Multi-region tree species identification and distribution prediction model based on transfer learning

Abstract

The invention relates to the technical field of data mining, in particular to a multi-region tree species identification and distribution prediction model based on transfer learning, and the model comprises the steps: collecting images of different tree species in multiple regions and geographic positions corresponding to the images, and forming a target domain data set; forming a source domain data set by the marked images of the plant boundaries in the public data set and the geographic positions corresponding to the images; pre-training the neural network model to obtain a source domain model; carrying out feature extraction on the image, mapping a feature vector into a hyperbolic space, and forming a target domain feature set and a source domain feature set; and calculating the fitness of each feature vector in the target domain feature set, corresponding to the corrected mutation operator and the corresponding optimized feature vector, performing fine tuning on the source domain model, and performing tree species identification and distribution prediction on the image in the target domain data set by using the fine-tuned source domain model. According to the invention, the transfer learning effect can be improved, and the accuracy of tree species identification and distribution prediction by the model is improved.

Description

Technical Field

[0001] This application relates to the field of data mining technology, specifically to a multi-region tree species identification and distribution prediction model based on transfer learning. Background Technology

[0002] In forestry management and ecological protection, accurate identification and prediction of tree species distribution in different regions are crucial for resource management, biodiversity conservation, and ecosystem research. With the development of deep learning technology, neural network models can be used in forestry to identify and predict the distribution of tree species in multiple regions. However, neural network models rely on large amounts of labeled data for training. Therefore, transfer learning is used to transfer knowledge learned from the source domain to the target domain, thereby improving the identification performance in the target domain.

[0003] Traditional transfer learning methods based on Euclidean space rely on the relatively flat geometry of Euclidean space. However, tree species classification has a complex hierarchical structure, which cannot be preserved in Euclidean space. This results in poor model performance when dealing with tree species that have multiple levels of features. Furthermore, due to the differences in feature distribution between the source and target domains, the shared low-level feature representations between the source and target tasks cannot be aligned during transfer learning, leading to poor feature alignment. This reduces the effectiveness of transfer learning and affects the accuracy of the neural network model in tree species identification and distribution prediction. Summary of the Invention

[0004] To address the aforementioned technical issues, a multi-region tree species identification and distribution prediction model based on transfer learning is provided to resolve existing problems.

[0005] The solution to the technical problem in this application is to provide a multi-region tree species identification and distribution prediction model based on transfer learning, including the following steps:

[0006] Images of different tree species and their corresponding geographical locations are collected from multiple regions to form a target domain dataset; labeled plant kingdom images and their corresponding geographical locations from public datasets are collected to form a source domain dataset; a neural network model is pre-trained based on the source domain dataset to obtain a source domain model;

[0007] Feature extraction is performed on images in the target domain dataset and the source domain dataset to extract the corresponding feature vectors. The feature vectors are then subjected to nonlinear transformation to construct a nonlinear mapping function. The feature vectors are mapped to the hyperbolic space of the Poincaré disk model through the nonlinear mapping function. The mapped feature vectors are then used to form the target domain feature set and the source domain feature set, respectively.

[0008] By analyzing the differences between each feature vector in the target domain feature set and all feature vectors in the source domain feature set, the fitness of each feature vector in the target domain feature set is calculated. The mutation operator in the differential evolution algorithm is then modified to determine the modified mutation operator corresponding to each feature vector in the target domain feature set.

[0009] Based on the modified mutation operator and fitness, the differential evolution algorithm is used to iteratively optimize the feature vectors in the target domain feature set, obtain the optimized feature vectors corresponding to the feature vectors in the target domain feature set, fine-tune the source domain model, and use the fine-tuned source domain model to identify tree species and predict the distribution of images in the target domain dataset.

[0010] Preferably, the feature vector extraction process corresponding to the image is as follows: using the source domain model to extract features from each image in the target domain dataset and the source domain dataset respectively, and extracting the feature vector corresponding to the image.

[0011] Preferably, the construction of the nonlinear mapping function specifically involves: Where g() represents the nonlinear mapping function, W and b are the learning parameters of the fully connected layer in the source domain model, where W represents the weight matrix, b represents the bias vector, f represents the extracted feature vector, ε represents a preset value greater than 0, ‖‖ represents the calculation of the Euclidean norm, and tanh() represents the hyperbolic tangent function.

