,如表3所示。表中各算法檢測(cè)單張圖片的時(shí)間均采用本地樹莓派4B為邊緣設(shè)備進(jìn)行推理得到。將SSD-Mobilenet與SSD-VGG16對(duì)比,模型參數(shù)量更小,雖然精度有所降低,但是單張圖片檢測(cè)時(shí)間由19.1 s減少至3.73 s;同理,YOLOV4-tiny在YOLOV4的基礎(chǔ)上進(jìn)一步降低參數(shù)量,在僅降低 mAP(mean Average Precision)23.43%的情況下檢測(cè)速度提高約7倍。YOLOV5-lite、YOLO-fastest、YOLOX-NANO雖然將網(wǎng)絡(luò)參數(shù)量降至10 MB以內(nèi),在樹莓派4B上單張圖片檢測(cè)時(shí)間仍然較高,這說明了現(xiàn)有的輕量化網(wǎng)絡(luò)依然無法滿足低算力邊緣計(jì)算設(shè)備的要求。因此在保證滿足模型精度要求的情況下實(shí)現(xiàn)對(duì)模型進(jìn)行一步壓縮優(yōu)化也是邊緣計(jì)算研究的重點(diǎn)之一,常見模型壓縮方法主要包括網(wǎng)絡(luò)剪枝、知識(shí)蒸餾、參數(shù)量化、結(jié)構(gòu)優(yōu)化。
表3 邊緣計(jì)算在農(nóng)業(yè)中應(yīng)用的主流深度學(xué)習(xí)算法
注:表中各算法的單張檢測(cè)時(shí)間為采用樹莓派4B進(jìn)行推理測(cè)速得到。以上各模型均在pytorch環(huán)境下測(cè)試,torch版本為1.5.0,torchvision版本為0.6.0,opencv版本為3.4.6。
Note: The detection time for each algorithm in the table is measured with Raspberry Pi 4B. All the above models are tested in pytorch environment, with torch 1.5.0, torchvision 0.6.0 and opencv 3.4.6.
1)網(wǎng)絡(luò)剪枝,通常網(wǎng)絡(luò)模型參數(shù)過多有些權(quán)重接近0,或者神經(jīng)元的輸出為0,可以將這些多余的參數(shù)從網(wǎng)絡(luò)中移除。具體步驟為預(yù)訓(xùn)練一個(gè)比較龐大的模型,評(píng)估每個(gè)權(quán)重和神經(jīng)元的重要性,按照參數(shù)重要性排序,刪除不重要的參數(shù),將縮小的模型用訓(xùn)練數(shù)據(jù)重新微調(diào)一次,可以減小損失,如果模型縮小之后仍然沒達(dá)到要求則重新評(píng)估權(quán)重和神經(jīng)元迭代操作[84]。
2)知識(shí)蒸餾,基本思想是可以先訓(xùn)練一個(gè)規(guī)模大的初始網(wǎng)絡(luò),再訓(xùn)練一個(gè)小的子網(wǎng)絡(luò)去學(xué)習(xí)大的初始網(wǎng)絡(luò)的行為。使用初始網(wǎng)絡(luò)的輸出來訓(xùn)練而不直接使用標(biāo)注數(shù)據(jù),是因?yàn)槌跏季W(wǎng)絡(luò)可以提供更多的信息,輸入一個(gè)樣本后初始網(wǎng)絡(luò)會(huì)輸出各種類別的概率值,這比單純的標(biāo)簽信息要更豐富[85]。
3)參數(shù)量化,如果說網(wǎng)絡(luò)剪枝是通過減少權(quán)重的數(shù)量來壓縮模型,那么量化則是通過減少權(quán)重的大小來壓縮模型。量化通常是將大集合值映射到小集合值的過程,這意味著輸出包含的可能值范圍比輸入小,理想情況下在該過程中不會(huì)丟失太多信息[86]。參數(shù)量化會(huì)使用更少的空間的來存儲(chǔ)一個(gè)參數(shù),然后使用聚類中心來代替整個(gè)類的值,這樣可以減少參數(shù)的儲(chǔ)存[87]。
4)結(jié)構(gòu)優(yōu)化,通過調(diào)整網(wǎng)絡(luò)結(jié)構(gòu)使得其只需要較少的參數(shù),常見方法為低秩近似與切除分離卷積。深層神經(jīng)網(wǎng)絡(luò)通常存在大量重復(fù)參數(shù),不同層或通道之間存在許多相似性或冗余性,低秩近似的目標(biāo)是使用較少濾波器的線性組合來近似一個(gè)層的大量冗余濾波器,以這種方式壓縮層減少了網(wǎng)絡(luò)的內(nèi)存占用以及卷積運(yùn)算的計(jì)算復(fù)雜性,實(shí)現(xiàn)加速。切除分離卷積方法則是將計(jì)算進(jìn)行拆分,共用部分參數(shù),最終實(shí)現(xiàn)參數(shù)規(guī)??s小[88]。
