赵晋陵 杜世州 黄林生

摘要：白粉病主要侵染小麦叶部，可利用卫星遥感技术进行大范围监测和评估。本研究利用多源多时相卫星遥感影像监测小麦白粉病并提升分类精度。使用四景 Landat-8的热红外传感器数据（Thermal Infrared Sensor ，TIRS ）和20景 MODIS 影像的 MOD11A1温度产品反演地表温度（Land Surface Temperature ， LST ），使用4景国产高分一号（ GF-1） 宽幅相机数据（Wide Field of View ，WFV ）提取小麦种植区和计算植被指数。首先，利用ReliefF算法优选对小麦白粉病敏感的植被指数，然后利用时空自适应反射率融合模型（Spa? tial and Temporal Adaptive Reflectance Fusion Model ，STARFM ）对 Landsat-8 LST 和 MOD11A1数据进行时空融合。利用 Z-score 标准化方法对植被指数和温度数据统一量度。最后，将处理和融合后的单一时项 Landsat-8 LST 、多时相 Landsat-8 LST 、累加 MODIS LST 和多时相Landsat-8 LST 与累加 MODIS LST 结合的数据分别输入支持向量机（Support Vector Machine ， SVM ）構建了四个分类模型，即 LST-SVM 、SLST-SVM 、MLST- SVM 和 SMLST-SVM ，利用用户精度、生产者精度、总体精度和 Kappa 系数对比四个模型的分类精度。结果显示，本研究构建的 SMLST-SVM 取得了最高分类精度，总体精度和 Kappa 系数分别为81.2%和0.67，而 SLST-SVM 则为76.8%和0.59。表明多源多时相的 LST 联合 SVM 能够提升小麦白粉病的识别精度。

关键词：小麦白粉病;高分一号;MODIS ;Landsat-8;地表温度;支持向量机

Monitoring Wheat Powdery Mildew （Blumeriagraminis f. sp. tritici） Using Multisource and Multitemporal SatelliteImages and Support Vector Machine Classifier

ZHAO Jinling1 ， DU Shizhou2* ， HUANG Linsheng1

（1. National Engineering Research Centerfor Analysis and Application of Agro-Ecological Big Data， Anhui Univer‐sity， Hefei 230601， China;2. Institute of Crops， Academy of Agricultural Sciences， Hefei 230031， China ）

Abstract： Since powdery mildew （Blumeriagraminis f. sp. tritici） mainly infects the foliar of wheat， satellite remote sensing technology can be used to monitor and assess it on a large scale. In this study， multisource and multitempo‐ral satellite images were used to monitor the disease and improve the classification accuracy. Specifically， four Land‐ sat-8 thermal infrared sensor （TIRS） and twenty MODerate-resolution imaging spectroradiometer （MODIS） temper‐ature product （MOD11A1） were used to retrieve the land surface temperature （LST）， and four Chinese Gaofen-1（GF-1） wide field of view （WFV） images was used to identify the wheat-growing areas and calculate the vegetation indices （VIs）. ReliefF algorithm was first used to optimally select the vegetation index （VIs） sensitive to wheat pow ‐dery mildew， spatial-temporal fusion between Landsat-8 LST and MOD11A1 data was performed using the spatial and temporal adaptive reflectance fusion model （STARFM）. The Z-score standardization method was then used to unify the VIs and LST data. Four monitoring models were then constructed through a single Landsat-8 LST， multi‐ temporal Landsat-8 LSTs （SLST）， cumulative MODIS LST （MLST） and the combination of cumulative Landsat-8 and MODIS LST （SMLST） using the Support Vector Machine （SVM） classifier， that were LST-SVM， SLST-SVM， MLST-SVM and SMLST-SVM. Four assessment indicators including user accuracy， producer accuracy， overall ac‐ curacy and Kappa coefficient were used to compare the four models. The results showed that， the proposed SMLST- SVM obtained the best identification accuracies. The overall accuracy and Kappa coefficient of the SMLST-SVM model had the highest values of 81.2% and 0.67， respectively， while they were respectively 76.8% and 0.59 for the SLST-SVM model. Consequently， multisource and multitemporal LSTs can considerably improve the differentiation accuracies of wheat powdery mildew.

Key words： wheat powdery mildew; GF-1; MODIS; Landsat-8; land surface temperature; support vector machine

CLC number：S435.121.4;TP79               Documents code：A          Article ID：SA202202009

Citation：ZHAO Jinling， DU Shizhou， HUANG Linsheng. Monitoring wheat powdery mildew （Blumeriagraminis f. sp. tritici） using multisource and multitemporal satellite images and support vector machine classifier[J]. Smart Agri‐ culture， 2022， 4（1）：17-28.（in English with Chinese abstract）

趙晋陵， 杜世州， 黄林生.联合多源多时相卫星影像和支持向量机的小麦白粉病监测方法[J].智慧农业（中英文）， 2022， 4（1）：17-28.

