Accelerating the discovery of biodiversity by detecting “new” species based on machine learning method

Accelerating the discovery of biodiversity by detecting “new” species based on machine learning method | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Accelerating the discovery of biodiversity by detecting “new” species based on machine learning method Yuanyuan Lu, Jing Li, Zhengyu Zhao, Yongchao Zhang, Yijie Tong, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3832815/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Recently, machine learning (ML) has been widely used in species auto-identification systems for multi-scene applications in biodiversity, while most of the existing ML systems relying on images are limited to identifying the species on which they are trained, and unknown species out of the system are normally incorrectly identified. Results Here, we propose a new workflow system based on the ML system and PERMANOVA analysis, named Taichi for the detection and confirmation of unknown species status, stimulated by the traditional biodiversity discovery process. First, we developed a series of high-throughput photography devices that could efficiently obtain aligned multiangle images for ML system training. Then the new analysis workflow was integrated in Python codes based on the convolutional neural networks: MobileNetV2 (selected from four available networks) and further PERMANOVA analysis of euclidean distances to detect 'new' species. Two newly established beetle datasets: Melanopopillia (4 species, 55 specimens, 42 450 images) and Hong Kong beetles (21 species, 206 specimens, 35 450 images) were constructed in this work to demonstrate the Taichi system. The diagnostic information for species is generated from the output results of the ML system (top-1 accuracy rate reached 97.76% and 96.22% respectively) naming as the artificial intelligence (AI) barcode, and extra analysis of AI barcodes could help visualization the diagnostic characters from multiangle images. Conclusions By comparing AI barcodes of different specimens outside the system, it is proved that the Taichi workflow can rapidly detect whether the input sample is a known species or possible 'new' species. This workflow provides a newly possible self-extensibility solution for the acceleration of biodiversity discovery. Additionally, it can also drive the development of data mining from the digitization of natural history collections around the world and has great potential to advance the field of biodiversity science. Unknown species Convolution neural networks Identification PERMANOVA Figures Figure 1 Figure 2 Figure 3 Figure 4 INTRODUCTION Discovering, recording, and monitoring species are regarded the key steps in exploring Earth’s biodiversity [ 1 – 3 ]. While the large number of unknown species within arthropods hinders the investigation of species’ biodiversity, especially faunal “new” species: new record species and new species [ 4 , 5 ]. At the same time, the reduced financial support for taxonomic research and the ever-growing pool of described species exacerbates the situation [ 6 – 8 ]. Therefore, there is an urgent need for technological innovation, such as machine learning (ML), to find, record, and monitor species efficiently and at high throughput [ 9 – 13 ]. ML has been widely used in species autoidentification systems with increased accuracy (top-5 accuracy greater than 99%) [ 14 – 19 ]. However, most of the autoidentification systems provide only the species names within the system [ 20 – 25 ], when new regional records or unknown species (out-of-distribution data) are put into, the system usually incorrectly identified [ 25 , 26 ]. This shortcoming is due to the limited interpretability and self-iteration of the identification system [ 27 ]. Some complex image-based ML systems have been developed to explore unknown group identification methods [ 28 ]. Probability values have recently been considered as a factor to evaluate out-of-distribution species, while single probability values have revealed less than ideal outcomes in identification systems [ 19 , 29 , 30 ]. And in some cases, DNA barcodes were combined with the image feature vectors for explore the out-of-distribution species [ 31 ]. Here, we propose a new workflow system derived from the procedure by which taxonomists discover new species [ 32 , 33 ], named Taichi , enabled by ML and new data collection devices and analysis methods (Fig. 1 ): 1) establishing a database of known species by an ML system based on image stacks from the newly designed high-throughput photography devices; 2) generating diagnostic information between species using the results generated by the ML system, which named the AI barcode; and 3) identifying unknown species as facilitated by the AI barcode. To demonstrate and test the effectiveness of the Taichi workflow, two real high-quality datasets were used to detect unknown species and confirm their status (new record or new species). This method provides a new perspective from the taxonomic view to detect the “new” species, and also visualizes the identification process of deep learning systems in some extant. MATERIALS AND METHODS 2.1 | Specimens and datasets Our study uses two datasets (Fig. 2 a, Table S1). The Melanopopillia dataset (image size: 224 x 224 pixels) consists of 4 species and 55 specimens, including all three known species in this genus and an undescribed new species (Fig. 2 a). The dataset Hong Kong beetles (image size: 224 x 224 pixels) includes 21 species and 206 specimens (Figs. 2 a, 2 b). Here, we assume that each species is represented by four or five specimens as a group to define the species within a certain range of variation. Dataset Melanopopillia (Fig. 2 a, Table S1): The genus Melanopopillia belongs to the subfamily Rutelinae (Coleoptera: Scarabaeidae) and was proposed by Lin, 1980 [ 34 , 35 ]. This data set is an example of known species and new species within the same genus. The genus Melanopopillia includes three species worldwide: Melanopopillia dinghuensis Lin, 1980 (M.ding), Melanopopillia hainanensis Lin, 1980 (M.hain) and Melanopopillia praefica (Machatschke, 1971) (M.prae). Since then, no other species of Melanopopillia have been described. An undescribed species, Melanopopillia ns., was included in this data set as a new species to demonstrate the application of workflow (M.sp.n) (Fig. 2 a). Five additional specimens of Melanopopillia praefica were used as contrast group (M.prae-con). The species in this genus are very similar in terms of their external morphology. Fourteen specimens of Melanopopillia dinghuensis were deposited at the Institute of Zoology, Guangdong Academy of Sciences (IZGAS), and the rest 41 specimens of Melanopopillia were deposited at the Institute of Zoology, Chinese Academy of Sciences (IZCAS). Dataset Hong Kong beetles (Figs. 2 a, 2 b, Table S1): These beetles (order Coleoptera) were collected in Hong Kong during 2017–2019 and deposited in IZCAS [ 36 ]. The dataset Hong Kong beetles is an example of 17 record species (Carab-1, Carab-2, Carab-3, Carab-4, Staph-1, Scara-1, Scara-2, Scara-3, Scara-4, Elate-1, Elate-2, Elate-3, Prion-1, Teneb-1, Morde-1, Ceram-1, and Ceram-2) and four new record species (Hydro-1, Hybos-1, Scara-5 and Scara-6) at the same locality, and four of the record species include extra specimens as contrast groups (Carab-3-con, Carab-4-con, Scara-3-con and Teneb-1-con). We selected all contrast specimens randomly from the known species. 2.2 | Photographing image stacks of the specimens We used an Olympus E-M5 Mark II to obtain the videos of the dorsal-lateral, lateral and ventral-lateral surfaces of the specimens from three angles: elevation of 45°, 0° and − 45°. In each video, the specimen was rotated 360 °. We use PotPlayer software (Daum Communications) to uniformly save frames of the image from videos at equal step sizes, which is equivalent to obtaining equal 360° camera azimuth images. In addition, we also took two pictures of the dorsal and ventral surfaces of the specimen. Each of the two images was rotated 360 ° at a synchronous length to obtain rotated images. This gave a series of images with different camera elevations and azimuth angles for each specimen. The process of obtaining these images was integrated into a serial of specially designed devices. To increase the efficiency of obtaining image stacks from dry pinned specimens, we designed three high-throughput multiangle photography mechanical devices. Specimens were placed on the platform designed to rotate them for a full 360 ° turn when videos and images were captured at 45°, 90°, and − 45°to the optic axis. To maintain stability and efficiency, the devices were updated from version one to version three. And the versions two and three got the Chinese patents respectively (Nos. CN202110042390.9 and PY23DX10783FNUM-CN). 2.3 | CNN models Four network models: AlexNet [ 37 ], ResNet-152 [ 38 ], DenseNet [ 39 ] and MobileNetV2 [ 40 ] were used as an alternative backbone to differentiate known species in the workflow. We removed the original fully connected layer of these networks and added a global average pooling layer and a fully connected layer with three or 17 nodes, respectively. Before training, the weights trained on ImageNet were loaded as initial weights and frozen, and only the fully connected layer was trained. Sparse categorical cross-entropy and Adam were selected as the loss function and optimizer, respectively. To ensure balanced training data, we selected 1000 images for each known species and constructed a dataset containing 3000 images for the three known species in dataset-1 and 17 000 images for the 17 known species in dataset-2. The images were processed by random shearing, changing brightness, and changing size. After sufficiently shuffling the order, the dataset was randomly divided into a training set and validation set in an 8: 2 ratio from the full image set. The learning rate is 0.0001. 2.5 | AI barcoding In this step, four or five specimens that were different from those used in the CNN step were selected from each of the known species, together with four or five specimens of unknown species (test and new). For each specimen, 1578 images (dataset-1) and 150 images (dataset-2) were taken at certain camera angles, and finally made up a total of 39 450 images in dataset-1 and 18450 images in dataset-2. All images were aligned to ensure the same image ID for different specimens at the same angle. The image in which the beetles’ head was in the middle which was coded as the first image (ID 1), and then the other images were aligned from the dorsal lateral view to the lateral view to the ventral lateral view. The calibrated error was approximately two or three images. These images were then inputted into the trained CNN models. After CNN calculation, each image was given a probability of belonging to all known species (dataset-1: three species, dataset-2: seventeen species), known as confidence. The results are represented as (Sp.1: ɑ , Sp.2: β , Sp.3: γ , …Sp.n: G n ; n = all known species in CNN), where 0<= ɑ , β , γ , … G n < = 1 and ɑ + β + γ +…G n = 1. The confidence of all images of one specimen that formed a matrix (number of image IDs x number of all known species) was taken as a quantitative descriptor of the morphological characteristics of this specimen, called the AI barcode. The t-SNE algorithm was employed to display the similarities and differences between species and specimens in the reduced-dimensionality feature space [ 41 ]. Other unsupervised dimensionality reduction algorithms, such as PCA, are also suitable here, but t-SNE is often more successful in visualizing high dimensional data by retraining the significant structure of the data [ 42 ]. To further explore whether the AI barcodes represent diagnostic characters to some extent, we highlighted the continuous angles with higher accuracy in the AI barcodes of dataset-1. Every 40 angles of the images were evaluated together by their average accuracy. Then the top continuous angles were selected and compared with their diagnostic characters that are recognized by traditional taxonomists. 