[0012] Preferably, the hyperbolic space of the Poincaré disk model is defined as follows: in, Let h represent an n-dimensional Poincaré disk model, where h represents a point on the Poincaré disk model, i.e., the mapped eigenvector. Let represent the set of all n-dimensional real vectors, and let ||| denote the computation of the Euclidean norm.

[0013] Preferably, the step of mapping the feature vectors to the hyperbolic space of the Poincaré disk model includes: mapping the feature vectors corresponding to each image in the target domain dataset and the source domain dataset respectively based on a nonlinear mapping function to obtain an intermediate vector, and performing feature cropping on the intermediate vector to obtain the mapped feature vector.

[0014] Preferably, the feature trimming process is as follows: Where h is the mapped feature vector, g i Let be the i-th element in the intermediate vector g, and Clip() represents the clipping operation.

[0015] Preferably, the j-th feature vector within the target domain feature set fitness The calculation formula is: in, in, For the source domain feature set, For the target domain feature set The j-th eigenvector For source domain feature set The i-th eigenvector within N s For source domain feature set The number of all eigenvectors in the matrix, |||2 represents the calculation of the L2 norm, α is the preset smoothing coefficient, and exp{} represents the exponential function with the natural constant as the base.

[0016] Preferably, the j-th feature vector in the target domain feature set corresponds to the modified mutation operator P. j The calculation formula is: Among them, P min As a predefined minimum mutation operator, P max To pre-determine the maximum mutation operator, Represents the j-th eigenvector within the target domain feature set. fitness, F max F represents the maximum fitness of all feature vectors within the target domain feature set. min It represents the minimum fitness of all feature vectors within the target domain feature set.

[0017] Preferably, the fine-tuning of the source domain model includes: selecting a portion of images from the target domain dataset for annotation; inputting the labels corresponding to the annotated images in the target domain dataset and the corresponding optimized feature vectors into the source domain model for fine-tuning, thereby obtaining the fine-tuned source domain model.

[0018] Preferably, the step of identifying tree species and predicting their distribution in the images within the target domain dataset includes: using the optimized feature vectors corresponding to the unlabeled images in the target domain dataset and their corresponding geographical locations as inputs to the fine-tuned source domain model to identify and predict the tree species in the images.

[0019] This application has at least the following beneficial effects:

[0020] This application obtains a source domain model by pre-training a neural network model. The advantage of this is that the model can learn knowledge from the source domain, enabling subsequent transfer of this knowledge to the target domain. Features are extracted from images in both the target and source domain datasets, and the extracted feature vectors are mapped onto the hyperbolic space of the Poincaré disk model to form target and source domain feature sets. This allows for the extraction of morphological features of tree species from the images, mapping these features onto the hyperbolic space to better reflect the hierarchical structure of tree species classification and improve the expressive power of subtle differences in tree species features within the hyperbolic space. The fitness of each feature vector in the target domain feature set is calculated, taking into account the similarity between the feature vectors in the target domain feature set and those in the source domain, reflecting the closeness of the feature distribution between the feature vectors in the target domain feature set and those in the source domain. The mutation operator in the differential evolution algorithm is modified, and the modified mutation operator corresponding to each feature vector in the target domain feature set is determined. This allows for dynamic adjustment of the differential evolution algorithm. The mutation operator in the algorithm allows for a finer search using a smaller mutation operator on high-fit feature vectors to ensure that the optimal solution is not destroyed, while a larger mutation operator is used on low-fit feature vectors for broader exploration, increasing the ability to explore the solution space and thus avoiding getting trapped in local optima, which leads to poor alignment between the feature vectors of the source and target domains. The differential evolution algorithm iteratively optimizes the feature vectors within the target domain feature set, obtaining optimized feature vectors corresponding to the feature vectors in the target domain feature set. Its benefit lies in continuously optimizing and iterating the feature vectors within the target domain feature set in hyperbolic space, making the optimized feature vectors as aligned as possible with the feature vectors in the source domain feature set in terms of distribution, reducing the feature distribution difference between the source and target domains, and improving the effect of transfer learning. Fine-tuning the source domain model and using the fine-tuned source domain model to perform tree species identification and distribution prediction on images in the target domain dataset has the benefit of allowing the source domain model to better adapt to the characteristics of the target domain, improving the model's performance in the target domain, and thus improving the accuracy of the model in tree species identification and distribution prediction. Attached Figure Description

[0021] The following section provides a more detailed description of the multi-region tree species identification and distribution prediction model based on transfer learning in this application, with reference to the accompanying drawings.