1  Introduction

Powdery mildew （Blumeriagraminis f. sp. trit‐ici） can occur at all stages of the wheat growth. It is a serious disease in some provinces of China， such as Sichuan， Guizhou， and Yunnan[1]. In recent years， it has become more severe in the wheat-growing ar‐eas of Northeastern China， North China， and North ‐ western China. When infected by the disease， some serious results will be caused such as early wither‐ing of leaves， decrease of panicle number， and de‐ crease of 1000-grain weight. Generally， the disease can lead to a 5%-10% reduction in yield and when it is seriously infected， a more serious loss of more than 20% can occur[2]. In view of the harmful effects on wheat production caused by the wheat powdery mildew，  it  is  of great  significance  to  improve  the monitoring efficiency and accuracy. However， it is difficult for traditional in-situ sampling and random investigation  methods  to  meet the needs  of large- scale monitoring due to the limitations in terms of timeliness，  economy，  and  accuracy[3]. Fortunately， the  modern  information  technology  has  facilitated the accurate and efficient identification of crop dis‐ eases to ensure food security[4–6].

In  recent  years，  the  development  of  remote sensing  technology  has  provided  an  important means  for  monitoring  and  forecasting  large-scale wheat diseases and insect pests. It obtains crop in ‐ formation quickly， accurately and objectively. Many scholars have  studied remote  sensing-based  moni‐toring of wheat diseases and insect pests by using ground-based， airborne and spaceborne remote sens‐ing data. For example， Huang et al.[7] showed that the  photochemical  reflectance  index （PRI） was strongly correlated with wheat yellow rust and the coefficient  of determination （R2） could  reach 0.97 for the PRI-based monitoring model. Zhang et al.[8] built a discriminant model for wheat powdery mil‐dew  severities  by  introducing  continuous  waveletanalysis based on the leaf-scale hyperspectral data.Luo  et  al.[9]  constructed  a  monitoring  model  ofwheat  aphid  in  two-dimensional  feature  space  de‐rived from the modified normalized difference wa‐ter  index （MNDWI） and  land  surface  temperature（LST） based on Landsat-5 TM imagery， which hada high discriminant precision. Zheng et al.[10] con ‐structed a red edge disease stress index （REDSI） us‐ing  three  red-edge  bands  from  Sentinel-2 satelliteimagery  for  monitoring  the  stripe  rust  on  winterwheat and got a satisfying result. The above studiesshowed that the spaceborne remote sensing imageshave greatly facilitated the monitoring and diagno‐sis of wheat diseases.

In previous  studies， multispectral satellite im‐ages were adopted to investigate the large-scale dis‐ease occurrence. Since the rapid development of hy‐perspectral remote sensing technology， some studieson wheat powdery mildew had been explored usingthe  hyperspectral  data. More  vegetation  indiceswere derived from the hundreds of spectral bands ofhyperspectral data. He et al.[11] improved the moni‐toring accuracy of wheat powdery mildew severityby selecting suitable observation angles and devel‐oping a novel vegetation index （VI）. Zhao et al.[12]identified the powdery mildew  severities of wheatleaves  quantitatively  on  hyperspectral  images  andimage  segmentation techniques. Khan  et  al.[13]  de‐tected the disease using a partial least-squares lineardiscrimination  analysis  and  the  combined  optimalfeatures （i. e.， normalized difference texture indices（NDTIs） and VIs）.

It  can  be  found  that  the  disease  monitoringmainly depends on the vegetation changes betweenhealthy and diseased wheat. Nevertheless， the com ‐mission and omission errors are usually caused dueto  the  influences  of  other  stress  types  such  asdrought，  inadequate  nutrition，  and  other  diseases. The  incidence  of  wheat  powdery  mildew  is  in ‐ volved in several affecting factors such as tempera‐ture， humidity and planting system. Considering the availability of temperature data of satellite images for large-scale monitoring， in this  study， particular attention was  given to the  contribution  of LST to the disease occurrence. As a key habitat factor， LST was included in the construction of monitoring mod‐el for the disease. LST remote sensing image was usually used in the previous monitoring of the dis‐ ease[14] ， but it has a cumulative effect on wheat pow ‐dery mildew. A single-phase LST image cannot ac‐curately represent the disease occurrence condition during the whole growth period. The primary objec‐tive of this study was to explore the availability and feasibility to identify wheat powdery mildew using a  combination  of  multisource  and  multitemporal spaceborne remote  sensing  imagery. More  specifi‐cally， three types of satellite images were adopted to identify  the  wheat-planting  areas  and  retrieve  the LST. Single and multitemporal LSTs were input in ‐ to  the  support vector machine （SVM） to  compare the monitoring effects.

