摘要:【目的】利用無人機圖像數(shù)據(jù)的顏色特征和形態(tài)特征構(gòu)建滴灌棉花苗期株數(shù)估算模型,為棉花田間精準管理提供理論依據(jù)。【方法】于2020―2021年開展試驗,以魯棉研24號為供試品種,設置3個不同種植密度,分別為:低密度(D1,6.9×104株·hm-2)、中密度(D2,13.8×104株·hm-2)、高密度(D3,24×104株·hm-2)。對出苗后25 d的無人機圖像提取基于紅綠藍(red, green, and blue, RGB)的植被指數(shù)和目標形態(tài)特征,構(gòu)建棉花株數(shù)估算特征集合;在自變量間的相關性分析的基礎上,利用逐步多元回歸的方法構(gòu)建棉苗株數(shù)的估算模型,并進行驗證?!窘Y(jié)果】(1)三角綠度指數(shù)(triangular greenness index,TGI)、超綠指數(shù)(excess green index,ExG)、綠-藍差值+修正超綠指數(shù)(green-blue difference + modified excess green index,GBDI + MExG)均對圖像有較好的分割效果,其中TGI對棉花目標的分割完整度最高。(2)對比2種特征參數(shù)構(gòu)建的棉花株數(shù)估算模型,基于目標形態(tài)特征的苗期棉花估算模型的擬合優(yōu)度(R2=0.935 5)要高于基于RGB植被指數(shù)的株數(shù)估算模型(R2= 0.903 6)。(3)基于RGB植被指數(shù)的株數(shù)估算模型在D1、D2、D3密度下估算精度分別為96.77%、99.55%和95.95%,整體估算精度為98.47%;基于目標形態(tài)特征的株數(shù)估算模型在D1、D2和D3密度下估算精度分別為99.98%、99.21%和97.92%,整體估算精度為99.21%?;谀繕诵螒B(tài)特征的株數(shù)估算模型的估算精度略高于基于RGB植被指數(shù)的株數(shù)估算模型,但2個模型在不同種植密度下均具有較好的估算效果?!窘Y(jié)論】利用集成高分辨率傳感器的低空無人機遙感平臺,通過顏色特征和目標形態(tài)特征構(gòu)建的滴灌棉花苗期株數(shù)估算模型均能有效、精準識別膜下滴灌棉花株數(shù),可為后續(xù)棉花田間精準管理提供技術(shù)支撐。
關鍵詞:棉花;無人機;顏色特征;形態(tài)特征;株數(shù)估算;模型
Abstract: [Objective] A model for estimating the quantity of seedlings in drip-irrigated cotton using color" characteristics and morphological characteristics of unmanned aerial vehicle (UAV) image data was constructed to provide a theoretical basis for accurate management in cotton field. [Methods] The experiment was carried out in 2020-2021 and the cultivar Lumianyan 24 was used in the experiment. Three different planting densities were set as follow: low density (D1, 6.9 × 104 plant·hm-2), medium density (D2, 13.8 × 104 plant·hm-2) and high density (D3, 24 × 104 plant·hm-2). The UAV images were obtained on the 25 days old cotton seedlings, and the vegetation indices (VIs) of red, green, and blue (RGB) and target morphological features were extracted from the acquired UAV images. Based on the selected independent variable according to the correlation analysis, the model to estimate the quantity of cotton seedlings was constructed using stepwise multiple regression, followed by the model validation. [Results] (1) Comparing the segmentation effects of extracting cotton targets by triangular greenness index (TGI), excess greenness index (ExG), and green-blue difference + modified excess greenness index (GBDI + MExG), all these three VIs had relatively good segmentation effects, while TGI showed the highest precision of segmentation of cotton targets. (2) Comparing the two cotton plant quantity estimation models constructed with the two feature parameters, the estimation model based on the target morphological features for cotton seedling (R2=0.935 5) is better than the estimation model based on the VI of RGB (R2=0.903 6). (3) The estimation accuracy of the VIs-based seedling quantity estimation model were 96.77%, 99.55%, and 95.95% at D1, D2 and D3 densities respectively, and the overall estimation accuracy was 98.47%; the estimation accuracy of the plant estimation model based on the target morphological features at D1, D2 and D3 densities were 99.98%, 99.21%, and 97.92% respectively, and the overall estimation accuracy was 99.21%. The accuracy of the plant number estimation model based on the target morphological characteristics was slightly higher than that of the plant number estimation model based on VIs, but both models had good estimation outcome under different planting densities. [Conclusion] Using the UAV based low-altitude remote sensing platform with the integration of high-resolution sensors, the quantity estimation models for the drip-irrigated cotton seedlings were constructed by color vegetation indices and morphological features of target plants. Both models can effectively and accurately identify and quantify the drip-irrigated cotton plants under mulching, providing technical support for subsequent precision management in cotton fields.