[0022] Figure 1 A flowchart illustrating the steps of a multi-region tree species identification and distribution prediction model based on transfer learning provided in this application embodiment;

[0023] Figure 2 A flowchart illustrating the steps of the method for obtaining the fine-tuned source domain model provided in this application embodiment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of the multi-region tree species identification and distribution prediction model based on transfer learning proposed in this application, in conjunction with the accompanying drawings and implementation examples, is provided. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit the scope of this application.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0026] Please see Figure 1 The diagram illustrates a flowchart of a multi-region tree species identification and distribution prediction model based on transfer learning provided in an embodiment of this application. The model includes the following steps:

[0027] Step 1: Collect images of different tree species and their corresponding geographical locations in multiple regions to form the target domain dataset; collect labeled plant kingdom images and their corresponding geographical locations from the public dataset to form the source domain dataset; pre-train the neural network model based on the source domain dataset to obtain the source domain model.

[0028] Tree species identification has always been a crucial method for forest resource management and tree species diversity assessment. High-precision tree species identification information can improve the efficiency of forest management, enabling the rational utilization and monitoring of forest resources, promoting their sustainable development, and fully realizing their social, economic, and ecological benefits. Traditional tree species identification methods mainly rely on manual observation of tree morphological characteristics, such as leaves, fruits, and flowers. However, these morphological characteristics are greatly affected by seasonal changes and require specialized knowledge and experience. The rapid development of remote sensing technology has provided a more convenient and efficient data source for forest type and tree species identification and classification. By identifying tree species in remote sensing images, continuous monitoring at multiple temporal and spatial scales can be achieved, thus providing a comprehensive understanding of the composition and distribution of vegetation tree species within the study area.

[0029] Based on the above analysis, a drone equipped with optical remote sensing equipment was used to fly in different geographical areas and collect images of a large number of tree species. When capturing each image, the geographic location of the image was obtained using satellite positioning technology. The collected images and geographic locations were then preprocessed, specifically including:

[0030] Radiometric and geometric corrections are performed on the images to correct brightness, contrast, and geometric distortion, ensuring consistent image quality; geographical locations are unified and standardized.

[0031] It should be noted that radiation correction and geometric correction are well-known techniques and will not be elaborated upon here.

[0032] All collected images and their geographical locations are combined to form a target domain dataset;

[0033] A subset of images from the target domain dataset are selected for annotation. It should be noted that the tree species in the annotated images correspond to complete classification level labels.

[0034] In this embodiment, 5% of all images in the target domain dataset are selected for annotation. As another implementation method, the implementer can set it according to the actual situation.

[0035] It should be noted that since the classification hierarchy of tree species is arranged from largest to smallest, including phylum, class, order, family, genus, and species, when labeling tree species images, the complete hierarchical labels of phylum, class, order, family, genus, and species must be included. For example, for Yunnan pine, the phylum is gymnosperms, the class is Pinales, the order is Pinales, the family is Pinaceae, the genus is Pinus, and the species is Yunnan pine. For birch, the phylum is angiosperms, the class is Magnoliopsida, the order is Fagales, the family is Betulaceae, the genus is Betulaceae, and the species is Betulaceae.

[0036] Images of the plant kingdom and their corresponding geographical locations, along with the labels for each image, are obtained from publicly available datasets to form the source domain dataset;

[0037] In this embodiment, the public datasets include the iNaturalist and GBIF datasets. All images of the plant kingdom, along with their geographical locations and complete classification hierarchical labels, are obtained from the public datasets. The acquisition of public datasets is a well-known technique and will not be described in detail here.

[0038] Based on the source domain dataset, the neural network model is pre-trained, and the pre-trained neural network model is denoted as the source domain model.

[0039] In this embodiment, a deep convolutional neural network is used for training. The feature extraction layer is built using a ResNet network model, the classifier and regressor are built using fully connected layers, the optimizer is the Adam optimizer, and the loss function is multi-class cross-entropy loss and regression loss. Deep convolutional neural networks are well-known technologies and will not be described in detail here.

[0040] At this point, we have obtained the target domain dataset, the source domain dataset, and the source domain model.

[0041] Step 2: Extract features from the images in the target domain dataset and the source domain dataset to obtain the feature vectors corresponding to the images; perform nonlinear transformation on the feature vectors to construct a nonlinear mapping function, and map the feature vectors to the hyperbolic space of the Poincaré disk model through the nonlinear mapping function. The mapped feature vectors are then used to form the target domain feature set and the source domain feature set, respectively.

