基于Fisher判别分析与随机森林的马尾松毛虫害检测
|
摘要
虫害检测算法的构建是耦合"地—天"特征的过程,是实现其遥感监测的重要保障。以福建省三明市、将乐县、沙县、南平市延平区等4个县(区、市)为试验区,收集182组马尾松毛虫害样本数据,随机划分为训练集与验证集,设置5次重复试验及1次指标筛除试验。结合马尾松毛虫危害下的寄主表征,获取松林叶面积指数LAI、叶面积指数标准误SEL、归一化差值植被指数NDVI、缨帽变换湿度轴WET及影像绿光波段B2、红光波段B3、近红外波段B4等7个地面与遥感特征指标,建立其危害等级的Fisher判别分析与随机森林模型,从检测精度、Kappa系数、ROC曲线等角度综合比较两种算法的检测效果,并给予配对t检验。结果表明:7个指标均具备虫害响应能力,SEL和NDVI相对较弱;Fisher判别分析6次试验的虫害平均检测精度为73.26%,Kappa系数为0.631 9,而RF法则分别为79.30%,0.715 1,显著优于前者(p<0.05);RF法对无危害、轻度危害、中度危害3个虫害等级的检测精度、Kappa系数、AUC均显著高于Fisher判别分析(p<0.05),对于重度危害等级,Fisher判别分析则占优。总体而言,RF法对马尾松毛虫害的检测效果优于Fisher判别分析,但Fisher判别分析对重度危害等级有更高准确性且模型明确、易于推广,可综合应用两种算法开展虫害监测工作。该成果为马尾松毛虫害及其他森林病虫害的有效检测提供技术参考,奠定其遥感监测的基础。
The construction of the pest detection algorithm is a process of coupling the"ground-space"features,which is an important guarantee to realize its remote sensing monitoring.Taking Sanming City,Jiangle County,Sha County and Yanping District in Nanping City in Fujian Province as the experimental areas,it gathered 182samples of Dendrolimus punctatus Walker damage and randomly divided them into training set and validation set,and 5repeated tests and 1test of index screening were performed.According the host representations damaged by Dendrolimus punctatus Walker,7ground and remote sensing characteristic indices including pine forest leaf area index(LAI),standard deviation of LAI(SEL),normalized difference vegetation index(NDVI),wetness from tasseled cap transformation(WET),green band(B_2),red band(B_3),near infrared band(B_4)were obtained,then the models of Fisher discriminant analysis and random forest for pest levels were constructed.The detection precision,Kappa coefficient and ROC curve were used to comprehensively compare the detection effects of these two algorithms,as well as the paired t-test.The results showed that all the 7indices have the pest responsiveness,while SEL and NDVI are relatively weak;the average detection precision of Fisher discriminant analysis in 6tests was 73.26%,Kappa coefficient was 0.6319,and 79.30%,0.715 1of RF respectively,indicating RF is significantly better than the former one(p<0.05);for the 3pest levels of non-damage,mild damage and moderate damage,the detection precision,Kappa coefficient and AUC of RF were all significantly higher than Fisher discriminant analysis(p<0.05),while for the severe damage,Fisher was better.On the whole,the Dendrolimus punctatus Walker damage detection effect of RF is better than Fisher discriminant analysis,but Fisher has more accurate for the severe damage and the mode is clear,easy to by promoted,so these two algorithms could be comprehensively utilized to put forward the pest monitoring work.The results can provide a technical reference for the effective detection of Dendrolimus punctatus Walker damage as well as other forest pests and diseases,and lay a foundation of the remote sensing monitoring.
The construction of the pest detection algorithm is a process of coupling the"ground-space"features,which is an important guarantee to realize its remote sensing monitoring.Taking Sanming City,Jiangle County,Sha County and Yanping District in Nanping City in Fujian Province as the experimental areas,it gathered 182samples of Dendrolimus punctatus Walker damage and randomly divided them into training set and validation set,and 5repeated tests and 1test of index screening were performed.According the host representations damaged by Dendrolimus punctatus Walker,7ground and remote sensing characteristic indices including pine forest leaf area index(LAI),standard deviation of LAI(SEL),normalized difference vegetation index(NDVI),wetness from tasseled cap transformation(WET),green band(B_2),red band(B_3),near infrared band(B_4)were obtained,then the models of Fisher discriminant analysis and random forest for pest levels were constructed.The detection precision,Kappa coefficient and ROC curve were used to comprehensively compare the detection effects of these two algorithms,as well as the paired t-test.The results showed that all the 7indices have the pest responsiveness,while SEL and NDVI are relatively weak;the average detection precision of Fisher discriminant analysis in 6tests was 73.26%,Kappa coefficient was 0.6319,and 79.30%,0.715 1of RF respectively,indicating RF is significantly better than the former one(p<0.05);for the 3pest levels of non-damage,mild damage and moderate damage,the detection precision,Kappa coefficient and AUC of RF were all significantly higher than Fisher discriminant analysis(p<0.05),while for the severe damage,Fisher was better.On the whole,the Dendrolimus punctatus Walker damage detection effect of RF is better than Fisher discriminant analysis,but Fisher has more accurate for the severe damage and the mode is clear,easy to by promoted,so these two algorithms could be comprehensively utilized to put forward the pest monitoring work.The results can provide a technical reference for the effective detection of Dendrolimus punctatus Walker damage as well as other forest pests and diseases,and lay a foundation of the remote sensing monitoring.
