An Automatic Method for Constructing Insect Phylogenetic Trees Based on Multimodal Large Model Morphological Representation Learning
By using multimodal large-scale collaborative modeling and generating continuous morphological feature vectors with image and text encoders, the problems of single modality and manual encoding in traditional methods are solved, and the automation and stability of insect phylogenetic trees are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-04-30
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies rely on single-modal data or manual coding when constructing phylogenetic trees, which is time-consuming and labor-intensive and makes it difficult to fully reflect complex morphological differences. Furthermore, there is a lack of application of multimodal large models in biological morphological representation learning and phylogenetic inference.
A multimodal large model is adopted, which generates continuous morphological feature vectors through collaborative modeling of image encoder and text encoder. Combined with supervised contrastive learning and image-text alignment optimization, it realizes the collaborative representation of image and text information, and is finally used for phylogenetic inference.
It improves the automation and generalization ability of morphological tree construction, enhances the integrity and semantic consistency of feature representation, can be directly used for phylogenetic inference, reduces the workload of manual coding, and improves the objectivity and stability of phylogenetic tree construction.
Smart Images
Figure CN122493490A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of bioinformatics, computational biology, computer vision, and artificial intelligence. Specifically, it relates to an automatic construction method for insect phylogenetic trees based on multimodal large model morphological representation learning, and more particularly to a method that uses an image encoder and a text encoder to collaboratively model biological morphological information and uses the learned continuous morphological features for phylogenetic inference. Background Technology
[0002] Phylogenetic tree construction is one of the fundamental tasks in phylogenetic research, playing a crucial role in revealing the kinship, evolutionary paths, and classification systems among biological groups. Traditional phylogenetic tree construction methods mainly rely on molecular sequence data, manual morphological feature encoding, or expert judgment. Although molecular data has high resolution in many studies, it is often insufficient to meet practical needs when ancient fossil groups, specimen data are missing, or molecular information is unavailable. Meanwhile, traditional morphological tree construction methods typically rely on manual extraction and discrete encoding of morphological traits. This approach is not only time-consuming and labor-intensive but also easily influenced by subjective experience, making it difficult to comprehensively reflect complex morphological differences.
[0003] With the development of machine learning and deep learning technologies, researchers have begun to explore using image features to replace manual morphological coding in order to improve the efficiency and objectivity of feature extraction. However, most existing deep learning-based morphological feature tree construction methods are still based on single-modal image features and fail to fully integrate the semantic information contained in textual morphological descriptions. Therefore, they still have limitations in terms of morphological interpretability, cross-class generalization ability, and feature semantic consistency. On the other hand, although multimodal image-text alignment technology has achieved good results in general visual semantic tasks, its application in biological morphological representation learning and phylogenetic inference is still limited, especially lacking a systematic approach to directly use the continuous representations learned by large multimodal models for phylogenetic tree construction.
[0004] Furthermore, existing encoding methods for morphological data largely rely on discrete trait matrices, which struggle to express the continuity and complex correlations of morphological changes. While phylogenetic inference based on continuous morphological features offers certain advantages, automatically learning continuous traits suitable for tree construction from raw images and text remains a pressing issue in this field. Therefore, there is an urgent need to propose a phylogenetic tree construction method that can jointly utilize both image and text modalities to automatically learn continuous biological morphological representations and further apply them to phylogenetic inference, thereby improving the automation, generalization ability, and interpretability of morphological tree construction. Summary of the Invention
[0005] This invention aims to overcome the problems of high manual coding costs, reliance on single data modalities for morphological representation, insufficient continuous morphological feature learning ability, and limited automation in existing morphological tree construction methods. It proposes an automatic construction method for insect phylogenetic trees based on multimodal large model morphological representation learning.
