Landslide identification method and system based on staged feature adaptive transfer learning

Abstract

The invention relates to the cross technical field of artificial intelligence and geological disaster monitoring and early warning, in particular to a landslide recognition method and system based on staged feature adaptive transfer learning, and the method comprises the steps: constructing an initial landslide recognition model F1 with an encoder-decoder; a universal visual pre-training weight is migrated, and encoder parameters of the model F1 are finely adjusted by using a remote sensing image data set DS of a source domain, so that a landslide identification optimization model F2 adaptive to the field is obtained; inputting the remote sensing image data set DS of the source domain and the unmanned aerial vehicle image data set DT of the target domain into the model F2, and processing based on a covariance alignment mechanism to obtain a domain adaptive model F3; inputting a small amount of labeled target domain small sample data D 'T into the model F3, and finely adjusting specified parameters in the model F3 through a supervised learning mechanism to obtain a target domain adaptive model F4; and based on the target domain adaptation model F4, executing a landslide identification task on an input to-be-detected unmanned aerial vehicle image.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary technical field of artificial intelligence and geological disaster monitoring and early warning, and in particular to a landslide identification method and system based on staged feature adaptive transfer learning. Background Art

[0002] Landslides are highly destructive geological hazards, and their accurate and rapid identification is crucial for disaster prevention and mitigation. Currently, drone remote sensing, due to its high resolution, timeliness, and flexibility, has become an important means of acquiring images of landslide areas. Although deep learning models based on convolutional neural networks or Transformer architectures have made significant progress in computer vision, using deep learning models to directly process drone imagery for landslide identification still faces the following significant challenges:

[0003] 1) Sample scarcity and high labeling costs: Acquiring and accurately labeling a large number of drone-based landslide image samples is costly and time-consuming, resulting in a severe shortage of training data (target domain data). This makes it difficult to fully train deep learning models, leading to overfitting and poor generalization. Consequently, recognition accuracy drops significantly under small sample sizes.

[0004] 2) Domain gap: Although transfer learning has been introduced to reduce reliance on target domain data, common transfer strategies (such as using ImageNet pre-trained models) still suffer from a significant domain gap due to fundamental differences in data distribution and feature space between the source domain (natural images) and the target domain (landscape remote sensing images).

[0005] 3) Cross-source domain feature mismatch: Even when more task-relevant source domain data (such as satellite / aerial remote sensing imagery) is introduced for migration, these data still differ significantly from the target domain drone imagery in terms of resolution, lighting, and terrain details. This makes it difficult to align the feature spaces of the source and target domains, and the model is prone to negative transfer during cross-domain migration.

[0006] 4) The model's ability to process complex features is insufficient: Landslides have diverse morphologies, varying scales, fuzzy boundaries, and are easily disturbed by vegetation, shadows, and other factors. Existing models (such as U-Net) use single-scale convolution kernels or fixed receptive fields, making it difficult to capture the multi-scale features of landslides (such as overall morphology and local cracks) and accurately delineate boundaries. They are also not robust enough to noise interference (such as vegetation occlusion). There is still room for improvement in capturing multi-scale features and accurately delineating boundaries.

[0007] Therefore, there is an urgent need for a new landslide identification method that can effectively utilize limited target domain (UAV) labeled samples, fully utilize relevant source domain (such as remote sensing) knowledge, accurately align cross-domain features (especially features that distinguish landslides from background), and have powerful multi-scale feature extraction capabilities to solve the problems of low recognition accuracy and low efficiency under small sample conditions. Summary of the Invention

[0008] In order to solve the problems of low recognition accuracy and low efficiency under small sample conditions, the purpose of the present invention is to provide a landslide recognition method and system based on staged feature adaptive transfer learning. The technical solutions adopted are as follows:

[0009] In a first aspect, the present application discloses a landslide identification method based on staged feature adaptive transfer learning, the method comprising:

[0010] S1, constructing an initial landslide identification model F1 with an encoder-decoder;

[0011] S2, by migrating general visual pre-training weights and utilizing the remote sensing image dataset D of the source domain S Fine-tune the encoder parameters of model F1 to obtain the domain-adapted landslide identification optimization model F2;

[0012] S3, the remote sensing image dataset D of the source domain S , and the target domain drone image dataset D T The input is fed into model F2, and the domain adaptation model F3 is obtained by reducing the distribution difference between the source domain and the target domain in the feature space based on the covariance alignment mechanism.

[0013] S4, a small amount of labeled target domain small sample data D′ T Input into model F3, fine-tune the specified parameters in model F3 through the supervised learning mechanism, and obtain the target domain adaptation model F4;

[0014] S5. Based on the target domain adaptation model F4, a landslide identification task is performed on the input UAV image to be detected.