Keywords: cotton; UAV; color characteristics; morphological characteristics; estimated number of plants; model
棉花是新疆重要的經(jīng)濟作物,其種植面積和產(chǎn)量均居全國首位。新疆棉花產(chǎn)業(yè)的發(fā)展關系到當?shù)剞r(nóng)業(yè)產(chǎn)業(yè)的發(fā)展,是當?shù)剞r(nóng)戶重要的經(jīng)濟收入來源之一[1-3]。棉花出苗數(shù)量和質(zhì)量是棉花生長發(fā)育和產(chǎn)量形成的基礎,同時也是棉花生長發(fā)育后期田間機械化調(diào)控管理的關鍵因素。棉花出苗數(shù)量的準確估算可為及時補苗以保證成苗數(shù)量和棉田因苗管理實現(xiàn)高產(chǎn)提供判斷依據(jù)[4-5]。
近年來,無人機低空遙感技術(shù)以其數(shù)據(jù)的時空分辨率高、測量機動靈活、觀測范圍大、低成本、易操作、無損檢測等優(yōu)勢在農(nóng)田信息獲取與監(jiān)測中廣泛應用[6-9]。作物監(jiān)測作為遙感應用的重要領域,已在作物長勢、病蟲害、產(chǎn)量、作物識別及密度估算等方面取得了顯著的成果[10-14]。Luthfan等[15]和Randelovi等[16]結(jié)合無人機遙感影像、衛(wèi)星影像、氣候數(shù)據(jù)分別在大豆的四葉期和結(jié)莢期建立了植株密度預測模型,發(fā)現(xiàn)在大豆四葉期密度估算模型精度更高。王偉等[17]研究表明基于無人機可見光影像提取的覆蓋度與小麥密度具有較好的相關性,模型反演結(jié)果與真實值的決定系數(shù)(coefficient of determination, R2)達到91.98%。Zhao等[18]和朱孟等[19]利用低空無人機遙感平臺搭載可見光傳感器獲取油菜和火龍果冠層可見光圖像,發(fā)現(xiàn)可見光圖像的顏色特征參數(shù)可以用于識別油菜和火龍果,其識別精度高達90%以上。戴建國等[20]利用低空無人機遙感平臺獲取棉花3~4葉期遙感影像,結(jié)合顏色植被指數(shù)和最大類間方差(Otsu)法自適應閾值法對棉花目標進行識別和圖像分割,并對棉花的出苗率做出了準確估計,預測結(jié)果與真實值的誤差僅為0.89%。但該研究的密度估算模型主要針對棉花寬窄行種植模式,忽略了同樣在新疆棉區(qū)大規(guī)模應用的等行距種植模式,且基于支持向量機的模型算法和程序運行復雜,模型普適性較差。
本文選取位于新疆石河子的棉田為研究區(qū)域,利用低空無人機遙感平臺搭載可見光傳感器獲取的高分辨率影像數(shù)據(jù),運用圖像處理技術(shù)提取影像植被指數(shù)和植株目標形態(tài)特征指數(shù),分別建立模型估算植株數(shù)量并比較其可行性,以期為實現(xiàn)早期棉田株數(shù)信息的快速提取、高效棉苗監(jiān)測、提高棉花栽培管理水平提供有效的技術(shù)支撐。
1 材料與方法
1.1 研究區(qū)域
試驗于2020―2021年在石河子大學試驗站二連棉花試驗基地(44°19′46″N,85°59′50″E)開展,以魯棉研24號為供試品種,設置了3種密度處理,分別為低密度(D1,6.9×104株·hm-2,株距19 cm)、中密度(D2,13.8×104株·hm-2,株距9.5 cm)和高密度(D3,24×104株·hm-2,株距11 cm)。其中,低密度和中密度種植模式均為等行距種植模式,行距為76 cm;高密度種植模式為寬窄行種植模式,行距為66 cm+10 cm(試驗點位置、研究區(qū)域和密度設置見圖1)。2020年每個密度處理種植4條膜,共12條膜,面積約為945 m2(35 m×27 m);2021年每個密度處理種植3條膜,共9條膜,面積約為840 m2(40 m×21 m)。
1.2 試驗設計與影像獲取
本研究的影像數(shù)據(jù)基于MAVIC PRO四旋翼無人機(MAVIC PRO,大疆,中國)平臺,搭載影像傳感器。該無人機平臺在無風環(huán)境下最高水平飛行速度為 65 km·h-1,最大飛行高度為5 km,最遠續(xù)航里程為13 km。平臺搭載的相機配置:1/2.3英寸CMOS傳感器,焦距5 mm,實際輸出分辨率為1 200萬像素(4 000像素×3 000像素),采用FOV 78.8° 28 mm(35 mm格式等效)f/2.2鏡頭,同時配備GPS模塊和無線觸發(fā)器。無人機影像于當?shù)?2:00-14:00天空晴朗無云時獲取。飛行高度為10 m,飛行時,航向和旁向重疊率設為85%。相機采用自動拍攝模式。
本研究分別于2020年5月12日和2021年5月25日獲取無人機影像數(shù)據(jù),2次影像數(shù)據(jù)獲取時間均為棉花播種后25 d(棉苗處于3~4葉期),單次無人機飛行過程共拍攝影像300幅,圖像記錄為真彩色JPG格式。無人機獲取的影像數(shù)據(jù)首先借助Agisoft PhotoScan Professional軟件進行無人機高清數(shù)碼影像的拼接處理,生成試驗區(qū)域的高清數(shù)字正射影像(digital orthophoto map, DOM)。從無人機航空定位定向系統(tǒng)(position and orientation system, POS)中導出POS數(shù)據(jù),每個POS數(shù)據(jù)與無人機獲取的高清數(shù)碼影像是一一對應的,包含每張影像拍攝時的經(jīng)度、緯度、高度、偏航角、俯仰角和旋轉(zhuǎn)角共6個元素以表征影像獲取時的空間位置和姿態(tài)信息。圖1A中區(qū)域a和b分別是2020年和2021年研究區(qū)域的正射影像,對其進行切割獲取研究樣方。2020年將每條膜均分為20個樣方,每個密度處理均分為80個樣方,3個密度共240個研究樣方,樣方大小為2.28 m×1.75 m;2021年將每條膜均分為30個樣方,每個密度處理均分為90個樣方,3個密度共270個研究樣方,樣方大小為2.28 m×1.33 m。從2年研究區(qū)域的正射影像中獲得510個研究樣方,不同的樣方尺寸不會對本研究造成影響。