2.6 | Distinguishing species PERMANOVA analysis (permutations = 1000) was used to test for significance between species (known species) and specimens (new species and known test species) based on calculated Euclidean distances between specimens [ 43 , 44 ]. PERMANOVA is a permutation-based extension of multivariate analysis of variance applied to a matrix of pairwise distances [ 45 ] that is suitable for analysis of the result matrix. RESULTS 3.1 | Multiangle photography mechanical devices The multiangle photography mechanical devices were utilized and updated from version one to version three (Fig. 2 c). In version 1, the specimens rotated 360 ° in one direction on a turnplate and the angle of camera adjusted by the manual. In version two, the specimens could be rotated in two directions, in version three, the platform for specimens and camera integrated together: the specimens could be rotated in several directions automatically and more stable, and the camera could move in the one direction (Video S1). The updated device version three is fast and stable, when capturing muti-angle images of specimens, the total handing time is about 2 min per specimen. 3.2 | Overview of the Taichi workflow The overall Taichi workflow can be subdivided into three main steps: species database construction, AI barcoding, and identification unknown species (Fig. 1 b). The first step is establishing a database of morphological characteristics for known species by high-throughput multiangle photography of specimens and ML system training based on image stacks. The second step is the generation of diagnostic information among species by the ML system, named the AI barcode. The aligned multiangle images of the specimens from all known species, which were not used in the first step, are input into the trained ML system. The probability values of the aligned image output from the ML system form an array that represents the specimen. The third step is the detection and confirmation of unknown species status facilitated by AI barcodes. We put aligned multiangle images of the unidentified specimens, including a mix of specimens of both known and unknown species, into the trained ML system and obtained AI barcodes. Based on the AI barcodes of these unidentified specimens, the differences among all specimens were calculated using permutational multivariate analysis of variance (PERMANOVA). These unidentified specimens were classified into known species and unknown species, which is the method used to detect unknown species. Then the status (new record or new species) of unknown species was confirmed. 3.3 | Taichi Step 1: Database of morphological characteristics of known species We used two data sets as examples to test the Taichi workflow (Fig. 2 a, Table S1). Dataset-1: Melanopopillia consists of 4 species and 50 specimens, including all three species known in this genus (M.prae, M.ding, and M.hain) and an undescribed new species (M.sp.n) as an example of finding a species previously undescribed (Fig. 2 a). Dataset-2: Hong Kong beetles consist of 21 species and 206 specimens, including 4 new record species for Hong Kong as an example of finding new records for a region (Figs. 2 a, 2 b). Both data sets included additional specimens of known species as contrast groups (species code plus '-con'). To collect characters from the specimens, we obtain a series of images for each specimen by multiangle photography mechanical devices (Fig. 2 c). A total of 1578 and 150 aligned images were collected per specimen for dataset-1 and dataset-2, respectively, which were collected in the same sequence of photography angles. Within each dataset, the images with the same aligned serial number from different specimens represent specimens photographed at the same angle (Fig. S1). To establish a database of known species, we use multiangle random images of specimens to train the ML system. After 30 epochs, the best top-1 score of CNN identification systems in dataset-1 reached 97.76% and in dataset-2 reached 96.22% (Table S2). Except for AlexNet, the other three CNNs all perform well. In this workflow, we have multiple selections for CNN models as long as they have better identification performance [ 38 , 46 ]. Here, the “MobileNetV2” network model was selected as the backbone of the integrated workflow because of the good performance in identifying highly similar species, the high convergence speed and the high precision [ 40 ]. 3.4 | Taichi Step 2: AI barcodes as diagnostic information In Step 2, we selected four or five specimens from each known species that differed from the specimens in the last step input into the trained ML system, and 1578 aligned images in dataset-1 or 150 aligned images in dataset-2 were used for every specimen. After CNN calculation, the probability of each aligned image belonging to all known species was determined (dataset-1: three species, dataset-2: seventeen species). The probabilities of all images of each specimen made up the AI barcode (Figs. 1 b, 3 a). To visualize the result of AI barcodes, the t-distributed stochastic neighbor embedding (t-SNE) algorithm was employed to display the similarities and differences between pairs of species and specimens in the reduced-dimensionality feature space [ 41 , 42 ]. We observed a noticeable variation in AI barcodes among species (Fig. 3 a). More precisely, in dataset-1, the five replicated specimens of M.prae, M.ding, and M.hain all showed a high probability for their own species (Fig. 3 b), and the three known species were separated from each other (Fig. 3 c). Dataset-2 displayed similar trends (Figs. 3 d, 3 e), except for a few specimens: specimen 5 of Carab-3, specimen 4 of Elate-2, specimen 1 of Elate-3, and specimen 5 of Scara-3 showed relatively dispersed AI barcodes, and specimen 5 of Carab-3-con showed a high probability of belonging to Carab-2 species (Fig. 3 d, Fig. S2). In the t-SNE plot, the 17 known species were still separated from each other (Fig. 3 e). Although several specimens showed a range of variations, the feature space of the species, including all specimens, still showed clear separation. In all angles of the AI barcodes of dataset-1, the top continuous angles were ID 185 to 224 which mainly contain the elytron and pygidium, and the latter was considered as the key characters to distinguish the species within the genus [ 34 , 35 ]. 3.5 | Taichi Step 3: Unknown species detection and status confirmation In Step 3, image stacks of unknown species (new species or new record species and known contrast species) were put into the trained ML system to obtain their AI barcodes. After comparing the resulting AI barcodes with those of known species determined in Step 2, we found that different from the AI barcodes of known species that concentrate high probability in one species, the new species or new record species usually showed a relatively dispersed distribution pattern. In the t-SNE plots, new species or new record species were clearly separated from all known species, while known species overlapped with their own species. In this way, the unknown species could be detected. Specifically, M.prae-con in dataset-1 showed an AI barcode spectrum similar to M.prae, and M.sp.n showed relatively inconsistent patterns (Fig. 3 b). In the t-SNE plot, M.prae and M.prae-con almost overlapped with each other, while M .sp.n was represented by the largest circle, which was very different from the small and concentrated circle of other known species (Fig. 3 c). The same pattern appeared in dataset-2: new record species (Hydro-1-n.r., Hybos-1-n.r., Scara-5-n.r. and Scara-6-n.r.) all displayed dispersed AI barcodes, but each species had its own distribution pattern. In the t-SNE plot, four new record species were separated from all known species, but showed different distribution patterns: Scara-5-n.r. and Scara-6-n.r. were located close to the known species Scara-2; Hybos-1-n.r. and Hydro-1-n.r were located close to the known species Scara-1 and/or Morde-1 (Fig. 3 e). At the same time, four contrast known species (Scara-3-con, Teneb-1-con, Carab-4-con, and Carab-3-con) all showed high probability with their own species and overlapped with their corresponding species in the t-SNE plot. The results of dataset-2 showed that the new record species have AI barcodes that differ not only from those of all known species but also from those of each other. To quantify the differences among specimens and species, we analyzed AI barcodes using PERMANOVA with 1, 000 permutations [ 43 – 45 ]. The adjusted p values are shown in Fig. 4 a. Overall, p .adj = 0.05 could be considered as the significance threshold to distinguish species groups. All known species and new species/new record species were significantly separated from each other, and the contrast known species were correctly grouped with their corresponding species. For dataset-1, M.sp.n group was significantly separated ( p .adj < 0.05) from all known species of Melanopopillia (M.prae, M.ding, and M.hain), and M.prae-con showed no difference ( p .adj = 0.15) from M.prae (Fig. 4 a). For dataset-2, all known species were separated ( p .adj < 0.05) from each other, and the “new” species Hydro-1, Hybos-1, Scara-5, and Scara-6 were separated ( p .adj < 0.05) from all known species and from each other. The contrast groups Carab-3-con, Carab-4-con, Scara-3-con, and Teneb-1-con showed no difference ( p .adj = 0.671, p .adj = 0.929, p .adj = 0.087, p .adj = 0.218) from their corresponding species (Fig. 4 b). 3.6 | Confirmation of new species or new records based on traditional examination The new inferred species in dataset-1 were later studied by morphological comparison and DNA barcoding methods. The results of morphological comparison suggested that the new species is indeed similar to its known congeners but differs in the punctation of the pronotum, elytra striae, and basic shape of the aedeagus in males. Among the three known species, M.hain shared the most characters with this new species. The results of DNA barcoding analysis also support this inferred new species (Text S1, Table S3, Fig. S3). The description of this new species and the revision of the genus Melanopopillia will be published in the future. In dataset-2, four new record species belonged to the same order (Hydro-1-n.r), superfamily, subfamily, and genus as known species; therefore, they probably have different morphological distances from all known species. Specifically, Hybos-1-n.r belongs to the superfamily Scarabaeoidea, as do Scara-1, Scara-2, Scara-3, and Scara-4; Scara-6-n.r belongs to the subfamily Rutelinae, as do Scara-3; and Scara-5-n.r belongs to the genus Sophrops , as do Scara-2. The true classification status and similarities of these new species and records are also reflected in the AI barcode results to a certain extent, such as M.sp.n being close to M.hain, and Scara-5-n.r. being close to Scara-2. DISCUSSION Here, we establish the feasibility of developing a CNN-based biodiversity discovery workflow ( Taichi ) that is capable of detecting potential new species in a highly morphologically conserved genus and detecting potential new records of species from a certain region. The core of this workflow is to obtain an aligned image stack through the newly designed photography devices, acquire AI barcodes, and qualitatively evaluate group differences by comparing AI barcodes. The new photography device developed here can automatically and rapidly obtain a series of multiangle aligned images from stereoscopic specimens. The aligned image stack makes it possible to compare different specimens in more