[0015] Furthermore, in step S1, the encoder part of the initial landslide recognition model F1 adopts the ResNet network as the backbone feature extraction network to utilize the powerful feature extraction capability of the ResNet network and alleviate the gradient vanishing problem.

[0016] Furthermore, in step S1, an ASSP module is integrated between the encoder and decoder of the initial landslide identification model F1, or in the decoder of the initial landslide identification model F1, so as to effectively expand the receptive field and obtain richer context information by using dilated convolutions with different dilation rates in parallel without increasing the amount of computation.

[0017] Furthermore, in step S2, the method transfers the general visual pre-training weights and utilizes the remote sensing image dataset D of the source domain. S Fine-tune the encoder parameters of model F1 to obtain the domain-adapted landslide identification optimization model F2, including:

[0018] S21, loading the encoder weights pre-trained on a large-scale general image dataset into the encoder part of model F1, and accelerating the landslide feature learning process by utilizing general visual knowledge to initialize the encoder parameters of model F1;

[0019] S22. Using the remote sensing image dataset D in the source domain S , and its corresponding landslide mask labels are used to fine-tune the encoder parameters of the model F1. Through pre-training in the intermediate domain to narrow the semantic gap between the semantic distribution of landslides in the target domain, a domain-adapted landslide recognition optimization model F2 is obtained.

[0020] Furthermore, in step S3, the remote sensing image dataset D of the source domain is S , and the target domain drone image dataset D T The input is fed into model F2, and the domain adaptation model F3 is obtained by reducing the distribution difference between the source domain and the target domain in the feature space based on the covariance alignment mechanism:

[0021] S31, the remote sensing image dataset D of the source domain S , and the target domain drone image dataset D T Input to model F2, through the feature extraction module of model F2, respectively from the dataset D S 、D T Extract the source domain features f s , and target domain features f T ;

[0022] S32, for each predefined semantic category, calculate its feature f in the source domain s The covariance matrix C in sc , and in the target domain feature f T The covariance matrix C in tc ;

[0023] S33, the covariance matrix C sc 、C tcSubstitute it into the predefined category-conditional covariance alignment loss function for backpropagation calculation, and obtain the domain adaptation model F3 by updating the parameters of model F2 to minimize the distribution difference between the source domain and the target domain in semantic categories.

[0024] Furthermore, in step S33, the category conditional covariance alignment loss function L C-CORAL As shown below:

[0025]

[0026] Where C represents the total number of predefined semantic categories, d represents the feature dimension, represents the square of the Frobenius norm.

[0027] Furthermore, in step S4, the specified parameters include all parameters or decoder parameters in model F3 and the last multi-layer encoder parameters.

[0028] Furthermore, in step S4, during fine-tuning of the specified parameters in model F3, the method further includes: performing pixel-level supervised optimization using the Logits binary cross entropy loss function suitable for pixel-level binary classification segmentation tasks, and dynamically setting the group learning rate based on the model structure using the Adam optimizer, or performing dynamic learning rate adjustment using a learning rate decay strategy to balance the convergence speed and model stability.

[0029] In a second aspect, the present application discloses a landslide identification system based on staged feature adaptive transfer learning, the system comprising a model building module, a basic feature transfer module, a category condition domain feature alignment module, a small sample fine-tuning module, and a landslide identification application module, wherein:

[0030] The model building module is used to build an initial landslide identification model F1 with an encoder-decoder;

[0031] The basic feature migration module is used to migrate the general visual pre-training weights and utilize the remote sensing image dataset D of the source domain S Fine-tune the encoder parameters of model F1 to obtain the domain-adapted landslide identification optimization model F2;

[0032] The category condition domain feature alignment module is used to align the remote sensing image dataset D S , and the target domain drone image dataset D T The input is fed into model F2, and the domain adaptation model F3 is obtained by reducing the distribution difference between the source domain and the target domain in the feature space based on the covariance alignment mechanism.

[0033] The small sample fine-tuning module is used to convert a small amount of labeled target domain small sample data D′T Input into model F3, fine-tune the specified parameters in model F3 through the supervised learning mechanism, and obtain the target domain adaptation model F4;

[0034] The landslide identification application module is used to perform a landslide identification task on the input UAV image to be detected based on the target domain adaptation model F4.

[0035] In a third aspect, the present application discloses a computer storage medium, which is used to store computer execution instructions, and the computer execution instructions are used to execute any of the aforementioned landslide identification methods based on staged feature adaptive transfer learning.