引文
[1]CUI Heng-jian,WU Hong-gan,QIAO Yan-you,et al(崔恒建,武红敢,乔彦友,等).Journal of Biomathematics(生物数学学报),1997,12(S1):611.
[2]XU Zhang-hua,YU Kun-yong,LIU Jian,et al(许章华,余坤勇,刘健,等).Acta Agriculturae Universitatis Jiangxiensis·Natural Sciences Edition(江西农业大学学报),2012,34(5):933.
[3]XIANG Chang-sheng,ZHOU Zi-ying(向昌盛,周子英).Acta Entomologica Sinica(昆虫学报),2010,53(9):1055.
[4]Peixoto M D S,Barros L C D,Bassanezi R C.Applied Mathematics,2014,5(5):1133.
[5]Bisht K S,Chaudhary S.Signal Image&Video Processing,2016,10(3):551.
[6]BAI Ji-heng,XU Gang,ZHOU Ting-ting,et al(白基恒,徐刚,周婷婷,等).Journal of Natural Disasters(自然灾害学报),2016,25(2):26.
[7]Park Y S,Chung Y J.Forest Ecology&Management,2006,222(1-3):222.
[8]XU Zhang-hua,LI Cong-hui,LIU Jian,et al(许章华,李聪慧,刘健,等).Transactions of the Chinese Society for Agricultural Machinery(农业机械学报),2014,45(6):275.
[9]Breiman L.Machine Learning,2001,45:5.
[10]LI Ting,TIAN Yuan,WU Lun,et al(李亭,田原,邬伦,等).Geography and Geo-Information Science(地理与地理信息科学),2014,30(6):25.
[11]PENG Ling,XU Su-ning,PENG Jun-huan(彭令,徐素宁,彭军还).Journal of Jilin University·Earth Science Edition(吉林大学学报·地球科学版),2016,46(1):175.
[12]XU Guang-yu,XU Wen,FANG Si-ai,et al(徐光余,徐文,方思爱,等).Journal of Hebei Agricultural Sciences(河北农业科学),2008,12(10):31.
[13]Martinez B,Camacho F,Verger A,et al.International Journal of Applied Earth Observation and Geoinformation,2013,(21):463.
[14]Wong M S,Sarker M L R,Nichol J,et al.International Journal of Applied Earth Observation and Geoinformation,2013,(21):66.
[15]Pu R L,Gong P,Biging G S,et al.IEEE Transactions on Geoscience and Remote Sensing,2003,41(4):916.
[16]Cho M A,Debba P,Mutanga O,et al.International Journal of Applied Earth Observation and Geoinformation,2012,(16):85.
[17]Prasad A M,Iverson L R,Liaw A.Ecosystems,2006,9(2):181.
[2]XU Zhang-hua,YU Kun-yong,LIU Jian,et al(许章华,余坤勇,刘健,等).Acta Agriculturae Universitatis Jiangxiensis·Natural Sciences Edition(江西农业大学学报),2012,34(5):933.
[3]XIANG Chang-sheng,ZHOU Zi-ying(向昌盛,周子英).Acta Entomologica Sinica(昆虫学报),2010,53(9):1055.
[4]Peixoto M D S,Barros L C D,Bassanezi R C.Applied Mathematics,2014,5(5):1133.
[5]Bisht K S,Chaudhary S.Signal Image&Video Processing,2016,10(3):551.
[6]BAI Ji-heng,XU Gang,ZHOU Ting-ting,et al(白基恒,徐刚,周婷婷,等).Journal of Natural Disasters(自然灾害学报),2016,25(2):26.
[7]Park Y S,Chung Y J.Forest Ecology&Management,2006,222(1-3):222.
[8]XU Zhang-hua,LI Cong-hui,LIU Jian,et al(许章华,李聪慧,刘健,等).Transactions of the Chinese Society for Agricultural Machinery(农业机械学报),2014,45(6):275.
[9]Breiman L.Machine Learning,2001,45:5.
[10]LI Ting,TIAN Yuan,WU Lun,et al(李亭,田原,邬伦,等).Geography and Geo-Information Science(地理与地理信息科学),2014,30(6):25.
[11]PENG Ling,XU Su-ning,PENG Jun-huan(彭令,徐素宁,彭军还).Journal of Jilin University·Earth Science Edition(吉林大学学报·地球科学版),2016,46(1):175.
[12]XU Guang-yu,XU Wen,FANG Si-ai,et al(徐光余,徐文,方思爱,等).Journal of Hebei Agricultural Sciences(河北农业科学),2008,12(10):31.
[13]Martinez B,Camacho F,Verger A,et al.International Journal of Applied Earth Observation and Geoinformation,2013,(21):463.
[14]Wong M S,Sarker M L R,Nichol J,et al.International Journal of Applied Earth Observation and Geoinformation,2013,(21):66.
[15]Pu R L,Gong P,Biging G S,et al.IEEE Transactions on Geoscience and Remote Sensing,2003,41(4):916.
[16]Cho M A,Debba P,Mutanga O,et al.International Journal of Applied Earth Observation and Geoinformation,2012,(16):85.
[17]Prasad A M,Iverson L R,Liaw A.Ecosystems,2006,9(2):181.