深度學(xué)習(xí)技術(shù)在農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估中的應(yīng)用*

黃睿茜，趙俊芳，霍治國(guó)，彭慧文，謝鴻飛

深度學(xué)習(xí)技術(shù)在農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估中的應(yīng)用*

黃睿茜，趙俊芳**，霍治國(guó)，彭慧文，謝鴻飛

（中國(guó)氣象科學(xué)研究院災(zāi)害天氣國(guó)家重點(diǎn)實(shí)驗(yàn)室，北京 100081）

人工智能技術(shù)的發(fā)展，特別是深度學(xué)習(xí)的出現(xiàn)，推進(jìn)了農(nóng)業(yè)新發(fā)展，是農(nóng)業(yè)現(xiàn)代化生產(chǎn)的新方向。深度學(xué)習(xí)具有學(xué)習(xí)能力強(qiáng)、覆蓋范圍廣、適應(yīng)力強(qiáng)、可移植性好等優(yōu)點(diǎn)，其開(kāi)發(fā)模擬數(shù)據(jù)集可以解決實(shí)際問(wèn)題，在農(nóng)業(yè)干旱的監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估中的應(yīng)用越來(lái)越廣泛。本文采用文獻(xiàn)綜述方法，歸納農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估技術(shù)的發(fā)展與應(yīng)用，總結(jié)深度學(xué)習(xí)模型的原理、優(yōu)勢(shì)和不足，概述深度學(xué)習(xí)模型在農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估方面的實(shí)際應(yīng)用，探討深度學(xué)習(xí)數(shù)據(jù)集要求大、數(shù)據(jù)預(yù)處理耗時(shí)長(zhǎng)、預(yù)定義類別范圍窄、遙感圖像復(fù)雜的問(wèn)題，并對(duì)未來(lái)研究方向進(jìn)行展望。結(jié)果表明，近年來(lái)農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估技術(shù)取得重要進(jìn)展，然而由于農(nóng)業(yè)系統(tǒng)的非線性及干旱災(zāi)害發(fā)生的復(fù)雜性，現(xiàn)有技術(shù)在適用地域、對(duì)象和精準(zhǔn)性等方面仍難以滿足新形勢(shì)下實(shí)際農(nóng)業(yè)生產(chǎn)的需求。深度學(xué)習(xí)方法為農(nóng)業(yè)干旱研究提供了新手段，但深度學(xué)習(xí)模型無(wú)法準(zhǔn)確表達(dá)作物生長(zhǎng)具體過(guò)程與機(jī)理，可嘗試探索通過(guò)深度學(xué)習(xí)模型和作物生長(zhǎng)模型的耦合來(lái)確保深度學(xué)習(xí)模型的可解釋性。對(duì)于修正預(yù)測(cè)序列中存在的誤差，可建立基于深度學(xué)習(xí)模型與大氣環(huán)流模式的耦合模型，進(jìn)一步提升模型對(duì)中長(zhǎng)期農(nóng)業(yè)干旱的預(yù)測(cè)能力。針對(duì)災(zāi)害樣本容量有限問(wèn)題，加強(qiáng)基于深度學(xué)習(xí)和遷移學(xué)習(xí)的農(nóng)業(yè)干旱監(jiān)測(cè)與評(píng)估研究，可進(jìn)一步提高農(nóng)業(yè)干旱精細(xì)化監(jiān)測(cè)與評(píng)估精度。針對(duì)影響農(nóng)業(yè)干旱形成的因子具有數(shù)據(jù)量大、類型多樣、非線性的特點(diǎn)，采用深度學(xué)習(xí)與信息融合相結(jié)合的方法，可進(jìn)一步提高區(qū)域農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估精度。因此，深度學(xué)習(xí)模型與作物生長(zhǎng)模型的耦合、融合深度學(xué)習(xí)模型和大氣環(huán)流模式的農(nóng)業(yè)干旱預(yù)測(cè)、基于深度學(xué)習(xí)與遷移學(xué)習(xí)的農(nóng)業(yè)干旱精細(xì)化監(jiān)測(cè)與評(píng)估、深度學(xué)習(xí)與信息融合技術(shù)相結(jié)合的區(qū)域農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估是未來(lái)深度學(xué)習(xí)技術(shù)在農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估中應(yīng)用的發(fā)展趨勢(shì)。

深度學(xué)習(xí)；農(nóng)業(yè)干旱；監(jiān)測(cè)預(yù)測(cè)；風(fēng)險(xiǎn)評(píng)估；精度

在過(guò)去的一百年里，地球氣候正經(jīng)歷著以全球變暖為特征的重大變化。全球變暖對(duì)自然生態(tài)和人類生存環(huán)境影響顯著，改變了區(qū)域降水、蒸散、土壤水分、徑流等水文因子的循環(huán)過(guò)程，導(dǎo)致干旱等極端天氣事件和重大氣象災(zāi)害頻繁發(fā)生[1]。干旱具有發(fā)生頻率高、時(shí)間周期長(zhǎng)、危害范圍廣的特點(diǎn)[2]，其嚴(yán)重性和不可控性常造成經(jīng)濟(jì)和財(cái)產(chǎn)損失，破壞生態(tài)環(huán)境，威脅國(guó)家糧食安全和可持續(xù)發(fā)展[3]。

農(nóng)業(yè)是受氣候和天氣限制最大的領(lǐng)域。中國(guó)是農(nóng)業(yè)大國(guó)，干旱是中國(guó)農(nóng)業(yè)領(lǐng)域主要發(fā)生的自然災(zāi)害。2022年夏季，中國(guó)多地發(fā)布了干旱預(yù)警。對(duì)1949-2015年自然災(zāi)害的回顧表明，各種自然災(zāi)害中旱災(zāi)位列首位，其次為洪災(zāi)、風(fēng)雹、低溫和臺(tái)風(fēng)[4]。根據(jù)《中國(guó)水旱災(zāi)害公報(bào)》的統(tǒng)計(jì)，近年來(lái)中國(guó)年均農(nóng)作物受災(zāi)面積呈逐年增加的趨勢(shì)，從20世紀(jì)50年代的531.7萬(wàn)hm2增至90年代的1384.2萬(wàn)hm2，每年因干旱造成的糧食損失基本保持在300億kg，造成工業(yè)和農(nóng)業(yè)直接經(jīng)濟(jì)損失近1000億元[5]。因此，能否對(duì)干旱等農(nóng)業(yè)重大氣象災(zāi)害進(jìn)行準(zhǔn)確監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估，減少其負(fù)面影響已成為當(dāng)前迫切需要解決的重大問(wèn)題，這對(duì)全面推進(jìn)鄉(xiāng)村振興、防御和減輕農(nóng)業(yè)氣象災(zāi)害、提升災(zāi)害性天氣監(jiān)測(cè)預(yù)測(cè)準(zhǔn)確率、健全農(nóng)業(yè)氣象防災(zāi)減災(zāi)體系、保障國(guó)家糧食生產(chǎn)安全等均具有重要的現(xiàn)實(shí)意義和社會(huì)價(jià)值。

近幾十年來(lái)，國(guó)內(nèi)外學(xué)者在深度學(xué)習(xí)技術(shù)在農(nóng)業(yè)干旱中的應(yīng)用方面開(kāi)展了大量的科研工作，特別是在農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估指標(biāo)、農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)技術(shù)、農(nóng)業(yè)干旱風(fēng)險(xiǎn)評(píng)估技術(shù)等方面，取得了一系列研究成果?；诖耍疚膹霓r(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估的角度對(duì)現(xiàn)有研究成果進(jìn)行概述，系統(tǒng)梳理了國(guó)內(nèi)外農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估技術(shù)的發(fā)展與應(yīng)用，分析傳統(tǒng)農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估技術(shù)的優(yōu)劣勢(shì)，總結(jié)深度學(xué)習(xí)模型用于干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估的優(yōu)勢(shì)，概述深度學(xué)習(xí)模型在農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估方面的實(shí)際應(yīng)用，探討深度學(xué)習(xí)數(shù)據(jù)集要求大、數(shù)據(jù)預(yù)處理耗時(shí)長(zhǎng)、預(yù)定義類別范圍窄、遙感圖像復(fù)雜的問(wèn)題，提出未來(lái)深度學(xué)習(xí)技術(shù)在農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估中應(yīng)用的發(fā)展趨勢(shì)，以期為農(nóng)業(yè)氣象災(zāi)害精細(xì)化監(jiān)測(cè)預(yù)測(cè)、農(nóng)業(yè)氣象災(zāi)害損失精準(zhǔn)評(píng)估及農(nóng)業(yè)應(yīng)對(duì)氣候變化等提供科學(xué)參考。