available morphological information. In previous studies, species identification systems were normally trained using images taken in dorsal view or at random angles [ 17 , 21 , 24 , 47 ]. While it is not suitable to the stereoscopic species, for example insects, whose diagnostic characters distribute in muti-angles. From the multiangle identification results we found that different parts of the specimen have different values for ML system which similar as the taxonomic view. In some degree, it visualizes the identification process of deep learning system. As shown here, using multiangle images can increase the power of the training in CNNs and make it possible to infer species by more characters. The analysis result indicates that AI barcodes can be used to measure similarity between specimens and species based on photos taken at different angles, especially between known species and 'new' species. And AI barcodes contain diagnostic information that could identify the species. After the analysis of AI barcodes, the Taichi workflow has the ability to detect new species/new record species and has good performance in detecting more than one new record species. This workflow can be expanded to a variety of scenarios in which detection of the 'new' or 'extra' taxa is needed. Furthermore, adding the 'new' species to the workflow makes self-renewal of the CNN system and its further application to monitoring biodiversity possible. The ML system in the Taichi workflow is similar to the other image-based AI identification systems in terms of autoidentification performance (top-1 accuracy greater than 95%). By analyzing all preliminary results of aligned multiangle images of the specimen yielded better performance in inferring new species, in a sense, which extracts studied knowledge within CNNs similar to vectors or embeddings. Compared to traditional identification methods, the Taichi workflow is also superior in terms of efficiency and throughput. This workflow is faster than the traditional method from image acquisition to subsequent analysis and is capable of high throughput, including a large number of known species for reference. In the Taichi workflow described here, two aspects should be considered with caution. First, we should be able to assign unknown specimens to the same taxonomic level, such as genus (dataset-1) or order (dataset-2). In this way, this method is similar to traditional taxonomy, where a given taxon is gradually diagnosed from higher to lower hierarchical levels (order, family, genus, and then species). In dataset-1, we trained a genus identification system to assign the new species to Melanopopillia (Text S2, Table S4, Fig. S4). Therefore, the CNN model for identifying taxa at higher levels, higher taxa could be added to the Taichi workflow to increase compatibility in more complex identification tasks. Second, the species should include at least four/five specimens to increase the precision of the judgment. In the AI barcode results, the dispersed distribution pattern (such as those for M.hain-3 in dataset-1 and EI.Me-1 in dataset-2, Figs. 3 a, 3 d) or the high probability of incorrect species assignment (Carab-3-con-5 in dataset-2, Fig. S2, Fig. 3 d) indicates that variation within species always exists. A similar situation also exists in traditional taxonomy in which the confirmation of new species should rely on a series of specimens, or when conducting DNA barcoding, it is desirable to obtain DNA sequence data from multiple specimens representing a hypothetical species entity. We have demonstrated the potential of the Taichi workflow to accelerate the discovery of new species/new record species by creating a single, reproducible repository for quantifying the ranges of morphological variation exhibited by known species, and this will reduce the workload of taxonomists in routine identifications for biodiversity assessment. In the past 30 years, the digitization of museum collections has expanded rapidly, and large numbers of specimen images have been generated [ 48 ], which can be used by ML technology in the future. Reciprocally, ML can also drive the development and data mining from digitization of natural history collections [ 49 ] and thus has considerable potential to advance the field of biodiversity science. Declarations CONFLICT OF INTEREST The authors declare no conflict of interest. AUTHOR CONTRIBUTIONS Y.Y. L. and M. B conceived the ideas, designed the methodology, and led the writing of the manuscript. J. L. and Z. Y. Z. trained and wrote the CNN parts. Y.C. Z. and Z.Y. Z. performed the bioinformatic analysis and wrote the PERMANOVA parts. Y.Y. L., Z.Y. Z., Y.J. T., B. T., N. L., and J.J. S. collected and interpreted the data. All authors contributed critically to the drafts and gave final approval for publication. DATA AVAILABILITY STATEMENT The code can be found on GitHub (https://github.com/LSDXBZHH/detecting-new-species). The data set links can be found in Science Data Bank (https://www.scidb.cn/s/2Mjuaq). FUNDING This research was supported by the National Natural Science Foundation of China (Nos. 32270468, 31961143002, 32200354), National Key R&D Program of China (No. 2022YFC2601200), National Science & Technology Fundamental Resources Investigation Program of China (No. 2022FY100500) and the Bureau of International Cooperation, Chinese Academy of Sciences. ACKNOWLEDGMENTS We thank Ping Yang from Institute of Zoology, Guangdong Academy of Sciences, IZGAS and Mingzhi Zhao from South China Agricultural University for loaning specimens of Melanopopillia to study. The authors thank Norman MacLeod from Nanjing University for his advice in manuscript writing. We thank Xinhai Li and Zhengting Zou from the Institute of Zoology, CAS (IZCAS), Roberta Eleanor Hunt from University of Copenhagen for useful advice in biostatistical analysis. The authors thank Zhonglin Xu, Xin Chen, Jun Wang from Cangzhou Normal University, Jiajia Zhai, and Zhibin Sun from National Space Science Center, CAS for designing and processing the high-throughput multiangle photography device, Rongrong Shen from China Agricultural University, Yingying Yu, Xiaoxuan Li, and Huanxi Cao from IZCAS for the useful advice and help in DNA parts. We also thank Hongbin Liang (IZCAS), Yang Liu (Northwest University), and Chuanbu Gao (Institute of Zoology, Guangdong Academy of Sciences) for identifying the Hong Kong species. References Cardinale BJ, Duffy JE, Gonzalez A, Hooper DU, Perrings C, Venail P, et al. Biodiversity loss and its impact on humanity. Nature. 2012;486:59–67. Costello MJ, May RM, Stork NE. Can We Name Earth’s Species Before They Go Extinct? Science. 2013;339:413–6. Sandifer PA, Sutton-Grier AE, Ward BP. Exploring connections among nature, biodiversity, ecosystem services, and human health and well-being: Opportunities to enhance health and biodiversity conservation. Iss Environ Sci Tech. 2015;12:1–15. Deans AR, Yoder MJ, Balhoff JP. Time to change how we describe biodiversity. Trends Ecol Evol. 2012;27:78–84. Stork NE. How Many Species of Insects and Other Terrestrial Arthropods Are There on Earth? Annu Rev Entomol. 2018;63:31–45. Wheeler QD, Raven PH, Wilson EO. Taxonomy: Impediment or Expedient? Science. 2004;303:285–285. Ebach MC, Valdecasas AG, Wheeler QD. Impediments to taxonomy and users of taxonomy: accessibility and impact evaluation. Cladistics. 2011;27:550–7. Orr MC, Ferrari RR, Hughes AC, Chen J, Ascher JS, Yan Y-H, et al. Taxonomy must engage with new technologies and evolve to face future challenges. Nat Ecol Evol. 2020;5:3–4. MacLeod N, Benfield M, Culverhouse P. Time to automate identification. Nature. 2010;467:154–5. Bisgin H, Bera T, Ding H, Semey HG, Wu L, Liu Z, et al. Comparing SVM and ANN based Machine Learning Methods for Species Identification of Food Contaminating Beetles. Sci Rep. 2018;8:6532. Song Y, He F, Zhang X. To Identify Tree Species With Highly Similar Leaves Based on a Novel Attention Mechanism for CNN. IEEE Access. 2019;7:163277–86. Wu L, Liu Z, Bera T, Ding H, Langley DA, Jenkins-Barnes A, et al. A deep learning model to recognize food contaminating beetle species based on elytra fragments. Comput Electron Agric. 2019;166:105002. Seeland M, Rzanny M, Boho D, Wäldchen J, Mäder P. Image-based classification of plant genus and family for trained and untrained plant species. BMC Bioinf. 2019;20:4. Fedor P, Vaňhara J, Havel J, Malenovský I, Spellerberg I. Artificial intelligence in pest insect monitoring. Syst Entomol. 2009;34:398–400. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436–44. Niemi J, Tanttu J. Deep Learning Case Study for Automatic Bird Identification. Appl Sci. 2018;8:2089. Valan M, Makonyi K, Maki A, Vondráček D, Ronquist F. Automated Taxonomic Identification of Insects with Expert-Level Accuracy Using Effective Feature Transfer from Convolutional Networks. Buckley T, editor. Syst Biol. 2019;68:876–95. Lu W, Chen X, Wang L, Li H, Fu YV. Combination of an Artificial Intelligence Approach and Laser Tweezers Raman Spectroscopy for Microbial Identification. Anal Chem. 2020;92:6288–96. Ferreira AC, Silva LR, Renna F, Brandl HB, Renoult JP, Farine DR, et al. Deep learning‐based methods for individual recognition in small birds. Codling E, editor. Methods Ecol Evol. 2020;11:1072–85. Carranza-Rojas J, Goeau H, Bonnet P, Mata-Montero E, Joly A. Going deeper in the automated identification of Herbarium specimens. BMC Evol Biol. 2017;17:181. Wäldchen J, Rzanny M, Seeland M, Mäder P. Automated plant species identification—Trends and future directions. Bucksch A, editor. PLOS Comput Biol. 2018;14:e1005993. Liu L, Wang R, Xie C, Yang P, Wang F, Sudirman S, et al. PestNet: An End-to-End Deep Learning Approach for Large-Scale Multi-Class Pest Detection and Classification. IEEE Access. 2019;7:45301–12. Mitra R, Marchitto TM, Ge Q, Zhong B, Kanakiya B, Cook MS, et al. Automated species-level identification of planktic foraminifera using convolutional neural networks, with comparison to human performance. Mar Micropaleontol. 2019;147:16–24. Thenmozhi K, Srinivasulu Reddy U. Crop pest classification based on deep convolutional neural network and transfer learning. Comput Electron Agric. 2019;164:104906. Molnar C, Casalicchio G, Bischl B. Interpretable machine learning–a brief history, state-of-the-art and challenges. Joint European conference on machine learning and knowledge discovery in databases. Springer; 2020. p. 417–31. Chandola V, Banerjee A, Kumar V. Anomaly Detection: A Survey. Acm Comput Surv. 2009;41:1–58. Zhao Z, Lu Y, Tong Y, Chen X, Bai M. PENet : A phenotype encoding network for automatic extraction and representation of morphological discriminative features. Methods Ecol Evol. 2023;2041-210X.14235. Pastore VP, Zimmerman TG, Biswas SK, Bianco S. Annotation-free learning of plankton for classification and anomaly detection. Sci Rep. 2020;10:12142. Khalighifar A, Brown RM, Goyes Vallejos J, Peterson AT. Deep learning improves acoustic biodiversity monitoring and new candidate forest frog species identification (genus Platymantis) in the Philippines. Biodivers Conserv. 2021;30:643–57. Khalighifar A, Jiménez-García D, Campbell LP, Ahadji-Dabla KM, Aboagye-Antwi F, Ibarra-Juárez LA, et al. Application of Deep Learning to Community-Science-Based Mosquito Monitoring and Detection of Novel Species. Yee D, editor. J Med Entomol. 2022;59:355–62. Badirli S, Picard CJ, Mohler G, Richert F, Akata Z, Dundar M. Classifying the unknown: Insect identification with deep hierarchical Bayesian learning. Methods Ecol Evol. 2023;14:1515–30. Dubois A. Describing a New Species. TAPROBANICA: J Asian Biodivers. 2011;2:6. Miralles A, Bruy T, Wolcott K, Scherz MD, Begerow D, Beszteri B, et al. Repositories for Taxonomic Data: Where We Are and What is Missing. Friedman M, editor. Syst Biol. 2020;69:1231–53. Lin P. A new genus, Melanopopillia, from China (Coleoptera: Rutelidae). Entomotaxonomia. 