2  Materials and methods

2.1 Study area

The study area is located in Jinzhou City， He‐bei province， China （114.97°-115.20° E， 37.80°-38.17° N）（Fig.1）. It has a warm-temperate conti‐nental monsoon  climate， with  a  flat  and  open ter‐ rain. There is a significant seasonal variation of sun radiation. Wheat is one of the important grain crops and widely planted in this area. The critical periods of wheat growth range from April to May. Due to its  flat  terrain，  appropriate  climate  conditions  and relatively  single  planting  structure，  the  region  issuitable for studying wheat powdery mildew usingremote sensing technology[15]. The historic statisticaldata  also  show  that  the  occurrence  frequency  ofwheat powdery mildew is high and hazard degree isserious[16]. The wheat-planting areas were extractedby combining the elevation data and reflectance ofthe near-infrared （NIR） band of Chinese Gaofen-1（GF-1） by  using  decision  tree  classification  tech ‐niques （Fig.1（b））. The extraction results were com ‐pared with the statistical data of Shijiazhuang City，which could fulfill the accuracy requirement of re‐mote sensing-based crop extraction.

2.2 Data collection and pre-processing

The data used in this study mainly were space‐borne remote sensing images and field survey dataof wheat powdery mildew. In-situ experiments werecarried out on 27 and 28 May 2014 in the central ar‐eas of Jinzhou City. Several typical experimental re‐gions were selected to collect the ground-truth data， where wheat was widely planted  and wheat pow ‐dery mildew occurs frequently. A total of 69 valid data were acquired using the method presented by Yuan et al.[17]. The wheat powdery mildew severities were  specified  according  to  the  Rules  for  the  in ‐vestigation and forecast of wheat powdery mildew [B. graminis （DC.） Speer]（NY/T 613-2002）. First‐ly，  the  average  severity  D  of diseased  leaves  for the  colony  leaves  was  calculated  according  to Equation （1）. Then， the disease index I could be de‐ rived from Equation （2）. The five levels were final‐ly  obtained  according  to  Table 1. To  increase  the comparability of remote sensing-based disease mon‐itoring， the five levels of severities were further di‐ vided into three levels of 0（healthy）， 1（mild） and 2（severe）.

where di  is the value of eight severity levels， which is 1%， 5%， 10%， 20%， 40%， 60%， 80%， 100%; li  is the number of diseased leaves for each level; L is the total number of investigated diseased leaves.

where F is the percentage of diseased leaf， which is ratio between the diseased leaves and total investi‐ gated leaves.

Multisource  remote  sensing  data  were  com ‐ posed of GF-1， Landsat-8 and MODerate-resolution imaging spectroradio-meter （MODIS） data products.

GF-1： Four scenes of GF-1 wide field of view （WFV） images were collected on 6， 17 and 26 May， and 7 June 2014，  respectively，  with  the  path/rownumbers of 3/93 and 4/92. There are four spectralbands for the GF-1 WFV sensor within the spectralrange of 0.45-0.89μm， with the spatial resolutionof 16 m. The primary preprocessing procedures wereperformed  including  orthorectification，  radiationcalibration， atmospheric correction and image sub ‐setting. The radiation brightness （Le ， W/（m2·sr ·μm））of a WFV image can be derived from Equation （3）.

where DN is the digital number of pixel. The abso‐lute  radiometric  calibration  coefficients （gain  andoffset） were derived from the China Center for Re‐sources Satellite Data and Application （http：//www.cresda. com/CN/）. After  finishing  the  radiometriccalibration，  the  Fast  Line-of-Sight  AtmosphericAnalysis of Spectral Hypercubes （FLAASH） mod‐ule in ENVI 5.3 software was used to complete theatmospheric  correction. The  geometric  correctionwas carried out using the second-order polynomialmodel with the accuracy of less than a pixel.

Landsat-8： The Landsat-8 operational land im‐ager （OLI） data were used to estimate the land re‐flectance and the thermal infrared sensor （TIRS） da‐ta were used for retrieving the LST[18]. Four scenesof cloud-free images were selected and their acquisi‐tion date was 4 April， 23 April， 15 May and 22 May2014， respectively. They were preprocessed includ‐ing accurate geometric correction and radiation cor‐rection.

MOD11A1 product： It provided daily per-pix ‐el  land  surface  temperature  and  emissivity（LST&E） with 1 km spatial resolution[19]. A total of20 images were  obtained  from  the  land processesdistributed active archive center （LPDAAC） from 1April to 27 May， 2014. The data were preprocessedusing the MRT （MODIS Reprojection Tool）.

2.3 Selection of vegetation indices

A total of eight VIs suitable for monitoring thewheat powdery mildew were selected in this studybased on the previous studies （Table 2）.