[0042] Furthermore, the classification of tree species has a hierarchical structure, namely multiple classification levels of phylum-class-order-family-genus-species. Each classification level represents a category of tree species with a set of common characteristics. The next classification level is further subdivided by individual morphological characteristics. For example, both Korean pine and Chinese pine belong to the genus Pinus and have the morphological characteristics of the genus Pinus, such as needle-shaped leaves, pine cone-shaped fruits, and scaly trunks. However, Korean pine and Chinese pine have different numbers of needle bundles and different sizes of pine cones. Therefore, the morphological characteristics of tree species exhibit complex similarities across different classification levels. Traditional transfer learning based on Euclidean space struggles to naturally represent such complex hierarchical relationships when dealing with data exhibiting complex hierarchical similarity structures due to the inherent dimensionality of Euclidean space. In Euclidean space, the features of all tree species are mapped onto the same plane, making it difficult to accurately capture such hierarchical similarity structures. For example, while Korean pine and Chinese pine (Pinus tabuliformis) differ morphologically but belong to the same genus and share common characteristics, the differences in the number of needle bundles and the size of pine cones between them in Euclidean space may cause them to be mapped to distant locations, resulting in poor transfer learning performance.

[0043] Hyperbolic space is a non-Euclidean geometric space with constant negative curvature. It can be viewed as a continuous representation of a discrete tree structure in space, thus naturally capable of depicting hierarchical structures. In hyperbolic space, straight lines are curved, and parallel lines intersect at infinity. While the volume of Euclidean space grows polynomially with radius, hyperbolic space grows exponentially, which matches tree-like data well. Furthermore, hyperbolic space has the ability to express hierarchical structures, allowing for the embedding of hierarchical data. By embedding complex tree structures in a low dimension, with parent nodes close to the center and child nodes radially distributed, it is suitable for representing hierarchical classification relationships among tree species. For example, while the morphological characteristics of Korean pine and Chinese pine differ in different regions, both belong to the genus *Pinus*, and they can still be mapped to the vicinity of the same branch in hyperbolic space, thus achieving better feature alignment.

[0044] The Poincaré disk model is an n-dimensional hyperbolic geometric model. By embedding a tree structure into the 2-dimensional Poincaré disk model, the line segments between any two points in the Poincaré disk model represent the same length. That is, in the Poincaré disk model, the farther away from the center, the longer the line segment in unit Euclidean space. Since the volume expands exponentially with the radius, the actual distance between points becomes infinitely large as they get closer to the boundary of the disk.

[0045] Based on the above analysis, feature extraction is performed on images in both the target domain dataset and the source domain dataset. The extracted features are then mapped from Euclidean space to the hyperbolic space where the Poincaré disk model resides. Specifically:

[0046] The source domain model is used to extract features from each image in both the target domain dataset and the source domain dataset, and the feature vectors corresponding to the images are extracted.

[0047] It should be noted that the feature extraction process is a well-known technique and will not be elaborated upon here. Secondly, the feature vector contains distinguishable morphological features of the tree, such as trunk texture and color, leaf size, color, shape, and leaf edges, and fruit color and shape. Let the feature vector be denoted as f. Assuming the extracted feature vector f is d-dimensional, then the feature vector f corresponding to the images in both the target domain dataset and the source domain dataset belongs to a d-dimensional Euclidean space. Right now In this embodiment, d is set to 1024. In other implementation methods, the implementer can set it according to the actual situation.

[0048] Mapping feature vectors to the hyperbolic space of the Poincaré disk model typically involves transforming high-dimensional feature vectors into hyperbolic space through a nonlinear mapping. Assuming there are n morphological features available for tree species identification and classification, and setting the dimension of the hyperbolic space to n, the feature vectors can be... Feature vectors mapped to the hyperbolic space of the Poincaré disk model Above, the hyperbolic space of the Poincaré disk model is defined as follows:

[0049]

[0050] in, Let h represent an n-dimensional Poincaré disk model, where h represents a point on the Poincaré disk model, i.e., the mapped eigenvector. Let represent the set of all n-dimensional real vectors, and let ||| denote the computation of the Euclidean norm.

[0051] It should be noted that the calculation of the Euclidean norm is a well-known technique and will not be elaborated upon here.