邊緣計(jì)算具有高實(shí)時(shí)、低成本、低能耗的優(yōu)勢(shì),為深度學(xué)習(xí)人工智能算法部署提供新的技術(shù)途徑,其在農(nóng)業(yè)中的應(yīng)用正處于快速發(fā)展階段,為多場(chǎng)景智慧農(nóng)業(yè)發(fā)展提供具有成本效益的智能解決方案?,F(xiàn)有邊緣計(jì)算設(shè)備主要包括樹莓派、英偉達(dá)小型計(jì)算平臺(tái)、現(xiàn)場(chǎng)可編程邏輯門陣列、和移動(dòng)智能終端,受算力限制部署的人工智能算法主要是輕量化深度學(xué)習(xí)網(wǎng)絡(luò),且模型壓縮是加速邊緣計(jì)算的重要途徑。結(jié)合邊緣計(jì)算的智能農(nóng)業(yè)裝備主要包括智能農(nóng)業(yè)無人機(jī)、智能農(nóng)業(yè)機(jī)器人以及農(nóng)業(yè)智能移動(dòng)終端,旨在實(shí)現(xiàn)動(dòng)植物生長動(dòng)態(tài)監(jiān)測(cè)、環(huán)境實(shí)時(shí)檢測(cè)、和農(nóng)業(yè)裝備作業(yè)智能決策。就目前文獻(xiàn)分析,邊緣計(jì)算的農(nóng)業(yè)應(yīng)用主要包括環(huán)境監(jiān)測(cè)與病蟲害識(shí)別、作物生長及產(chǎn)量預(yù)測(cè)、農(nóng)業(yè)偵察與路徑規(guī)劃等方面,有效提升了工作效率。邊緣計(jì)算為農(nóng)業(yè)領(lǐng)域的各種復(fù)雜問題提供了高精度、實(shí)時(shí)性、低成本的解決方案,推動(dòng)邊緣計(jì)算在農(nóng)業(yè)中的應(yīng)用將進(jìn)一步促進(jìn)農(nóng)業(yè)數(shù)字化、智能化,為智慧農(nóng)業(yè)發(fā)展提供助力。隨著邊緣計(jì)算在農(nóng)業(yè)中的深入應(yīng)用,未來將面臨重大的挑戰(zhàn)與機(jī)遇。
1)多場(chǎng)景、多功能邊緣計(jì)算智能農(nóng)業(yè)裝備亟待開發(fā)
隨著人口老齡化加劇和城市化發(fā)展,越來越多的農(nóng)業(yè)生產(chǎn)環(huán)節(jié),需要智能農(nóng)業(yè)裝備來替代傳統(tǒng)人工,而邊緣計(jì)算將為農(nóng)業(yè)裝備提供高精度、低時(shí)延、低成本人工智能計(jì)算平臺(tái)?,F(xiàn)有的邊緣計(jì)算智能農(nóng)業(yè)裝備主要應(yīng)用于作物病蟲害識(shí)別與動(dòng)態(tài)生長監(jiān)測(cè),未來在動(dòng)物飼養(yǎng)管控,如疾病診斷、生長狀態(tài)監(jiān)測(cè)、智能飼喂;作物種植管控,如多功能表型檢測(cè)、精準(zhǔn)除草、變量施肥、智能采摘等領(lǐng)域亟待開發(fā)相關(guān)智能農(nóng)業(yè)裝備。
2)輕量化、高精度的邊緣計(jì)算人工智能算法亟待發(fā)展
隨著人工智能高速發(fā)展,深度學(xué)習(xí)在眾多領(lǐng)域得到了廣泛應(yīng)用,而近年來摩爾定律的逐步放緩,使得邊緣計(jì)算設(shè)備很難依靠硬件升級(jí)滿足復(fù)雜模型的需求,如何將人工智能模型前端化、輕量化,如何保證高精度的前提下盡可能壓縮模型提升效率,成為亟待解決的問題。因此,為實(shí)現(xiàn)邊緣計(jì)算農(nóng)業(yè)應(yīng)用大規(guī)模落地,發(fā)展輕量化、高精度的邊緣計(jì)算專用人工智能算法,實(shí)現(xiàn)模型精度與速度的平衡,是開發(fā)智能農(nóng)業(yè)裝備的重要前提。
3)云-邊緣協(xié)同、多機(jī)協(xié)作智能管控方法亟待研究
隨著邊緣計(jì)算節(jié)點(diǎn)數(shù)量增加,對(duì)異構(gòu)、分散的邊緣計(jì)算資源管理是未來將面臨的主要挑戰(zhàn)。隨著各種智能農(nóng)業(yè)裝備的研發(fā)與應(yīng)用,以云平臺(tái)為中心創(chuàng)建云邊協(xié)同、多機(jī)協(xié)作智能工作模式,對(duì)邊緣智能農(nóng)業(yè)裝備進(jìn)行統(tǒng)一管理,從數(shù)據(jù)、模型、應(yīng)用、安全等方面實(shí)現(xiàn)云端與邊緣設(shè)備之間的協(xié)同;制定相關(guān)的標(biāo)準(zhǔn)規(guī)范和通訊協(xié)議實(shí)現(xiàn)異構(gòu)邊緣設(shè)備之間交流,按照指定任務(wù)開展多機(jī)互助協(xié)作;建立統(tǒng)一的數(shù)據(jù)命名和標(biāo)注規(guī)范,開展云邊數(shù)據(jù)協(xié)同分析,進(jìn)一步提升數(shù)據(jù)處理效率。
[1] Armbrust M, Fox A, Griffith R, et al. A view of cloud computing[J]. Communications of the ACM, 2010, 53(4): 50-58.
[2] 李瑾,郭美榮,高亮亮. 農(nóng)業(yè)物聯(lián)網(wǎng)技術(shù)應(yīng)用及創(chuàng)新發(fā)展策略[J]. 農(nóng)業(yè)工程學(xué)報(bào),2015,31(增刊2):200-209.
Li Jin, Guo Meirong, Gao Liangliang. Application and innovation strategy of agricultural Internet of Things[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2015, 31(Suppl.2): 200-209. (in Chinese with English abstract)