[0036] The present invention has the following beneficial effects:

[0037] 1) By migrating general vision pre-trained weights, which contain a large amount of general vision knowledge, such as basic features such as edges, textures, and shapes, the model has a certain feature extraction capability at the initial stage, avoiding training from scratch, greatly improving the performance of the model in the initial stage, and helping the model better generalize on different datasets;

[0038] 2) Since source domain remote sensing image datasets typically contain rich landslide-related features and annotation information, by fine-tuning the encoder parameters, the model can better adapt to the distribution characteristics of the source domain data and learn a representation that is more consistent with the source domain landslide characteristics, thereby improving the recognition accuracy on the source domain data and making full use of the existing data resources in the source domain;

[0039] 3) Since source and target domain data may differ in terms of acquisition method, resolution, and viewing angle, resulting in different data distributions, the covariance alignment mechanism adjusts the feature distribution to make the features learned by the model more similar in the source and target domains. This allows the model to better utilize the knowledge learned in the source domain when processing target domain data, improving the model's adaptability and recognition capabilities for data from different domains and achieving domain adaptation.

[0040] 4) Combining a small amount of labeled data from the target domain not only utilizes a large amount of data and previously learned knowledge from the source domain, but also fully considers the particularity of the target domain. Through effective knowledge transfer and feature alignment, the model converges extremely quickly in the small sample fine-tuning stage, achieving stable performance in only 5-10 training rounds, which is much faster than the 40-50 rounds required for traditional training from scratch, saving a lot of training time and computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 A flow chart of a landslide identification method based on staged feature adaptive transfer learning provided by one embodiment of the present invention;

[0043] Figure 2 A system structure diagram of a landslide identification system based on staged feature adaptive transfer learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0044] To further illustrate the technical means and effectiveness of the present invention to achieve its intended objectives, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, structure, features, and effectiveness of a landslide identification method and system based on staged feature adaptive transfer learning. In the following description, references to "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0045] Unless defined otherwise, 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 invention belongs.

[0046] The following describes in detail a specific solution of a landslide identification method and system based on staged feature adaptive transfer learning provided by the present invention with reference to the accompanying drawings.

[0047] See also Figure 1 , which shows a flow chart of a landslide identification method based on staged feature adaptive transfer learning provided by one embodiment of the present invention, the method comprising:

[0048] Step S1: construct an initial landslide identification model F1 with an encoder-decoder.

[0049] Specifically, this application chooses to build a semantic segmentation network with an encoder-decoder structure as the basic model, in which the encoder uses a pre-trained convolutional neural network to extract multi-scale features, and the decoder fuses low-level and high-level features through jump connections to restore spatial details.

[0050] Step S2, by migrating the general visual pre-training weights and using the remote sensing image dataset D of the source domainS Fine-tune the encoder parameters of model F1 to obtain the domain-adapted landslide identification optimization model F2.

[0051] Specifically, this application loads the encoder weights pre-trained on a large-scale general image dataset (such as ImageNet) into the encoder part of model F1, and freezes some of the underlying weights to retain general feature extraction capabilities (such as edges and textures). The purpose is to use general visual knowledge to migrate and initialize the underlying parameters of the encoder. Afterwards, in order to adapt the model to the spatial resolution and spectral characteristics unique to remote sensing images, this application further selects a source domain remote sensing image dataset related to the landslide identification task and with relatively sufficient data volume (such as a medium-resolution satellite or aerial remote sensing landslide image dataset). Using this dataset and its corresponding Mask landslide label, the encoder parameters of model F1 are further pre-trained or fine-tuned to enhance the model's ability to understand the spatial semantics of landslide areas.

[0052] Step S3: The remote sensing image dataset D of the source domain S , and the target domain drone image dataset D T The input is input into model F2, and the domain adaptation model F3 is obtained by reducing the distribution difference between the source domain and the target domain in the feature space based on the covariance alignment mechanism.

[0053] Specifically, in order to solve the domain differences between the source domain remote sensing images and the target domain drone images, especially the differences in feature distribution for distinguishing landslide and non-landslide categories, this application adopts the category-conditional covariance alignment loss. By calculating the covariance matrix of the landslide / non-landslide category features in the source domain and the target domain, and minimizing the weighted Frobenius norm distance between the two, the back propagation optimizes the network parameters, so that the second-order statistics (covariance) of the features of the corresponding categories in the source domain and the target domain are as close as possible.

[0054] Step S4: a small amount of labeled target domain small sample data D′ T The input is sent to model F3, and the specified parameters in model F3 are fine-tuned through the supervised learning mechanism to obtain the target domain adaptation model F4.