[0052] By employing a nonlinear mapping function in the fully connected layer of a neural network model, the feature vector f is mapped to the feature vector h in the hyperbolic space of the Poincaré disk model. Specifically, the nonlinear mapping function is:

[0053]

[0054] Where g() represents the nonlinear mapping function, W and b are the learning parameters of the fully connected layer in the source domain model, where W represents the weight matrix, b represents the bias vector, f represents the extracted feature vector, ε represents a preset value greater than 0, ‖‖ represents the calculation of the Euclidean norm, and tanh() represents the hyperbolic tangent function.

[0055] In this embodiment, the preset value ε, which is greater than 0, is 0.001. In other implementation methods, the implementer can set it according to the actual situation.

[0056] It should be noted that the weight matrix W and bias vector b are obtained by training the neural network model, i.e., the learning parameters of the fully connected layers in the source domain model; secondly, W×f+b is a linear transformation of the feature vector. The vector represents the unit vector corresponding to the vector after linear transformation, that is, the vector after linear transformation is normalized. tanh(‖Wf+b‖) is a nonlinear transformation of the eigenvector.

[0057] Based on a nonlinear mapping function, the feature vectors corresponding to each image in the target domain dataset and the source domain dataset are mapped to obtain intermediate vectors. The intermediate vectors are then cropped to obtain the mapped feature vectors.

[0058] It should be noted that the output of the nonlinear mapping function is a vector, denoted as the intermediate vector g. Feature clipping is then performed on the intermediate vector g. The specific clipping process is as follows:

[0059]

[0060] Where h is the mapped feature vector, g i Let be the i-th element in the intermediate vector g.

[0061] Therefore, through a nonlinear mapping function, the extracted feature vectors are transformed from Euclidean space. Mapped to the hyperbolic space of the Poincaré disk model

[0062] The feature vectors corresponding to all images in the target domain dataset are mapped to form the target domain feature set;

[0063] The feature vectors corresponding to all images in the source domain dataset are mapped to form the source domain feature set;

[0064] Thus, the target domain feature set and source domain feature set on the Poincaré disk model are obtained.

[0065] Step 3: Calculate the fitness of each feature vector in the target domain feature set by comparing the differences between each feature vector in the target domain feature set and all feature vectors in the source domain feature set. Modify the mutation operator in the differential evolution algorithm and determine the modified mutation operator corresponding to each feature vector in the target domain feature set.

[0066] Furthermore, in transfer learning, prior knowledge learned from the source domain is transferred to the target domain to improve the performance of transfer learning. However, the distribution of feature vectors corresponding to images in the source and target domain datasets often differs. For example, the morphology of tree species in different regions may vary due to environmental factors, which can lead to a decrease in the performance of transfer learning. Therefore, in the hyperbolic space where the Poincaré disk is located, it is necessary to optimize the feature vectors mapped into the target domain feature set in the hyperbolic space so that they are as close as possible to the feature vectors mapped into the source domain feature set in the hyperbolic space. This achieves the alignment process of feature vectors between the source and target domains, thereby improving the performance of transfer learning.

[0067] Differential evolution is an algorithm that initializes individuals using real-number vectorized encoding and uses a differential strategy to determine the optimal process of individual mutation. Specifically, it selects two different individuals, subtracts their vectors to generate a difference vector, selects another individual and sums it with the difference vector to generate an experimental individual, then performs a crossover operation between the parent individual and the experimental individual to generate a new individual, and selects individuals from the parent and new individuals to keep in the next generation of the population, continuously iterating to find the optimal individual.

[0068] Secondly, the differential evolution algorithm uses a real-number vectorization encoding method, mapping the feature vectors in the target domain feature set and the source domain feature set onto the Poincaré disk model. It satisfies the format of real number vectorization encoding. Therefore, the feature vector h in the target domain feature set can be directly used as the encoded individual in the differential evolution algorithm, without the need for additional encoding representation of the feature vector. This avoids the possibility that the feature vector representation may be biased after multiple encodings, introducing unnecessary complexity.