To find out the most sensitive features， the Re‐liefF  algorithm was used  due to  its  advantages  of dealing with multi-class classification problem and having no restrictions on the data types. A sample R was  randomly  taken  from  the  training  sample  set each  time，  and  then  k  nearest  neighbor  samples （near Hits） of R were found from the sample set of the same class as R， and k nearest neighbor samples （near Misses） were found from each sample set of different  classes  of  R. The  input  features  were ranked according to the weights from large to small. Then  the  correlation  analysis  was  carried  out  for each feature and the combination with the smallest correlation coefficient was selected as the best com ‐ bination  for  model  construction. The  superiority and efficiency have been illustrated in remote sens‐ing-based  classification  and  object recognition[20-23]. Consequently， it was adopted to perform the feature selection， which gave different weights to the fea‐tures in terms of the correlations between features and various disease samples[24]. Specifically， accord‐ing  to  the  ReliefF  algorithm，  all  the  VI  variables were  sorted  in  descending  order  of  weight，  and eight VIs were selected with the weight of 0.075 as the threshold value. Then， the  correlation  analysisamong the selected features were conducted. Whenthe correlation coefficient （r） of the feature owingthe highest weight that was greater than 0.9， it waseliminated，  and  then  that  of  the  second  highestweight with a high r was eliminated， and so on. Inaddition， there was a close relationship between thedisease  incidence  and  meteorological  factors  suchas  temperature，  precipitation，  humidity，  etc. Thechanges of VIs calculated at different growth stagesalso affected the sensitivity to the disease. Consider‐ing the temporal features and accumulative effect oftemperature， three VIs were finally selected， namelythe SAVI on 26 May 2014 and the SIPI and EVI on17 May 2014.

2.4 Estimation of LST

Two-channel  nonlinear  splitter  algorithm  wasadopted to increase the information amount and re‐duce the influence of errors. The equation is shownas （12）.

where ε and Δε respectively  represents  the  meanand difference values of the two channels ′emissivi- ties depending on the land type and coverage; Ti  and Tj  are the  observed brightness temperatures  of the two channels; bi （i =0， 1， 2， … ， 7） is the simulation dataset of various coefficients which can be derived from  laboratory  data，  atmospheric  parameters  and the atmospheric radiation transmission equation. In order to modify the calculation precision， the coeffi- cient depends on the atmospheric water vapour col- umn[33]. Fig.2 is the LST of 22 May 2014 retrieved from Landsat-8 TIRS image. In general， the LSTs of crop planting areas are lower than other regions. To calculate the land surface emissivity， the weight- ing  method  of vegetation  coverage  was  adopted， which bases on the NDVI and vegetation coverageretrieved from the visible and NIR bands of Land- sat-8 imagery [34].

The  occurrence  and  development  of  wheat powdery mildew is a relatively long process during which the LST has a significant cumulative effect.For example， temperature conditions which are fa-vorable for wheat powdery mildew infection in ear-ly April will aggravate the disease. As a result， thedisease  at  the  end  of May  has  a  high  correlationwith previous LST. The Landsat-8 LST data at fourgrowth stages were selected including the standingstage， jointing  stage，  flowering  stage  and  milkingstage， for the retrieval of LST and their summationSLST is shown in Equation （13）.

where SLST denotes the cumulative effect of multi-temporal  Landsat-8 LST  data; LST  is  calculatedfrom a single Landsat-8 TIRS image; i （1， 2， 3， 4）represents the four stages， and the 20 means the up-per limit of temperature suitable for the incidence ofwheat  powdery  mildew. Similarly，  the  summationof 20 MOD11A1（MLST） can be also obtained ac-cording to the Equation （14）.

2.5 Spatial-temporal fusion of MOD11A1and Landsat-8 LST

Since the four Landsat-8 LST data cannot fullyreflect the variation trend of LST during the wholewheat growth， the daily MOD11A1 was introduced.In order to obtain a data sequence that have enoughspatial and temporal resolutions， the widely appliedspatial  and  temporal  adaptive  reflectance  fusionmodel （STARFM）[35] was used to conduct the spa-tial-temporal  fusion  of  Landsat-8  LST  andMOD11A1. The algorithm ignores the spatial posi-tion registration and atmospheric correction errors，so the pixel values of low spatial resolution （LSR）remote sensing data at the moment t can be calculat-ed using the weighted  sum  of that of high  spatialresolution （HSR） data （Equation （15））.

Ct =∑（ Fti× A t（i））                  （15）

where Fti  is the pixel value of HSR data for the posi‐tioni at the time t; A t（i） shows the weight of coverage area for each pixel; Ct  represents the pixel value of LSR data at the corresponding time.

The  STARFM  algorithm  first  obtains  the MOD11A1 and Landsat-8 LST data and their devia‐tion at the time tk . The deviation is caused by the systematic  errors  and  land  cover  changes. Mean ‐ while， the  Landsat-8 LST  of the time  t0  in  accor‐ dance with MOD11A1 were predicted. The devia‐tion remains constant as time changes was assumed， and the pixel values of Landsat-8 LST at the time t0 are Equation （16）.