[3] Cisco. 2022 Global Hybrid Cloud Trends Report[R]. USA. S&P Market Intelligence, 2022.
[4] Ashton K. That ‘internet of things’ thing[J]. RFID Journal, 2009, 22(7): 97-114.
[5] Zhang Y, Liu J. Prediction of overall energy consumption of data centers in different locations[J]. Sensors, 2022, 22(10): 3704.
[6] Zhu Y, Zhang Y, Wang J, et al. From data-driven to intelligent-driven: Technology evolution of network security in big data era[C]//2019 IEEE 43rd Annual Computer Software and Applications Conference (COMPSAC). IEEE, 2019, 2: 103-109.
[7] Fan Z, Yang W, Tian K. An edge computing service model based on information-centric networking[C]//2019 IEEE 25th International Conference on Parallel and Distributed Systems (ICPADS). IEEE, 2019: 498-505.
[8] 邊緣計(jì)算產(chǎn)業(yè)聯(lián)盟. 邊緣計(jì)算參考架構(gòu)2. 0(中)[J]. 自動(dòng)化博覽,2018,35(2):54-56.
[9] Zheng Z , Xie D , Pu J , et al. MELODY: Adaptive task definition of COP prediction with metadata for HVAC control and electricity saving[C]//e-Energy '20: The Eleventh ACM International Conference on Future Energy Systems. ACM, 2020
[10] 趙春江.智慧農(nóng)業(yè)發(fā)展現(xiàn)狀及戰(zhàn)略目標(biāo)研究[J].智慧農(nóng)業(yè),2019,1(1):1-7.
Zhao Chunjiang. State-of-the-art and recommended developmental strategic objectivs of smart agriculture[J]. Smart Agriculture, 2019, 1(1): 1-7. (in Chinese with English abstract)
[11] Chen Y. Discussing on relation between agricultural internet of things and agriculture complex large system intelligent control[J]. Agriculture Network Information, 2012, 2: 8-12.
[12] Lavanya G, Rani C, GaneshKumar P. An automated low cost IoT based fertilizer intimation system for smart agriculture[J]. Sustainable Computing: Informatics and Systems, 2020, 28: 100300.
[13] Chen X, Shi Q, Yang L, et al. ThriftyEdge: Resource-efficient edge computing for intelligent IoT applications[J]. IEEE Network, 2018, 32(1): 61-65.
[14] Jaiganesh S, Gunaseelan K, Ellappan V. IOT agriculture to improve food and farming technology[C]//2017 Conference on Emerging Devices and Smart Systems (ICEDSS). IEEE, 2017: 260-266.
[15] Henriksen A, Selmer-Olsen A R. Automatic methods for determining nitrate and nitrite in water and soil extracts[J]. Analyst, 1970, 95(1130): 514-518.