1 深度學(xué)習(xí)技術(shù)用于農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估的適用性

1.1 農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估指標(biāo)

干旱可分為氣象、水文、農(nóng)業(yè)和社會(huì)經(jīng)濟(jì)干旱。農(nóng)業(yè)干旱是指在作物生長(zhǎng)期間，作物缺水影響作物生長(zhǎng)發(fā)育的現(xiàn)象。這是一種范圍最廣、頻率最高、災(zāi)害和影響最嚴(yán)重的干旱類型，也是對(duì)農(nóng)業(yè)生產(chǎn)影響最為嚴(yán)重的氣象災(zāi)害[6]。據(jù)統(tǒng)計(jì)，中國(guó)農(nóng)作物旱災(zāi)成災(zāi)面積大約占總成災(zāi)面積的52.53%，受災(zāi)頻率約為14.25a一遇，且周期呈逐漸縮小的趨勢(shì)[7]。

農(nóng)業(yè)干旱的發(fā)生受大氣、作物、土壤等有關(guān)因素的影響。因此，農(nóng)業(yè)干旱的監(jiān)測(cè)預(yù)測(cè)評(píng)估指標(biāo)大致可分為三類，即水分指標(biāo)、溫度指標(biāo)和綜合指標(biāo)。其中，水分指標(biāo)常用的有降水距平百分比、標(biāo)準(zhǔn)化降水指數(shù)[8]、連續(xù)無(wú)雨天數(shù)等降水指標(biāo)，土壤有效水分貯存量、綜合旱澇指標(biāo)等土壤水分指標(biāo)，以及葉水勢(shì)、氣孔開(kāi)度、光合等作物水分指標(biāo)。溫度指標(biāo)常用的是植物冠層的溫度。綜合指標(biāo)常用的是植被指數(shù)及作物干旱指數(shù)和水分指數(shù)等[9]。在農(nóng)業(yè)干旱的監(jiān)測(cè)預(yù)測(cè)評(píng)估指標(biāo)中，標(biāo)準(zhǔn)化降水指數(shù)受地理、氣候條件限制較小[10?11]，常用于深度學(xué)習(xí)評(píng)估農(nóng)業(yè)干旱[12]。

1.2 傳統(tǒng)農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估技術(shù)的優(yōu)劣勢(shì)

農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估技術(shù)主要包括模糊綜合評(píng)價(jià)、信息擴(kuò)散技術(shù)、統(tǒng)計(jì)預(yù)報(bào)、數(shù)值預(yù)報(bào)、隨機(jī)森林模型、作物生長(zhǎng)模型等方法[13?18]。其中，模糊綜合評(píng)價(jià)法主要以模糊推理為主，結(jié)合定性和定量及精確與非精確開(kāi)展綜合分析[15]。運(yùn)用模糊隸屬函數(shù)計(jì)算單個(gè)樣本在整個(gè)論域區(qū)間上的“隸屬度”分布，可克服傳統(tǒng)概率統(tǒng)計(jì)法對(duì)樣本數(shù)量要求高，以及對(duì)總體分布做假設(shè)所帶來(lái)的誤差等缺點(diǎn)[19]。因此，利用模糊數(shù)學(xué)理論，如信息擴(kuò)散方法[20?21]，結(jié)合干旱評(píng)價(jià)指標(biāo)和干旱評(píng)價(jià)理論，可以對(duì)干旱進(jìn)行更科學(xué)、合理的評(píng)價(jià)[18]。統(tǒng)計(jì)預(yù)報(bào)方法是基于數(shù)理統(tǒng)計(jì)理論揭示氣象要素的變化規(guī)律及與預(yù)報(bào)因子、預(yù)報(bào)量等之間的關(guān)系，并利用數(shù)學(xué)模式對(duì)未來(lái)的農(nóng)業(yè)干旱進(jìn)行預(yù)測(cè)，如運(yùn)用灰色-Markov鏈模型對(duì)旱澇情況開(kāi)展預(yù)測(cè)以及基于經(jīng)驗(yàn)正交函數(shù)的農(nóng)業(yè)干旱預(yù)測(cè)[22]。數(shù)值預(yù)報(bào)方法主要是在一定初值和邊界條件下，運(yùn)用天氣預(yù)報(bào)數(shù)值計(jì)算來(lái)開(kāi)展農(nóng)業(yè)干旱程度的預(yù)報(bào)，如MOS預(yù)報(bào)模型的運(yùn)用。統(tǒng)計(jì)建模使用氣象觀測(cè)數(shù)據(jù)和作物資料，通過(guò)統(tǒng)計(jì)回歸量化它們之間的關(guān)聯(lián)，便于大規(guī)模應(yīng)用[23]。隨機(jī)森林模型的高準(zhǔn)確性是基于聚集大量決策樹(shù)，已有研究表明隨機(jī)森林模型在預(yù)測(cè)建模變量間復(fù)雜相互作用的能力較強(qiáng)，處理非線性關(guān)系、高階相關(guān)、評(píng)估變量等方面具有高精度的預(yù)測(cè)能力[24]。作物生長(zhǎng)模型是利用計(jì)算機(jī)來(lái)定量表達(dá)作物生長(zhǎng)發(fā)育過(guò)程及與環(huán)境的動(dòng)態(tài)關(guān)系，從而開(kāi)展作物的生長(zhǎng)發(fā)育、干物質(zhì)累積與分配、產(chǎn)量等的監(jiān)測(cè)預(yù)測(cè)，被認(rèn)為是目前定量評(píng)估氣候變化對(duì)農(nóng)業(yè)影響研究方面較理想的方法[9,18,25]。

近幾十年來(lái)，國(guó)內(nèi)外農(nóng)業(yè)干旱研究特別是干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估在技術(shù)的進(jìn)一步更新、新方法應(yīng)用和系統(tǒng)平臺(tái)建設(shè)等方面均取得了明顯進(jìn)展。20世紀(jì)80年代美國(guó)已領(lǐng)先進(jìn)行了大量干旱災(zāi)害管理研究，對(duì)區(qū)域季節(jié)性干旱風(fēng)險(xiǎn)展開(kāi)了深入研究，并創(chuàng)建了農(nóng)業(yè)旱災(zāi)風(fēng)險(xiǎn)評(píng)估模型和旱災(zāi)風(fēng)險(xiǎn)指標(biāo)，由此也實(shí)現(xiàn)了對(duì)干旱災(zāi)害危險(xiǎn)區(qū)域的地理空間范圍劃分標(biāo)準(zhǔn)與災(zāi)害層次界定[14]。澳大利亞應(yīng)急管理部門(mén)和氣象部門(mén)合作建立了旱災(zāi)損失系統(tǒng)[12]。意大利Todisco等提出了干旱經(jīng)濟(jì)風(fēng)險(xiǎn)評(píng)估（DERA）方法，強(qiáng)調(diào)了一般干旱指數(shù)（量化缺水量）與無(wú)法滿足需水量的經(jīng)濟(jì)影響之間關(guān)系的重要性。利用綜合嚴(yán)重度－持續(xù)時(shí)間－頻率（SDF）曲線，這種關(guān)系可以繪制干旱嚴(yán)重程度和相應(yīng)影響。該程序適用于意大利中部翁布里亞地區(qū)的農(nóng)業(yè)干旱（向日葵作物）[17]。Shahid等結(jié)合降水標(biāo)準(zhǔn)化指數(shù)與地理信息系統(tǒng)技術(shù)評(píng)價(jià)孟加拉西部不同時(shí)間尺度的干旱風(fēng)險(xiǎn)[18]。中國(guó)學(xué)者建立了以人工控制和大田試驗(yàn)以及災(zāi)害資料為基礎(chǔ)的農(nóng)業(yè)氣象災(zāi)害監(jiān)測(cè)預(yù)測(cè)及災(zāi)損評(píng)價(jià)的數(shù)學(xué)模型[18,25]。黃崇福等應(yīng)用概率統(tǒng)計(jì)和模糊風(fēng)險(xiǎn)模型相結(jié)合的方法，定量評(píng)估了湖南省農(nóng)業(yè)干旱災(zāi)害的風(fēng)險(xiǎn)，并達(dá)到了較好的成效[15]。Zhao等研發(fā)了基于TIGGE集合預(yù)報(bào)、衛(wèi)星遙感和分布式水文模型XXT相結(jié)合的農(nóng)業(yè)干旱動(dòng)態(tài)監(jiān)測(cè)預(yù)警技術(shù)[26]，進(jìn)一步提高了復(fù)雜地形條件下農(nóng)業(yè)干旱災(zāi)害的預(yù)測(cè)預(yù)警能力，并在重慶地區(qū)的農(nóng)業(yè)氣象業(yè)務(wù)服務(wù)中得到驗(yàn)證，應(yīng)用效果良好，促進(jìn)了省級(jí)農(nóng)業(yè)氣象業(yè)務(wù)發(fā)展，但是該研究沒(méi)有考慮具體的農(nóng)作物種類，在實(shí)際生產(chǎn)中仍無(wú)法滿足作物干旱預(yù)報(bào)的需求。