1980;2:297–301. Lu Y, Yang H, Bai M. Micro CT approach applied in taxonomy: An example on the species Melanopopillia hainanensis (Coleoptera: Scarabaeidae). Zoological Systematics. 2019;44:294–303. Zhao S, Tong Y, Teng B, Chen X, Yang X, Li J, et al. A species diversity dataset of beetles by three passive acquisition methods in Tei Tong Tsai (Hong Kong). Sci Data. 2022;9:210. Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems. 2012;25:1097–105. He K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. Las Vegas, NV, USA: IEEE; 2016 [cited 2023 Nov 25]. p. 770–8. Available from: http://ieeexplore.ieee.org/document/7780459/ Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely Connected Convolutional Networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. Honolulu, HI: IEEE; 2017 [cited 2023 Nov 25]. p. 2261–9. Available from: https://ieeexplore.ieee.org/document/8099726/ Sandler M, Howard A, Zhu M, Zhmoginov A, Chen L-C. MobileNetV2: Inverted Residuals and Linear Bottlenecks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition [Internet]. Salt Lake City, UT: IEEE; 2018 [cited 2023 Nov 25]. p. 4510–20. Available from: https://ieeexplore.ieee.org/document/8578572/ Laurens Van Der Maaten, Hinton G. Visualizing data using t-SNE. J Mach Learn Res. 2008;9:2579–605. Anowar F, Sadaoui S, Selim B. Conceptual and empirical comparison of dimensionality reduction algorithms (PCA, KPCA, LDA, MDS, SVD, LLE, ISOMAP, LE, ICA, t-SNE). Comput Sci Rev. 2021;40:100378. Anderson MJ. A new method for non‐parametric multivariate analysis of variance. Austral Ecol. 2001;26:32–46. Anderson MJ, Walsh DCI. PERMANOVA, ANOSIM, and the Mantel test in the face of heterogeneous dispersions: What null hypothesis are you testing? Ecol Monogr. 2013;83:557–74. Kelly BJ, Gross R, Bittinger K, Sherrill-Mix S, Lewis JD, Collman RG, et al. Power and sample-size estimation for microbiome studies using pairwise distances and PERMANOVA. Bioinformatics. 2015;31:2461–8. Zoph B, Vasudevan V, Shlens J, Le QV. Learning Transferable Architectures for Scalable Image Recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition [Internet]. Salt Lake City, UT: IEEE; 2018 [cited 2023 Nov 25]. p. 8697–710. Available from: https://ieeexplore.ieee.org/document/8579005/ Rawat W, Wang Z. Deep Convolutional Neural Networks for Image Classification: A Comprehensive Review. Neural Comput. 2017;29:2352–449. Nelson G, Ellis S. The history and impact of digitization and digital data mobilization on biodiversity research. Philos Trans R Soc B: Biol Sci. 2019;374:20170391. Hedrick BP, Heberling JM, Meineke EK, Turner KG, Grassa CJ, Park DS, et al. Digitization and the Future of Natural History Collections. Bioscience. 2020;70:243–51. Video S1 Video S1 is not available with this version Supplementary Files Supplementaryinformation.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-3832815","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":265340872,"identity":"0b1f3232-8997-48b9-85e3-7346328d0790","order_by":0,"name":"Yuanyuan Lu","email":"","orcid":"","institution":"Institute of Zoology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Yuanyuan","middleName":"","lastName":"Lu","suffix":""},{"id":265340873,"identity":"8f22d562-2505-4c3e-a5b7-1ff1985c0189","order_by":1,"name":"Jing Li","email":"","orcid":"","institution":"Institute of Zoology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Li","suffix":""},{"id":265340874,"identity":"216d02af-1481-419c-baf8-f72479e8ae26","order_by":2,"name":"Zhengyu Zhao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6UlEQVRIiWNgGAWjYDACCSBmbAAzGR8kkKqF2YBkLWwSROmQn9388OHPHTZ58hE5ZhUPKuoY+NsPMH4uwKOFcc4xYwPJM2nFhjdyzG4knGFjkDiTwCw9A48WZokEMwnDtsOJG2cAtSS28TAw3GBgY+bBo4VNIv2bRGLbf7CWgsQ2CQZ5Qlp4JHLMJA62HUicD2QwJLYZMBgQ0iIhkVNs2NiWnLiB51mxRMKZBB7DM4nN0vi0yM9I3/jwZ5td4vz25I0ff1TUyckdP3zwMz4tcGBwIQHiUng0EQTy/QeIUzgKRsEoGAUjDwAArANIZjYuTHwAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0003-2335-447X","institution":"Institute of Zoology Chinese Academy of Sciences","correspondingAuthor":true,"prefix":"","firstName":"Zhengyu","middleName":"","lastName":"Zhao","suffix":""},{"id":265340875,"identity":"a36f5205-73ed-49cf-aa45-e117be68c7c3","order_by":3,"name":"Yongchao Zhang","email":"","orcid":"","institution":"National Institute of Biological Sciences Beijing","correspondingAuthor":false,"prefix":"","firstName":"Yongchao","middleName":"","lastName":"Zhang","suffix":""},{"id":265340876,"identity":"bb800fdb-b7b1-47cd-97e3-445086f7b6b7","order_by":4,"name":"Yijie Tong","email":"","orcid":"","institution":"Institute of Zoology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Yijie","middleName":"","lastName":"Tong","suffix":""},{"id":265340877,"identity":"cb61015a-841b-4d30-ab9c-a000efb11e9a","order_by":5,"name":"Bei Teng","email":"","orcid":"","institution":"Institute of Zoology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Bei","middleName":"","lastName":"Teng","suffix":""},{"id":265340878,"identity":"03ccb403-c565-4239-a050-f8ca08dc1fc4","order_by":6,"name":"Ning Liu","email":"","orcid":"","institution":"Institute of Zoology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Ning","middleName":"","lastName":"Liu","suffix":""},{"id":265340879,"identity":"761f83cf-5ed2-4e01-96d1-8d768773ceb7","order_by":7,"name":"Josh Jenkins Shaw","email":"","orcid":"","institution":"Institute of Zoology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Josh","middleName":"Jenkins","lastName":"Shaw","suffix":""},{"id":265340880,"identity":"3a781664-162e-454b-8b09-45d5a8664850","order_by":8,"name":"Ming Bai","email":"","orcid":"https://orcid.org/0000-0001-9197-5900","institution":"Institute of Zoology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Ming","middleName":"","lastName":"Bai","suffix":""}],"badges":[],"createdAt":"2024-01-03 21:24:52","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3832815/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3832815/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":49382466,"identity":"b215739a-dd5b-486b-a9d0-57c5c8197ca8","added_by":"auto","created_at":"2024-01-09 19:24:55","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":881951,"visible":true,"origin":"","legend":"\u003cp\u003eBiodiversity discovery empowered by the \u003cem\u003eTaichi\u003c/em\u003e workflow. a) Fundamental steps for identifying a 'new' species by human expertise (orange) and machine learning (green). b) Workflow of \u003cem\u003eTaichi\u003c/em\u003e. The data set illustrated here is similar to dataset-1, including three known species (Sp.1: blue, Sp.2: purple, Sp.3: green) and one unknown species (Sp.unk: red). Sig.testing: significance testing.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-3832815/v1/1eb254a47dfd2a74c952354a.png"},{"id":49381732,"identity":"90f031ab-c7ab-4b7b-be3d-c03e0dc45dad","added_by":"auto","created_at":"2024-01-09 19:16:55","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1228650,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic diagram of the species database established in the \u003cem\u003eTaichi\u003c/em\u003eworkflow. a) Examples of species and 'new' species from the two case studies. Dataset-1: \u003cem\u003eMelanopopillia\u003c/em\u003e, all known species, and one new species belonging to the genus \u003cem\u003eMelanopopillia\u003c/em\u003e (Coleoptera: Scarabaeidae: Rutelinae). Dataset-2: \u003cem\u003eHong Kong beetles\u003c/em\u003e, species collected from Hong Kong (Coleoptera). b) Species included in dataset-2. c) Abridged general view of the integrated equipment for collecting the multiangle images of specimens (version one to version three).\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-3832815/v1/46b5c9befac71f68d50abd60.png"},{"id":49382467,"identity":"64768871-8974-4438-88bc-13f4bbe368e6","added_by":"auto","created_at":"2024-01-09 19:24:55","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1173055,"visible":true,"origin":"","legend":"\u003cp\u003eAI barcodes of specimens in the two datasets. a) Two heatmaps showing the construction of the AI barcode (M.hain: specimen-5 and M.sp.n: specimen-5): probability values of each image resulting from the CNN belonging to all known species (dataset-1: M.prae, M.ding, and M.hain), ranging from zero to one (dataset-1: images 1 to 1578). b) AI barcodes of all specimens in dataset-1: M.prae(specimens 1–5), M.prae-con (specimens 1–5), M.ding (specimens 1–5), M.hain (specimens 1–5) and M.sp.n (specimens 1–5). c) Ordination space for AI barcodes of specimens included in dataset-1 formed by the first two t-SNE dimensionality reduction axes. d) AI barcodes of the first specimens of all species (17 known species, 4 test species, and 4 new record species) in dataset-2: Carab-1, Carab-2, Carab-3, Carab-3-con, Carab-4, Carab-1-con, Staph-1, Scara-1, Scara-2, Scara-3, Scara-3-con, Scara-4, Elate-1, Elate-2, Elate-3, Prion-1, Teneb-1, Teneb-1-con, Morde-1, Ceram-1, Ceram-2, Hydro-1-n.r, Hybos-1-n.r, Scara-5-n.r and Scara-6-n.r. The probability results (ranging from zero to one) of each image of the first specimens belonging to all known species in dataset-2: Carab-1, Carab-2, Carab-3, Carab-4, Staph-1, Scara-1, Scara-2, Scara-3, Scara-4, Elate-1, Elate-2, Elate-3, Prion-1, Teneb-1, Morde-1, Ceram-1 and Ceram-2. For AI barcodes of all specimens, see Fig. S2. e) Ordination space for AI barcodes of the specimens included in dataset-2 formed by the first two t-SNE dimensionality reduction axes.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-3832815/v1/73029b46ddff732a3350df7c.png"},{"id":49381733,"identity":"93cd1095-7008-43ec-94f7-af9d18e12e1d","added_by":"auto","created_at":"2024-01-09 19:16:55","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1189911,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow and results showing the statistical analysis of group differences using AI barcodes based on PERMANOVA (permutations = 1 000). a) Workflow and results of dataset-1. b) Heatmap showing the results of the statistical analysis of dataset-2. The green groups are test species, which are grouped (\u003cem\u003ep\u003c/em\u003e.adj\u0026gt;0.05) with their corresponding species. Red groups are 'new' species that are separated (\u003cem\u003ep\u003c/em\u003e.adj\u0026lt;0.05) from all known species.