HSR （ xi，yi，t0）= LSR （ xi，yi，t0）+

HSR （ xi，yi，tk ）- LSR （ xi，yi，tk ）    （16）

Considering the edge effect of pixels， a cloud- free  pixel  was  selected  in  the  moving  window which was similar to the spectrum of central pixel when the pixel value was calculated. The calculat‐ing  the  central  pixel  value  is  shown  as  Equa‐tion （17）.

HSR （ xw/2，yw/2，t0）=

Wijk×（ LSR （ xi，yi，t0）+

HSR （ xi，yi，tk ）- LSR （ xi，yi，tk ））  （ 17）

where  w  is  the  size  of  the  moving  window;（ xw/2 ， yw/2） represents  the  position  of central  pixel; Wijk  denotes the weight coefficient of a pixel similar to  the  central pixel. The  spectral  distance  weight， temporal  distance  weight，  and  spatial  distance weight of a similar pixel were obtained in the win ‐dow by a normalization method. The three weight coefficients were taken by referring to the study of Gao  et  al. in 2006[35]. Four  Landsat-8 LST （LSTi） and 20 MOD11A1（LSTj） were used for the spatial- temporal  fusion. The  fusion  data  sequences  were summed up as the SMLST in Equation （18）.

SMLST = LSTi + LSTj                   （ 18）

2.6 Construction of monitoring models

Support vector machine （SVM） has the advan‐tages  of  a  simple  structure，  strong  generalizationability，  and high  accuracy， which has been widelyused  in the  classification  of remote  sensing  imag‐es[36]. The  discriminant  function  of  the  model  isshown as Equation （19）.

f （ x）= sgn （x ai yi k （ xi，x）+ b）

where ai  is the Lagrange multiplier; Sv  is the supportvector; xi  and yi  represent two kinds of support vec‐tors; b is the threshold; k（xi ， x） represents a positivedefinite kernel function which satisfies the Mercertheorem.

The  Z-score  method  was  used  to  standardizeVIs and temperature data according to Equation （20）due to their different units. The training and valida‐tion  datasets  were  divided  using  the  ratio  of 7：3.Four  SVM-based models （LST-SVM，  SLST-SVM，MLST-SVM and SMLST-SVM） were trained usingthe optimally selected three VIs. In addition to theVIs， these models were also involved in the Land‐sat-8 LST on 22 May （LST）， four cumulative Land‐sat-8 LST data （SLST）， 20 cumulative MOD11A1（MLST） and  cumulative  spatial-temporal  fusionLST  combing  Landsat-8 LST  and  MOD11A1（SMLST）， respectively. These models will explore abetter  application  of remote  sensing-based LST tothe monitoring of wheat powdery mildew.

x '=（x -μ）ρ                      （20）

where x ' is the standardized data;μ is the mean oforiginal data;ρ is the standard deviation.

3 Results and discussion

3.1 Validation of the monitoring results

The  cross-validation  was  adopted  to  estimatethe  monitoring  accuracies. As  shown  in  Table 3，four indicators， namely the user accuracy （UA）， pro ‐ducer  accuracy （PA），  overall  accuracy （OA） and Kappa coefficient （k） were used to assess the four SVM-based models. It could be seen that the SLST- SVM and SMLST-SVM achieved better classifica‐tion  performance. In  terms  of  OA，  the  SMLST- SVM  model  obtained the best result，  followed by the SLST-SVM and MLST-SVM models， while the LST-SVM model got the lowest value. The OAs of models indicated that the LST had a cumulative ef‐fect on the wheat powdery mildew infection. The k values  of  SMLST-SVM，  MLST-SVM  SLST-SVM and LST-SVM were 0.67， 0.54， 0.59 and 0.38， re‐spectively， which also showed the similar trend withOAs. From the perspective of UA， the ability of thefour  models  to  distinguish  diseased  and  healthywheat  was  strong. The  differentiation  ability  ofSLST-SVM and SMLST-SVM models for mild andsevere  wheat  powdery  mildew  were  significantlyhigher than that  of LST-SVM model. The  accura‐cies of MLST-SVM model were slightly lower thanthose  of  SLST-SVM  and  SMLST-SVM  models，mainly  due  to  the  low  spatial  resolution  ofMOD11A1 data on the city scale. The above resultsshow that the introduction of multi-temporal and cu ‐mulative LST can effectively improve the monitor‐ing and identification of wheat powdery mildew se‐verities.

Huang et al.[37] identified wheat powdery mil‐ dew  of the  study  area using 30 m-resolution  Chi‐nese HJ-1A/1B data to inverse LST and extract four-band reflectance data and build seven VIs. A combi‐ nation  method （GaborSVM） of  SVM  and  Gabor wavelet features were proposed to obtain the OA of 86.7% that was higher than 81.2% of this study， but the OA of SVM-based method was 80% and lower than this study. The primary reason is that the spa‐tial resolution of HJ-1B IRS is 300 m， but it is just 1000 m  for  MOD11A1. The  comparison  studyshows that spatial resolution of optical and thermalinfrared satellite remote sensing images is an impor‐tant factor of affect the accuracy of wheat powderymildew.