干旱發(fā)生過(guò)程是個(gè)非常復(fù)雜的非線性過(guò)程，干旱災(zāi)害具有動(dòng)態(tài)性、復(fù)雜性、緊迫性和不確定性等特征，現(xiàn)有干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估技術(shù)在適用地域、對(duì)象和精準(zhǔn)性等方面仍難以滿足新形勢(shì)下實(shí)際農(nóng)業(yè)生產(chǎn)的需求。多因子協(xié)同作用和多尺度疊加效應(yīng)引起的非線性問(wèn)題一直是農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估中的“卡脖子”技術(shù)瓶頸，迫切需要引入新的技術(shù)方法來(lái)提高農(nóng)業(yè)干旱災(zāi)害監(jiān)測(cè)預(yù)測(cè)評(píng)估精度。

1.3 深度學(xué)習(xí)技術(shù)用于干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估的優(yōu)勢(shì)

深度學(xué)習(xí)是一種利用人工智能技術(shù)來(lái)處理新型圖像和分析數(shù)據(jù)的方法，與人工神經(jīng)網(wǎng)絡(luò)有類似之處[27]。其中常見(jiàn)的有卷積神經(jīng)網(wǎng)絡(luò)、自編碼神經(jīng)網(wǎng)絡(luò)、深度置信網(wǎng)絡(luò)三類方法[28]。深度學(xué)習(xí)技術(shù)是一種特征表示技術(shù)，其核心從人工設(shè)計(jì)升級(jí)為學(xué)習(xí)系統(tǒng)自主獲得，以將原始數(shù)據(jù)從基本的非線性模型組合向更高層次轉(zhuǎn)型，對(duì)于影響因子復(fù)雜但要求農(nóng)業(yè)干旱精準(zhǔn)的監(jiān)測(cè)預(yù)測(cè)評(píng)估有優(yōu)勢(shì)[29?30]。農(nóng)業(yè)干旱的監(jiān)測(cè)預(yù)測(cè)評(píng)估所需數(shù)據(jù)集巨大，深度學(xué)習(xí)可以在不損失精度的情況下，將這些模型的尺寸壓縮數(shù)十倍。深度學(xué)習(xí)模型具有優(yōu)良的自動(dòng)特征提取功能，圖像處理時(shí)特征工程的繁瑣步驟被極大縮減，從而縮短了建模訓(xùn)練時(shí)間，使模型分析結(jié)果更加精確，運(yùn)行效率也相應(yīng)較高[31]；具有更好的信息效應(yīng)，提供了更好的分類效果[32]；深度學(xué)習(xí)模型在分類方面優(yōu)于統(tǒng)計(jì)計(jì)算模型方法；在影響因素方面，深度學(xué)習(xí)的性能與數(shù)據(jù)采集的多樣性、數(shù)據(jù)規(guī)模和完整性有重要關(guān)系。深度學(xué)習(xí)模型通過(guò)增加模型的復(fù)雜性，并通過(guò)多個(gè)抽象層次，使用以分層方式表示數(shù)據(jù)的各種函數(shù)轉(zhuǎn)換數(shù)據(jù)，擴(kuò)展了經(jīng)典機(jī)器學(xué)習(xí)[33]，可以快速解決更復(fù)雜的問(wèn)題，且允許大規(guī)模并行化[34?36]]。在識(shí)別農(nóng)業(yè)干旱圖像方面，深度學(xué)習(xí)可以廣泛識(shí)別并開(kāi)展圖像分析，提高了圖像識(shí)別和目標(biāo)檢測(cè)的質(zhì)量[37]。深度學(xué)習(xí)的圖像識(shí)別主要使用兩種分類過(guò)程，即基于像素的分類（PBC）和面向?qū)ο蠓诸悾∣OC）。面向?qū)ο蠓诸愂菍?duì)圖片上顯示的對(duì)象進(jìn)行分類，其主要特征可以從高分辨率衛(wèi)星數(shù)據(jù)的空間光譜特征信息中提取?；谙袼氐姆诸悾雎粤藖?lái)自訓(xùn)練的光譜響應(yīng)像素?cái)?shù)據(jù)集混合的影響，無(wú)法識(shí)別大于一個(gè)像素的對(duì)象[38]。相比之下，面向?qū)ο蠓诸愅ㄟ^(guò)使用光譜信息，如形狀、紋理等，可增加對(duì)象檢測(cè)精度。該方法通過(guò)非線性尺度空間濾波對(duì)圖像進(jìn)行分析，可提供不同種類的圖片。

總體來(lái)說(shuō)，深度學(xué)習(xí)技術(shù)可以將長(zhǎng)時(shí)間序列的氣象、土壤和遙感等多源、非線性數(shù)據(jù)融合為具有時(shí)空一致性的數(shù)據(jù)集，從數(shù)據(jù)中挖掘出以往未知的新信息，對(duì)于農(nóng)業(yè)干旱預(yù)測(cè)的復(fù)雜函數(shù)及圖像識(shí)別具有良好的效果[39]，已被廣泛應(yīng)用于農(nóng)業(yè)領(lǐng)域，如農(nóng)業(yè)氣象災(zāi)害監(jiān)測(cè)預(yù)測(cè)、風(fēng)險(xiǎn)評(píng)估、作物產(chǎn)量預(yù)估和農(nóng)業(yè)病蟲(chóng)害監(jiān)測(cè)等[40?41]。

2 深度學(xué)習(xí)技術(shù)在農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估中的應(yīng)用

2.1 用于農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)

對(duì)農(nóng)業(yè)氣象災(zāi)害開(kāi)展及時(shí)準(zhǔn)確的監(jiān)測(cè)預(yù)測(cè)可以有效降低農(nóng)業(yè)氣象災(zāi)害帶來(lái)的損失。農(nóng)業(yè)干旱形成原因涉及氣候特征、作物種植結(jié)構(gòu)、土壤特性、生產(chǎn)力水平、抗旱能力和管理水平等多方面的因素，具有數(shù)據(jù)量大、類型多樣和非線性等特點(diǎn)[19]。由于農(nóng)業(yè)干旱形成過(guò)程中多因子協(xié)同作用和多尺度疊加效應(yīng)引起的非線性等問(wèn)題[3]，使目前農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)研究還有很大的提升空間[11]。近年來(lái)，深度學(xué)習(xí)等人工智能技術(shù)的出現(xiàn)，推進(jìn)了農(nóng)業(yè)新發(fā)展，為農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)提供了新手段[31]。

Rhee等基于降水量、日間地表溫度、夜間地表溫度和歸一化差異植被指數(shù)等遙感數(shù)據(jù)，利用深度學(xué)習(xí)模型預(yù)測(cè)了未來(lái)6個(gè)月農(nóng)業(yè)干旱的發(fā)展趨勢(shì)[42]。Dikshit等在澳大利亞?wèn)|南部新南威爾士州進(jìn)行的一項(xiàng)研究發(fā)現(xiàn)，采用深度學(xué)習(xí)模型提高了該地農(nóng)業(yè)干旱的預(yù)測(cè)能力[43]。Lee等使用深度神經(jīng)網(wǎng)絡(luò)模型估算出的土壤濕度RMSE較低，表明深度學(xué)習(xí)模型可提高農(nóng)業(yè)干旱監(jiān)測(cè)的可靠性[44]。Darwin等基于機(jī)器視覺(jué)和深度學(xué)習(xí)模型，提出了虛擬分析和分類器相結(jié)合的作物產(chǎn)量檢測(cè)技術(shù)的多種自動(dòng)化方法[26]。Feng等使用三種機(jī)器學(xué)習(xí)模型，如BRF、SVM和MLPNN（多層感知器神經(jīng)網(wǎng)絡(luò)）制作農(nóng)業(yè)干旱分布圖，并將土壤濕度作為植被水分脅迫的重要指標(biāo)[45]。此外，學(xué)者們分析了農(nóng)業(yè)干旱和作物產(chǎn)量的關(guān)系及農(nóng)業(yè)干旱特征，將深度學(xué)習(xí)訓(xùn)練平臺(tái)搭建在本地GNU/Linux操作系統(tǒng)服務(wù)器中，基于Tensorflow、keras和theano、python等環(huán)境，采用卷積神經(jīng)網(wǎng)絡(luò)對(duì)作物產(chǎn)量進(jìn)行預(yù)測(cè)，用交叉驗(yàn)證和多模型比較的方法來(lái)測(cè)試模型。胡小楓等基于標(biāo)準(zhǔn)化降水、地表溫度、歸一化植被距平指數(shù)、土壤可用含水量AWC和氣象干旱指數(shù)SPEI等多源數(shù)據(jù)，利用深度學(xué)習(xí)技術(shù)構(gòu)建了京津冀地區(qū)綜合干旱評(píng)估模型，并輸出月尺度的綜合干旱指數(shù)CDI[46]。