1.3 研究方法
1.3.1 技術(shù)路線。本研究中,利用集成高分辨率傳感器的低空無人機遙感平臺獲取影像數(shù)據(jù),基于Python語言并利用顏色植被指數(shù)和Otsu法對原始圖像的植株目標進行分割識別,獲得真彩色掩碼圖像(圖像植株部分不變;其余部分所有像素點R=0、G=0、B=0)和二值圖像。通過RGB提取算法直接提取真彩色掩碼圖像的RGB值,基于形態(tài)特征提取算法從二值圖像中獲取目標的周長、面積等特征值,從而構(gòu)建模型并檢驗。圖2為本研究的技術(shù)路線。
1.3.2 植被指數(shù)篩選。植株目標的提取即根據(jù)植株的某些特征將其從復雜的田間環(huán)境中分離出來。樣方圖像主要包含棉苗、地膜、土壤3種元素,其顏色具有顯著的差異,可作為分離植株目標的依據(jù)。根據(jù)前人研究結(jié)果[21-25],本研究基于三角綠度指數(shù)(triangular greenness index, TGI)、超綠指數(shù)(excess green index, ExG)、修正超綠指數(shù)(modified excess green index, MExG)、綠-藍差值指數(shù)(green-blue difference index, GBDI)和綠藍差值+修正超綠指數(shù)(GBDI + MExG,根據(jù)GBDI和MExG在作物目標識別中的特性提出的1種綜合指數(shù))進行棉花植株目標的識別與提取,篩選對棉花植株目標分割效果較好的顏色植被指數(shù)。表1列舉了各候選顏色植被指數(shù)的基本信息。
1.3.3 Otsu閾值分割。基于圖像的植物分割技術(shù)是指將圖像分為植物和非植物的過程,目前常用的植物提取算法有基于顏色植被指數(shù)、基于閾值和基于機器學習的分割方法[29-30]。Otsu法是圖像分割中常用的1種閾值確定方法,根據(jù)圖像的灰度特性將目標和背景分割為差異明顯的兩部分[31-32]。圖像的像素點為I(x, y),設背景和目標的分割閾值為灰度級T,背景的像素點數(shù)量占整幅圖像像素點數(shù)量的比例為p0,目標的像素點數(shù)量占整幅圖像像素點數(shù)量的比例為p1,整幅圖像的平均灰度為g,背景的平均灰度為g0,目標的平均灰度為g1,類間方差為d。當類間方差d取最大值時的閾值為最佳閾值T。公式如下:
1.3.4 圖像噪聲去除(中值濾波法)。圖像的前端采集往往會受器件或環(huán)境影響而使獲得的圖像含有噪聲,因而在圖像處理前需要進行去噪聲處理,中值濾波是1種非線性空域濾波方法,可以有效抑制圖像噪聲、提高圖像信噪比,能夠高效去除圖像中目標外的微小噪點,完整提取目標邊緣[33]。本研究在使用Otsu閾值分割算法提取植株目標時加入中值濾波算法去除圖像中的噪聲以便于后期掩碼圖像顏色植被指數(shù)和棉苗目標形態(tài)特征的準確提取。中值濾波算法如下:
1.3.5 基于真彩色掩碼圖像的顏色植被指數(shù)的提取。設真彩色掩碼圖像總像素點數(shù)量為n,每個像素點紅色、綠色、藍色的分量分別為IR、IG、IB,圖像的紅色、綠色、藍色的平均分量為R、G、B,計算公式如下:
基于樣方掩碼圖像提取的R、G、B數(shù)值,通過公式變換獲取棉苗密度估測所需的顏色植被指數(shù),獲取公式見表2。
1.3.6 植株目標形態(tài)特征的提取。從二值圖像中提取每個植株目標的形態(tài)特征,對表3中所列的9種植株的形態(tài)特征進行研究對比,從中篩選出合適的建模參數(shù),形態(tài)特征參數(shù)的計算公式見表3。
1.3.7 多元線性回歸建模。多元線性回歸是對具有多個自變量,且因變量和自變量之間是線性關系的1種分析方法[38]。在棉苗密度預測模型和棉苗目標株數(shù)預測模型構(gòu)建中,隨機選擇1.3.5和1.3.6兩種特征參數(shù)中的75%的數(shù)據(jù)作為建模數(shù)集,剩余25%的數(shù)據(jù)作為驗證數(shù)集。在模型構(gòu)建中首先對變量之間進行相關性分析,篩選出適當?shù)慕W宰兞?,排除具有共線性(即相關系數(shù)|r|>0.8)的自變量;再對所選自變量進行逐步回歸模型構(gòu)建。
1.3.8 模型精度檢驗。本研究對多元線性回歸模型的精度評價主要通過25%驗證數(shù)集的觀測值與預測值的比較、研究區(qū)域不同種植密度下棉苗株數(shù)及棉苗總株數(shù)調(diào)查值和預測值的比較,檢驗指標主要有決定系數(shù)(R2)、均方根誤差(root mean squared error, RMSE)、平均絕對誤差(mean absolute error, MAE)、平均百分比誤差(mean absolute percentage error, MAPE),計算公式如下:
