Remote Sensing Identification And Monitoring Of Larch Needle Pests Based On Ground Hyperspectral Data

Conifer pests threaten the health of Larch and damage the forest ecosystem by changing the number of needles,canopy color and internal biochemical components of Larch,and the harmfulness of different pests to trees is different.Therefore,it is very important to distinguish different types of pests and monitor its severity.In recent years,the research on remote sensing monitoring of forest pests has made great progress,but there are few studies on remote sensing discrimination of different pest types.However,the remote sensing monitoring of pest severity mainly focuses on the use of a single pest indicator,while the multi indicator monitoring is rarely reported.At present,the accuracy of hyperspectral remote sensing monitoring is high,but its process is complex,calculation is large and universality is poor.By contrast,the process of multispectral remote sensing monitoring is simple and the calculation is small,but its accuracy needs to be improved.In this study,two typical conifer pests in Mongolian Plateau,Erannis Jacobsoni Djak.(EJD)and Pendrolimus Sibiricus Tschtv.(PST),are selected as examples,taking Binder and Barenburen of Mongolia as the experimental areas.Based on ground non-imaging hyperspectral data,UAV RGB image data,multispectral image data of Sentinel-2A satellite and measured data of pest indicators on the ground,choosing forest canopy color(FCC),leaf loss rate(LLR),chlorophyll content(CHLC),leaf weight content dry(LWCD),leaf weight content fresh(LWCF)and population density(POPD)as indicators of pest damage.On the canopy scale,the sensitive hyperspectral features of smooth spectral reflectance(SSR),differential spectral reflectance(DSR),ground spectral index(GSI)and spectral continuous wavelet coefficient(CWC)are selected by Findpeaks-SPA pattern.Based on the partial least square regression(PLSR),support vector machine regression(SVMR),stepwise multiple linear regression(SMLR),random forest classification(FR),support vector classification(SVC)and Fisher discrimination,established the estimation(discrimination)models of six monitoring indicators and the discrimination models of different pest types.On the regional scale,based on the star ground combination model,the spectral reflectance are simulated from Sentinel-2A image,and the spectral index(SI)and spectral derivative feature(SDF)are calculated,the sensitive spectral features of SI and SDF are extracted by combining threshold method and SPA algorithm,and then establish the discrimination model of different pest types based on RF.Finally,the EJD and PST were distinguished.At the same time,the monitoring models of pest indicators were constructed,and the severity of pests were identified by FCM fuzzy clustering.The following conclusions are obtained in this study:(1)Hyperspectral features were sensitive to the six pest indicators,among which continuous wavelet coefficient was the most sensitive,followed by differential spectral reflectance,ground spectral index,and smooth spectral reflectance.FP-SPA pattern can effectively extract the spectral features that are sensitive to pests.Compared the two pests,the sensitivity of the sensitive spectral features to the PST was more obvious.(2)We can distinguish different types of pests and get better results by using hyperspectral and multispectral features.On the canopy scale,the discrimination model accuracy of RF and SVC are higher than that of Fisher.The feature with best discrimination ability is continuous wavelet coefficient,followed by ground spectral index,differential spectral reflectance and the smooth spectral reflectance.On the regional scale,the accuracy of RF model based on Sentinel-2A remote sensing simulation data is significantly improved.PSRI,PSSR and IRECI have better discrimination ability.(3)On the canopy scale,the continuous wavelet coefficient hyperspectral feature has the best ability to identify conifer pests.The accuracy of leaf loss rate,chlorophyll content,leaf weight content dry,leaf weight content fresh,population density estimation models based on PLSR and SVMR,and the forest canopy color discrimination models based on RF and SVC were all improved.PLSR and SVMR models can effectively identify the degree of forest damage and pest occurrence.RF and SVC can accurately distinguish the color change of forest canopy caused by conifer pests,and can describe the degree of forest damage.(4)On the regional scale,the spectral index and derivative spectral features of Sentinel-2A remote sensing simulation data have significant sensitivity to the two pest indicators.Using the spectral features of remote sensing simulation data,the indicators of conifer pest can be identified by RF and PLSR algorithm.In the identification of conifer pests based on the non-simulated Sentinel-2A remote sensing data,the estimation accuracy of the two pests' leaf loss rate is the highest.(5)On the regional scale,the FCM fuzzy clustering method is used to automatically classify the severity of pests(light,medium and heavy)by using several indicators such as forest canopy color,leaf loss rate,chlorophyll content,leaf weight content dry leaf weight content fresh and population density,and good results have been achieved.Compared with the results of pest severity identification based on single indicator,the results of using population density and leaf loss rate were similar to those of multiple indicators.It can be seen that population density or leaf loss rate can reliably monitor the severity of conifer pests.In a word,this study proposes a set of methods from the extraction of sensitive spectral features to the construction of pest monitoring model,which not only can be realized by hyperspectral features on the canopy scale,but also can be effectively used on the regional scale.The accuracy of multispectral features was lower than that of hyperspectral features,while the multispectral features based on Sentinel-2A remote sensing simulation data can significantly improve the accuracy of conifer pest monitoring,it provides a way to improve the accuracy of forest pest monitoring by multispectral remote sensing.In addition,the research results showed that the leaf loss rate estimation accuracy based on Sentinel-2A remote sensing data can meet the general needs of conifer pest severity monitoring,which has important reference value for forest pest control and decision-making.