將深度學(xué)習(xí)方法與支持向量機(jī)、邏輯回歸、隨機(jī)森林和決策樹(shù)等預(yù)測(cè)結(jié)果進(jìn)行對(duì)比，結(jié)果表明深度學(xué)習(xí)模型總體表現(xiàn)更好[47]。Agana和Homaifar在美國(guó)西南部Gunnison河流域進(jìn)行的一項(xiàng)案例研究表明，與其他模型如多層感知器（MLP）和支持向量回歸（SVR）相比，基于深度信念網(wǎng)絡(luò)（DBN）的深度學(xué)習(xí)方法在預(yù)測(cè)具有不同時(shí)間尺度的長(zhǎng)期干旱方面優(yōu)于其他模型，在預(yù)測(cè)長(zhǎng)期干旱時(shí)誤差最小[48]。此外，深度學(xué)習(xí)方法在處理衛(wèi)星遙感方面的數(shù)據(jù)是有效的，能夠識(shí)別和分類目標(biāo)以及探測(cè)環(huán)境和結(jié)構(gòu)特征[49]。遙感技術(shù)在作物產(chǎn)量預(yù)測(cè)中具有準(zhǔn)確性和可靠性，為具有計(jì)算機(jī)視覺(jué)和深度學(xué)習(xí)模型的圖像分析中的自動(dòng)化提供了精確的場(chǎng)和產(chǎn)量圖。隨著中國(guó)衛(wèi)星遙感技術(shù)與計(jì)算機(jī)人工智能技術(shù)應(yīng)用的迅速發(fā)展[50]，遙感影像將更加豐富多彩，深度學(xué)習(xí)技術(shù)也越來(lái)越完善[51]。融合遙感數(shù)據(jù)和深度學(xué)習(xí)技術(shù)在農(nóng)作物災(zāi)情圖像識(shí)別方面顯示出巨大的潛力[52]。

2.2 用于農(nóng)業(yè)干旱風(fēng)險(xiǎn)評(píng)估

深度學(xué)習(xí)技術(shù)為農(nóng)業(yè)干旱風(fēng)險(xiǎn)評(píng)估研究提供了新的視角。深度學(xué)習(xí)方法在農(nóng)業(yè)干旱風(fēng)險(xiǎn)評(píng)估中的應(yīng)用越來(lái)越受到關(guān)注[53?54]，常用的有決策樹(shù)、隨機(jī)森林和人工神經(jīng)網(wǎng)絡(luò)等深度學(xué)習(xí)算法。薄乾禎等通過(guò)對(duì)基于決策樹(shù)方法的湖北水稻旱情評(píng)估，表明決策樹(shù)分類法對(duì)水稻旱災(zāi)監(jiān)測(cè)的總體精度為93.1%，旱情監(jiān)測(cè)效果顯著[55]。侯陳瑤等基于隨機(jī)森林方法建立了多尺度標(biāo)準(zhǔn)化降水蒸散指數(shù)與小麥損失率關(guān)系模型，并對(duì)河南小麥旱災(zāi)損失進(jìn)行了定量評(píng)估[56]。劉冀等基于多源遙感數(shù)據(jù)和隨機(jī)森林算法構(gòu)建了農(nóng)業(yè)干旱監(jiān)測(cè)模型BRFDC，定量評(píng)估了2001-2014年淮河流域的農(nóng)業(yè)干旱[57]。馮嶺等采用支持向量機(jī)和長(zhǎng)短期記憶神經(jīng)網(wǎng)絡(luò)相結(jié)合的方法，建立了一套旱災(zāi)風(fēng)險(xiǎn)等級(jí)評(píng)估流程，即首先構(gòu)建旱災(zāi)風(fēng)險(xiǎn)等級(jí)評(píng)估模型訓(xùn)練集，應(yīng)用支持向量機(jī)算法開(kāi)展旱災(zāi)風(fēng)險(xiǎn)等級(jí)評(píng)估模型訓(xùn)練，再運(yùn)用長(zhǎng)短期記憶神經(jīng)網(wǎng)絡(luò)算法預(yù)測(cè)未來(lái)氣象因子的特征值，以此開(kāi)展未來(lái)旱災(zāi)風(fēng)險(xiǎn)等級(jí)的定量評(píng)估[58]。

3 深度學(xué)習(xí)方法用于農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估時(shí)存在的問(wèn)題

3.1 深度學(xué)習(xí)對(duì)表達(dá)復(fù)雜函數(shù)的能力有限

農(nóng)業(yè)干旱是一種復(fù)雜的現(xiàn)象，單個(gè)因子難以描述其發(fā)生、發(fā)展過(guò)程和影響范圍。復(fù)雜函數(shù)將多種影響因子綜合，可對(duì)農(nóng)業(yè)干旱進(jìn)行監(jiān)測(cè)預(yù)測(cè)評(píng)估，但深度學(xué)習(xí)對(duì)表達(dá)復(fù)雜函數(shù)的能力有限[59]。盡管以往研究取得了令人滿意的預(yù)測(cè)結(jié)果，但卷積神經(jīng)網(wǎng)絡(luò)（convolutional neural net-work，CNN）無(wú)法處理干旱監(jiān)測(cè)預(yù)測(cè)估計(jì)中氣象因素隨時(shí)間而變化造成的非平穩(wěn)性[60]。現(xiàn)有數(shù)據(jù)在訓(xùn)練集上表現(xiàn)很好，但是在遇到新的數(shù)據(jù)后，泛化能力降低，這種現(xiàn)象叫過(guò)擬合，時(shí)間序列數(shù)據(jù)易出現(xiàn)過(guò)擬合的問(wèn)題。因此，在深度學(xué)習(xí)中還需引入其它方法優(yōu)化過(guò)擬合，如Dropout、數(shù)據(jù)增強(qiáng)、早停法和標(biāo)簽平滑等。此外，在自適應(yīng)神經(jīng)模糊推理系統(tǒng)（ANFIS）的情況下，使用特定輸入無(wú)法優(yōu)化建模[61]。該模型在訓(xùn)練期間可能會(huì)過(guò)度學(xué)習(xí)，導(dǎo)致測(cè)試期間的性能降低、訓(xùn)練速度下降、易陷入局部極小值等[62]。

3.2 深度學(xué)習(xí)對(duì)設(shè)備的存儲(chǔ)容量要求較大

深度學(xué)習(xí)模型的瓶頸是存儲(chǔ)空間[63]。開(kāi)展深度學(xué)習(xí)模型訓(xùn)練的前提是如何獲得大量高精度的訓(xùn)練樣本[64]。但實(shí)際由于所研究的問(wèn)題本身的高度復(fù)雜性要求（分類數(shù)、所需精度等）限制，需要大量?jī)?nèi)存和很高的計(jì)算能力來(lái)進(jìn)行培訓(xùn)、測(cè)試和部署[65]。然而，這些要求使得整個(gè)測(cè)試的任務(wù)很難在資源有限的低成本設(shè)備上完成[66?67]，需要在可運(yùn)行的內(nèi)存大小和存儲(chǔ)性能要求實(shí)時(shí)進(jìn)行。若數(shù)據(jù)采樣集不全面，會(huì)削弱對(duì)總體數(shù)據(jù)分析結(jié)果的可靠性。此外，由于當(dāng)前的國(guó)內(nèi)農(nóng)業(yè)氣象與干旱災(zāi)害相關(guān)的公開(kāi)數(shù)據(jù)集相對(duì)較少，導(dǎo)致研發(fā)設(shè)計(jì)人員需要花費(fèi)大量的時(shí)間來(lái)獲取這些數(shù)據(jù)，降低了效率。

3.3 深度學(xué)習(xí)法數(shù)據(jù)預(yù)處理耗時(shí)較長(zhǎng)