[0069] Based on the above analysis, in the hyperbolic space of the Poincaré disk model, the feature vectors in the target domain feature set are treated as individuals to be optimized. In subsequent iterations, they continuously mutate and cross over, generating new individuals, thereby continuously optimizing the feature vectors in the target domain feature set to minimize the difference in feature distribution between the source and target domains. Specifically, in hyperbolic space feature alignment based on differential evolution, fitness is used to quantify the similarity of feature vectors between the target and source domains. Therefore, through the concept of maximum mean difference (MMD), the difference between the feature vectors in the target domain feature set and the feature vectors in the source domain feature set is analyzed, and fitness is constructed as follows:

[0070] The formula for calculating the maximum mean difference (MMD) is:

[0071]

[0072] in, For the source domain feature set, For the target domain feature set, This represents the maximum mean difference between the feature sets of the source domain and the feature sets of the target domain. For source domain feature set The i-th eigenvector within the matrix, For the target domain feature set The j-th eigenvector within N s For source domain feature set The number of all eigenvectors in N t For the target domain feature set The number of all eigenvectors in the L2 norm, and |||2 represents the calculation of the L2 norm.

[0073] It should be noted that the maximum mean difference is a well-known technique and will not be elaborated upon here. This represents the mean vector of all eigenvectors within the feature set of the source domain. This represents the mean vector of all feature vectors within the feature set of the target domain.

[0074] Fitness is determined by analyzing the difference between each feature vector in the target domain feature set and the mean vector of all feature vectors in the source domain feature set. The calculation formula is as follows:

[0075]

[0076]

[0077] in, For the source domain feature set, For the target domain feature set The j-th eigenvector For source domain feature set The i-th eigenvector within Ns For source domain feature set The number of all eigenvectors within the range, where |||2 represents the L2 norm. For the target domain feature set The j-th eigenvector The fitness of α is a preset smoothing coefficient, and exp{} represents an exponential function with the natural constant as the base.

[0078] In this embodiment, if the smoothing coefficient is too large, the fitness becomes overly sensitive to small differences, potentially leading to local optima. Conversely, if the smoothing coefficient is too small, the fitness changes gradually, requiring more iterations to converge. Therefore, to control the fitness's sensitivity to feature differences... The sensitivity is set to a preset smoothing coefficient α of 5. As for other implementation methods, the implementer can set the value according to the actual situation.

[0079] It should be noted that, The larger the value, the more likely the feature vectors are to be within the target domain feature set. The greater the difference in distribution between the feature vector and the source domain feature set, the stronger the feature vector. The smaller the fitness, the better the eigenvector. A lower degree of matching with the source domain feature set indicates a higher degree of matching with the target domain feature set. With source domain feature set The similarity in the characteristic distributions reflects... It is closer to the source domain distribution.

[0080] Furthermore, in traditional differential evolution algorithms, the mutation operators for different individuals are fixed during mutation operations. This increases the risk of encountering local optima during the iteration process of the differential evolution algorithm. Moreover, for high-dimensional feature vectors, excessively high mutation operators may destroy the semantic information of the feature vectors. Secondly, the lower the fitness of an individual, the higher the mutation operator is required to randomly mutate and update the individual, generating new individuals and thus increasing the ability to explore the solution space. For individuals with high fitness, which represent the current better solution, the mutation operator for that individual should be reduced to ensure that the better solution is not destroyed, thereby improving the algorithm's ability to explore the optimal solution.

[0081] It should be noted that all feature vectors within the target domain feature set are used as the initial population, and the feature vectors are used as individuals in the initial population.

[0082] Therefore, in tree species identification scenarios, the differences in tree species characteristics between the source and target domains mainly lie in morphological differences with discernible characteristics. If the feature distribution of individuals within a population is relatively concentrated, it indicates that the feature vectors corresponding to individuals in the population can match the key morphological features of the tree species well. In this case, the mutation operator should be reduced to avoid disrupting the already found stable feature combinations. Conversely, if the feature distribution of the population is relatively dispersed, and the morphological differences between individuals are large, it indicates high population diversity. In this case, the mutation rate should be appropriately increased, and the mutation operator should be added to enhance the ability to explore potential cross-domain features. Therefore, the mutation operator of the feature vector is modified through fitness, specifically as follows:

[0083] Obtain the maximum and minimum fitness of all feature vectors within the target domain feature set;

[0084] The formula for calculating the modified mutation operator corresponding to the j-th eigenvector is:

[0085]

[0086] Among them, P j For the modified mutation operator corresponding to the j-th eigenvector in the target domain feature set, P min As a predefined minimum mutation operator, P max To pre-determine the maximum mutation operator, Represents the j-th eigenvector within the target domain feature set. fitness, F max For maximum fitness, F min Minimum fitness;

[0087] In this embodiment, a maximum mutation operator P is preset. max The value is set to 1, and the minimum mutation operator P is preset. min The value is 0.5. As for other implementation methods, the implementer can set it according to the actual situation.