[16] Garcia Lopez P, Montresor A, Epema D, et al. Edge-centric computing: Vision and challenges[J]. ACM SIGCOMM Computer Communication Review, 2015, 45(5): 37-42.
[17] Shi W, Dustdar S. The promise of edge computing[J]. Computer, 2016, 49(5): 78-81.
[18] Chen J, Ran X. Deep learning with edge computing: A review[J]. Proceedings of the IEEE, 2019, 107(8): 1655-1674.
[19] Colbert I, Daly J, Kreutz-Delgado K, et al. A competitive edge: Can FPGAs beat GPUs at DCNN inference acceleration in resource-limited edge computing applications?[J/OL]. (2021-03-09) [2022-5-23].https://arxiv.org/abs/2102. 00294.
[20] 陳天華,盧思翰. 基于DSP的小型農(nóng)用無人機(jī)導(dǎo)航控制系統(tǒng)設(shè)計(jì)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2012,28(21):164-169.
Chen Tianhua, Lu Sihan. Autonomous navigation control system of agricultural mini-unmaned aerial vehicles based on DSP[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(21): 164-169. (in Chinese with English abstract)
[21] 孟曉麗,牛煥娜,賈東梨,等. 基于儲(chǔ)能Soc日前計(jì)劃的微電網(wǎng)實(shí)時(shí)能量優(yōu)化調(diào)度方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2016,32(8): 155-161.
Meng Xiaoli, Niu Huanna, Jia Dongli, et al. Real-time energy optimal dispatch for microgrid based on day-ahead scheduling of charge state[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(8): 155-161. (in Chinese with English abstract)
[22] Jolles J W. Broad-scale applications of the Raspberry Pi: A review and guide for biologists[J]. Methods in Ecology and Evolution, 2021, 12(9): 1562-1579.
[23] Wang X, Zhang W, Wu X, et al. Real-time vehicle type classification with deep convolutional neural networks[J]. Journal of Real-Time Image Processing, 2019, 16(1): 5-14.
[24] Avelar V. Cost benefit analysis of edge micro data center deployments[J]. Schneider Electric, 2017, 223: 1-8.
[25] 王梓儒. 深度目標(biāo)檢測(cè)模型的邊緣計(jì)算方案研究[D]. 北京:北京交通大學(xué),2020.
Wang Ziru. Research on Edge Caculation Scheme of Object Detection Model[D]. Beijing: Beijing Jiaotong University, 2020. (in Chinese with English abstract)
[26] 張釗. 面向邊緣計(jì)算的目標(biāo)檢測(cè)算法研究與應(yīng)用[D]. 大連:大連理工大學(xué),2021.
Zhang Zhao. Research and Application of Target Detection Algorithm for Edge Computing[D]. Dalian: Dalian University of Technology, 2021. (in Chinese with English abstract)
[27] Ma F, Wang B, Li M, et al. Edge intelligent perception method for power grid icing condition based on multi-scale feature fusion target detection and model quantization[J]. Frontiers in Energy Research, 2021,9: 591-602.
[28] Kim J, Lee J, Kim T. AdaMM: Adaptive object movement and motion tracking in hierarchical edge computing system[J]. Sensors, 2021, 21(12): 4089.
[29] 樊琦,李卓,陳昕. 基于邊緣計(jì)算的分支神經(jīng)網(wǎng)絡(luò)模型推斷延遲優(yōu)化[J]. 計(jì)算機(jī)應(yīng)用,2020,40(2):342-346.
Fan Qi, Li Zhuo, Chen Xin. Inference delay optimization of branchy neural network model based on edge computing[J]. Journal of Computer Applications, 2020, 40(2): 342-346. (in Chinese with English abstract)
[30] Oteros-Rozas E, Ruiz-Almeida A, Aguado M, et al. A social–ecological analysis of the global agrifood system[J]. Proceedings of the National Academy of Sciences, 2019, 116(52): 26465-26473.
[31] 聶鵬程,張慧,耿洪良,等. 農(nóng)業(yè)物聯(lián)網(wǎng)技術(shù)現(xiàn)狀與發(fā)展趨勢(shì)[J]. 浙江大學(xué)學(xué)報(bào)(農(nóng)業(yè)與生命科學(xué)版),2021,47(2):135-146.