雖然深度學(xué)習(xí)的測(cè)試時(shí)間通常比其他基于機(jī)器學(xué)習(xí)的方法快，但通常訓(xùn)練時(shí)間較長(zhǎng)。由于數(shù)據(jù)存在的低辨識(shí)度、低準(zhǔn)確度形式帶來(lái)的噪音等諸多問(wèn)題，所以數(shù)據(jù)必須通過(guò)歸一化處理或進(jìn)行離散化處理，來(lái)增強(qiáng)數(shù)據(jù)有效性。深度學(xué)習(xí)模型由于過(guò)度消耗而導(dǎo)致的功耗很多，對(duì)于連續(xù)監(jiān)控中使用的設(shè)備，運(yùn)行時(shí)間可能非常長(zhǎng)。此外，在偏遠(yuǎn)地區(qū)的電源接入并不總是得到保證。因此，電池的有限能量需要深度學(xué)習(xí)模型的功耗最小化。在資源有限的情況下開(kāi)展數(shù)據(jù)處理需要壓縮深度學(xué)習(xí)模型。數(shù)據(jù)預(yù)處理耗時(shí)較長(zhǎng)可以用CPU并行的方法加快計(jì)算速度[68]，或使用修剪參數(shù)的方法通過(guò)刪減模型數(shù)據(jù)來(lái)減小模型的大小，減少來(lái)自神經(jīng)網(wǎng)絡(luò)的不必要的連接。修剪有助于減少計(jì)算量、成本和存儲(chǔ)空間，同時(shí)保持其性能。

3.4 深度學(xué)習(xí)法預(yù)定義類別范圍窄且圖像處理復(fù)雜

給定一張衛(wèi)星遙感圖像用于對(duì)象檢測(cè)時(shí)，通常有許多預(yù)類別，檢測(cè)對(duì)象是否為來(lái)自預(yù)定義類別的對(duì)象時(shí)，對(duì)象的空間位置和范圍可以使用邊界框（與對(duì)象緊密綁定的軸對(duì)齊矩形）、精確的像素分割掩碼或封閉邊界，其中可能只存在范圍較窄的預(yù)定義。與自然場(chǎng)景圖像不同，衛(wèi)星遙感圖像包括各種類型的物體，如一個(gè)文件中的不同大小、顏色、旋轉(zhuǎn)和位置場(chǎng)景，而屬于不同種類的不同場(chǎng)景可能在許多方面相似[69]。雖然深度學(xué)習(xí)技術(shù)可以達(dá)到的準(zhǔn)確率較高，但由于圖像的復(fù)雜性，其預(yù)測(cè)效果可能很差，很難用深度學(xué)習(xí)區(qū)分場(chǎng)景和物體，需要提高深度學(xué)習(xí)的魯棒性，即提升控制系統(tǒng)在一定的參數(shù)攝動(dòng)下的維穩(wěn)特性[70]。

4 未來(lái)研究方向

4.1 深度學(xué)習(xí)模型與作物生長(zhǎng)模型的耦合

深度學(xué)習(xí)模型的本質(zhì)是從數(shù)據(jù)到數(shù)據(jù)的特征提取，卻無(wú)法準(zhǔn)確表達(dá)作物生長(zhǎng)具體過(guò)程與機(jī)理，也無(wú)法學(xué)習(xí)到干旱如何影響農(nóng)業(yè)生產(chǎn)。因此，未來(lái)可嘗試探索深度學(xué)習(xí)模型和作物生長(zhǎng)模型的耦合，以提高對(duì)作物生長(zhǎng)過(guò)程的理解及干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估精度，從而提高深度學(xué)習(xí)模型的合理性和可解釋性，進(jìn)一步加深對(duì)農(nóng)業(yè)生產(chǎn)過(guò)程的理解和認(rèn)識(shí)。作物生長(zhǎng)模型參數(shù)往往存在不確定性，從而導(dǎo)致模型輸出出現(xiàn)誤差，基于實(shí)際觀測(cè)和其他輔助數(shù)據(jù)，通過(guò)使用深度學(xué)習(xí)模型可校準(zhǔn)作物生長(zhǎng)模型的輸出。因此，氣候變化背景下進(jìn)一步加強(qiáng)深度學(xué)習(xí)模型與作物生長(zhǎng)模型的耦合研究將是未來(lái)研究的重要課題之一。

4.2 融合大氣環(huán)流模式和深度學(xué)習(xí)模型的農(nóng)業(yè)干旱預(yù)測(cè)

深度學(xué)習(xí)模型屬于非線性統(tǒng)計(jì)模型，其時(shí)序變量間的相互關(guān)系決定了時(shí)間序列數(shù)據(jù)的建模，因此，深度學(xué)習(xí)模型預(yù)測(cè)值可能會(huì)與實(shí)際觀測(cè)值有所偏差，隨著預(yù)測(cè)天數(shù)的增加，會(huì)使預(yù)測(cè)序列的誤差積累，從而降低了預(yù)測(cè)準(zhǔn)確率[71]。大氣環(huán)流模式是基于基本的物理定律模擬大氣環(huán)流要素變化動(dòng)態(tài)，能較為準(zhǔn)確地預(yù)測(cè)某些氣象要素的未來(lái)變化。因此，未來(lái)研究可以建立基于大氣環(huán)流模式與深度學(xué)習(xí)模型的耦合模型，修正預(yù)測(cè)序列中存在的誤差，進(jìn)一步提升模型對(duì)中長(zhǎng)期農(nóng)業(yè)干旱的預(yù)測(cè)能力。

4.3 基于深度學(xué)習(xí)與遷移學(xué)習(xí)的農(nóng)業(yè)干旱精細(xì)化監(jiān)測(cè)與評(píng)估

由于農(nóng)業(yè)干旱災(zāi)害的樣本容量有限，而深度學(xué)習(xí)模型創(chuàng)建于大樣本之上才能保證其監(jiān)測(cè)評(píng)估精度，因此，可以使用有限的樣本在大型數(shù)據(jù)集上預(yù)訓(xùn)練的模型參數(shù)進(jìn)行微調(diào)的遷移學(xué)習(xí)的方法來(lái)改善小樣本的限制，同時(shí)采用按比例分層抽樣、數(shù)據(jù)增強(qiáng)等方法，提高弱樣本的應(yīng)用能力[63]。遷移學(xué)習(xí)是一種新型學(xué)習(xí)方法，旨在遷移現(xiàn)有的知識(shí)來(lái)解決目標(biāo)領(lǐng)域內(nèi)標(biāo)簽樣本數(shù)據(jù)量少的問(wèn)題[72]。遷移學(xué)習(xí)利用微調(diào)預(yù)學(xué)習(xí)訓(xùn)練模型來(lái)達(dá)到更優(yōu)的學(xué)習(xí)效果，常見(jiàn)的遷移學(xué)習(xí)技術(shù)是使用預(yù)訓(xùn)練的神經(jīng)網(wǎng)絡(luò)模型。預(yù)先培訓(xùn)的模型之前接受過(guò)大型數(shù)據(jù)集的培訓(xùn)，可以提供相應(yīng)的用于深度學(xué)習(xí)的響應(yīng)結(jié)構(gòu)和權(quán)重[73]。因此，未來(lái)的研究應(yīng)該加強(qiáng)基于深度學(xué)習(xí)模型與預(yù)訓(xùn)練遷移學(xué)習(xí)的農(nóng)業(yè)干旱監(jiān)測(cè)與評(píng)估研究，進(jìn)一步提高農(nóng)業(yè)干旱精細(xì)化監(jiān)測(cè)與評(píng)估精度。

4.4 深度學(xué)習(xí)與信息融合技術(shù)相結(jié)合的區(qū)域農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估

農(nóng)業(yè)生產(chǎn)體系是一個(gè)非常復(fù)雜的非線性系統(tǒng)、干旱發(fā)生過(guò)程也是個(gè)非常復(fù)雜的非線性過(guò)程，干旱災(zāi)害發(fā)生具有動(dòng)態(tài)性、復(fù)雜性、緊迫性和不確定性等特征[74]。影響農(nóng)業(yè)干旱形成的因子具有數(shù)據(jù)量大、類型多樣、非線性等特點(diǎn)。此外，目前農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估主要采用經(jīng)驗(yàn)法、統(tǒng)計(jì)學(xué)方法和模式模擬法等，但是不同方法選取的干旱視角和指標(biāo)等不同，得出的干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估結(jié)果也不盡相同。因此，采用深度學(xué)習(xí)與信息融合相結(jié)合的方法，可將長(zhǎng)時(shí)間序列的氣象、土壤、農(nóng)業(yè)等多源、非線性數(shù)據(jù)融合為具有時(shí)空一致性的數(shù)據(jù)集，從而綜合性地建立干旱監(jiān)測(cè)預(yù)測(cè)評(píng)估模型，進(jìn)一步提高區(qū)域農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估精度，這也是未來(lái)農(nóng)業(yè)氣象災(zāi)害領(lǐng)域最為重要的研究方向之一。

[1] Tian H R,Wang P X,Tansey K,et al.A deep learning framework under attention mechanism for wheat yield estimation using remotely sensed indices in the Guanzhong plain,P R China[J].International Journal of Applied Earth Observation and Geoinformation,2021,102:102375.