2 結(jié)果與分析
2.1 棉苗的識別與分割
基于綠-藍差值指數(shù)(GBDI)、修正超綠指數(shù)(MExG)、綠-藍差值+修正超綠指數(shù)(GBDI+MExG)、超綠指數(shù)(ExG)和三角綠度指數(shù)(TGI)進行Otsu閾值化分割的結(jié)果如圖3所示。由圖3B和3C發(fā)現(xiàn),GBDI只能完整去除圖像中的地膜背景,MExG只能去除圖像中的土壤背景和棉苗陰影,而GBDI+MExG、ExG和TGI的顏色植被指數(shù)對棉苗具有較好的分割效果。
為從GBDI+MExG、ExG和TGI這3種顏色植被指數(shù)中篩選最優(yōu)分割效果的指數(shù),對這3種顏色植被指數(shù)所提取的棉苗目標進行了疊加分析,結(jié)果(圖4)表明,TGI對棉苗目標的分割完整度高于ExG和GBDI+MExG。因此,本研究選擇三角綠度指數(shù)(TGI)用于棉苗目標的識別和分割。
2.2 特征參數(shù)的提取
根據(jù)圖像的空間分辨率,結(jié)合棉苗的生長情況及對真彩色原圖的預處理結(jié)果,選擇4 cm2(約70個像素)為閾值,在植株目標分割過程中直接從圖像中剔除小于4 cm2的植株目標。最終,基于TGI從510幅真彩色掩碼圖像中提取了510組R、G、B值,并計算出9種常見的顏色植被指數(shù);從510幅二值圖像中分割提取了18 316個棉苗目標,圖5展示了用于棉苗目標形態(tài)特征提取的外接矩形邊界、目標邊界和目標。
2.3 棉花株數(shù)估測模型
2.3.1 基于圖像顏色植被指數(shù)的棉苗密度預測模型。為防止無效特征的引入,首先對自變量進行了篩選。表4所示為基于圖像顏色植被指數(shù)的建模變量之間的相關系數(shù),所有顏色植被指數(shù)與樣方棉苗密度y1均表現(xiàn)為正相關關系,其中a2(G)與y1的相關系數(shù)最大,達到0.948。排除與a2相關系數(shù)的絕對值大于0.8的自變量:a1(R)、a3(B)、a4(GRDI)、a5(GBDI)、a6(ExG)、a7(ExR)、a10(TGI)、a11(MExG)、a12(GBDI + MExG),初步選擇a2(G)、a8(ExB)、a9(ExGR)作為建模參數(shù),基于所選參數(shù)進行逐步多元回歸建模。由于在加入變量a9后模型精度變化微小,但因增加了模型的變量,使得模型預測變得復雜,最終選擇a2(G)和a8(ExB)為模型的自變量。表5展示了a2(G)和a8(ExB)建模數(shù)據(jù)與總體數(shù)據(jù)的統(tǒng)計分析結(jié)果,這2種顏色特征參數(shù)的數(shù)據(jù)符合正態(tài)分布且各統(tǒng)計指標差異很小,表明抽樣所得數(shù)據(jù)能夠較好反應總體數(shù)據(jù)的特征,能夠應用于建模當中?;谘诖a圖像的a2(G)值和a8(ExB)值與樣方棉苗密度y1構(gòu)建的多元線性回歸模型的決定系數(shù)R2為0.903 6,P<0.05,估計標準誤差為1.637 3,因此樣方棉苗密度與樣方掩碼圖像的G值、ExB存在線性關系,回歸方程(模型Ⅰ)為:y1=3.002 44×a2+0.768 14× a8-0.127 8。
2.3.2 基于目標形態(tài)特征的棉苗株數(shù)預測模型。為防止無效特征的引入,首先對自變量進行了篩選。表6所示為基于棉苗目標形態(tài)特征參數(shù)的建模變量之間的相關系數(shù),所有形態(tài)特征參數(shù)與因變量y2均表現(xiàn)為正相關關系,其中b2(目標面積)與y2的相關系數(shù)最大,達到0.946。排除與b2存在嚴重的共線性的自變量(相關系數(shù)的絕對值大于0.8):b1(周長)、b4(外接矩形的長)、b7(外接矩形的周長)、b8(外接矩形的面積)、b9(外接矩形面積周長比),初步選擇b2(目標面積)、b3(面積周長比)、b5(外接矩形的寬)、b6(外接矩形的長寬比)作為建模參數(shù)?;谒x參數(shù)進行逐步多元回歸建模,在引入變量b3和b6后,模型精度變化不大,但因增加了模型的變量,使得模型預測變得復雜,最終選擇b2(目標面積)和b5(目標外接矩形的寬)為模型的自變量。表5展示了b2(目標面積)和b5(目標外接矩形的寬)建模數(shù)據(jù)與總體數(shù)據(jù)的統(tǒng)計分析結(jié)果,這2種形態(tài)特征參數(shù)的數(shù)據(jù)符合正態(tài)分布且各統(tǒng)計指標差異很小,表明抽樣所得數(shù)據(jù)能夠較好反應總體數(shù)據(jù)的特征,能夠應用于建模當中?;赽2(目標面積)和b5(目標外接矩形的寬)與棉苗目標株數(shù)(y2)構(gòu)建的多元線性回歸模型的決定系數(shù)R2為0.935 5,P<0.05,估計標準誤差為0.233 0,因此棉苗目標與其面積、外接矩形的寬存在線性關系,回歸方程(模型Ⅱ)為:y2=0.024 14×b2-0.117 97×b5+0.978 92。
2.4 模型精度評價
樣方棉苗密度調(diào)查值與模型I預測值的擬合效果見圖6A:R2為0.889 8;RMSE為1.743 9;MAE為1.301 1;MAPE為10.10%;目標棉苗株數(shù)觀測值與模型II預測值的擬合效果見圖6B:R2為0.927 7;RMSE為0.028 7;MAE為0.171 0;MAPE為14.49%。各項指標顯示,2個模型均有較好的預測效果。