3.2 Mapping of wheat powdery mildew

Based on multisource and multitemporal Land‐sat-8， GF-1 and MODIS data， three VIs of SIPI， SA ‐VI and EVI were optimally selected by the ReliefFalgorithm  and  correlation  analysis. The  VIs  andfour  temperature  data  of LST，  SLST，  MLST  andSMLST  were  respectively  used  to  construct  four monitoring  models  through  the  SVM，  namely  the LST-SVM，  SLST-SVM， MLST-SVM  and  SMLST- SVM. For example， the severity distribution on 26 May 2014 in Jinzhou was shown in Fig.3 using the SLST-SVM and SMLST-SVM models. The overall spatial distribution of wheat powdery mildew usingthe two monitoring models were similar. Neverthe‐less，  there  were  also  some  obvious  differences  asshown in the red boxes. It was more serious in theeastern part than in the western part of the study ar‐ea. It was also obvious that the wheat powdery mil‐dew mainly occurred in the areas where wheat waswidely planted.

In the  central regions where the  ground-truth points  were  located，  the  monitoring  results  of the two models were also similar by visual observation. In the two figures， the distribution trends of wheat powdery mildew were both relatively concentrated. In  comparison  with  the 32% of healthy  samples， 55% of mild  samples  and 13% of severe  samples for the  in-situ  investigation， they were 37%， 49% and 14% for the SLST-SVM model and 31%， 55% and 14% for  SMLST-SVM model， respectively. It can be seen that the SMLST-SVM model had a bet‐ter result than SLST-SVM model.

3.3 Analysis of influence factors

Temperature is one of the key factors to affectthe incidence of wheat powdery mildew， however， itis difficult and inaccurate to identify the disease insmall and medium-sized regions， due to constraintof low  spatial resolution of regular meteorologicaldata. Ma et al.[38] combined meteorological and re‐mote  sensing data to monitor wheat powdery mil‐dew， because the distribution of meteorological sta‐tions was too sparse. The primary objective of thisstudy is to  compare the relationship between LST and  the  disease  severities  of wheat  powdery  mil‐ dew. It can be found that different LST data have significant impact on the accuracies of the disease diagnosis. For example， in the retrieval of LST from Landsat-8 data，  the  two-channel  nonlinear  splitter used in this paper was more reliable than the single- channel method[39]. The  spatial resolution of Land‐ sat-8 TIRS  data  and  the  temporal  resolution  of MOD11A1 data  can  both  meet  the  requirements. Therefore， the combination and fusion of both the images were performed to acquire a better perfor‐mance. The accumulated temperature and effective accumulated  temperature  are  key  factors  to  affect wheat  powdery  mildew[40]. Multitemporal  LST  is better to show the temperature influence on the dis‐ ease monitoring than a single one.

4 Conclusion

It is a progressive process for the occurrence of wheat powdery mildew， the cumulative effect must be considered with the disease development. To ac‐curately identify the disease， multisource and multi‐ temporal  GF-1，  Landsat-8 and  MOD11A1 were jointly utilized at a city  scale. The vegetation fea‐tures and LSTs are simultaneously adopted to mon‐itor the disease for improving the classification ac‐ curacy.

To find out the most sensitive vegetation fea‐tures to wheat powdery mildew， four most impor‐ tant growth stages were selected including the stand‐ing stage， jointing stage， flowering stage and milk‐ing  stage， which  are  also the  criteria  for  selecting satellite  images. Additionally，  it  has  a  progressive process for the disease incidence with the growth of wheat  and  changes  of meteorological  factors. As one of the most essential influencing factors， the ac‐ cumulative  effects  in  LSTs  must  be  considered when  identifying  the  disease. Landsat-8 TIRS hashigh spatial resolution but low temporal resolution，however， it is quite contrary to MOD11A1. As a re‐sult， the STARFM was selected to perform the spa‐tial-temporal fusion of both data. Four SVM modelswere  respectively  constructed  through  a  singleLandsat-8 LST （LST），  multitemporal  Landsat-8LSTs （SLST），  cumulative  MODIS  LST （MLST）and  the  combination  of cumulative  Landsat-8 andMODIS LST （SMLST）.

The OA of LST-SVM model was improved by13% using the four temporal LSTs. In addition， theKappa coefficient also increased from 0.38 to 0.59，indicating that LST is a key habitat factor for the oc‐currence  and  development  of wheat powdery mil‐dew， due to a cumulative effect. On the other hand，the accuracies of MLST-SVM model were smallerthan that of SLST-SVM and SMLST-SVM models，indicating that it is inappropriate to apply MOD11A1data  directly  to  the  monitoring  of wheat powderymildew  in  a relatively  small  area. Conversely， theperformance of SMLST-SVM is slightly better thanthat  of SLST-SVM， indicating that the monitoringperformance can be enhanced by using a combina‐tion  and  fusion  of high-resolution  spatial-temporalremote sensing data.