[0006] This invention constructs an image-text pairing sample set of images and morphological description text. Based on the image encoder and text encoder in a pre-trained multimodal large-scale model, visual and morphological semantic features are extracted respectively. Category-level image-text alignment is performed in a shared embedding space. Through supervised contrastive learning and a joint optimization strategy of image-text alignment, a continuous morphological feature vector representing the morphological differences of the target insect group is generated. This continuous morphological feature vector is further converted into a continuous trait data format required for phylogenetic inference and input into a Bayesian phylogenetic inference process based on a continuous trait evolution model. This achieves joint representation learning and phylogenetic inference from multimodal morphological information to phylogenetic topology. This method can automatically construct a phylogenetic tree without relying on molecular sequence data, fully utilizing the complementary information in images and text, and is applicable to insect groups at different taxonomic levels.
[0007] This invention relates to an automatic method for constructing insect phylogenetic trees based on multimodal large model morphological representation learning, comprising: S1. Acquire image data of the target insect group and morphological description text data corresponding to the target insect group, standardize the image data and morphological description text data, and construct an image-text pairing sample set; S2. A multimodal morphological representation learning framework is constructed based on the image data and morphological description text data. The multimodal morphological representation learning framework includes an image encoder, a text encoder, an image projector, a text projector, and a shared embedding space. The image encoder encodes the feature sequence of the input image and obtains a visual representation vector through global feature extraction. The visual representation vector represents the overall morphological information of the target object. The text encoder extracts morphological semantic representations. The image projector and text projector map the image representation and text representation to the shared embedding space, respectively. The multimodal morphological representation learning framework is jointly optimized based on supervised contrast loss and image-text alignment loss to generate continuous morphological feature vectors that can represent the morphological differences of the target insect group. S3. The continuous morphological feature vectors are converted into a continuous trait data matrix required for phylogenetic inference. A Bayesian phylogenetic inference process is constructed based on the continuous trait evolution model. The candidate set of phylogenetic trees and their corresponding posterior probability distributions are obtained through Markov chain Monte Carlo sampling. S4. Based on the candidate phylogenetic tree set and its posterior probability distribution, statistically summarize the phylogenetic trees to generate the optimal phylogenetic tree. The optimal phylogenetic tree is then compared and evaluated based on the topological difference index between the reference phylogenetic tree and the optimal phylogenetic tree.
[0008] In step S1, the morphological description text data is automatically generated from the image data using a pre-trained text generation model based on preset rules. The initial morphological description text is then formed by combining manual verification and expert review to create a structured text library. The entries in the structured text library are associated one-to-one with the image data of the corresponding category, and the text entries contain descriptive information of the visible morphological parts of that category.
[0009] The morphological description text data includes seven morphological elements: color, pronotum, elytra, legs, head and neck, eyes, and antennae. It is organized using a standardized description template to form category-level text input.
[0010] The processing in step S1 includes performing size unification, brightness normalization, color normalization, and data augmentation on the image data; wherein, the data augmentation includes random flipping, rotation, translation, and scaling; and performing word segmentation, terminology normalization, noise reduction, and text length normalization on the morphological description text data.
[0011] The multimodal morphological representation learning framework described in step S2 includes: fine-tuning and updating relevant parameters such as attention in the image encoder to enhance the adaptability of the target task while retaining the pre-trained representation capability; inputting multiple morphological description information of the same category into the text encoder after structural integration, extracting category-level text semantic features, and constructing a category-level text feature library; wherein the image projection head and the text projection head project the image features and text features to a unified low-dimensional representation space, respectively.
[0012] In step S2, the supervised contrastive loss is used to minimize the feature distance between samples of the same class in the shared embedding space and to maximize the feature distance between samples of different classes; the image-text alignment loss is used to minimize the semantic distance between image features of the same class and text features of the same class; a weighted joint objective function is constructed by supervised contrastive loss and image-text alignment loss to jointly optimize the multimodal morphological representation learning framework to generate continuous morphological feature vectors for phylogenetic analysis.