比較2個模型對2年來研究區(qū)域不同密度及總體株數(shù)的預測精度(表7)。其中,2個模型對D1、D2密度的預測精度均高于D3。模型Ⅱ?qū)Σ煌芏鹊念A測精度均高于模型Ⅰ,但2個模型對不同密度及總體株數(shù)都具有較高的預測精度,均達95%以上。綜上所述,2類模型對不同密度下膜下滴灌棉花苗期株數(shù)均具有較好的估算預測效果,模型Ⅱ略優(yōu)于模型Ⅰ。
3 討論
近年來,利用低空遙感技術(shù)獲取作物冠層影像,通過對目標作物或病蟲害的特征對作物冠層影像進行顏色特征參數(shù)或者形態(tài)特征參數(shù)的提取與計算的方法已用于作物分類和病蟲害識別[39]。Li等[23]和武威等[40]基于超綠指數(shù)(ExG)對作物冠層圖像進行分割處理,均有效提取了植株目標。García-Martínez等[11]基于三角綠度指數(shù)(TGI)估算玉米冠層覆蓋度,并結(jié)合作物密度及超綠指數(shù)(ExG)、三角綠度指數(shù)(TGI)等植被指數(shù)對玉米產(chǎn)量進行估算,R2達到0.97。Abrantes等[41]的研究結(jié)果表明,超綠指數(shù)(ExG)能高效地對大豆植物損傷和大豆籽粒產(chǎn)量進行評估。本研究利用低空遙感技術(shù)在棉花苗期(3~4葉期)獲取棉花冠層影像,對比分析了三角綠度指數(shù)(TGI)、超綠指數(shù)(ExG)和綠-藍差值+修正超綠指數(shù)(GBDI+MExG)等3種顏色植被指數(shù)對不同種植密度下棉花識別提取的適用性,發(fā)現(xiàn)三角綠度指數(shù)(TGI)、超綠指數(shù)(ExG)、綠-藍差值+修正超綠指數(shù)(GBDI+MExG)均對圖像有較好的分割效果,但TGI的分割效果最好。
當前,不少學者利用無人機低空遙感技術(shù)進行作物株數(shù)估算[42-43]。劉帥兵等[42]基于無人機遙感影像,利用骨架提取算法結(jié)合Harris角點檢測提取了玉米苗期株數(shù),總體識別率達97.8%,這種方法的識別率很高,但算法復雜,且對密度大和重疊率高的作物效果不佳,具有一定的局限性。付虹雨等[43]基于無人機遙感影像,利用面向?qū)ο蠖喑叨确指钏惴▽崿F(xiàn)了對劍麻株數(shù)的自動提取,識別精度達87.1%,這種方法分割效果好且操作簡單,但多尺度分割易出現(xiàn)過分割現(xiàn)象,使株數(shù)識別精度下降。本研究利用顏色特征和形態(tài)特征2種方法對苗期棉花目標進行提取估算,結(jié)果表明:(1)基于棉花冠層植被指數(shù)結(jié)合Otsu閾值分割方法[20, 23, 35-36, 40-41],并加入中值濾波去噪,能夠精確并完整地分割圖像背景與植株目標,獲得掩碼圖像,根據(jù)掩碼圖像單位面積的顏色植被指數(shù)總值可對植株密度進行估算,分割效果好,株數(shù)估算準確率高,操作簡單。(2)對掩碼圖像的棉花目標進行形態(tài)學分析,利用形態(tài)特征參數(shù)對目標對應的株數(shù)進行估算,準確率較基于顏色植被指數(shù)的株數(shù)估算模型更高,且在高密度下仍能準確估算棉苗株數(shù)。
本研究中的2種株數(shù)估算模型在3種試驗密度下均獲得了較好的結(jié)果,表明這2種方法在不同種植密度的大田中棉花株數(shù)估算的普遍適用性。但是,本研究中影像數(shù)據(jù)的采集在棉花的苗期且棉花為膜下滴灌種植方式,背景中大的干擾物較少,小的干擾物容易去除,植株目標識別提取準確,因此,本研究的2種方法僅適用于在背景影響較小的種植環(huán)境中進行作物株數(shù)的估算,下一步將針對作物行間大的干擾物的準確識別去除開展進一步的研究。
4 結(jié)論
本研究利用無人機低空遙感技術(shù),通過對目標植株的特征值進行提取,構(gòu)建了2種不同特征參數(shù)的棉花苗期株數(shù)估算模型,結(jié)論如下:(1)顏色植被指數(shù)TGI相較于ExG和GBDI+MExG對棉花苗期(3~4葉期)研究區(qū)域的棉苗目標的識別效果最好;(2)整體來看,基于顏色植被指數(shù)的棉花株數(shù)估算模型和基于植株目標形態(tài)特征參數(shù)的棉花株數(shù)估算模型均能對棉花苗期(3~4葉期)株數(shù)進行精確估算,預測精度均達到95%以上,但基于植株目標形態(tài)特征參數(shù)的棉花株數(shù)估算模型的精度更高。
參考文獻:
[1] 張賀軒, 徐愛武. 新疆棉花生產(chǎn)地位分析[J/OL]. 中國棉花加工, 2020(4): 4-7[2022-04-21]. https://doi.org/10.3969/j.issn.1003-0662.2020.04.002.
Zhang Hexuan, Xu Aiwu. Analysis of cotton production status in Xinjiang[J/OL]. China Cotton Processing, 2020(4): 4-7[2022-04-21]. https://doi.org/10.3969/j.issn.1003-0662.2020.04.002.
[2] 喻樹迅, 周亞立, 何磊. 新疆兵團棉花生產(chǎn)機械化的發(fā)展現(xiàn)狀及前景[J/OL]. 中國棉花, 2015, 42(8): 1-4[2022-04-21]. https://doi.org/10.11963/issn.1000-632X.201508001.