Nie Pengcheng, Zhang Hui, Geng Hongliang, et al. Current situation and development trend of agricultural Internet of Things technology[J]. Journal of Zhejiang University (Agriculture and Life Sciences), 2021, 47(2): 135-146. (in Chinese with English abstract)
[32] 曹冰雪,李瑾,馮獻(xiàn),等. 我國智慧農(nóng)業(yè)的發(fā)展現(xiàn)狀、路徑與對(duì)策建議[J]. 農(nóng)業(yè)現(xiàn)代化研究,2021,42(5):785-794.
Cao Bingxue, Li Jin, Feng Xian, et al. Development status, path, and countermeasures of smart agriculture in China[J]. Research of Agricultural Modernization, 2021, 42(5): 785-794. (in Chinese with English abstract)
[33] 葛文杰,趙春江. 農(nóng)業(yè)物聯(lián)網(wǎng)研究與應(yīng)用現(xiàn)狀及發(fā)展對(duì)策研究[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2014,45(7):222-230.
Ge Wenjie, Zhao Chunjiang. State-of-the-art and developing trategies of agricultural internet of things[J]. Transactions of the Chinese Society for Agricultural Machinery, 2014, 45(7): 222-230. (in Chinese with English abstract)
[34] 鄭紀(jì)業(yè),阮懷軍,封文杰,等. 農(nóng)業(yè)物聯(lián)網(wǎng)體系結(jié)構(gòu)與應(yīng)用領(lǐng)域研究進(jìn)展[J]. 中國農(nóng)業(yè)科學(xué),2017,50(4):657-668.
Zheng Jiye, Ruan Huaijun, Feng Wenjie, et al. Agricultural lOT architecture and application model research[J]. Scientia Agricultura Sinica, 2017, 50(4): 657-668. (in Chinese with English abstract)
[35] Buyya R, Yeo C S, Venugopal S, et al. Cloud computing and emerging IT platforms: Vision, hype, and reality for delivering computing as the 5th utility[J]. Future Generation Computer Systems, 2009, 25(6): 599-616.
[36] Varghese B, Wang N, Barbhuiya S, et al. Challenges and opportunities in edge computing[C]//IEEE International Conference on Smart Cloud. Piscataway. NewYork: IEEE, 2016: 20-26.
[37] Alharbi H A, Elgorashi T E H, Elmirghani J M H. Energy efficient virtual machines placement over cloud-fog network architecture[J]. IEEE Access, 2020, 8: 94697-94718.
[38] 葉惠卿. 基于邊緣計(jì)算的農(nóng)業(yè)物聯(lián)網(wǎng)系統(tǒng)的研究[J]. 無線互聯(lián)科技,2019,16(10):30-32.
Ye Huiqing. Research on agricultural lnternet of Things system based on edge computing[J]. Wireless Internet Technology, 2019, 16(10): 30-32. (in Chinese with English abstract)
[39] Zheng B G. Study on the agricultural internet of things key technology of the intelligent control of sunlight greenhouse complex system[C]//Advanced Materials Research. Trans Tech Publications Ltd, 2013, 756: 2369-2373.
[40] 劉蘇偉. 基于邊緣計(jì)算的玉米病害識(shí)別深度學(xué)習(xí)應(yīng)用研究[D]. 哈爾濱:東北農(nóng)業(yè)大學(xué),2020.
Liu Suwei. The Study on Deep Learning of Maize Disease Recognition Based on Edge Computing[D]. Harbin: Northeast Agricultural University, 2020. (in Chinese with English abstract)
[41] 牛愷銳,張正華,包飛霞,等. 基于邊緣計(jì)算的作物病蟲害監(jiān)測(cè)嵌入式系統(tǒng)設(shè)計(jì)[J]. 計(jì)算機(jī)與網(wǎng)絡(luò),2021,47(14):61-65.
Niu Kairui, Zhang Zhenghua, Bao Feixia, et al. Design on Embedded System for Crop Diseases and Pests Monitoring Based on Edge Computing[J]. Computer & Network, 2021, 47(14): 61-65. (in Chinese with English abstract)
[42] 李鳳迪. 基于深度學(xué)習(xí)和無人機(jī)遙感的松材線蟲病樹監(jiān)測(cè)系統(tǒng)研究[D]. 泰安:山東農(nóng)業(yè)大學(xué),2020.
Li Fengdi. A Study on Monitoring System of Pine Nematode Tree Based on Deep Learning and UAV Remote Sensing[D]. Taian: Shandong Agricultural University, 2020. (in Chinese with English abstract)
[43] 孫志朋. 基于物聯(lián)網(wǎng)技術(shù)的水稻生長環(huán)境監(jiān)測(cè)系統(tǒng)及關(guān)鍵技術(shù)研究[D]. 長春:吉林大學(xué),2021.