[0088] It should be noted that if the feature vector When the morphological features in the eigenvectors differ significantly from those in the source domain, the fitness is low. A larger mutation operator increases the mutation rate, thus enhancing the ability to explore the solution space. Conversely, if the eigenvectors... When the morphological features in the tree differ less from those in the source domain, the fitness is greater, the resulting mutation operator is smaller, and the better solution of the individual is preserved. This strategy not only retains the global search capability of the traditional differential evolution algorithm, but also improves the performance and efficiency of feature optimization between the source and target domains of the tree species by dynamically adjusting the mutation operator based on feature differences.

[0089] Step 4: Based on the corrected mutation operator and fitness, the differential evolution algorithm is used to iteratively optimize the feature vectors in the target domain feature set, obtain the optimized feature vectors corresponding to the feature vectors in the target domain feature set, fine-tune the source domain model, and use the fine-tuned source domain model to perform tree species identification and distribution prediction on the images in the target domain dataset.

[0090] Furthermore, based on the modified mutation operator, mutation, crossover, and selection operations are performed using the differential evolution algorithm. The specific process of the differential evolution algorithm is as follows:

[0091] Treat each feature vector within the target domain feature set as an individual;

[0092] Based on the modified mutation operator corresponding to each individual, and by randomly selecting two distinct individuals from the target domain feature set, performing mutation operations to generate mutated individuals, specifically:

[0093]

[0094] Among them, v r For the generated mutant individuals, and They are two randomly selected distinct individuals, that is... and They are not the same, and respectively with They are also different, P j For the j-th individual The corresponding modified mutation operator;

[0095] For each mutated individual v r and individuals After performing the crossover operation, a selection operation is performed based on fitness to select individuals with better fitness to enter the next generation. This process is repeated iteratively to obtain the optimal individual for each individual.

[0096] The optimal individual is taken as the optimized feature vector corresponding to each feature vector in the feature set of the target domain.

[0097] It should be noted that the differential evolution algorithm is a well-known technique and will not be elaborated here. During the crossover operation of the differential evolution algorithm, the crossover probability is set to 0.3 and the number of iterations is 50. As other implementation methods, implementers can set them according to the actual situation. Secondly, by continuously selecting individuals with higher fitness for iteration, the maximum mean difference (MMD) between the optimized feature vector and the feature vector in the source domain feature set is continuously reduced. By continuously optimizing the feature vector of the target domain feature set, the feature distribution difference between the source domain and the target domain is minimized, thereby aligning the feature vectors between the source domain feature set and the target domain feature set in the hyperbolic space where the Poincaré disk model is located, and improving the performance of subsequent transfer learning.

[0098] Furthermore, based on the optimized feature vectors corresponding to each feature vector in the target domain feature set, the source domain model is fine-tuned, specifically as follows:

[0099] The complete classification hierarchical labels corresponding to the labeled images in the target domain dataset and the optimized feature vectors corresponding to the labeled images are input into the source domain model for fine-tuning to obtain the fine-tuned source domain model.

[0100] It should be noted that a hierarchical learning rate strategy is adopted in the fine-tuning stage: the feature extraction layer uses a lower learning rate to retain general features, while the fully connected layer uses a higher learning rate to quickly adapt to the characteristics of the target domain. The process of fine-tuning the source domain model is a well-known technique and will not be described in detail here. The flowchart of the method for obtaining the fine-tuned source domain model provided in this embodiment is as follows: Figure 2 As shown.

[0101] The optimized feature vectors and corresponding geographical locations of unlabeled images in the target domain dataset are used as inputs to the fine-tuned source domain model to identify and predict the tree species and their distribution in the images.

[0102] Thus, the construction of a multi-region tree species identification and distribution prediction model is completed, improving the transfer learning performance of the model.

[0103] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this application, without departing from the content of the technical solution of this application, shall fall within the protection scope of the technical solution of this application.