[2] 江懿,陳磊,陳娟,等.安徽省農(nóng)業(yè)旱災(zāi)風(fēng)險(xiǎn)評(píng)估與區(qū)劃[J].中國(guó)農(nóng)學(xué)通報(bào),2016,32(5):178-183.

Jiang Y,Chen L,Chen J,et al.Agricultural drought risk assessment and zoning in Anhui province[J].Chinese Agricultural Science Bulletin,2016,32(5):178-183.(in Chinese)

[3] Prodhan F A,Zhang J H,Hasan S S,et al.A review of machine learning methods for drought hazard monitoring and forecasting:current research trends,challenges,and future research directions[J].Environmental Modelling and Software,2022,149:105327.

[4] 趙映慧,郭晶鵬,毛克彪,等.1949-2015年中國(guó)典型自然災(zāi)害及糧食災(zāi)損特征[J].地理學(xué)報(bào),2017,72(7):1261-1276.

Zhao Y H,Guo J P,Mao K B,et al.Spatio-temporal distribution of typical natural disasters and grain disaster losses in China from 1949 to 2015[J].Acta Geographica Sinica,2017,72(7):1261-1276.(in Chinese)

[5] 張玨,郝鵬,孫曉懿.國(guó)內(nèi)外干旱評(píng)估體系研究進(jìn)展及展望[J].陜西水利,2018(3):54-57.

Zhang J,Hao P,Sun X Y.Research progress and prospects of drought assessment systems at home and abroad[J].Shaanxi Water Resources,2018(3):54-57.(in Chinese)

[6] Dikshit A,Pradhan B,Huete A,et al.Spatial based drought assessment: where are we heading? a review on the current status and future[J].Science of the Total Environment, 2022,844:157239.

[7] 呂彥彬,王玉斌.我國(guó)旱災(zāi)變遷及對(duì)糧食產(chǎn)量的影響[J].河南農(nóng)業(yè)科學(xué),2014(10):19-22.

Lv Y B,Wang Y B.Change of drought and its impact on grain production in China[J].Journal of Henan Agricultural Science,2014(10):19-22.(in Chinese)

[8] Vicente-Serrano S M,Begueria S,Lopez-Moreno J I.A multi-scalar drought index sensitive to global warming:the stand-ardized precipitation evapotranspiration index[J]. Journal of Climate,2010,23(7):1696-1718．

[9] 趙俊芳,李寧,侯英雨,等.基于APSIM模型評(píng)估北方八省春玉米生產(chǎn)對(duì)氣候變化的響應(yīng)[J].中國(guó)農(nóng)業(yè)氣象,2018,39 (2):108-118.

Zhao J F,Li N,Hou Y Y,et al.Evaluation of response of spring maize production to climate change in the eight provinces of northern China based on APSIM model[J]. Chinese Journal of Agrometeorology,2018,39(2):108-118. (in Chinese)

[10] Ding Y B,Gong X L,Xing Z X,et a1.Attribution of meteorological,hydrological and agricultural drought propagation in different climatic regions of China[J]. Agriculture Water Management,2021,255(1):106996.

[11] Hao Z,Singh V P.Drought characterization from a multivariate perspective:a review[J].J.Hydroloy,2015,527: 668-678.

[12] Lan Y B,Huang Z X,Deng X L,et al.Comparison of machine learning methods for citrus greening detection on UAV multispectral images[J].Computers and Electronics in Agriculture,2020,171:105234.

[13] Jaihuni M,Khan F,Lee D,et al.Determining spatiotemporal distribution of macronutrients in a cornfield using remote sensing and a Deep Learning model[J].IEEE Access,2021 (9):30256.

[14] 陳海濤.農(nóng)業(yè)干旱風(fēng)險(xiǎn)分析方法及應(yīng)用研究[D].鄭州:華北水利水電大學(xué),2020.

Chen H T.Research on agricultural drought risk analysis method and application[D].Zhengzhou:North China University of Water Resources and Electric Power,2020.(in Chinese)

[15] 李阿龍.河南省農(nóng)業(yè)旱災(zāi)風(fēng)險(xiǎn)分析[D].鄭州:華北水利水電大學(xué),2016.

Li A L.Risk analysis of agricultural drought in Henan province[D].Zhengzhou:North China University of Water Resources and Electric Power,2016.(in Chinese)

[16] 杜棟,劉夢(mèng)雨.基于信息融合與深度學(xué)習(xí)的區(qū)域干旱致災(zāi)預(yù)警方法研究[J].海河水利,2020(1):18.

Du D,Liu M Y.Research on regional drought disaster early warning method based on information fusion and deep learning[J].Haihe Water Resources,2020(1):18.(in Chinese)

[17] Todisco F,Mannocchi F,Vergni L.Severity-duration-frequency curves in the mitigation of drought impact:an agricultural case study[J].Natural Hazards,2013,65(3):1863-1881.

[18] 王連喜,肖瑋鈺,李琪,等.中國(guó)北方地區(qū)主要農(nóng)作物氣象災(zāi)害風(fēng)險(xiǎn)評(píng)估方法綜述[J].災(zāi)害學(xué),2013,28(2):114-119.

Wang L X,Xiao W Y,Li Q.Summary of risk assessment methods of main crops meteorological disasters in north China[J].Journal of Catastrophology,2013,28(2):114-119. (in Chinese)

[19] 閆超君,歐陽(yáng)蔚,金菊良,等.基于信息擴(kuò)散和頻率曲線適線的農(nóng)業(yè)旱災(zāi)風(fēng)險(xiǎn)評(píng)估方法[J].水利水電技術(shù),2014,45 (7):107-111.

Yan C J,Ouyang W,Jin J L,et al.Agricultural drought risk assessment and zoning in Anhui province[J].Water Resources and Hydropower Engineering,2014,45(7):107- 111.(in Chinese)

[20] 張冬有,張麗娟,萬(wàn)魯河.基于信息擴(kuò)散理論的自然災(zāi)害風(fēng)險(xiǎn)評(píng)估系統(tǒng)設(shè)計(jì)與實(shí)現(xiàn)[J].自然災(zāi)害學(xué)報(bào),2013,22(4): 234-238.

Zhang D Y,Zhang L J,Wan L H.Design and implementation of natural disaster risk assessment system based on information diffusion theory[J].Journal of Natural Disasters, 2013,22(4):234-238.(in Chinese)

[21] 李暉,尹輝,白旸,等.基于信息擴(kuò)散理論的湖南省農(nóng)業(yè)水旱災(zāi)害風(fēng)險(xiǎn)損失情景分析[J].廣東農(nóng)業(yè)科學(xué),2013,40(7): 223-226.

Li H,Yin H,Bai Y,et al.Scenario analysis of agricultural flood and drought disaster risk loss in Hunan province based on information diffusion theory[J].Guangdong Agricultural Sciences,2013,40(7):223-226.(in Chinese)

[22] 崔修來(lái),孫瑤,王東.干旱監(jiān)測(cè)預(yù)報(bào)研究綜述[J].南方農(nóng)業(yè),2019,13(20):151-152.

Cui X L,Sun Y,Wang D.A summary of drought monitoring and forecasting research[J].South China Agriculture,2019, 13(20):151-152.(in Chinese)

[23] Lin T,Zhong R H,Wang Y D,et al.Deep crop net:a deep spatial-temporal learning framework for county-level corn yield estimation[J].Environmental Research Letters,2020, 15(3):34016.

[24] 施光耀,周宇,桑玉強(qiáng),等.基于隨機(jī)森林方法分析環(huán)境因子對(duì)空氣負(fù)離子的影響[J].中國(guó)農(nóng)業(yè)氣象,2021,42(5): 390-401.

Shi G Y,Zhou Y,Sang Y Q,et al.Influence of environmental factors on negative air ion using random forest algorithm[J].Chinese Journal of Agrometeorology,2021,42 (5):390-401.(in Chinese)

[25] 王連喜,孟丹,耿秀華,等.基于GIS的寧夏農(nóng)業(yè)干旱風(fēng)險(xiǎn)評(píng)價(jià)與區(qū)劃[J].自然資源學(xué)報(bào),2013,22(5):213-220.

Wang L X,Meng D,Geng X H,et al.GIS based drought risk assessment and zoning in Ningxia[J].Journal of Natural Disasters,2013,22(5):213-220.(in Chinese)

[26] Zhao J F,Xu J W,Xie X M,et al.Drought monitoring based on TIGGE and distributed hydrological model in Huaihe river basin[J].Science of the Total Environment,2016, 553:358-365.