Sun Zhipeng. Research on Monitoring System and Key Technologies Ofrice growth Environment Based on Internet of Things[D]. Changchun: Jilin University, 2021. (in Chinese with English abstract)
[44] Guillén M A, Llanes A, Imbernón B, et al. Performance evaluation of edge-computing platforms for the prediction of low temperatures in agriculture using deep learning[J]. The Journal of Supercomputing, 2021, 77(1): 818-840.
[45] Park J, Choi J H, Lee Y J, et al. A layered features analysis in smart farm environments[C]//Proceedings of the International Conference on Big Data and Internet of Thing. 2017: 169-173.
[46] Coviello L, Cristoforetti M, Jurman G, et al. GBCNet: In-field grape berries counting for yield estimation by dilated CNNs[J]. Applied Sciences, 2020, 10(14): 4870.
[47] Yang M D, Boubin J G, Tsai H P, et al. Adaptive autonomous UAV scouting for rice lodging assessment using edge computing with deep learning EDANet[J]. Computers and Electronics in Agriculture, 2020, 179: 105817
[48] Chen C J, Huang Y Y, Li Y S, et al. Identification of fruit tree pests with deep learning on embedded drone to achieve accurate pesticide spraying[J]. IEEE Access, 2021, 9: 21986-21997.
[49] Nilay K, Mandal S, Agarwal Y, et al. A proposal of FPGA-based low cost and power efficient autonomous fruit harvester[C]//2020 6th International Conference on Control, Automation and Robotics (ICCAR). IEEE, 2020: 324-329.
[50] Chechliński ?, Siemi?tkowska B, Majewski M. A system for weeds and crops identification-reaching over 10 FPS on raspberry pi with the usage of MobileNets, DenseNet and custom modifications[J]. Sensors, 2019, 19(17): 3787.
[51] Meng L, Hirayama T, Oyanagi S. Underwater-drone with panoramic camera for automatic fish recognition based on deep learning[J]. Ieee Access, 2018, 6: 17880-17886.
[52] Seo J, Ahn H, Kim D, et al. Embedded pig det-fast and accurate pig detection for embedded board implementations[J]. Applied Sciences, 2020, 10(8): 2878.
[53] Deng J, Zhong Z, Huang H, et al. Lightweight semantic segmentation network for real-time weed mapping using unmanned aerial vehicles[J]. Applied Sciences, 2020, 10(20): 7132.
[54] Wang J, Gao Z, Zhang Y, et al. Real-time detection and location of potted flowers based on a ZED camera and a YOLO V4-Tiny deep learning algorithm[J]. Horticulturae, 2021, 8(1): 21.
[55] Camargo T D, Schirrmann M, Landwehr N, et al. Optimized deep learning model as a basis for fast UAV mapping of weed species in winter wheat crops[J]. Remote Sensing, 2021, 13(9): 1-19.
[56] Liu Y, Su J, Shen L, et al. Development of a mobile application for identification of grapevine (Vitis vinifera L. ) cultivars via deep learning[J]. International Journal of Agricultural and Biological Engineering, 2021, 14(5): 172-179.
[57] He Y, Zeng H, Fan Y, et al. Application of deep learning in integrated pest management: A real-time system for detection and diagnosis of oilseed rape pests[J]. Mobile Information Systems, 2019, 2019: 1-14.
[58] Ahmed A A, Reddy G H. A mobile-based system for detecting plant leaf diseases using deep learning[J]. AgriEngineering, 2021, 3(3): 478-493.
[59] Ai Y, Sun C, Tie J, et al. Research on recognition model of crop diseases and insect pests based on deep learning in harsh environments[J]. IEEE Access, 2020, 8: 171686-171693.
[60] Buzzy M, Thesma V, Davoodi M, et al. Real-time plant leaf counting using deep object detection networks[J]. Sensors, 2020, 20(23): 6896.
[61] Mishra S, Sachan R, Rajpal D. Deep convolutional neural network based detection system for real-time corn plant disease recognition[J]. Procedia Computer Science, 2020, 167: 2003-2010.
[62] Partel V, Kakarla S C, Ampatzidis Y. Development and evaluation of a low-cost and smart technology for precision weed management utilizing artificial intelligence[J]. Computers and Electronics in Agriculture, 2019, 157: 339-350.
[63] Tarek H, Aly H, Eisa S, et al. Optimized deep learning algorithms for tomato leaf disease detection with hardware deployment[J]. Electronics, 2022, 11(1): 140.
[64] Emebo O, Fori B, Victor G, et al. Development of tomato septoria leaf spot and tomato mosaic diseases detection device using raspberry Pi and deep convolutional neural networks[C]//Journal of Physics: Conference Series. IOP Publishing, 2019, 1299(1): 012118.