[0013] The joint optimization of the supervised contrast loss and image-text alignment loss in step S2 is expressed in the following form: The overall loss function is: (1) in For supervised comparative loss, For image-text alignment loss, These are the weighting coefficients; The supervised comparison loss Defined as: (2) Where zi represents the normalized embedding vector of the i-th sample, τ is the temperature coefficient, and P(i) represents the set of positive samples of the same class as sample i. The image-text alignment loss Defined as: (3) in, Indicates the total number of categories. The mean value of the features of image class C is used to represent the feature value of image class C. This represents the text feature representation of the corresponding category.
[0014] The continuous morphological feature vector is a global visual representation vector output by the image encoder after feature encoding of the input image. The global visual representation vector corresponds to the overall morphological information of the target insect and is used to construct a continuous trait matrix.
[0015] The Bayesian phylogenetic inference described in step S3 is based on the Brownian motion model of continuous traits, and a set of candidate phylogenetic trees is obtained through tree topology search, branch length optimization, and parameter posterior sampling.
[0016] The topological difference index in step S4 includes a distance metric for measuring the topological structural differences between phylogenetic trees and a similarity evaluation index for assessing the topological consistency between phylogenetic trees; the distance metric is used to characterize the structural differences between the inferred phylogenetic tree and the reference phylogenetic tree, and the similarity index is used to characterize the degree of consistency between the two phylogenetic trees in their splitting structure.
[0017] The target insect group is classified at the family, subfamily, genus, species, or other taxonomic levels, and the method is applicable to any insect taxonomic unit that has image data and morphological description text data.
[0018] The working principle of this invention is: This invention first acquires image data and morphological description text data for the target insect group, and then standardizes the data to form an image-text pairing sample set with category labels.
[0019] Subsequently, the image encoder in the pre-trained multimodal large model is used to model the features of the image, extracting high-dimensional visual feature representations and generating global semantic representation vectors. Simultaneously, the morphological description text is encoded using a text encoder to extract structured morphological semantic features, and a category-level text feature library is constructed according to categories to ensure that text semantic information of the same category has a consistent structural expression.
[0020] In the multimodal representation learning stage, image features and category-level text features are mapped to a shared embedding space through their respective projection modules. Supervised contrastive learning and image-text alignment are used to jointly optimize the embedding space, minimizing the distance between samples of the same category and maximizing the distance between samples of different categories. At the same time, the semantic distance between image features and corresponding category text features is constrained to be minimized, thereby generating a continuous morphological feature vector that can represent the morphological differences of the target insect group.
[0021] In the phylogenetic inference stage, the continuous morphological feature vectors are extracted and converted into the continuous trait data format required for phylogenetic inference, a continuous trait matrix is constructed, and after being organized according to the standard format, it is input into the Bayesian phylogenetic inference model. Based on the continuous trait evolution process, the phylogenetic tree topology is subjected to posterior sampling and optimization search to obtain a set of candidate phylogenetic trees and generate the optimal phylogenetic tree.
[0022] Finally, the consistency between the inferred phylogenetic tree and the reference phylogenetic tree is quantitatively analyzed by using the topological difference index, thereby completing the automatic construction and evaluation of the phylogenetic tree.
[0023] The innovation of this invention is: This invention proposes a morphological representation learning framework based on a multimodal large model. This framework utilizes an image encoder to extract visual morphological information of individual insects, while simultaneously introducing a text encoder to semantically encode category-level morphological descriptions, achieving collaborative modeling of image and text information and overcoming the problem of insufficient information in single-modal morphological modeling.
[0024] This invention constructs a category-level text feature library and introduces it as a stable semantic anchor into the image-text alignment process, so that image features of the same category correspond to a unified semantic representation in a shared embedding space, thereby enhancing the category consistency and morphological interpretability of feature learning.
[0025] This invention employs a joint optimization mechanism of supervised contrast loss and image-text alignment loss, which enhances inter-class separability while improving the responsiveness of image features to text semantics, enabling the model to generate continuous morphological feature vectors that are more suitable for phylogenetic analysis.
[0026] This invention directly uses the continuous morphological features learned from multimodal large model learning to construct phylogenetic trees, realizing the connection from morphological representation learning to continuous trait phylogenetic inference, and breaking through the limitations of traditional discrete trait coding or single image feature tree construction methods.