References：

[1] ZHANG J， WANG N， YUAN L， et al. Discriminationof winter wheat disease and insect stresses using con‐tinuous wavelet features extracted from foliar spectralmeasurements[J]. Biosystems Engineering， 2017， 162：20-29.

[2] FENG W， SHEN W Y， HE L， et al. Improved remotesensing detection of wheat powdery mildew using dual-green  vegetation  indices[J]. Precision  Agriculture，2016， 17（5）：608-627.

[3] GALLEGO F J， KUSSUL N， SKAKUN S， et al. Effi‐ciency assessment of using satellite data for crop areaestimation in Ukraine[J]. International Journal of Ap‐plied Earth Observation and Geoinformation， 2014， 29：22-30.

[4] SETHY P K， BARPANDA N K， RATH A K， et al. Im‐age processing techniques for diagnosing rice plant dis‐ease： A survey[J]. Procedia  Computer  Science， 2020， 167：516-530.

[5] YANG  C. Remote  sensing  and  precision  agriculturetechnologies  for  crop  disease  detection  and  manage‐ment with a practical application example[J]. Engineer‐ing， 2020， 6（5）：528-532.

[6] ZHENG Q， YE H， HUANG W， et al. Integrating spec‐tral  information  and  meteorological  data  to  monitor wheat yellow rust at a regional scale： A case study[J]. Remote Sensing， 2021， 13（2）： ID 278.

[7] HUANG W J， LAMB D W， NIU Z， et al. Identificationof yellow  rust  in  wheat using  in-situ  spectral  reflec‐tance measurements and airborne hyperspectral imag‐ing[J]. Precision Agriculture， 2007， 8（4-5）：187-197.

[8] ZHANG  J  C， PU  R L， WANG  J H，  et  al. Detectingpowdery mildew of winter wheat using leaf level hy‐perspectral measurements[J]. Computers and Electron‐ics in Agriculture， 2012， 85：13-23.

[9] LUO J H， ZHAO C J， HUANG W J， et al. Discriminat‐ing  wheat  aphid  damage  degree  using 2-dimensional feature  space  derived  from  Landsat 5 TM [J]. Sensor Letters， 2012， 10（1-2）：608-614.

[10] ZHENG Q， HUANG W， CUI X， et al. New spectral in‐dex  for  detecting  wheat  yellow  rust  using  sentinel-2 multispectral imagery[J]. Sensors， 2018， 18（3）： ID 868.

[11] HE L， QI S， DUAN J， et al. Monitoring of Wheat pow ‐dery mildew disease  severity using multiangle hyper‐ spectral remote sensing[J]. IEEE Transactions on Geo‐ science and Remote Sensing， 2020， 59（2）：979-990.

[12] ZHAO J， FANG Y， CHU G， et al. Identification of leaf-scale wheat powdery mildew （Blumeriagraminis f. sp. Tritici） combining hyperspectral imaging and an SVM classifier[J]. Plants， 2020， 9（8）： ID 936.

[13] KHAN I H， LIU H， LI W， et al. Early detection of pow ‐dery mildew disease and accurate quantification of its severity  using  hyperspectral  images  in  wheat[J]. Re‐ mote Sensing， 2021， 13（18）： ID 3612.

[14] ZHAO  J， XU  C， XU  J，  et  al. Forecasting  the wheatpowdery mildew （Blumeriagraminis f. Sp. tritici） us‐ing a remote sensing-based decision-tree classification at  a provincial  scale[J]. Australasian Plant Pathology， 2018， 47（1）：53-61.

[15] LIANG W， CARBERRY P， WANG G， et al. Quantify‐ing the yield gap in wheat-maize cropping systems of the Hebei Plain， China[J]. Field Crops Research， 2011， 124（2）：180-185.

[16] CHEN H， ZHANG H， LI Y. Review on research of me‐teorological conditions and prediction methods of crop disease and insect pest[J]. Chinese Journal of Agrome‐teorology， 2007， 28（2）：212-216.

[17] YUAN L， PU R L， ZHANG J C， et al. Using high spa‐tial resolution  satellite imagery  for mapping powderymildew  at  a  regional  scale[J]. Precision  Agriculture，2016， 17（3）：332-348.

[18] LOVELAND T R， IRONS J R. Landsat 8： The plans，the reality， and the legacy[J]. Remote Sensing of Envi‐ronment， 2016， 185：1-6.

[19] WANG W， LIANG S， MEYERS T. Validating MODISland  surface  temperature  products  using  long-termnighttime ground measurements[J]. Remote Sensing ofEnvironment， 2008， 112（3）：623-635.

[20] HUANG L， JIANG J， HUANG W， et al. Wheat yellowrust monitoring based on Sentinel-2 Image and BPNNmodel[J]. Transactions of the Chinese Society of Agri‐cultural Engineering， 2019， 35（17）：178-185.