[0055] Specifically, here is a small amount of labeled target domain small sample data D′ T It can be considered as covering 100 accurately labeled drone images and their landslide masks. “A small amount” is not limited to the absolute number (such as 100 samples), but the relative sparsity relative to the scale of the source domain dataset (such as 10,000 remote sensing images in the source domain). TAfter inputting into model F3, by fine-tuning all or part of the parameters of model F3 (such as the decoder and the last few layers of encoders), combining it with a loss function suitable for the pixel-level binary classification task, and setting the group learning rate and using the learning rate decay mechanism to prevent overfitting, the target domain adaptation model F4 can be obtained.

[0056] Step S5: Based on the target domain adaptation model F4, a landslide identification task is performed on the input UAV image to be detected.

[0057] Specifically, the final model F4 obtained after the above three-stage training is applied to the new drone image to be detected. The model outputs a pixel-level probability map. By setting an adaptive threshold (such as 0.5) or morphological post-processing, the final landslide area segmentation mask (i.e., recognition result) can be obtained.

[0058] As can be seen from the above, the present application discloses a landslide recognition method based on staged feature adaptive transfer learning. By transferring general visual pre-training weights, these pre-training weights contain a large amount of general visual knowledge, such as basic features such as edges, textures, and shapes, so that the model has a certain feature extraction capability in the initial stage, avoiding training from scratch, greatly improving the performance of the model in the initial stage, and helping the model to better generalize on different data sets; since the source domain remote sensing image dataset usually contains rich landslide-related features and annotation information, by fine-tuning the encoder parameters, the model can better adapt to the distribution characteristics of the source domain data and learn a representation that is more consistent with the source domain landslide characteristics, thereby improving the recognition accuracy on the source domain data and making full use of the existing data resources in the source domain; since the source domain and target domain data may differ in acquisition method, resolution, viewing angle, etc., resulting in different data distribution. Based on the covariance alignment mechanism, by adjusting the feature distribution, the features learned by the model are made more similar in the source domain and the target domain, so that the model can better utilize the knowledge learned in the source domain when processing the target domain data, improving the model's adaptability and recognition ability on data in different domains, and realizing domain adaptation; combined with a small amount of labeled data in the target domain, it not only utilizes a large amount of data and previously learned knowledge in the source domain, but also fully considers the particularity of the target domain. Through effective knowledge transfer and feature alignment, the model converges extremely quickly in the small sample fine-tuning stage, and only 5-10 training rounds are required to achieve stable performance, which is much faster than the 40-50 rounds required for traditional training from scratch, saving a lot of training time and computing resources.

[0059] In one embodiment, in step S1, the encoder part of the initial landslide recognition model F1 uses a ResNet network as a backbone feature extraction network to utilize the powerful feature extraction capability of the ResNet network and alleviate the gradient vanishing problem.

[0060] Specifically, the ResNet network introduces a residual block structure that enables learning the residual (i.e., the difference) between input and output, rather than directly learning the input-to-output mapping. This design allows the network to be stacked deeper while still maintaining good feature extraction capabilities. In addition, ResNet uses residual connections to transfer shallow layer features directly to deep layers, allowing gradients to propagate more efficiently through these connections, thereby alleviating the gradient vanishing problem and making it possible to train deep networks.

[0061] In one embodiment, in step S1, an ASSP module is integrated between the encoder and decoder of the initial landslide identification model F1, or in the decoder of the initial landslide identification model F1, so as to effectively expand the receptive field and obtain richer context information by using dilated convolutions with different dilation rates in parallel without increasing the amount of computation.

[0062] Specifically, in the task of landslide identification, the identification of landslides often depends on the surrounding contextual information (such as terrain, vegetation, water bodies, etc.). The ASPP module expands the receptive field, allowing the model to focus on local details and global contextual information at the same time, thereby improving the accuracy of landslide identification. It should be noted that traditional pyramid pooling methods (such as SPP) require downsampling and upsampling operations to capture multi-scale information, which increases the amount of computation. The ASPP module operates directly on the original image size through dilated convolution, avoiding the complex downsampling and upsampling process, thereby achieving the capture of multi-scale information while maintaining computational efficiency.

[0063] In one embodiment, in step S2, the method transfers the general visual pre-training weights and utilizes the remote sensing image dataset D of the source domain. S Fine-tune the encoder parameters of model F1 to obtain the domain-adapted landslide identification optimization model F2, including:

[0064] In step S21 , the encoder weights pre-trained on a large-scale general image dataset are loaded into the encoder part of the model F1 , and the encoder parameters of the model F1 are initialized by utilizing general visual knowledge to accelerate the landslide feature learning process.