[27] Egger J,Pepe A,Gsaxner C,et al.Deep learning a first meta survey of selected reviews across scientific disciplines,their commonalities, challenges and research impact[J].PeerJ Computer Science,2021,11(17):773.

[28] Aya F,Ali B A,Vincent B,et al.Forecasting vegetation indices from spatio-temporal remotely sensed data using deep learning-based approaches:a systematic literature review[J].Ecological Informatics,2022,68:101552.

[29] Liu L,Ouyang W L,Wang X G,et al.Deep learning for generic object detection:a survey[J].International Journal of Computer Vision,2020,128(2):261-318.

[30] Anna C,Salah S,Brett W.Machine learning approaches for crop yield prediction and nitrogen status estimation in precision agriculture:a review[J].Computers and Electronics in Agriculture,2018,151:61-69.

[31] 呂盛坪,李燈輝,冼榮亨.深度學(xué)習(xí)在我國(guó)農(nóng)業(yè)中的應(yīng)用研究現(xiàn)狀[J].計(jì)算機(jī)工程與應(yīng)用,2019,55(20):24-33.

Lv S P,Li D H,Xian R H.Research status of deep learning in agriculture of China[J].Computer Engineering and Application,2019,55(20):24-33.(in Chinese)

[32] Kuwata K,Shibasaki R.Estimating crop yields with deep learning and remotely sensed data[C].Milan,Italy: IEEE International Geoscience and Remote Sensing Symposium, 2015:858-861.

[33] 吳運(yùn)強(qiáng).基于深度學(xué)習(xí)的農(nóng)業(yè)遙感語(yǔ)義分割算法的研究[D].長(zhǎng)春:吉林大學(xué),2021.

Wu Y Q.Semantic segmentation algorithm of agricultural remote sensing based on deep learning[D].Changchun:Jilin University,2021.(in Chinese)

[34] Kim N,Ha K J,Park N W,et al.A comparison between major artificial intelligence models for crop yield prediction:case study of the midwestern United States,2006-2015[J]. International Journal of Geo-information,2019,8(5):240- 262.

[35] Wan J,Wang D,Hoi C H,et al.Deep learning for content- based image retrieval comprehensive study[C]//Proceedings of the 22nd ACM International Conference on Multimedia, 2014:157-166.

[36] Lecu Y,Bengio Y,Hinton G.Deep learning[J].Nature,2015, 521(7553):436.

[37] Jawak S D,Devliyal P,Luis A J.A comprehensive review on pixel oriented and object oriented methods for information extraction from remotely sensed images with a special emphasis on cryospheric applications[J].Environmental Science,2015(4):177-195.

[38] Corcoran P,Winstanley A,Mooney P.Segmentation performance evaluation for object based remotely sensed image analysis[J]. International Journal of Remote Sensing,2010,31:617-645.

[39] Sun J,Lai Z L,Di L P,et al.Multilevel deep learning network for county-level corn yield estimation in the US corn belt[J].IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,2020(13): 5048-5060.

[40] 陳文龍,楊云麗,張煜,等.基于天地協(xié)同與深度學(xué)習(xí)的灌區(qū)地下水位模擬研究[J].長(zhǎng)江科學(xué)院院報(bào),2022(6):18.

Chen W L,Yang Y L,Zhang Y,et al.Research on the simulation of groundwater level in irrigation area based on sky-earth cooperation and deep learning[J].Journal of Yangtze River Scientific Research Institute,2022(6):18.(in Chinese)

[41] Shen Y F,Zhou H L,Li J T,et al.Detection of stored-grain insects using deep learning[J].Computers and Electronics in Agriculture,2018,145:319-325.

[42] Rhee J,Im J,Park S.Drought forecasting based on machine learning of remote sensing and long-range forecast data[J]. ISPRS-International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences,2016, XLI-B8:157-158.

[43] Dikshit A,Pradhan B,Alamri A M.Long lead time drought forecasting using lagged climate variables and a stacked long short-term memory model[J].Science of the Total Environment,2021,755:142638.

[44] Lee C S,Sohn E,Park J D,et al.Estimation of soil moisture using deep learning based on satellite data:a case study of South Korea[J].GIScience and Remote Sensing,2018, 56(1):43-67.

[45] Feng P,Wang B,Li L D,et al.Machine learning-based integration of remotely-sensed drought factors can improve the estimation of agricultural drought in south-eastern Australia[J].Agricultural Systems,2019,173:303-316.

[46] 胡小楓,王冬利,趙安周,等.基于深度學(xué)習(xí)算法的京津冀地區(qū)綜合干旱評(píng)估模型構(gòu)建[J].中國(guó)農(nóng)業(yè)氣象,2021,42 (9):775-787.

Hu X F,Wang D L,Zhao A Z,et al.A comprehensive drought evaluation model in Beijing-Tianjin-Hebei region based on deep learning algorithm[J].Chinese Journal of Agrometeorology, 2021,42(9):775-787.(in Chinese)

[47] Jiang H,Hu H,Zhong R H,et al.A deep learning approach to conflating heterogeneous geospatial data for corn yield estimation:a case study of the US corn belt at the county level[J].Global Change Biology,2020,26(3):1754-1766.

[48] Agana N A,Homaifar A.A deep learning-based approach for long-term drought prediction[J].IEEE:Southeast Con, 2017:1-8.

[49] Johnson M D,Hsieh W,Cannon A J,et al.Crop yield forecasting on the Canadian Prairies by remotely sensed vegetation indices and machine learning methods[J]. Agricultural and Forest Meteorology,2016,218-219:74-84.

[50] 田甜,王迪,王珍,等.基于深度學(xué)習(xí)模型的種植結(jié)構(gòu)復(fù)雜區(qū)農(nóng)作物精細(xì)分類研究[J].中國(guó)農(nóng)業(yè)資源與區(qū)劃,2022,43 (12):12.

Tian T,Wang D,Wang Z,et al.Research on fine classification of crops in complex planting structure based on deep learning model[J].Chinese Journal of Agricultural Resources and Regional Planning,2022,43(12):12.(in Chinese)

[51] Karakizi C,Oikonomou M,Karantzalosa K.Vineyard detection and vine variety discrimination from very high resolution satellite data[J].Remote Sens,2017,8(3):235.

[52] 張勞模.基于深度學(xué)習(xí)的林地分類與動(dòng)態(tài)變化模擬技術(shù)研究[D].北京:中國(guó)林業(yè)科學(xué)研究院,2022.

Zhang L M.Research on woodland classification and dynamic change simulation technology based on deep learning[D].Beijing:Chinese Academy of Forestry Research, 2022.(in Chinese)

[53] 鄧智予.基于深度學(xué)習(xí)模型的根系微生物組數(shù)據(jù)和植物產(chǎn)量的關(guān)聯(lián)分析[D].武漢:中國(guó)科學(xué)院大學(xué)(中國(guó)科學(xué)院武漢植物園),2021.

Deng Z Y.Association analysis of root microbiome data and plant yield based on deep learning model[D]. Wuhan:University of Chinese Academy of Sciences (Wuhan Botanical Garden, Chinese Academy of Sciences), 2021.(in Chinese)

[54] Reichstein M,Camps-Valls G,Stevens B,et al.Deep learning and process understanding for data-driven earth system science[J].Nature,2019,566(7743):195-204.

[55] 薄乾禎,陳志杰,汪權(quán)方.基于決策樹(shù)分類的水稻干旱災(zāi)損評(píng)估[J].湖北大學(xué)學(xué)報(bào)(自然科學(xué)版),2021,43(2):155-162.

Bo Q Z,Chen Z J,Wang Q F.Drought damage evaluation of rice based on decision tree classification[J].Journal of Hubei University(Natural Science),2021,43(2):155-162.(in Chinese)

[56] 侯陳瑤,朱秀芳,孫丹一,等.基于隨機(jī)森林的旱災(zāi)損失模型構(gòu)建與應(yīng)用:以河南省雨養(yǎng)小麥為例[J].災(zāi)害學(xué),2019, 34(4):210-215.

Hou C Y,Zhu X F,Sun D Y,et al.Construction and application of drought loss model based on random forest:taking rain-fed wheat in Henan province as an example[J].Journal of Catastrophology,2019,34(4):210-215. (in Chinese)

[57] 劉冀,張?zhí)?魏榕,等.基于隨機(jī)森林偏差校正的農(nóng)業(yè)干旱遙感監(jiān)測(cè)模型研究[J].農(nóng)業(yè)機(jī)械學(xué)報(bào),2020,51(7):170-177.