Yu Shuxun, Zhou Yali, He Lei. Development status and prospect of cotton production mechanization in Xinjiang Corps[J/OL]. China Cotton, 2015, 42(8): 1-4[2022-04-21]. https://doi.org/10.11963/issn.1000-632X.201508001.
[3] Li H M , Ma J H, Wang P. Cotton bollworm resistance to the Bt cotton and management strategy in Xinjiang, China[J]. Egyptian Journal of Biological Pest Control, 2014, 24(2): 533-541.
[4] 楊長琴, 張國偉, 劉瑞顯, 等. 種植密度和縮節(jié)胺調(diào)控對麥后直播棉產(chǎn)量和冠層特征的影響[J/OL]. 棉花學報, 2016, 28(4): 331-338[2022-04-21]. https://doi.org/10.11963/issn.1002-7807.201604003.
Yang Changqin, Zhang Guowei, Liu Ruixian, et al. Effects of planting density and growth regulator mepiquat chloride on yields and canopy architecture of cotton sown after harvesting barley[J/OL]. Cotton Science, 2016, 28(4): 331-338[2022-04-21]. https://doi.org/10.11963/issn.1002-7807.201604003.
[5] 支曉宇, 韓迎春, 王國平, 等. 不同密度下棉花群體光輻射空間分布及生物量和纖維品質(zhì)的變化[J/OL]. 棉花學報, 2017, 29(4): 365-373[2022-04-21]. https://doi.org/10.11963/1002-7807.zxylyb.20170407.
Zhi Xiaoyu, Han Yingchun, Wang Guoping, et al. Spatial distribution of optical radiation and changes in biomass and fiber quality of cotton populations at different densities[J/OL]. Cotton Science, 2017, 29(4): 365-373[2022-04-21]. https://doi.org/ 10.11963/1002-7807.zxylyb.20170407.
[6] 紀景純, 趙原, 鄒曉娟, 等. 無人機遙感在農(nóng)田信息監(jiān)測中的應用進展[J/OL]. 土壤學報, 2019, 56(4): 773-784[2022-04-21]. https://doi.org/10.11766/trxb201811190508.
Ji Jingchun, Zhao Yuan, Zou Xiaojuan, et al. Application progress of UAV remote sensing in farmland information monitoring[J/OL]. Acta Metallurgica Sinica, 2019, 56(4): 773-784[2022-04-21]. https://doi.org/10.11766/trxb201811190508.
[7] Chen P C, Chiang Y C, Weng P Y. Imaging using unmanned aerial vehicles for agriculture land use classification[J/OL]. Agriculture, 2020, 10(9): 416[2022-04-21]. https://doi.org/10.3390/agriculture10090416.
[8] 姚立民, 李明, 謝景鑫, 等. 無人機低空遙感在作物育種中的應用研究進展[J/OL]. 湖南農(nóng)業(yè)科學, 2020(11): 108-112[2022-04-21]. https://doi.org/10.16498/j.cnki.hnnykx.2020.011.028.
Yao Limin, Li Ming, Xie Jingxin, et al. Research progress on the application of low altitude remote sensing by unmanned aerial vehicles in crop breeding[J/OL]. Hunan Agricultural Sciences, 2020(11): 108-112[2022-04-21]. https://doi.org/10.16498/j.cnki.hnnykx.2020.011.028.
[9] Olson D, Anderson J. Review on unmanned aerial vehicles, remote sensors, imagery processing, and their applications in agriculture[J/OL]. Agronomy Journal, 2021, 113: 971-992[2022-04-21]. https://doi.org/10.1002/agj2.20595.
[10] 田明璐, 班松濤, 常慶瑞, 等. 基于低空無人機成像光譜儀影像估算棉花葉面積指數(shù)[J/OL]. 農(nóng)業(yè)工程學報, 2016, 32(21): 102-108[2022-04-21]. https://doi.org/10.11975/j.issn.1002-6819.2016.21.014.
Tian Minglu, Ban Songtao, Chang Qingrui, et al. Estimation of cotton leaf area index based on low-altitude UAV imaging spectrometer image[J/OL]. Transactions of the Chinese Society of Agricultural Engineering, 2016, 32(21): 102-108[2022-04-21]. https://doi.org/10.11975/j.issn.1002-6819.2016.21.014.
[11] García-Martínez H, Flores-Magdaleno H, Ascencio-Hernández R, et al. Corn grain yield estimation from vegetation indices, canopy cover, plant density, and a neural network using multispectral and RGB images acquired with unmanned aerial vehicles[J/OL]. Agriculture, 2020, 10(7): 277[2022-04-21]. https://doi.org/10.3390/agriculture10070277.
[12] Abdullah M M, Al-Ali Z M, Abdullah M T, et al. Investigating the applicability of UAVs in characterizing desert shrub biomass and developing biological indicators for the selection of suitable revegetation sites[J/OL]. Journal of Environmental Management, 2021, 288: 112416[2022-04-21]. https://doi.org/10.1016/j.jenvman.2021.112416.
[13] Wengert M, Piepho H P, Astor T, et al. Assessing spatial variability of barley whole crop biomass yield and leaf area index in silvoarable agroforestry systems using UAV-borne remote sensing[J/OL]. Remote Sensing, 2021, 13(14): 2751[2022-04-21]. https://doi.org/10.3390/rs13142751.
[14] 張凝, 楊貴軍, 趙春江, 等. 作物病蟲害高光譜遙感進展與展望[J/OL]. 遙感學報, 2021, 25(1): 403-422[2022-04-21]. https://doi.org/10.11834/jrs.20210196.
