Classification Research On Moths Based On Digital Images

The digital classification and identification of thirty nine moths (five superfamilies,eight familys) were studied by using the methods of mathematical morphology andGeometric morphometry, taking the moths which are harmful to agriculture and forestryas material. Each method was evaluated on the feasibility and reliability, and theclassificatory importance of each morphological character was analyzed in classificationof the moths at superfamily level, family level and species level.At first, digital technology was used to get images of wing and vein of moths andpretreatments such as threshold segmentation and smooth denoising were applied toimages of moth wing. Then twenty six math-morphological characters(MMCs) such aseccentricity, sphericity, lobation, roundness, rectangularity, elongation and movementinvariants including Hu1and Hu2were selected for being invariant to the image size anddirection by software BugShape1.0. These MMCs which were used for classification anddiscrimination of the moths at each level were extracted from the images of rightforewing and right hindwing of thirty nine moths. For the species which had relativelylower classification accuracy resulting from the method, Geometric morphometry wasused to make further analysis on their images.The analytic results showed that by using MMCs of moth wings, classification anddiscrimination of the moths at each level were realized successfully. High classificationaccuracies were got for all moths at superfamily level, family level and species levelexcept for Sphingidae insects with accuracies of regression and intersecting discriminantanalysis87.7%and84.0%respectivelyThe accuracies of regression and intersecting discriminant analysis were100.0%and97.4%respectively, and six MMCs were selected as the classification variables atsuperfamily level. The contribution of these variables were ranked as follows:(FW-rectangularity, FW-Hu5, HW-eccentricity)＞HW-roundness＞(HW-rectangularity,HW-Hu5). Whereas, at family level, the accuracies of regression and intersectingdiscriminant analysis were100.0%and92.1%respectively, and the contribution of sixselected variables were ranked as follows:(FW-Hu6, FW-rectangularity, FW-Hu5)＞HW-Hu2＞(HW-roundness, FW-roundness).The accuracies of regression and intersecting discriminant analysis were both higherthan93.3%at species level. The contribution of eleven selected variables used toNoctuidae insects were ranked as follows:(FW-eccentricity, FWHu5, FWHu7）＞FWHu2＞FW-roundness＞FW-sphericity＞FWHu3＞(FW-lobation, FWHu1, FWHu6) ＞FWHu4; for Eucleidae insects:(HW-lobation, HWHu7)＞(HWHu6, FWHu5,FW-lobation)＞(FW-rectangularity, HW-elongation); for Arctiidae insects:(FWHu3,FWHu7, HWHu7, HW-eccentricity)＞(HWHu4, HWHu3, FW-roundness, FW-sphericity,FWHu4)＞(FW-elongation, FWHu5, FW-lobation, FWHu1); for Lasiocampidae insects:(HWHu3, FW-eccentricity, FWHu6, FW-sphericity)＞(FW-roundness,HW-rectangularity, FWHu4, FWHu3, FWHu1)＞(FWHu2, HW-elongation,FW-rectangularity, HWHu1); for Notodontidae insects:(FW-eccentricity, HWHu1,FWHu5, FW-rectangularity)＞(FW-sphericity, FW-roundness, FW-lobation,HW-rectangularity)＞(HWHu7, FWHu1, FWHu2); for Lymantriidae insects:(FW-roundness, FWHu4, HW-rectangularity)＞(FW-elongation, FWHu1, HWHu6)＞(HW-lobation, FW-rectangularity); and for Geometridae insects:(FWHu2, FW-lobation,HW-roundness)＞(HW-elongation, HWHu6, FWHu1)＞(FWHu3, FW-eccentricity).Differences were existed for the effectiveness and importance of MMCs of mothwings used to different taxonomic category or different species. However, at each level,the selected MMCs could be used for classification and discrimination well, which meantthat each MMC played varied role at different level.The method of geometric morphometry was used to analyze the vein images ofSphingidae insects which had relatively lower classification accuracy. Seventeenintersection points of the veins were selected as landmarks by using TpsDig software insequence of the intersection points. Procrustes superimposition and relative warpanalyses were done by applying TpsSuper and TpsRelw softwares. Finally, the accuraciesof regression and intersecting discriminant analysis reached to100.0%and99.7%respectively. The research indicated that both the MMCs of the moth wings and theGeometric morphological characters of the veins had good function for classification anddiscrimination and therefore, laid foundation for gradually realizing the automaticrecognition of moths in the future.Finally, artificial neural network method was applied to do further study. The resultsindicated that BP neural network was able for classification and identification ofNoctuidae insects and identification accuracy could be enhanced by combining withmathematical method such as principal component analysis which was used for datapreprocessing of the mathematical morphological characteristics in this study. The studyplayed an elementary and exploratory role in forming automatic identification system.