Liu J,Zhang T,Wei R,et al.Development of agricultural drought monitoring model using remote sensing based on bias-correcting random forest[J].Journal of Agricultural Machinery,2020,51(7):170-177.(in Chinese)

[58] 馮嶺,宋文輝,陳繼坤.基于深度學(xué)習(xí)的旱災(zāi)風(fēng)險(xiǎn)評(píng)估方法[J].人民長(zhǎng)江,2022,53(6):111-118．

Feng L,Song W H,Chen J K.Drought risk assessment method based on deep learning[J].Yangtze River,2022, 53(6):111-118.(in Chinese)

[59] Shen R,Huang A,Li B,et al.Construction of a drought monitoring model using deep learning based on multi-source remote sensing data[J].Int J Appl Earth Obs Geoinf,2019,79:48-57.

[60] 顧成奎,王正歐.基于神經(jīng)網(wǎng)絡(luò)動(dòng)態(tài)非線性非平穩(wěn)經(jīng)濟(jì)系統(tǒng)預(yù)測(cè)[J].系統(tǒng)工程學(xué)報(bào),2002,17(2):133-136.

Gu C K,Wang Z O.Prediction of dynamic nonlinear non-stationary economic system based on neural network[J].Journal of Systems Engineering,2002,17(2): 133-136.(in Chinese)

[61] Alizadeh M R,Nikoo M R.A fusion-based methodology for meteorological drought estimation using remote sensing data[J].Remote Sens Environ,2018,211:229-247.

[62] Mokhtarzad M,Eskandari F,Vanjani N J,et al.Drought forecasting by ANN,ANFIS,and SVM and comparison of the models[J].Environ Earth Sci,2017,76(21):729.

[63] 馬瑋良,彭軒,熊倩,等.深度學(xué)習(xí)中的內(nèi)存管理問(wèn)題研究綜述[J].大數(shù)據(jù),2020,6(4):56-68.

Ma W L,Peng X,Xiong Q,et al.Memory management in deep learning:a survey[J].Big Data Research,2020,6(4):56- 68.(in Chinese)

[64] Choubin B,Khalighi-Sigaroodi S,Malekian A,et al.Drought forecasting in a semi-arid watershed using climate signals:a neuro-fuzzy modeling approach[J].J Mountain Sci,2014,11 (6):1593-1605.

[65] 許晴,張錦水,張鳳,等.深度學(xué)習(xí)農(nóng)作物分類的弱樣本適用性[J].遙感學(xué)報(bào),2022,26(7):1395-1409.

Xu Q,Zhang J S,Zhang F,et al.Applicability of weak samples to deep learning crop classification[J].National Remote Sensing Bulletin,2022,26(7):1395-1409.(in Chinese)

[66] 陳希亮,曹雷,何明,等.深度逆向強(qiáng)化學(xué)習(xí)研究綜述[J].計(jì)算機(jī)工程與應(yīng)用,2018,54(5):24-35.

Chen X L,Cao L,He M,et al.Overview of deep inverse reinforcement learning[J].Computer Engineering and Applications,2018,54(5):24-35.(in Chinese)

[67] Fountsop A N,Ebongue K F J L,Atemkeng M.Deep learning models compression for agricultural plants[J]. Applied Sciences,2020,10(19):6866.

[68] 孫萍,胡旭東,張永軍.結(jié)合注意力機(jī)制的深度學(xué)習(xí)圖像目標(biāo)檢測(cè)[J].計(jì)算機(jī)工程與應(yīng)用,2019,55(17):180-184.

Sun P,Hu X D,Zhang Y J.Object detection based on deep learning and attention mechanism[J].Computer Engineering and Applications,2019,55(17):180-184.(in Chinese)

[69] Zhang L P,Zhang L F,Du B.Deep learning for remote sensing data: a technical tutorial on the state of the art[J].IEEE Geoscience and Remote Sensing Magazine, 2016,4(2):22-40.

[70] 張榮,李偉平,莫同.深度學(xué)習(xí)研究綜述[J].信息與控制,2018,47(4):385-397.

Zhang R,Li W P,Mo T.Review of deep learning[J]. Information and Control,2018,47(4):385-397.(in Chinese)

[71] 米前川,高西寧,李玥,等.深度學(xué)習(xí)方法在干旱預(yù)測(cè)中的應(yīng)用[J].應(yīng)用氣象學(xué)報(bào),2022,33(1):104-114.

Mi Q C,Gao X N,Li Y,et al.Application of deep learning method to drought prediction[J].Journal of Applied Meteorological Science,2022,33(1):104-114.(in Chinese)

[72] Punithavathi R,RaniA D C,Sughashini K R,et al.Computer vision and deep learning-enabled weed detection model for precision agriculture[J].Computer Systems Science and Engineering,2022,44(3):2759-2774.

[73] Wang K L,Chen K Y,Du H Y,et al.New image dataset and new negative sample judgment method for crop pest recognition based on deep learning models[J].Ecological Informatics,2022,69(7):101620.

[74] Konstantinos P F.Deep learning models for plant disease detection and diagnosis[J].Computers and Electronics in Agriculture,2018,145:311-318.

Application of Deep Learning Technology in Monitoring, Forecasting and Risk Assessment of Agricultural Drought

HUANG Rui-xi, ZHAO Jun-fang, HUO Zhi-guo, PENG Hui-wen, XIE Hong-fei

(State Key Laboratory of Severe Weather, Chinese Academy of Meteorological Sciences, Beijing 100081, China)

The development of artificial intelligence technology, especially the emergence of deep learning, has promoted new developments of agriculture, and is regarded as a new direction of modern agricultural production. Deep learning has the advantages of strong learning ability, wide coverage, strong adaptability, and great portability. Considering that its development of simulated datasets can solve real-world problems, it is more and more widely used in monitoring, forecasting and risk assessment of agricultural drought. This paper used the method of literature review to summarize the development and application of monitoring, forecasting and risk assessment of agricultural drought, and summarized the principles, advantages and disadvantages of the deep learning model. The practical applications of depth learning model in monitoring, prediction and risk assessment of agricultural drought were systematically summarized. The existing problems of large dataset requirements, long data preprocessing time, narrow predefined category range, and complex remote sensing images were discussed, and the future research directions were prospected. The results showed that in recent years, the technologies of monitoring, prediction and risk assessment of agricultural drought had made important progress. However, due to the nonlinearity of agricultural system and the complexity of disasters, existing technologies were still difficult to meet the needs of actual agricultural production in the new situation in terms of applicable regions, objects and accuracies. The deep learning technology provided a new means for agricultural drought research. However, the deep learning model could not accurately express the specific process and mechanism of crop growth, so coupling of crop growth model with deep learning model could ensure the interpretability of deep learning model. For correcting the prediction sequence, coupling models based on general circulation model and depth learning model could be established to further improve the prediction ability of deep learning model for medium and long-term agricultural drought. Aiming at the problem of limited disaster sample size, strengthening the research on agricultural drought monitoring and evaluation based on migration learning could further improve the precisions in fine monitoring and evaluation of agricultural drought. In view of the fact that the factors affecting agricultural drought formation was characterized by large amount of data, diverse types and nonlinearity, the method of combining deep learning and information fusion was adopted to further improve the accuracies in regional monitoring, prediction and risk assessment of agricultural drought. Therefore, the coupling of deep learning models and crop growth models, agricultural drought prediction by integrating deep learning models and general circulation models, fine monitoring and evaluation of agricultural drought based on deep learning and migration learning, regional monitoring, prediction and risk assessment of agricultural drought based on deep learning and information fusion were considered as the development trends of applicating deep learning technologies in monitoring, prediction and risk assessment of agricultural drought in the future.

Deep learning; Agricultural drought; Monitoring and prediction; Risk assessment; Accuracy

10.3969/j.issn.1000-6362.2023.10.007

收稿日期：2022?11?04

國(guó)家重點(diǎn)研發(fā)計(jì)劃項(xiàng)目02課題“黃淮海小麥干旱和春季凍害監(jiān)測(cè)評(píng)估及預(yù)警預(yù)測(cè)研究”（2022YFD2300202）

通訊作者：趙俊芳，博士，研究員，主要從事全球變化與農(nóng)業(yè)氣象研究，E-mail：zhaojf@cma.gov.cn

黃睿茜，E-mail：huangruiqian22@mails.ucas.ac.cn

黃睿茜,趙俊芳,霍治國(guó),等.深度學(xué)習(xí)技術(shù)在農(nóng)業(yè)干旱監(jiān)測(cè)預(yù)測(cè)及風(fēng)險(xiǎn)評(píng)估中的應(yīng)用[J].中國(guó)農(nóng)業(yè)氣象,2023,44(10):943-952