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-3832815/v1/bc26b3dcbcc69ae07c565e99.png"},{"id":51091648,"identity":"1cf793af-15f5-41c2-9803-e24c2f6b594c","added_by":"auto","created_at":"2024-02-14 01:38:23","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2421407,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3832815/v1/023b468d-b378-46b1-bd8f-e7961ab06c30.pdf"},{"id":49381739,"identity":"79c51ebb-af19-4ea2-8294-143ed2df3235","added_by":"auto","created_at":"2024-01-09 19:16:55","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1169521,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryinformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-3832815/v1/af1c09a847f77f1b6a2c555f.docx"}],"financialInterests":"","formattedTitle":"Accelerating the discovery of biodiversity by detecting “new” species based on machine learning method","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eDiscovering, recording, and monitoring species are regarded the key steps in exploring Earth\u0026rsquo;s biodiversity [\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. While the large number of unknown species within arthropods hinders the investigation of species\u0026rsquo; biodiversity, especially faunal \u0026ldquo;new\u0026rdquo; species: new record species and new species [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. At the same time, the reduced financial support for taxonomic research and the ever-growing pool of described species exacerbates the situation [\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Therefore, there is an urgent need for technological innovation, such as machine learning (ML), to find, record, and monitor species efficiently and at high throughput [\u003cspan additionalcitationids=\"CR10 CR11 CR12\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eML has been widely used in species autoidentification systems with increased accuracy (top-5 accuracy greater than 99%) [\u003cspan additionalcitationids=\"CR15 CR16 CR17 CR18\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. However, most of the autoidentification systems provide only the species names within the system [\u003cspan additionalcitationids=\"CR21 CR22 CR23 CR24\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], when new regional records or unknown species (out-of-distribution data) are put into, the system usually incorrectly identified [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. This shortcoming is due to the limited interpretability and self-iteration of the identification system [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSome complex image-based ML systems have been developed to explore unknown group identification methods [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Probability values have recently been considered as a factor to evaluate out-of-distribution species, while single probability values have revealed less than ideal outcomes in identification systems [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. And in some cases, DNA barcodes were combined with the image feature vectors for explore the out-of-distribution species [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Here, we propose a new workflow system derived from the procedure by which taxonomists discover new species [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e], named \u003cem\u003eTaichi\u003c/em\u003e, enabled by ML and new data collection devices and analysis methods (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e): 1) establishing a database of known species by an ML system based on image stacks from the newly designed high-throughput photography devices; 2) generating diagnostic information between species using the results generated by the ML system, which named the AI barcode; and 3) identifying unknown species as facilitated by the AI barcode. To demonstrate and test the effectiveness of the \u003cem\u003eTaichi\u003c/em\u003e workflow, two real high-quality datasets were used to detect unknown species and confirm their status (new record or new species). This method provides a new perspective from the taxonomic view to detect the \u0026ldquo;new\u0026rdquo; species, and also visualizes the identification process of deep learning systems in some extant.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 | Specimens and datasets\u003c/h2\u003e \u003cp\u003eOur study uses two datasets (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea, Table S1). The \u003cem\u003eMelanopopillia\u003c/em\u003e dataset (image size: 224 x 224 pixels) consists of 4 species and 55 specimens, including all three known species in this genus and an undescribed new species (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). The dataset \u003cem\u003eHong Kong beetles\u003c/em\u003e (image size: 224 x 224 pixels) includes 21 species and 206 specimens (Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea, \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). Here, we assume that each species is represented by four or five specimens as a group to define the species within a certain range of variation.\u003c/p\u003e \u003cp\u003eDataset \u003cem\u003eMelanopopillia\u003c/em\u003e (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea, Table S1): The genus \u003cem\u003eMelanopopillia\u003c/em\u003e belongs to the subfamily Rutelinae (Coleoptera: Scarabaeidae) and was proposed by Lin, 1980 [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. This data set is an example of known species and new species within the same genus. The genus \u003cem\u003eMelanopopillia\u003c/em\u003e includes three species worldwide: \u003cem\u003eMelanopopillia dinghuensis\u003c/em\u003e Lin, 1980 (M.ding), \u003cem\u003eMelanopopillia hainanensis\u003c/em\u003e Lin, 1980 (M.hain) and \u003cem\u003eMelanopopillia praefica\u003c/em\u003e (Machatschke, 1971) (M.prae). Since then, no other species of \u003cem\u003eMelanopopillia\u003c/em\u003e have been described. An undescribed species, \u003cem\u003eMelanopopillia\u003c/em\u003e ns., was included in this data set as a new species to demonstrate the application of workflow (M.sp.n) (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). Five additional specimens of \u003cem\u003eMelanopopillia praefica\u003c/em\u003e were used as contrast group (M.prae-con). The species in this genus are very similar in terms of their external morphology. Fourteen specimens of \u003cem\u003eMelanopopillia dinghuensis\u003c/em\u003e were deposited at the Institute of Zoology, Guangdong Academy of Sciences (IZGAS), and the rest 41 specimens of \u003cem\u003eMelanopopillia\u003c/em\u003e were deposited at the Institute of Zoology, Chinese Academy of Sciences (IZCAS).\u003c/p\u003e \u003cp\u003eDataset \u003cem\u003eHong Kong beetles\u003c/em\u003e (Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea, \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb, Table S1): These beetles (order Coleoptera) were collected in Hong Kong during 2017\u0026ndash;2019 and deposited in IZCAS [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. The dataset \u003cem\u003eHong Kong\u003c/em\u003e beetles is an example of 17 record species (Carab-1, Carab-2, Carab-3, Carab-4, Staph-1, Scara-1, Scara-2, Scara-3, Scara-4, Elate-1, Elate-2, Elate-3, Prion-1, Teneb-1, Morde-1, Ceram-1, and Ceram-2) and four new record species (Hydro-1, Hybos-1, Scara-5 and Scara-6) at the same locality, and four of the record species include extra specimens as contrast groups (Carab-3-con, Carab-4-con, Scara-3-con and Teneb-1-con). We selected all contrast specimens randomly from the known species.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 | Photographing image stacks of the specimens\u003c/h2\u003e \u003cp\u003eWe used an Olympus E-M5 Mark II to obtain the videos of the dorsal-lateral, lateral and ventral-lateral surfaces of the specimens from three angles: elevation of 45\u0026deg;, 0\u0026deg; and \u0026minus;\u0026thinsp;45\u0026deg;. In each video, the specimen was rotated 360 \u0026deg;. We use PotPlayer software (Daum Communications) to uniformly save frames of the image from videos at equal step sizes, which is equivalent to obtaining equal 360\u0026deg; camera azimuth images. In addition, we also took two pictures of the dorsal and ventral surfaces of the specimen. Each of the two images was rotated 360 \u0026deg; at a synchronous length to obtain rotated images. This gave a series of images with different camera elevations and azimuth angles for each specimen. The process of obtaining these images was integrated into a serial of specially designed devices.\u003c/p\u003e \u003cp\u003eTo increase the efficiency of obtaining image stacks from dry pinned specimens, we designed three high-throughput multiangle photography mechanical devices. Specimens were placed on the platform designed to rotate them for a full 360 \u0026deg; turn when videos and images were captured at 45\u0026deg;, 90\u0026deg;, and \u0026minus;\u0026thinsp;45\u0026deg;to the optic axis. To maintain stability and efficiency, the devices were updated from version one to version three. And the versions two and three got the Chinese patents respectively (Nos. CN202110042390.9 and PY23DX10783FNUM-CN).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 | CNN models\u003c/h2\u003e \u003cp\u003eFour network models: AlexNet [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], ResNet-152 [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e], DenseNet [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e] and MobileNetV2 [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e] were used as an alternative backbone to differentiate known species in the workflow. We removed the original fully connected layer of these networks and added a global average pooling layer and a fully connected layer with three or 17 nodes, respectively. Before training, the weights trained on ImageNet were loaded as initial weights and frozen, and only the fully connected layer was trained. Sparse categorical cross-entropy and Adam were selected as the loss function and optimizer, respectively. To ensure balanced training data, we selected 1000 images for each known species and constructed a dataset containing 3000 images for the three known species in dataset-1 and 17 000 images for the 17 known species in dataset-2. The images were processed by random shearing, changing brightness, and changing size. After sufficiently shuffling the order, the dataset was randomly divided into a training set and validation set in an 8: 2 ratio from the full image set. The learning rate is 0.0001.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.5 | AI barcoding\u003c/h2\u003e \u003cp\u003eIn this step, four or five specimens that were different from those used in the CNN step were selected from each of the known species, together with four or five specimens of unknown species (test and new). For each specimen, 1578 images (dataset-1) and 150 images (dataset-2) were taken at certain camera angles, and finally made up a total of 39 450 images in dataset-1 and 18450 images in dataset-2. All images were aligned to ensure the same image ID for different specimens at the same angle. The image in which the beetles\u0026rsquo; head was in the middle which was coded as the first image (ID 1), and then the other images were aligned from the dorsal lateral view to the lateral view to the ventral lateral view. The calibrated error was approximately two or three images.\u003c/p\u003e \u003cp\u003eThese images were then inputted into the trained CNN models. After CNN calculation, each image was given a probability of belonging to all known species (dataset-1: three species, dataset-2: seventeen species), known as confidence. The results are represented as (Sp.1: \u003cem\u003eɑ\u003c/em\u003e, Sp.2: \u003cem\u003eβ\u003c/em\u003e, Sp.3: \u003cem\u003eγ\u003c/em\u003e, \u0026hellip;Sp.n: G\u003cem\u003en\u003c/em\u003e; n\u0026thinsp;=\u0026thinsp;all known species in CNN), where 0\u0026lt;=\u003cem\u003eɑ\u003c/em\u003e, \u003cem\u003eβ\u003c/em\u003e, \u003cem\u003eγ\u003c/em\u003e, \u0026hellip; G\u003cem\u003en\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;1 and \u003cem\u003eɑ\u003c/em\u003e+\u003cem\u003eβ\u003c/em\u003e\u0026thinsp;+\u0026thinsp;\u003cem\u003eγ\u003c/em\u003e+\u0026hellip;G\u003cem\u003en\u0026thinsp;=\u003c/em\u003e\u0026thinsp;1. The confidence of all images of one specimen that formed a matrix (number of image IDs x number of all known species) was taken as a quantitative descriptor of the morphological characteristics of this specimen, called the AI barcode.\u003c/p\u003e \u003cp\u003eThe t-SNE algorithm was employed to display the similarities and differences between species and specimens in the reduced-dimensionality feature space [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]. Other unsupervised dimensionality reduction algorithms, such as PCA, are also suitable here, but t-SNE is often more successful in visualizing high dimensional data by retraining the significant structure of the data [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTo further explore whether the AI barcodes represent diagnostic characters to some extent, we highlighted the continuous angles with higher accuracy in the AI barcodes of dataset-1. Every 40 angles of the images were evaluated together by their average accuracy. Then the top continuous angles were selected and compared with their diagnostic characters that are recognized by traditional taxonomists.