Zhang Ning, Yang Guijun, Zhao Chunjiang, et al. Progress and prospect of hyperspectral remote sensing of crop diseases and insect pests[J/OL]. Journal of Remote Sensing, 2021, 25(1): 403-422[2022-04-21]. https://doi.org/10.11834/jrs.20210196.
[15] Luthfan N H, Tomoya W, Tsutomu M, et al. Machine learning techniques to predict soybean plant density using UAV and satellite-based remote sensing[J/OL]. Remote Sensing, 2021, 13: 2548[2022-04-21]. https://doi.org/10.3390/rs13132548.
[16] Randelovi"P, Dordevi"V, Milic S, et al. Prediction of soybean plant density using a machine learning model and vegetation indices extracted from RGB images taken with a UAV[J/OL]. Agronomy, 2020, 10: 1108[2022-04-21]. https://doi.org/10.3390/agronomy10081108.
[17] 王偉, 王新盛, 姚嬋, 等. 基于無人機影像的小麥植株密度估算方法研究[J/OL]. 國土資源遙感, 2020, 32(4): 111-119[2022-04-21]. https://doi.org/10.6046/gtzyyg.2020.04.16.
Wang Wei, Wang Xinsheng, Yao Chan, et al. Study on wheat plant density estimation method based on drone image[J/OL]. Remote Sensing for Land amp; Resources, 2020, 32(4): 111-119[2022-04-21]. https://doi.org/10.6046/gtzyyg.2020.04.16.
[18] Zhao B Q, Zhang J, Yang C H, et al. Rapeseed seedling stand counting and seeding performance evaluation at two early growth stages based on unmanned aerial vehicle imagery[J/OL]. Frontiers in Plant Science, 2018, 9: 1362[2022-04-21]. https://doi.org/10.3389/fpls.2018.01362.
[19] 朱孟, 周忠發(fā), 趙馨, 等. 基于無人機遙感的喀斯特高原峽谷區(qū)火龍果單株識別提取方法[J/OL]. 熱帶地理, 2019, 39(4): 502-511[2022-04-21]. https://doi.org/10.13284/j.cnki.rddl.003146.
Zhu Meng, Zhou Zhongfa, Zhao Xin, et al. Identification and extraction method of dragon fruit per plant in karst plateau canyon area based on drone remote sensing[J/OL]. Tropical Geography, 2019, 39(4): 502-511[2022-04-21]. https://doi.org/10.13284/j.cnki.rddl.003146.
[20] 戴建國, 薛金利, 趙慶展, 等. 利用無人機可見光遙感影像提取棉花苗情信息[J/OL]. 農(nóng)業(yè)工程學報, 2020, 36(4): 63-71[2022-04-21]. https://doi.org/10.11975/j.issn.1002-6819.2020.04.008.
Dai Jianguo, Xue Jinli, Zhao Qingzhan, et al. Extraction of cotton seedling information using UAV visible light remote sensing images[J/OL]. Transactions of the Chinese Society of Agricultural Engineering, 2020, 36(4): 63-71[2022-04-21]. https://doi.org/10.11975/j.issn.1002-6819.2020.04.008.
[21] Guo Z C, Wang T, Liu S L, et al. Biomass and vegetation coverage survey in the Mu Us sandy land-based on unmanned aerial vehicle RGB images[J/OL]. International Journal of Applied Earth Observations and Geoinformation, 2021, 94: 102239[2022-04-21]. https://doi.org/10.1016/j.jag.2020.102239.
[22] Sánchez-Sastre L F, Alte da Veiga N M S, Ruiz-Potosme N M, et al. Assessment of RGB vegetation indices to estimate chlorophyll content in sugar beet leaves in the final cultivation stage[J/OL]. AgriEngineering, 2020, 2(1): 128-149[2022-04-21]. https://doi.org/10.3390/agriengineering2010009.
[23] Li B, Xu X M, Han J W, et al. The estimation of crop emergence in potatoes by UAV RGB imagery[J/OL]. Plant Methods, 2019, 15(1): 15[2022-04-21]. https://doi.org/10.1186/s13007-019-0399-7.
[24] Elsayed S, El-Hendawy S, Khadr M, et al. Combining thermal and RGB imaging indices with multivariate and data-driven modeling to estimate the growth, water status, and yield of potato under different drip irrigation regimes[J/OL]. Remote Sensing, 2021, 13: 1679[2022-04-21]. https://doi.org/10.3390/rs13091679.
[25] Hasan U, Sawut M, Chen S. Estimating the leaf area index of winter wheat based on unmanned aerial vehicle RGB-image parameters[J/OL]. Sustainability, 2019, 11: 6829[2022-04-21]. https://doi.org/10.3390/su11236829.
[26] Mao W, Wang Y, Wang Y. Real-time detection of between-row weeds using machine vision[C/OL]//American Society of Agricultural and Biological Engineers: 2003 ASAE Annual Meeting. Michigan: American Society of Agricultural and Biological Engineers, 2003[2022-04-21]. https://doi.org/10.13031/2013.15381.
[27] Marques M G, Cunha J, Lemes E M. Dicamba injury on soybean assessed visually and with spectral vegetation index[J/OL]. AgriEngineering, 2021, 3(2): 240-250[2022-04-21]. https://doi.org/10.3390/agriengineering3020016.
[28] Hamuda E, Glavin M, Jones E. A survey of image processing techniques for plant extraction and segmentation in the field[J/OL]. Computers and Electronics in Agriculture, 2016, 125:184-199[2022-04-21]. https://doi.org/10.1016/j.compag.2016.04.024.