[21] LIU R， ZAHNG S， JIA R. Application of feature selec‐tion  method  in  building  information  extracting  fromhigh  resolution  remote  sensing  image[J]. Bulletin  ofSurveying and Mapping， 2018， （2）：126-130.

[22] BAO W， ZHAO J， HU G， et al. Identification of wheatleaf diseases and their severity based on elliptical-max‐imum margin criterion metric learning[J]. SustainableComputing： Informatics  and  Systems， 2021， 30： ID100526.

[23] QIN F， LIU D， SUN B， et al. Identification of alfalfaleaf  diseases  using  image  recognition  technology[J].PLoS One， 2016， 11（12）： ID e0168274.

[24] ROBNIK- ?IKONJA M， KONONENKO I. Theoreticaland empirical analysis of ReliefF and RReliefF[J]. Ma‐chine Learning， 2003， 53（1）：23-69.

[25] JORDAN C F. Derivation of leaf‐ area index from qual ‐ity of light on the forest floor[J]. Ecology， 1969， 50（4）：663-666.

[26] ROUSE J W， HAAS R H， SCHELL J A， et al. Moni‐toring  vegetation  systems  in  the  great  plains  withERTS[C]// In Third ERTS Symposium Volume 1： Tech‐nical  Presentations. Washington，  DC，  USA： NASA，1973：309-317.

[27] GAMON J A， PENUELAS J， FIELD C B. A narrow-waveband spectral index that tracks diurnal changes inphotosynthetic efficiency[J]. Remote Sensing of Envi‐ronment， 1992， 41：35-44.

[28] HUETE A  R. A  soil-adjusted  vegetation  index （SA ‐VI）[J]. Remote Sensing of Environment， 1988， 25（3）：295-309.

[29] LIU H， HUETE A. A feedback based modification ofthe NDVI to minimize canopy background and atmo‐spheric noise[J]. IEEE Transactions on Geoscience andRemote Sensing， 1995， 33（2）：457-465.

[30] BROGE  N  H，  LEBLANC  E. Comparing  predictionpower  and  stability  of broadband  and  hyperspectralvegetation indices for estimation of green leaf area in‐dex and canopy chlorophyll density[J]. Remote  Sens‐ing of Environment， 2000， 76（2）：156-172.

[31] RICHARDSON A J， WIEGAND C L. Distinguishingvegetation from soil background information[J]. Photo‐grammetric Engineering and Remote Sensing， 1977， 43（12）：1541-1552.

[32] PENUELAS J， BARET F， FILELLA I. Semi-empiricalindices to  assess  carotenoids/chlorophyll  a ratio  from leaf spectral reflectance[J]. Photosynthetica， 1995， 31（2）：221-230.

[33] DU C， REN H， QIN Q， et al. A practical split-windowalgorithm for estimating land surface temperature from Landsat 8  data[J]. Remote  Sensing， 2015， 7（1）：647-665.

[34] REN H， DU C， LIU R， et al. Atmospheric water vaporretrieval  from  Landsat 8 thermal  infrared  images[J]. Journal of Geophysical Research： Atmospheres， 2015， 120（5）：1723-1738.

[35] GAO  F，  MASEK  J，  SCHWALLER  M，  et  al. On  theblending  of the  Landsat  and  MODIS  surface  reflec‐tance： Predicting  daily Landsat  surface reflectance[J]. IEEE  Transactions  on  Geoscience  and  Remote  Sens‐ing， 2006， 44（8）：2207-2218.

[36] OLATOMIWA L， MEKHILEF  S，  SHAMSHIRBANDS，  et  al. A  support  vector  machine-firefly  algorithm-based  model  for  global  solar  radiation  prediction[J].Solar Energy， 2015， 115：632-644.

[37] HUANG L， LIU W， HUANG W， et al. Remote sensingmonitoring of winter wheat powdery mildew based onwavelet analysis and support vector machine[J]. Trans‐actions  of the  Chinese  Society  of Agricultural  Engi‐neering， 2017， 33（14）：188-195.

[38] MA H， HUANG W， JING Y. Wheat powdery mildewforecasting  in  filling  stage  based  on  remote  sensingand  meteorological  data[J]. Transactions  of the  Chi‐nese Society of Agricultural Engineering， 2016， 32（9）：165-172.

[39] XU  H. Retrieval  of the  reflectance  and  land  surfacetemperature  of  the  newly-launched  Landsat 8 satel‐lite[J]. Chinese  Journal  of  Geophysics-Chinese  Edi‐tion， 2015， 58（3）：741-747.

[40] SHARMA A K， SHARMA R K， BABU K S， et al. Ef‐fect of planting options and irrigation schedules on de‐velopment of powdery mildew  and yield of wheat inthe North Western plains of India[J]. Crop Protection，2004， 23（3）：249-253.