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.6 | Distinguishing species\u003c/h2\u003e \u003cp\u003ePERMANOVA analysis (permutations\u0026thinsp;=\u0026thinsp;1000) was used to test for significance between species (known species) and specimens (new species and known test species) based on calculated Euclidean distances between specimens [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e, \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. PERMANOVA is a permutation-based extension of multivariate analysis of variance applied to a matrix of pairwise distances [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e] that is suitable for analysis of the result matrix.\u003c/p\u003e \u003c/div\u003e"},{"header":"RESULTS","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.1 | Multiangle photography mechanical devices\u003c/h2\u003e \u003cp\u003eThe multiangle photography mechanical devices were utilized and updated from version one to version three (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec). In version 1, the specimens rotated 360 ° in one direction on a turnplate and the angle of camera adjusted by the manual. In version two, the specimens could be rotated in two directions, in version three, the platform for specimens and camera integrated together: the specimens could be rotated in several directions automatically and more stable, and the camera could move in the one direction (Video S1). The updated device version three is fast and stable, when capturing muti-angle images of specimens, the total handing time is about 2 min per specimen.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.2 | Overview of the \u003cem\u003eTaichi\u003c/em\u003e workflow\u003c/h2\u003e \u003cp\u003eThe overall \u003cem\u003eTaichi\u003c/em\u003e workflow can be subdivided into three main steps: species database construction, AI barcoding, and identification unknown species (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb). The first step is establishing a database of morphological characteristics for known species by high-throughput multiangle photography of specimens and ML system training based on image stacks. The second step is the generation of diagnostic information among species by the ML system, named the AI barcode. The aligned multiangle images of the specimens from all known species, which were not used in the first step, are input into the trained ML system. The probability values of the aligned image output from the ML system form an array that represents the specimen.\u003c/p\u003e \u003cp\u003eThe third step is the detection and confirmation of unknown species status facilitated by AI barcodes. We put aligned multiangle images of the unidentified specimens, including a mix of specimens of both known and unknown species, into the trained ML system and obtained AI barcodes. Based on the AI barcodes of these unidentified specimens, the differences among all specimens were calculated using permutational multivariate analysis of variance (PERMANOVA). These unidentified specimens were classified into known species and unknown species, which is the method used to detect unknown species. Then the status (new record or new species) of unknown species was confirmed.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.3 | \u003cem\u003eTaichi\u003c/em\u003e Step 1: Database of morphological characteristics of known species\u003c/h2\u003e \u003cp\u003eWe used two data sets as examples to test the \u003cem\u003eTaichi\u003c/em\u003e workflow (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea, Table S1). Dataset-1: \u003cem\u003eMelanopopillia\u003c/em\u003e consists of 4 species and 50 specimens, including all three species known in this genus (M.prae, M.ding, and M.hain) and an undescribed new species (M.sp.n) as an example of finding a species previously undescribed (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). Dataset-2: \u003cem\u003eHong Kong beetles\u003c/em\u003e consist of 21 species and 206 specimens, including 4 new record species for Hong Kong as an example of finding new records for a region (Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea, \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). Both data sets included additional specimens of known species as contrast groups (species code plus '-con'). To collect characters from the specimens, we obtain a series of images for each specimen by multiangle photography mechanical devices (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec). A total of 1578 and 150 aligned images were collected per specimen for dataset-1 and dataset-2, respectively, which were collected in the same sequence of photography angles. Within each dataset, the images with the same aligned serial number from different specimens represent specimens photographed at the same angle (Fig. S1).\u003c/p\u003e \u003cp\u003eTo establish a database of known species, we use multiangle random images of specimens to train the ML system. After 30 epochs, the best top-1 score of CNN identification systems in dataset-1 reached 97.76% and in dataset-2 reached 96.22% (Table S2). Except for AlexNet, the other three CNNs all perform well.\u003c/p\u003e \u003cp\u003eIn this workflow, we have multiple selections for CNN models as long as they have better identification performance [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e, \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. Here, the “MobileNetV2” network model was selected as the backbone of the integrated workflow because of the good performance in identifying highly similar species, the high convergence speed and the high precision [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.4 | \u003cem\u003eTaichi\u003c/em\u003e Step 2: AI barcodes as diagnostic information\u003c/h2\u003e \u003cp\u003eIn Step 2, we selected four or five specimens from each known species that differed from the specimens in the last step input into the trained ML system, and 1578 aligned images in dataset-1 or 150 aligned images in dataset-2 were used for every specimen. After CNN calculation, the probability of each aligned image belonging to all known species was determined (dataset-1: three species, dataset-2: seventeen species). The probabilities of all images of each specimen made up the AI barcode (Figs.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb, \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). To visualize the result of AI barcodes, the t-distributed stochastic neighbor embedding (t-SNE) algorithm was employed to display the similarities and differences between pairs of species and specimens in the reduced-dimensionality feature space [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWe observed a noticeable variation in AI barcodes among species (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). More precisely, in dataset-1, the five replicated specimens of M.prae, M.ding, and M.hain all showed a high probability for their own species (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb), and the three known species were separated from each other (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec). Dataset-2 displayed similar trends (Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed, \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ee), except for a few specimens: specimen 5 of Carab-3, specimen 4 of Elate-2, specimen 1 of Elate-3, and specimen 5 of Scara-3 showed relatively dispersed AI barcodes, and specimen 5 of Carab-3-con showed a high probability of belonging to Carab-2 species (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed, Fig. S2). In the t-SNE plot, the 17 known species were still separated from each other (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ee). Although several specimens showed a range of variations, the feature space of the species, including all specimens, still showed clear separation.\u003c/p\u003e \u003cp\u003eIn all angles of the AI barcodes of dataset-1, the top continuous angles were ID 185 to 224 which mainly contain the elytron and pygidium, and the latter was considered as the key characters to distinguish the species within the genus [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.5 | \u003cem\u003eTaichi\u003c/em\u003e Step 3: Unknown species detection and status confirmation\u003c/h2\u003e \u003cp\u003eIn Step 3, image stacks of unknown species (new species or new record species and known contrast species) were put into the trained ML system to obtain their AI barcodes. After comparing the resulting AI barcodes with those of known species determined in Step 2, we found that different from the AI barcodes of known species that concentrate high probability in one species, the new species or new record species usually showed a relatively dispersed distribution pattern. In the t-SNE plots, new species or new record species were clearly separated from all known species, while known species overlapped with their own species. In this way, the unknown species could be detected.\u003c/p\u003e \u003cp\u003eSpecifically, M.prae-con in dataset-1 showed an AI barcode spectrum similar to M.prae, and M.sp.n showed relatively inconsistent patterns (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). In the t-SNE plot, M.prae and M.prae-con almost overlapped with each other, while \u003cem\u003eM\u003c/em\u003e.sp.n was represented by the largest circle, which was very different from the small and concentrated circle of other known species (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec). The same pattern appeared in dataset-2: new record species (Hydro-1-n.r., Hybos-1-n.r., Scara-5-n.r. and Scara-6-n.r.) all displayed dispersed AI barcodes, but each species had its own distribution pattern. In the t-SNE plot, four new record species were separated from all known species, but showed different distribution patterns: Scara-5-n.r. and Scara-6-n.r. were located close to the known species Scara-2; Hybos-1-n.r. and Hydro-1-n.r were located close to the known species Scara-1 and/or Morde-1 (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ee). At the same time, four contrast known species (Scara-3-con, Teneb-1-con, Carab-4-con, and Carab-3-con) all showed high probability with their own species and overlapped with their corresponding species in the t-SNE plot. The results of dataset-2 showed that the new record species have AI barcodes that differ not only from those of all known species but also from those of each other.\u003c/p\u003e \u003cp\u003eTo quantify the differences among specimens and species, we analyzed AI barcodes using PERMANOVA with 1, 000 permutations [\u003cspan additionalcitationids=\"CR44\" citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e–\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. The adjusted \u003cem\u003ep\u003c/em\u003e values are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea. Overall, \u003cem\u003ep\u003c/em\u003e.adj = 0.05 could be considered as the significance threshold to distinguish species groups. All known species and new species/new record species were significantly separated from each other, and the contrast known species were correctly grouped with their corresponding species.