[65] Tufail M, Iqbal J, Tiwana M I, et al. Identification of tobacco crop based on machine learning for a precision agricultural sprayer[J]. IEEE Access, 2021, 9: 23814-23825.
[66] Xu W C, Chen K, Zhang M, et al. Life lifting scheme for plant protection UAV based on solar[J]. Journal of Agricultural Science and Technology (Beijing), 2018, 20(11): 62-68.
[67] Wang R, Sun B Y. Development status and expectation of agricultural robot[J]. Bulletin of Chinese Academy of Sciences, 2015, 30(6): 803-809.
[68] Alzarliani W D, Arsyad M, Salam M, et al. Effect of App utilization on the agricultural efficiency of rural communities[C]//IOP Conference Series: Earth and Environmental Science. IOP Publishing, 2019, 235(1): 012101.
[69] Yue J, Lei T, Li C, et al. The application of unmanned aerial vehicle remote sensing in quickly monitoring crop pests[J]. Intelligent Automation & Soft Computing, 2012, 18(8): 1043-1052.
[70] Ukaegbu U F, Tartibu L K, Okwu M O, et al. Development of a light-weight unmanned aerial vehicle for precision agriculture[J]. Sensors, 2021, 21(13): 4417.
[71] Kun B, Xin Z, Ruifeng H. The trend of application and development of robot technology in agriculture[J]. China Agricultural Science Bulletin, 2011, 27(4): 469-473.
[72] Laamrani A, Pardo Lara R, Berg A A, et al. Using a mobile device “app” and proximal remote sensing technologies to assess soil cover fractions on agricultural fields[J]. Sensors, 2018, 18(3): 708.
[73] Filho C P, Marques Jr E, Chang V, et al. A systematic literature review on distributed machine learning in edge computing[J]. Sensors, 2022, 22(7): 2665.
[74] Huang Z, Yang S, Zhou M C, et al. Making accurate object detection at the edge: Review and new approach[J]. Artificial Intelligence Review, 2021: 1-30.
[75] Sheng Q, Sheng H, Gao P, et al. Real-time detection of cook assistant overalls based on embedded reasoning[J]. Sensors, 2021, 21(23): 8069.
[76] Su M, Tan J, Lin C Y, et al. Constructing a mobility and acceleration computing platform with NVIDIA Jetson TK1[C]//2015 IEEE 17th International Conference on High Performance Computing and Communications, 2015 IEEE 7th International Symposium on Cyberspace Safety and Security, and 2015 IEEE 12th International Conference on Embedded Software and Systems. IEEE, 2015: 1854-1858.
[77] Kundu N, Rani G, Dhaka V S, et al. IoT and interpretable machine learning based framework for disease prediction in pearl millet[J]. Sensors, 2021, 21(16): 5386.
[78] Fontaine J, Shahid A, Elsas R, et al. Multi-band sub-GHz technology recognition on NVIDIA’s Jetson Nano[C]//2020 IEEE 92nd Vehicular Technology Conference (VTC2020-Fall). IEEE, 2020: 1-7.
[79] Liu L, Yu L, Wu D, et al. PocketMaize: An android-smartphone application for maize plant phenotyping[J]. Frontiers in Plant Science, 2021, 12: 770217.
[80] Schmidhuber J. Deep learning in neural networks: An overview[J]. Neural Networks, 2015, 61: 85-117.
[81] Singh D, Banyal R K, Sharma A K. Cloud computing research issues, challenges, and future directions[C]//Emerging Trends in Expert Applications and Security. Singapore, Springer, 2019: 617-623.
[82] Wang Q, Jin G, Li Q, et al. Industrial edge computing: Vision and challenges[J]. Information and Control, 2021, 50(3): 257-274.
[83] Islam A, Debnath A, Ghose M, et al. A survey on task offloading in multi-access edge computing[J]. Journal of Systems Architecture, 2021, 118: 102225.
[84] 林景棟,吳欣怡,柴毅,等. 卷積神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)優(yōu)化綜述[J].自動(dòng)化學(xué)報(bào),2020,46(1):24-37.
Lin Jingdong, Wu Xinyi, Chai Yi, et al. Structure optimization of convolutional neural networks: A survey[J]. Acta Automatica Sinica, 2020, 46(1): 24-37. (in Chinese with English abstract)
[85] Gou J, Yu B, Maybank S J, et al. Knowledge distillation: A survey[J]. International Journal of Computer Vision, 2021, 129(6): 1789-1819.