[0027] The advantages of this invention are: This invention can effectively reduce the workload of manual coding in traditional morphological tree construction and improve the automation of morphological feature extraction; by fusing image data and morphological description text data for multimodal joint modeling, it improves the integrity and semantic consistency of morphological feature representation, enabling the model to characterize more fine-grained group differences; the learned continuous morphological features are both interpretable and computable, and can be directly used for continuous trait phylogenetic inference, improving the objectivity and stability of phylogenetic tree construction.
[0028] This invention is applicable to insect groups at different taxonomic levels, exhibiting good generalization and expansion capabilities at the genus, species, and higher taxonomic levels. Furthermore, based on multimodal large model construction, this method can adapt to data from different sources, under different imaging conditions, and with different descriptive styles, providing reliable technical support for insect taxonomy, phylogenetic analysis, and biodiversity research. Attached Figure Description
[0029] Figure 1 This is a flowchart of the method of the present invention.
[0030] Figure 2 This is a schematic diagram of the overall architecture of the multimodal large model of the present invention.
[0031] Figure 3 This is a flowchart of the phylogenetic tree construction process of the present invention. Detailed Implementation
[0032] The present invention will be further described below with reference to specific embodiments. Those skilled in the art will be able to implement the present invention based on these descriptions. Furthermore, the embodiments of the present invention described below are generally only some, not all, of the embodiments of the present invention. Therefore, all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.
[0033] This embodiment relates to an automatic construction method for insect phylogenetic trees based on multimodal large model morphological representation learning, such as... Figure 1 It includes the following steps: S1: Image Data Acquisition and Preprocessing First, standardized beetle image data was acquired during the input phase. The images used were from the Rove-Tree-11 dataset, which consists of images of the dorsal side of rove beetle pinpricks from the insect collection of the Natural History Museum of Denmark. This dataset contains 13,887 segmented images with species-level classification labels and a reference phylogenetic tree. Originally intended to support research combining deep learning and phylogenetic inference in digitized natural history specimens, this dataset covers 44 genera within three subfamilies of the family Staphylinidae. The sample size distribution is significantly unbalanced, with some genera containing only dozens of images while others contain thousands.
[0034] Since the original dataset is labeled by species, but the number of species within the same genus varies greatly, training directly at the species level can easily lead to a highly imbalanced class structure, which would affect the stability of the feature learning process and subsequent phylogenetic inference results. Therefore, this embodiment uses the genus level as a unified classification unit for multimodal representation learning and phylogenetic inference to ensure the consistency of training categories and the repeatability of comparative analysis.
[0035] To ensure input consistency and model adaptability, each image needs to undergo image preprocessing before entering the network; the data preprocessing includes the following steps: S11: Unified resolution: The original image is padded with a white background to form a square canvas, and scaled to the model input resolution of 518×518 pixels to match the input requirements of the backbone representation model; S12: Category Balancing and Data Augmentation: For genera with fewer than 50 images, geometric transformations such as rotation, translation, and flipping are used for augmentation until each genera has at least 50 independent images. Then, 50 images are randomly sampled from each genera as basic training samples, and the same augmentation operation is performed on the selected samples to build a balanced training dataset, ultimately obtaining 28,600 images, or 650 images per genera. S13: Data Partitioning: The augmented dataset is divided into training, validation, and test sets in a 7:2:1 ratio for representation learning and model evaluation. The processed images are used as network input in the form of standardized tensors.
[0036] S2: Morphological Text Construction and Multimodal Representation Learning
[0037] like Figure 2 In this invention, a multimodal large-scale model morphological representation framework based on collaborative learning of an image encoder and a text encoder is proposed. This framework maps image visual information and morphological description semantic information to a shared embedding space for learning continuous features that can characterize the morphological differences of rove beetles. The model design is as follows: S21: Generating and Organizing Morphological Description Text To incorporate morphological semantic information into image representation learning, this embodiment constructs category-level morphological description text for each genus. Specifically, for each representative image, a large language model is first used to generate an initial morphological description text. This text template contains seven morphological region fields: color, pronotum, elytra, legs, head and neck, eyes, and antennae. These seven regions correspond to key distinguishing features commonly used in rove beetle morphological classification and are used to characterize the main morphological differences between different genera.