\u003c/p\u003e \u003cp\u003eFor dataset-1, M.sp.n group was significantly separated (\u003cem\u003ep\u003c/em\u003e.adj \u0026lt; 0.05) from all known species of \u003cem\u003eMelanopopillia\u003c/em\u003e (M.prae, M.ding, and M.hain), and M.prae-con showed no difference (\u003cem\u003ep\u003c/em\u003e.adj = 0.15) from M.prae (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea). For dataset-2, all known species were separated (\u003cem\u003ep\u003c/em\u003e.adj \u0026lt; 0.05) from each other, and the “new” species Hydro-1, Hybos-1, Scara-5, and Scara-6 were separated (\u003cem\u003ep\u003c/em\u003e.adj \u0026lt; 0.05) from all known species and from each other. The contrast groups Carab-3-con, Carab-4-con, Scara-3-con, and Teneb-1-con showed no difference (\u003cem\u003ep\u003c/em\u003e.adj = 0.671, \u003cem\u003ep\u003c/em\u003e.adj = 0.929, \u003cem\u003ep\u003c/em\u003e.adj = 0.087, \u003cem\u003ep\u003c/em\u003e.adj = 0.218) from their corresponding species (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.6 | Confirmation of new species or new records based on traditional examination\u003c/h2\u003e \u003cp\u003eThe new inferred species in dataset-1 were later studied by morphological comparison and DNA barcoding methods. The results of morphological comparison suggested that the new species is indeed similar to its known congeners but differs in the punctation of the pronotum, elytra striae, and basic shape of the aedeagus in males. Among the three known species, M.hain shared the most characters with this new species. The results of DNA barcoding analysis also support this inferred new species (Text S1, Table S3, Fig. S3). The description of this new species and the revision of the genus \u003cem\u003eMelanopopillia\u003c/em\u003e will be published in the future.\u003c/p\u003e \u003cp\u003eIn dataset-2, four new record species belonged to the same order (Hydro-1-n.r), superfamily, subfamily, and genus as known species; therefore, they probably have different morphological distances from all known species. Specifically, Hybos-1-n.r belongs to the superfamily Scarabaeoidea, as do Scara-1, Scara-2, Scara-3, and Scara-4; Scara-6-n.r belongs to the subfamily Rutelinae, as do Scara-3; and Scara-5-n.r belongs to the genus \u003cem\u003eSophrops\u003c/em\u003e, as do Scara-2. The true classification status and similarities of these new species and records are also reflected in the AI barcode results to a certain extent, such as M.sp.n being close to M.hain, and Scara-5-n.r. being close to Scara-2.\u003c/p\u003e \u003c/div\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eHere, we establish the feasibility of developing a CNN-based biodiversity discovery workflow (\u003cem\u003eTaichi\u003c/em\u003e) that is capable of detecting potential new species in a highly morphologically conserved genus and detecting potential new records of species from a certain region. The core of this workflow is to obtain an aligned image stack through the newly designed photography devices, acquire AI barcodes, and qualitatively evaluate group differences by comparing AI barcodes.\u003c/p\u003e\u003cp\u003eThe new photography device developed here can automatically and rapidly obtain a series of multiangle aligned images from stereoscopic specimens. The aligned image stack makes it possible to compare different specimens in more available morphological information. In previous studies, species identification systems were normally trained using images taken in dorsal view or at random angles [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. While it is not suitable to the stereoscopic species, for example insects, whose diagnostic characters distribute in muti-angles. From the multiangle identification results we found that different parts of the specimen have different values for ML system which similar as the taxonomic view. In some degree, it visualizes the identification process of deep learning system. As shown here, using multiangle images can increase the power of the training in CNNs and make it possible to infer species by more characters.\u003c/p\u003e\u003cp\u003eThe analysis result indicates that AI barcodes can be used to measure similarity between specimens and species based on photos taken at different angles, especially between known species and 'new' species. And AI barcodes contain diagnostic information that could identify the species. After the analysis of AI barcodes, the \u003cem\u003eTaichi\u003c/em\u003e workflow has the ability to detect new species/new record species and has good performance in detecting more than one new record species. This workflow can be expanded to a variety of scenarios in which detection of the 'new' or 'extra' taxa is needed. Furthermore, adding the 'new' species to the workflow makes self-renewal of the CNN system and its further application to monitoring biodiversity possible.\u003c/p\u003e\u003cp\u003eThe ML system in the \u003cem\u003eTaichi\u003c/em\u003e workflow is similar to the other image-based AI identification systems in terms of autoidentification performance (top-1 accuracy greater than 95%). By analyzing all preliminary results of aligned multiangle images of the specimen yielded better performance in inferring new species, in a sense, which extracts studied knowledge within CNNs similar to vectors or embeddings. Compared to traditional identification methods, the \u003cem\u003eTaichi\u003c/em\u003e workflow is also superior in terms of efficiency and throughput. This workflow is faster than the traditional method from image acquisition to subsequent analysis and is capable of high throughput, including a large number of known species for reference.\u003c/p\u003e\u003cp\u003eIn the \u003cem\u003eTaichi\u003c/em\u003e workflow described here, two aspects should be considered with caution. First, we should be able to assign unknown specimens to the same taxonomic level, such as genus (dataset-1) or order (dataset-2). In this way, this method is similar to traditional taxonomy, where a given taxon is gradually diagnosed from higher to lower hierarchical levels (order, family, genus, and then species). In dataset-1, we trained a genus identification system to assign the new species to \u003cem\u003eMelanopopillia\u003c/em\u003e (Text S2, Table S4, Fig. S4). Therefore, the CNN model for identifying taxa at higher levels, higher taxa could be added to the \u003cem\u003eTaichi\u003c/em\u003e workflow to increase compatibility in more complex identification tasks.\u003c/p\u003e\u003cp\u003eSecond, the species should include at least four/five specimens to increase the precision of the judgment. In the AI barcode results, the dispersed distribution pattern (such as those for M.hain-3 in dataset-1 and EI.Me-1 in dataset-2, Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea, \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed) or the high probability of incorrect species assignment (Carab-3-con-5 in dataset-2, Fig. S2, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed) indicates that variation within species always exists. A similar situation also exists in traditional taxonomy in which the confirmation of new species should rely on a series of specimens, or when conducting DNA barcoding, it is desirable to obtain DNA sequence data from multiple specimens representing a hypothetical species entity.\u003c/p\u003e\u003cp\u003eWe have demonstrated the potential of the \u003cem\u003eTaichi\u003c/em\u003e workflow to accelerate the discovery of new species/new record species by creating a single, reproducible repository for quantifying the ranges of morphological variation exhibited by known species, and this will reduce the workload of taxonomists in routine identifications for biodiversity assessment. In the past 30 years, the digitization of museum collections has expanded rapidly, and large numbers of specimen images have been generated [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e], which can be used by ML technology in the future. Reciprocally, ML can also drive the development and data mining from digitization of natural history collections [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e] and thus has considerable potential to advance the field of biodiversity science.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eCONFLICT OF INTEREST\u003c/p\u003e\n\u003cp\u003eThe authors declare no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;AUTHOR CONTRIBUTIONS\u003c/p\u003e\n\u003cp\u003eY.Y. L. and M. B conceived the ideas, designed the methodology, and\u0026nbsp;led the writing of the manuscript.\u0026nbsp;J. L. and\u0026nbsp;Z. Y. Z.\u0026nbsp;trained and wrote the CNN parts. Y.C. Z. and Z.Y. Z. performed the bioinformatic analysis and wrote the PERMANOVA parts. Y.Y. L., Z.Y. Z., Y.J. T., B. T., N. L., and J.J. S. collected and interpreted the data. All authors contributed critically to the drafts and gave final approval for publication.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;DATA AVAILABILITY STATEMENT\u003c/p\u003e\n\u003cp\u003eThe code can be found on GitHub (https://github.com/LSDXBZHH/detecting-new-species). The data set\u0026nbsp;links can be found in Science Data Bank (https://www.scidb.cn/s/2Mjuaq).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;FUNDING\u003c/p\u003e\n\u003cp\u003eThis research was supported by the National Natural Science Foundation of China (Nos. 32270468, 31961143002,\u0026nbsp;32200354),\u0026nbsp;National Key R\u0026amp;D Program of China\u0026nbsp;(No.\u0026nbsp;2022YFC2601200), National Science \u0026amp; Technology Fundamental Resources Investigation Program of China (No. 2022FY100500) and the Bureau of International Cooperation, Chinese Academy of Sciences.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;ACKNOWLEDGMENTS\u003c/p\u003e\n\u003cp\u003eWe thank Ping Yang from Institute of Zoology, Guangdong Academy of Sciences, IZGAS and Mingzhi Zhao from South China Agricultural University for loaning specimens of \u003cem\u003eMelanopopillia\u003c/em\u003e to study. The authors thank Norman MacLeod from Nanjing University for his advice in manuscript writing. We thank Xinhai Li and Zhengting Zou from the Institute of Zoology, CAS (IZCAS), Roberta Eleanor Hunt from University of Copenhagen for useful advice in biostatistical analysis. The authors thank Zhonglin Xu, Xin Chen, Jun Wang from Cangzhou Normal University, Jiajia Zhai, and Zhibin Sun from National Space Science Center, CAS for designing and processing the high-throughput multiangle photography device, Rongrong Shen from China Agricultural University, Yingying Yu, Xiaoxuan Li, and Huanxi Cao from IZCAS for the useful advice and help in DNA parts. We also thank Hongbin Liang (IZCAS), Yang Liu (Northwest University), and Chuanbu Gao (Institute of Zoology, Guangdong Academy of Sciences) for identifying the Hong Kong species.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCardinale BJ, Duffy JE, Gonzalez A, Hooper DU, Perrings C, Venail P, et al. Biodiversity loss and its impact on humanity. Nature. 2012;486:59\u0026ndash;67.\u003c/li\u003e\n\u003cli\u003eCostello MJ, May RM, Stork NE. Can We Name Earth\u0026rsquo;s Species Before They Go Extinct? Science. 2013;339:413\u0026ndash;6.\u003c/li\u003e\n\u003cli\u003eSandifer PA, Sutton-Grier AE, Ward BP. Exploring connections among nature, biodiversity, ecosystem services, and human health and well-being: Opportunities to enhance health and biodiversity conservation. Iss Environ Sci Tech. 2015;12:1\u0026ndash;15.\u003c/li\u003e\n\u003cli\u003eDeans AR, Yoder MJ, Balhoff JP. Time to change how we describe biodiversity. Trends Ecol Evol. 2012;27:78\u0026ndash;84.\u003c/li\u003e\n\u003cli\u003eStork NE. How Many Species of Insects and Other Terrestrial Arthropods Are There on Earth? Annu Rev Entomol. 2018;63:31\u0026ndash;45.\u003c/li\u003e\n\u003cli\u003eWheeler QD, Raven PH, Wilson EO. Taxonomy: Impediment or Expedient? Science. 2004;303:285\u0026ndash;285.\u003c/li\u003e\n\u003cli\u003eEbach MC, Valdecasas AG, Wheeler QD. Impediments to taxonomy and users of taxonomy: accessibility and impact evaluation. Cladistics. 2011;27:550\u0026ndash;7.\u003c/li\u003e\n\u003cli\u003eOrr MC, Ferrari RR, Hughes AC, Chen J, Ascher JS, Yan Y-H, et al. Taxonomy must engage with new technologies and evolve to face future challenges. Nat Ecol Evol. 2020;5:3\u0026ndash;4.\u003c/li\u003e\n\u003cli\u003eMacLeod N, Benfield M, Culverhouse P. Time to automate identification. Nature. 2010;467:154\u0026ndash;5.