[86] 彭紅星,徐慧明,劉華鼐. 基于改進(jìn)ShuffleNet V2的輕量化農(nóng)作物害蟲識(shí)別模型[J]. 農(nóng)業(yè)工程學(xué)報(bào),2022,38(11):161-170.
Peng Hongxing, Xu Huiming, Liu Huanai. Light weight agricultural crops pest identification model using improved ShuffleNet V2[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2022, 38(11): 161-170. (in Chinese with English abstract)
[87] 李江昀,趙義凱,薛卓爾,等. 深度神經(jīng)網(wǎng)絡(luò)模型壓縮綜述[J]. 工程科學(xué)學(xué)報(bào),2019,41(10):1229-1239.
Li Jiangyun, Zhao Yikai, Xue Zhuoer, et al. A survey of model compression for deep neural networks[J]. Chinese Journal of Engineering, 2019, 41(10): 1229-1239. (in Chinese with English abstract)
[88] Ji R, Lin S, Chao F, Wu Y, et al. Deep neural network compression and acceleration: A review[J]. Journal of Computer Research and Development, 2018, 55(9): 1871-1888.
Application status and prospect of edge computing in smart agriculture
Huang Chenglong1, Ke Yuxi1, Hua Xiangdong1, Yang Junya1, Sun Mengyu1, Yang Wanneng2
(1.,,430070,; 2.,,430070,)
A large amount of data has been produced with the rapid development of internet technology. The commonly-used centralized processing has posed rigorous challenges to real-time performance, low energy consumption, and data security. Alternatively, edge computing combined with Artificial Intelligence (AI) can be used to reduce the cost and energy consumption for real-time detection of complex data processing in various industries. Nowadays, agricultural applications combined with deep learning have been widely reported, such as crop disease detection, growth monitoring, yield prediction, and automated management. Edge computing can also be expected to provide more efficient solutions with the rapid development of smart agriculture. In this review, the history, concept, and architecture of edge computing were firstly introduced to evaluate the performance in intelligent agriculture. Specifically, the statistical analysis of the literature volume was carried out until May 2022, including the most reported disease identification and environmental monitoring. Secondly, the main devices of edge computing were summarized, including the Raspberry Pi, FPGA devices, NVIDIA Jetson, and smartphones. The performances of edge computing devices were also compared under different scenarios. Besides, the commonly-used deep learning was selected to promote efficiency and accuracy using the Raspberry pie 4B. Some model acceleration methods were also introduced, including network pruning, knowledge distillation, parameter quantification, and structure optimization. Then, the AI agricultural equipment with edge computing was divided into unmanned aerial vehicle (UAV), ground robots, and portable devices. Three scenarios were considered in the agriculture application, such as environmental monitoring and pest identification, crop growth and yield prediction, and variable operation of intelligent agricultural equipment. Finally, the prospects and key issues were proposed for the edge computing applied in agriculture. Several suggestions were also drawn during this time. Specifically, the edge computing application should be developed with high efficiency and accuracy. The model compression and acceleration can be the key research direction in the model deployment of deep learning. Edge computing devices can greatly contribute to smart agriculture. The cost-saving AI agricultural equipment with edge computing can also be expected to develop for much more application scenarios. The communication protocols and standards between edge devices should be established to realize the cooperative operation of multiple machines. In conclusion, edge computing was still in the initial and rapid development stage in smart agriculture. Edge computing can also provide vital opportunities and challenges for the development of smart agriculture, due to the better real-time, lower cost, and energy consumption, compared with the current cloud computing.
internet of things; edge computing; cloud computing; smart agriculture; deep learning; model compression; model deployment
10.11975/j.issn.1002-6819.2022.16.025
S126
A
1002-6819(2022)-16-0224-11
黃成龍,柯宇曦,華向東,等. 邊緣計(jì)算在智慧農(nóng)業(yè)中的應(yīng)用現(xiàn)狀與展望[J]. 農(nóng)業(yè)工程學(xué)報(bào),2022,38(16):224-234.doi:10.11975/j.issn.1002-6819.2022.16.025 http://www.tcsae.org
Huang Chenglong, Ke Yuxi, Hua Xiangdong, et al. Application status and prospect of edge computing in smart agriculture[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2022, 38(16): 224-234. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2022.16.025 http://www.tcsae.org
2022-05-12
2022-08-11
國家自然科學(xué)基金項(xiàng)目(32270431,U21A20205);中央高校基本科研業(yè)務(wù)費(fèi)項(xiàng)目(2662022YJ018)
黃成龍,博士,副教授,研究方向?yàn)檗r(nóng)業(yè)技術(shù)與裝備/植物表型。Email:hcl@mail.hzau.edu.cn