[29] Arroyo J, Guijarro M, Pajares G. An instance-based learning approach for thresholding in crop images under different outdoor conditions[J/OL]. Computers and Electronics in Agriculture, 2016, 127: 669-679[2022-04-21]. https://doi.org/10.1016/j.compag.2016.07.018.
[30] Wang A C, Zhang W, Wei X H. A review on weed detection using ground-based machine vision and image processing techniques[J/OL]. Computers and Electronics in Agriculture, 2019, 158: 226-240[2022-04-21]. https://doi.org/10.1016/j.compag.2019.02.005.
[31] Bezdek J C. Pattern recognition with fuzzy objective function algorithms[M/OL]. New York: Plenum Press, 1981[2022-04-21]. https://doi.org/10.1007/978-1-4757-0450-1.
[32] Wang C, Yang J, Lü H. Otsu multi-threshold image segmentation algorithm based on improved particle swarm optimization[C/OL]//2019 IEEE 2nd International Conference on Information Communication and Signal Processing (ICICSP). IEEE, 2019[2022-04-21]. https://doi.org/10.1109/icicsp48821.2019.8958573.
[33] 祁欣, 陳劍鋒, 羅偉林. 圖像預處理算法的FPGA實現(xiàn)[J/OL]. 國外電子測量技術(shù), 2021, 40(2): 102-107[2022-04-21]. https://doi.org/10.19652/j.cnki.femt.2002395.
Qi Xin, Chen Jianfeng, Luo Weilin. FPGA implementation of image preprocessing algorithm[J/OL]. Electronic Measurement Technology, 2021, 40(2): 102-107[2022-04-21]. https://doi.org/10.19652/j.cnki.femt.2002395.
[34] Motohka T, Nasahara K N, Oguma H, et al. Applicability of green-red vegetation index for remote sensing of vegetation phenology[J/OL]. Remote Sensing, 2010, 2(10): 2369-2387[2022-04-21]. https://doi.org/10.3390/rs2102369.
[35] Liu H J, Sun H, Li M Z, et al. Application of color featuring and deep learning in maize plant detection[J/OL]. Remote Sensing, 2020, 12(14): 2229[2022-04-21]. https://doi.org/10.3390/rs12142229.
[36] Guijarro M, Pajares G, Riomoros I, et al. Automatic segmentation of relevant textures in agricultural images[J/OL]. Computers and Electronics in Agriculture, 2011, 75(1): 75-83[2022-04-21]. https://doi.org/10.1016/j.compag.2010.09.013.
[37] Liu X. Support vector data description for weed/corn image recognition[J/OL]. Journal of Food Agriculture and Environment, 2010, 8(1): 214-219[2022-04-21]. https://doi.org/10.1023/B:MACH.0000008084.60811.49.
[38] Denis D J. Univariate, bivariate, and multivariate statistics using R: quantitative tools for data analysis and data science[M/OL]. New York: John Wiley amp; Sons, 2020[2022-04-21]. https://doi.org/10.1002/9781119549963.
[39] Sabzi S, Abbaspour-Gilandeh Y, García-Mateos G. A fast and accurate expert system for weed identification in potato crops using metaheuristic algorithms[J/OL]. Computers in Industry, 2018, 98: 80-89[2022-04-21]. https://doi.org/10.1016/j.compind.2018.03.001.
[40] 武威, 劉濤, 孫成明,等. 基于圖像處理的小麥種植密度估算研究[J/OL]. 揚州大學學報(農(nóng)業(yè)與生命科學版), 2017, 38(1): 89-93[2022-04-21]. https://doi.org/10.16872/j.cnki.1671-4652.2017.01.017.
Wu Wei, Liu Tao, Sun Chengming, et al. Research on estimation of wheat planting density based on image processing[J/OL]. Journal of Yangzhou University (Agriculture and Life Sciences Edition), 2017, 38(1): 89-93[2022-04-21]. https://doi.org/10.16872/j.cnki.1671-4652.2017.01.017.
[41] Abrantes T C , Queiroz A R S, Lucio F R, et al. Assessing the effects of dicamba and 2,4 Dichlorophenoxyacetic acid (2,4 D) on soybean through vegetation indices derived from Unmanned Aerial Vehicle (UAV) based RGB imagery[J/OL]. International Journal of Remote Sensing, 2021, 42(7): 2740-2758[2022-04-21]. https://doi.org/10.1080/01431161.2020.1832283.
[42] 劉帥兵, 楊貴軍, 周成全, 等. 基于無人機遙感影像的玉米苗期株數(shù)信息提取[J/OL]. 農(nóng)業(yè)工程學報, 2018(22): 69-75[2022-04-21]. https://doi.org/10.11975/j.issn.1002-6819.2018.22.009.
Liu Shuaibing, Yang Guijun, Zhou Chengquan, et al. Extraction of plant number information at maize seedling stage based on UAV remote sensing image[J/OL]. Transactions of the Chinese Society of Agricultural Engineering, 2018(22): 69-75[2022-04-21]. https://doi.org/10.11975/j.issn.1002-6819.2018.22.009.
[43] 付虹雨, 崔國賢, 崔丹丹, 等. 基于無人機遙感影像的劍麻株數(shù)識別[J/OL]. 中國麻業(yè)科學, 2020, 42(6): 249-256[2022-04-21]. https://doi.org/10.3969/j.issn.1671-3532.2020.06.001.
Fu Hongyu, Cui Guoxian, Cui Dandan, et al. Sisal strain number recognition based on UAV remote sensing image[J/OL]. Chinese Journal of Hemp Industry, 2020, 42(6): 249-256[2022-04-21]. https://doi.org/10.3969/j.issn.1671-3532.2020.06.001.
(責任編輯:付毓" " 責任校對:秦凡)