[0038] The generated initial text is manually revised by the training personnel to correct any inaccuracies that may have occurred during the model generation process. It is then further reviewed by experts in the relevant domain to ensure that the text content is morphologically consistent with the specimen images. The manually corrected and expert-reviewed descriptive texts are organized into structured morphological description data, and a text feature library is constructed according to genus-level categories. Each genus corresponds to a set of structured morphological description texts, which serve as input for subsequent text encoding.
[0039] S22: Image Feature Extraction and Multimodal Alignment Structure
[0040] The preprocessed image is first input into the DINO-V2 image encoder for feature extraction. In this embodiment, the image backbone network adopts a pre-trained visual transformer structure to perform block embedding and multi-layer feature encoding on the input image, outputting a global visual feature vector. To enhance the model's adaptability to the target task, a LoRA low-rank adaptation module is introduced into the attention projection layer of the image backbone network, enabling the model to efficiently adapt parameters to downstream tasks using only a very small number of trainable parameters while freezing the backbone parameters.
[0041] Specifically, a LoRA structure is inserted into the linear transformation layer of the attention module to perform low-rank incremental updates on the QKV projection. Trainable low-rank matrices are introduced into the Q and V branches to enhance the modeling ability for fine-grained features of beetle morphology while maintaining the capabilities of the original large model. The output global features of the image backbone network are mapped to the shared embedding space via the image projector to obtain the image embedding vector.
[0042] S23: Category-level text feature coding
[0043] Corresponding to image feature extraction, the genus-level morphological description text is input into a text encoder for encoding to extract semantic features. In this embodiment, BioClinical-ModernBERT-base is used as the text encoding model to semantically represent the structured morphological description and output the text feature vector corresponding to each genus. Subsequently, the text feature vector is mapped to the same shared embedding space as the image features via a text projection head to form a category-level text embedding vector representation.
[0044] During training, the text feature library remains fixed and serves as a stable reference feature in the image-text alignment optimization process; the text projection head is used to map the text semantic features to a representation space consistent with the image features, so as to achieve category-level image-text feature alignment.
[0045] S24: Joint Optimization of Text and Graph and Learning of Continuous Morphological Representations
[0046] In this embodiment, image features and text features are jointly optimized and trained in a shared embedding space. To balance category discrimination capability and image-text semantic consistency, a weighted combination function of supervised contrastive loss and image-text alignment loss is used for model training.
[0047] The supervised contrastive loss minimizes the distance between genus-related samples in the shared space; the image-text alignment loss constrains the semantic consistency between image embeddings and text embeddings of the same genus in the shared space, thus ensuring the model learns a continuous representation corresponding to the morphological description. Taking all factors into consideration, the overall loss function is defined as follows: (1) in For supervised comparative loss, For image-text alignment loss, These are the weighting coefficients; The supervised comparison loss Defined as: (2) Where zi represents the normalized embedding vector of the i-th sample, τ is the temperature coefficient, and P(i) represents the set of positive samples of the same class as sample i. The image-text alignment loss Defined as: (3) in, Indicates the total number of categories. The mean value of the features of image class C is used to represent the feature value of image class C. This represents the text feature representation of the corresponding category.
[0048] During training, the parameters of the image-side LoRA module, image projection head, and text projection head are updated to achieve multimodal semantic alignment. After the above training, the model can learn continuous morphological feature vectors that combine morphological semantics and visual structural information.
[0049] S3: Derivation of Continuous Morphological Features
[0050] After model training is completed, the global feature vector output by the image backbone network is extracted as a continuous morphological feature representation for the corresponding category. The global feature vector is a high-dimensional continuous representation learned by the DINO-V2 image encoder, used to characterize the feature differences of different groups in multiple morphological attribute dimensions, and therefore can be used as a continuous trait input for phylogenetic analysis.