\u003c/li\u003e\n\u003cli\u003eBisgin H, Bera T, Ding H, Semey HG, Wu L, Liu Z, et al. Comparing SVM and ANN based Machine Learning Methods for Species Identification of Food Contaminating Beetles. Sci Rep. 2018;8:6532.\u003c/li\u003e\n\u003cli\u003eSong Y, He F, Zhang X. To Identify Tree Species With Highly Similar Leaves Based on a Novel Attention Mechanism for CNN. IEEE Access. 2019;7:163277\u0026ndash;86.\u003c/li\u003e\n\u003cli\u003eWu L, Liu Z, Bera T, Ding H, Langley DA, Jenkins-Barnes A, et al. A deep learning model to recognize food contaminating beetle species based on elytra fragments. Comput Electron Agric. 2019;166:105002.\u003c/li\u003e\n\u003cli\u003eSeeland M, Rzanny M, Boho D, W\u0026auml;ldchen J, M\u0026auml;der P. Image-based classification of plant genus and family for trained and untrained plant species. BMC Bioinf. 2019;20:4.\u003c/li\u003e\n\u003cli\u003eFedor P, Vaňhara J, Havel J, Malenovsk\u0026yacute; I, Spellerberg I. Artificial intelligence in pest insect monitoring. Syst Entomol. 2009;34:398\u0026ndash;400.\u003c/li\u003e\n\u003cli\u003eLeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436\u0026ndash;44.\u003c/li\u003e\n\u003cli\u003eNiemi J, Tanttu J. Deep Learning Case Study for Automatic Bird Identification. Appl Sci. 2018;8:2089.\u003c/li\u003e\n\u003cli\u003eValan M, Makonyi K, Maki A, Vondr\u0026aacute;ček D, Ronquist F. Automated Taxonomic Identification of Insects with Expert-Level Accuracy Using Effective Feature Transfer from Convolutional Networks. Buckley T, editor. Syst Biol. 2019;68:876\u0026ndash;95.\u003c/li\u003e\n\u003cli\u003eLu W, Chen X, Wang L, Li H, Fu YV. Combination of an Artificial Intelligence Approach and Laser Tweezers Raman Spectroscopy for Microbial Identification. Anal Chem. 2020;92:6288\u0026ndash;96.\u003c/li\u003e\n\u003cli\u003eFerreira AC, Silva LR, Renna F, Brandl HB, Renoult JP, Farine DR, et al. Deep learning‐based methods for individual recognition in small birds. Codling E, editor. Methods Ecol Evol. 2020;11:1072\u0026ndash;85.\u003c/li\u003e\n\u003cli\u003eCarranza-Rojas J, Goeau H, Bonnet P, Mata-Montero E, Joly A. Going deeper in the automated identification of Herbarium specimens. BMC Evol Biol. 2017;17:181.\u003c/li\u003e\n\u003cli\u003eW\u0026auml;ldchen J, Rzanny M, Seeland M, M\u0026auml;der P. Automated plant species identification\u0026mdash;Trends and future directions. Bucksch A, editor. PLOS Comput Biol. 2018;14:e1005993.\u003c/li\u003e\n\u003cli\u003eLiu L, Wang R, Xie C, Yang P, Wang F, Sudirman S, et al. PestNet: An End-to-End Deep Learning Approach for Large-Scale Multi-Class Pest Detection and Classification. IEEE Access. 2019;7:45301\u0026ndash;12.\u003c/li\u003e\n\u003cli\u003eMitra R, Marchitto TM, Ge Q, Zhong B, Kanakiya B, Cook MS, et al. Automated species-level identification of planktic foraminifera using convolutional neural networks, with comparison to human performance. Mar Micropaleontol. 2019;147:16\u0026ndash;24.\u003c/li\u003e\n\u003cli\u003eThenmozhi K, Srinivasulu Reddy U. Crop pest classification based on deep convolutional neural network and transfer learning. Comput Electron Agric. 2019;164:104906.\u003c/li\u003e\n\u003cli\u003eMolnar C, Casalicchio G, Bischl B. Interpretable machine learning\u0026ndash;a brief history, state-of-the-art and challenges. Joint European conference on machine learning and knowledge discovery in databases. Springer; 2020. p. 417\u0026ndash;31.\u003c/li\u003e\n\u003cli\u003eChandola V, Banerjee A, Kumar V. Anomaly Detection: A Survey. Acm Comput Surv. 2009;41:1\u0026ndash;58.\u003c/li\u003e\n\u003cli\u003eZhao Z, Lu Y, Tong Y, Chen X, Bai M. PENet : A phenotype encoding network for automatic extraction and representation of morphological discriminative features. Methods Ecol Evol. 2023;2041-210X.14235.\u003c/li\u003e\n\u003cli\u003ePastore VP, Zimmerman TG, Biswas SK, Bianco S. Annotation-free learning of plankton for classification and anomaly detection. Sci Rep. 2020;10:12142.\u003c/li\u003e\n\u003cli\u003eKhalighifar A, Brown RM, Goyes Vallejos J, Peterson AT. Deep learning improves acoustic biodiversity monitoring and new candidate forest frog species identification (genus Platymantis) in the Philippines. Biodivers Conserv. 2021;30:643\u0026ndash;57.\u003c/li\u003e\n\u003cli\u003eKhalighifar A, Jim\u0026eacute;nez-Garc\u0026iacute;a D, Campbell LP, Ahadji-Dabla KM, Aboagye-Antwi F, Ibarra-Ju\u0026aacute;rez LA, et al. Application of Deep Learning to Community-Science-Based Mosquito Monitoring and Detection of Novel Species. Yee D, editor. J Med Entomol. 2022;59:355\u0026ndash;62.\u003c/li\u003e\n\u003cli\u003eBadirli S, Picard CJ, Mohler G, Richert F, Akata Z, Dundar M. Classifying the unknown: Insect identification with deep hierarchical Bayesian learning. Methods Ecol Evol. 2023;14:1515\u0026ndash;30.\u003c/li\u003e\n\u003cli\u003eDubois A. Describing a New Species. TAPROBANICA: J Asian Biodivers. 2011;2:6.\u003c/li\u003e\n\u003cli\u003eMiralles A, Bruy T, Wolcott K, Scherz MD, Begerow D, Beszteri B, et al. Repositories for Taxonomic Data: Where We Are and What is Missing. Friedman M, editor. Syst Biol. 2020;69:1231\u0026ndash;53.\u003c/li\u003e\n\u003cli\u003eLin P. A new genus, Melanopopillia, from China (Coleoptera: Rutelidae). Entomotaxonomia. 1980;2:297\u0026ndash;301.\u003c/li\u003e\n\u003cli\u003eLu Y, Yang H, Bai M. Micro CT approach applied in taxonomy: An example on the species Melanopopillia hainanensis (Coleoptera: Scarabaeidae). Zoological Systematics. 2019;44:294\u0026ndash;303.\u003c/li\u003e\n\u003cli\u003eZhao S, Tong Y, Teng B, Chen X, Yang X, Li J, et al. A species diversity dataset of beetles by three passive acquisition methods in Tei Tong Tsai (Hong Kong). Sci Data. 2022;9:210.\u003c/li\u003e\n\u003cli\u003eKrizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems. 2012;25:1097\u0026ndash;105.\u003c/li\u003e\n\u003cli\u003eHe K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. Las Vegas, NV, USA: IEEE; 2016 [cited 2023 Nov 25]. p. 770\u0026ndash;8. Available from: http://ieeexplore.ieee.org/document/7780459/\u003c/li\u003e\n\u003cli\u003eHuang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely Connected Convolutional Networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. Honolulu, HI: IEEE; 2017 [cited 2023 Nov 25]. p. 2261\u0026ndash;9. Available from: https://ieeexplore.ieee.org/document/8099726/\u003c/li\u003e\n\u003cli\u003eSandler M, Howard A, Zhu M, Zhmoginov A, Chen L-C. MobileNetV2: Inverted Residuals and Linear Bottlenecks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition [Internet]. Salt Lake City, UT: IEEE; 2018 [cited 2023 Nov 25]. p. 4510\u0026ndash;20. Available from: https://ieeexplore.ieee.org/document/8578572/\u003c/li\u003e\n\u003cli\u003eLaurens Van Der Maaten, Hinton G. Visualizing data using t-SNE. J Mach Learn Res. 2008;9:2579\u0026ndash;605.\u003c/li\u003e\n\u003cli\u003eAnowar F, Sadaoui S, Selim B. Conceptual and empirical comparison of dimensionality reduction algorithms (PCA, KPCA, LDA, MDS, SVD, LLE, ISOMAP, LE, ICA, t-SNE). Comput Sci Rev. 2021;40:100378.\u003c/li\u003e\n\u003cli\u003eAnderson MJ. A new method for non‐parametric multivariate analysis of variance. Austral Ecol. 2001;26:32\u0026ndash;46.\u003c/li\u003e\n\u003cli\u003eAnderson MJ, Walsh DCI. PERMANOVA, ANOSIM, and the Mantel test in the face of heterogeneous dispersions: What null hypothesis are you testing? Ecol Monogr. 2013;83:557\u0026ndash;74.\u003c/li\u003e\n\u003cli\u003eKelly BJ, Gross R, Bittinger K, Sherrill-Mix S, Lewis JD, Collman RG, et al. Power and sample-size estimation for microbiome studies using pairwise distances and PERMANOVA. Bioinformatics. 2015;31:2461\u0026ndash;8.\u003c/li\u003e\n\u003cli\u003eZoph B, Vasudevan V, Shlens J, Le QV. Learning Transferable Architectures for Scalable Image Recognition. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition [Internet]. Salt Lake City, UT: IEEE; 2018 [cited 2023 Nov 25]. p. 8697\u0026ndash;710. Available from: https://ieeexplore.ieee.org/document/8579005/\u003c/li\u003e\n\u003cli\u003eRawat W, Wang Z. Deep Convolutional Neural Networks for Image Classification: A Comprehensive Review. Neural Comput. 2017;29:2352\u0026ndash;449.\u003c/li\u003e\n\u003cli\u003eNelson G, Ellis S. The history and impact of digitization and digital data mobilization on biodiversity research. Philos Trans R Soc B: Biol Sci. 2019;374:20170391.\u003c/li\u003e\n\u003cli\u003eHedrick BP, Heberling JM, Meineke EK, Turner KG, Grassa CJ, Park DS, et al. Digitization and the Future of Natural History Collections. Bioscience. 2020;70:243\u0026ndash;51.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Video S1","content":"\u003cp\u003eVideo S1 is not available with this version\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Unknown species, Convolution neural networks, Identification, PERMANOVA","lastPublishedDoi":"10.21203/rs.3.rs-3832815/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3832815/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eRecently, machine learning (ML) has been widely used in species auto-identification systems for multi-scene applications in biodiversity, while most of the existing ML systems relying on images are limited to identifying the species on which they are trained, and unknown species out of the system are normally incorrectly identified.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eHere, we propose a new workflow system based on the ML system and PERMANOVA analysis, named \u003cem\u003eTaichi\u003c/em\u003e for the detection and confirmation of unknown species status, stimulated by the traditional biodiversity discovery process. First, we developed a series of high-throughput photography devices that could efficiently obtain aligned multiangle images for ML system training. Then the new analysis workflow was integrated in Python codes based on the convolutional neural networks: MobileNetV2 (selected from four available networks) and further PERMANOVA analysis of euclidean distances to detect 'new' species. Two newly established beetle datasets: \u003cem\u003eMelanopopillia\u003c/em\u003e (4 species, 55 specimens, 42 450 images) and \u003cem\u003eHong Kong beetles\u003c/em\u003e (21 species, 206 specimens, 35 450 images) were constructed in this work to demonstrate the \u003cem\u003eTaichi\u003c/em\u003e system. The diagnostic information for species is generated from the output results of the ML system (top-1 accuracy rate reached 97.76% and 96.22% respectively) naming as the artificial intelligence (AI) barcode, and extra analysis of AI barcodes could help visualization the diagnostic characters from multiangle images.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eBy comparing AI barcodes of different specimens outside the system, it is proved that the \u003cem\u003eTaichi\u003c/em\u003e workflow can rapidly detect whether the input sample is a known species or possible 'new' species. This workflow provides a newly possible self-extensibility solution for the acceleration of biodiversity discovery. Additionally, it can also drive the development of data mining from the digitization of natural history collections around the world and has great potential to advance the field of biodiversity science.\u003c/p\u003e","manuscriptTitle":"Accelerating the discovery of biodiversity by detecting “new” species based on machine learning method","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-01-09 19:16:43","doi":"10.21203/rs.3.rs-3832815/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"9112ed05-e888-476e-81cd-e1f3979c40e2","owner":[],"postedDate":"January 9th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-02-14T01:30:14+00:00","versionOfRecord":[],"versionCreatedAt":"2024-01-09 19:16:43","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-3832815","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3832815","identity":"rs-3832815","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}