Claims

1. A multi-region tree species identification and distribution prediction model based on transfer learning, characterized in that, The model includes the following steps: Images of different tree species and their corresponding geographical locations are collected from multiple regions to form a target domain dataset; labeled plant kingdom images and their corresponding geographical locations from public datasets are collected to form a source domain dataset; a neural network model is pre-trained based on the source domain dataset to obtain a source domain model; Feature extraction is performed on images in the target domain dataset and the source domain dataset to extract the corresponding feature vectors. The feature vectors are then subjected to nonlinear transformation to construct a nonlinear mapping function. The feature vectors are mapped to the hyperbolic space of the Poincaré disk model through the nonlinear mapping function. The mapped feature vectors are then used to form the target domain feature set and the source domain feature set, respectively. By analyzing the differences between each feature vector in the target domain feature set and all feature vectors in the source domain feature set, the fitness of each feature vector in the target domain feature set is calculated. The mutation operator in the differential evolution algorithm is then modified to determine the modified mutation operator corresponding to each feature vector in the target domain feature set. Based on the modified mutation operator and fitness, the differential evolution algorithm is used to iteratively optimize the feature vectors in the target domain feature set, obtain the optimized feature vectors corresponding to the feature vectors in the target domain feature set, fine-tune the source domain model, and use the fine-tuned source domain model to identify tree species and predict the distribution of images in the target domain dataset.

2. The multi-region tree species identification and distribution prediction model based on transfer learning as described in claim 1, characterized in that, The process of extracting the feature vector corresponding to the image is as follows: the source domain model is used to extract features from each image in the target domain dataset and the source domain dataset respectively, and the feature vector corresponding to the image is extracted.

3. The multi-region tree species identification and distribution prediction model based on transfer learning as described in claim 1, characterized in that, The construction of the nonlinear mapping function is specifically as follows: Where g() represents the nonlinear mapping function, W and b are the learning parameters of the fully connected layer in the source domain model, where W represents the weight matrix, b represents the bias vector, f represents the extracted feature vector, ε represents a preset value greater than 0, ‖‖ represents the calculation of the Euclidean norm, and tanh() represents the hyperbolic tangent function.

4. The multi-region tree species identification and distribution prediction model based on transfer learning as described in claim 1, characterized in that, Define the hyperbolic space of the Poincaré disk model as: in, Let h represent an n-dimensional Poincaré disk model, where h represents a point on the Poincaré disk model, i.e., the mapped eigenvector. Let represent the set of all n-dimensional real vectors, and let ||| denote the computation of the Euclidean norm.

5. The multi-region tree species identification and distribution prediction model based on transfer learning as described in claim 1, characterized in that, The step of mapping the feature vectors to the hyperbolic space of the Poincaré disk model includes: mapping the feature vectors corresponding to each image in the target domain dataset and the source domain dataset respectively based on a nonlinear mapping function to obtain an intermediate vector, and performing feature cropping on the intermediate vector to obtain the mapped feature vector.

6. The multi-region tree species identification and distribution prediction model based on transfer learning as described in claim 5, characterized in that, The feature clipping process is as follows: Where h is the mapped feature vector, g i Let be the i-th element in the intermediate vector g, and Clip() represents the clipping operation.

7. The multi-region tree species identification and distribution prediction model based on transfer learning as described in claim 1, characterized in that, The j-th feature vector in the target domain feature set fitness The calculation formula is: in, in, For the source domain feature set, For the target domain feature set The j-th eigenvector For source domain feature set The i-th eigenvector within N s For source domain feature set The number of all eigenvectors in the matrix, |||2 represents the calculation of the L2 norm, α is the preset smoothing coefficient, and exp{} represents the exponential function with the natural constant as the base.

8. The multi-region tree species identification and distribution prediction model based on transfer learning as described in claim 1, characterized in that, The j-th eigenvector in the target domain feature set corresponds to the corrected mutation operator P. j The calculation formula is: Among them, P min As a predefined minimum mutation operator, P max To pre-determine the maximum mutation operator, Represents the j-th eigenvector within the target domain feature set. fitness, F max F represents the maximum fitness of all feature vectors within the target domain feature set. min It represents the minimum fitness of all feature vectors within the target domain feature set.

9. The multi-region tree species identification and distribution prediction model based on transfer learning as described in claim 1, characterized in that, The fine-tuning of the source domain model includes: selecting a portion of images from the target domain dataset for annotation; inputting the labels corresponding to the annotated images in the target domain dataset and the corresponding optimized feature vectors into the source domain model for fine-tuning, thereby obtaining the fine-tuned source domain model.

10. The multi-region tree species identification and distribution prediction model based on transfer learning as described in claim 9, characterized in that, The method of identifying tree species and predicting their distribution in images within the target domain dataset includes: using the optimized feature vectors corresponding to unlabeled images within the target domain dataset and their corresponding geographical locations as inputs to a fine-tuned source domain model to identify and predict the tree species and their distribution in the images.

Images