[0051] In this embodiment, the continuous morphological feature vector has a dimension of 768 and is organized into a continuous trait matrix. Subsequently, the continuous trait matrix is exported as a NEXUS format file to adapt to the input requirements of Bayesian phylogenetic inference software such as RevBayes.
[0052] S4: Phylogenetic Tree Inference and Phylogenetic Analysis
[0053] like Figure 3 As shown, the exported continuous trait matrix is input into RevBayes, and phylogenetic tree analysis is performed using a Bayesian phylogenetic framework based on continuous trait evolution. In this embodiment, a Brownian motion model is used to describe the evolution of continuous morphological characteristics on the phylogenetic tree, assuming that the changes in continuous morphological traits follow a random diffusion mechanism, and that their variance accumulates with the branch length.
[0054] In this embodiment, the tree topology is set to a uniform prior, and the topology search employs two move methods: Nearest Neighbor Exchange (NNI) and Subtree Pruning and Regrafting (SPR). Branch length and rate parameters are assigned exponential priors, and the posterior distribution of the tree is approximated using Markov Chain Monte Carlo sampling. The MCMC run includes a warm-up phase and a formal sampling phase, with the warm-up step count set to 1000 generations, followed by 500,000 generations of formal sampling to obtain a tree sample set.
[0055] After summarizing the posterior tree samples, a maximum cladecredibility tree (MCC tree) is generated using the maximum cladecredibility criterion, serving as the final phylogenetic inference result. The MCC tree represents the phylogenetic tree topology with the highest posterior probability given continuous trait data, and can be used for subsequent comparative analysis and biological interpretation.
[0056] S5: Phylogenetic Tree Comparison and Result Evaluation
[0057] To quantify the topological differences between the inference tree and the reference tree, this embodiment uses the Robinson-Foulds distance and the normalized Align Score to quantitatively evaluate the results. For two unrooted trees with the same set of taxonomic units, the Robinson-Foulds distance is defined as the symmetric difference in the split sets between the two trees; the Robinson-Foulds distance is defined as follows: (4) in, Let n be the total number of topological differences between the two trees, and n be the number of taxons. The maximum possible value for an unrooted binary tree distance. The value range is from 0 to 1, where 0 indicates that the two trees have completely identical topologies, and 1 indicates that the two trees have completely different topologies.
[0058] To further evaluate the consistency between the inference tree and the reference tree in nontrivial splits, this embodiment calculates the normalized Align Score, which is defined as follows: (5) in, and Let represent the sets of nontrivial binary trees, namely the reference tree and the inference tree, respectively. The maximum number of internal splits in an unrooted binary tree. The higher the value, the greater the consistency between the inference tree and the reference tree in terms of topology.
[0059] Through the steps described above, this embodiment achieves automatic learning of continuous morphological features from dual-modal data of images and text, and further completes the automatic construction and phylogenetic evaluation of rove beetle groups. The final output is an MCC tree and its corresponding topological comparison index, which can be used to analyze the phylogenetic relationships and evolutionary patterns among different genera.
Claims
1. An automatic method for constructing insect phylogenetic trees based on multimodal large-scale model morphological representation learning, characterized in that, Includes the following steps: S1. Acquire image data of the target insect group and morphological description text data corresponding to the target insect group, standardize the image data and morphological description text data, and construct an image-text pairing sample set; S2. A multimodal morphological representation learning framework is constructed based on the image data and morphological description text data. The multimodal morphological representation learning framework includes an image encoder, a text encoder, an image projector, a text projector, and a shared embedding space. The image encoder encodes the feature sequence of the input image and obtains a visual representation vector through global feature extraction. The visual representation vector represents the overall morphological information of the target object. The text encoder extracts morphological semantic representations. The image projector and text projector map the image representation and text representation to the shared embedding space, respectively. The multimodal morphological representation learning framework is jointly optimized based on supervised contrast loss and image-text alignment loss to generate continuous morphological feature vectors that can represent the morphological differences of the target insect group. S3. The continuous morphological feature vectors are converted into a continuous trait data matrix required for phylogenetic inference. A Bayesian phylogenetic inference process is constructed based on the continuous trait evolution model. The candidate set of phylogenetic trees and their corresponding posterior probability distributions are obtained through Markov chain Monte Carlo sampling. S4. Based on the candidate phylogenetic tree set and its posterior probability distribution, statistically summarize the phylogenetic trees to generate the optimal phylogenetic tree; compare and evaluate the optimal phylogenetic tree based on the topological difference index between the reference phylogenetic tree and the optimal phylogenetic tree.
2. The method as described in claim 1, characterized in that, In step S1, the morphological description text data is automatically generated from the image data using a pre-trained text generation model based on preset rules. The initial morphological description text is then formed by combining manual verification and expert review to create a structured text library. The entries in the structured text library are associated one-to-one with the image data of the corresponding category, and the text entries contain descriptive information of the visible morphological parts of that category.
3. The method as described in claim 2, characterized in that, The morphological description text data includes seven morphological elements: color, pronotum, elytra, legs, head and neck, eyes, and antennae. It is organized using a standardized description template to form category-level text input.
4. The method as described in claim 1, characterized in that, The processing in step S1 includes performing size unification, brightness normalization, color normalization, and data augmentation on the image data; wherein, the data augmentation includes random flipping, rotation, translation, and scaling; and performing word segmentation, terminology normalization, noise reduction, and text length normalization on the morphological description text data.
5. The method as described in claim 1, characterized in that, The multimodal morphological representation learning framework described in step S2 includes: fine-tuning and updating relevant parameters such as attention in the image encoder to enhance the adaptability of the target task while retaining the pre-trained representation capability; inputting multiple morphological description information of the same category into the text encoder after structural integration, extracting category-level text semantic features, and constructing a category-level text feature library; wherein the image projection head and the text projection head project the image features and text features to a unified low-dimensional representation space, respectively.
6. The method as described in claim 1, characterized in that, In step S2, the supervised contrastive loss is used to minimize the feature distance between samples of the same class in the shared embedding space and to maximize the feature distance between samples of different classes; the image-text alignment loss is used to minimize the semantic distance between image features of the same class and text features of the same class; a weighted joint objective function is constructed by supervised contrastive loss and image-text alignment loss to jointly optimize the multimodal morphological representation learning framework to generate continuous morphological feature vectors for phylogenetic analysis.
7. The method as described in claim 1, characterized in that, The joint optimization of the supervised contrast loss and image-text alignment loss in step S2 is expressed in the following form: The overall loss function is: （1） in For supervised comparative loss, For image-text alignment loss, These are the weighting coefficients; The supervised comparison loss Defined as: （2） Where zi represents the normalized embedding vector of the i-th sample, τ is the temperature coefficient, and P(i) represents the set of positive samples of the same class as sample i. The image-text alignment loss Defined as: （3） in, Indicates the total number of categories. The mean value of the features of image class C is used to represent the feature value of image class C. This represents the text feature representation of the corresponding category.
8. The method as described in claim 1, characterized in that, The continuous morphological feature vector in step S2 is a global visual representation vector output by the image encoder after feature encoding of the input image. The global visual representation vector corresponds to the overall morphological information of the target insect and is used to construct a continuous trait matrix.
9. The method as described in claim 1, characterized in that, The Bayesian phylogenetic inference described in step S3 is based on the Brownian motion model of continuous traits, and a set of candidate phylogenetic trees is obtained through tree topology search, branch length optimization, and parameter posterior sampling.
10. The method as described in claim 1, characterized in that, The topological difference index in step S4 includes a distance metric for measuring the topological structural differences between phylogenetic trees and a similarity evaluation index for assessing the topological consistency between phylogenetic trees; the distance metric is used to characterize the structural differences between the inferred phylogenetic tree and the reference phylogenetic tree, and the similarity index is used to characterize the degree of consistency between the two phylogenetic trees in their splitting structure.