[0065] Specifically, this application chooses to load the encoder weights pre-trained on a large-scale general image dataset (such as ImageNet) into the encoder part of model F1, and utilizes the rich low-level and mid-level visual features (such as edges, textures, simple shapes, etc.) learned in the pre-trained weights to quickly initialize the encoder parameters of model F1, so that it has better feature extraction capabilities at the beginning of training.

[0066] Step S22: using the remote sensing image dataset D in the source domainS , and its corresponding landslide mask labels are used to fine-tune the encoder parameters of the model F1. Through pre-training in the intermediate domain to narrow the semantic gap between the semantic distribution of landslides in the target domain, a domain-adapted landslide recognition optimization model F2 is obtained.

[0067] Specifically, this application selects a medium-resolution satellite or aerial remote sensing landslide image dataset as the remote sensing image dataset D of the source domain. S , leveraging its large-scale data samples and rich feature information covering a variety of terrains, landforms, and landslide types, intermediate domain pre-training is used to narrow the semantic gap between the target domain and the landslide semantic distribution. It should be noted that large-scale data samples can provide more diverse landslide manifestations, and rich feature information helps the model learn more comprehensive and generalizable landslide features. This allows the model to better adapt to landslide semantics in different scenarios during intermediate domain pre-training, narrowing the gap with the target domain semantic distribution.

[0068] In one embodiment, in step S3, the remote sensing image dataset D of the source domain is S , and the target domain drone image dataset D T The input is fed into model F2, and the domain adaptation model F3 is obtained by reducing the distribution difference between the source domain and the target domain in the feature space based on the covariance alignment mechanism:

[0069] Step S31: The remote sensing image dataset D of the source domain S , and the target domain drone image dataset D T Input to model F2, through the feature extraction module of model F2, respectively from the dataset D S 、D T Extract the source domain features f s , and target domain features f T .

[0070] Specifically, in the feature extraction process, the feature extraction module of model F2 first performs feature extraction on the input source domain remote sensing image dataset D S And the target domain drone image dataset D T Perform preliminary pixel-level processing to convert the original image data into a numerical form suitable for subsequent feature learning. Then, through layer-by-layer convolution operations, the features of different levels in the image are gradually mined. Finally, the dataset D S Extract the source domain features f s , from the dataset D T Extract the target domain features f T , which reflect the unique characteristics of landslides in source and target domain images, respectively.

[0071] Step S32: for each predefined semantic category, calculate and obtain its feature f in the source domain. s The covariance matrix C in sc , and in the target domain feature f T The covariance matrix C in tc .

[0072] Specifically, calculate the source domain feature f for each predefined semantic category s The covariance matrix C in sc And in the target domain feature f T The covariance matrix C in tc The purpose is to quantify the difference in feature distribution between the source domain and the target domain under this semantic category. The covariance matrix can reflect the correlation between different dimensions in the feature vector. By comparing C sc and C tc , we can understand whether the changing trends and correlations of features under the same semantic category in different dimensions in the source domain and the target domain are consistent.

[0073] Step S33: convert the covariance matrix C sc 、C tc Substitute it into the predefined category-conditional covariance alignment loss function for backpropagation calculation, and obtain the domain adaptation model F3 by updating the parameters of model F2 to minimize the distribution difference between the source domain and the target domain in semantic categories.

[0074] Specifically, the predefined category-conditional covariance alignment loss function is constructed based on the difference measure of the covariance matrix. Its core idea is to calculate the source domain covariance matrix C sc and the target domain covariance matrix C tc The distance metric between them is used to measure the distribution difference between the two in semantic categories.

[0075] In the above embodiment, by inputting source domain remote sensing images and target domain drone images into the model to extract features, and calculating the covariance matrix of each semantic category in the source domain and target domain features, and then using the category conditional covariance alignment loss function to perform backpropagation to update the model parameters, the feature distribution difference between the source domain and the target domain in semantic categories is effectively narrowed, so that the model learns a more domain-invariant feature representation, thereby improving the generalization ability of the model in the target domain, enhancing the model's recognition accuracy and robustness for targets such as landslides in image data from different sources (such as remote sensing images and drone images), and reducing the performance degradation problem caused by domain differences.

[0076] In one embodiment, in step S33, the class conditional covariance alignment loss function L C-TORAL As shown below:

[0077]

[0078] Where C represents the total number of predefined semantic categories, d represents the feature dimension, represents the square of the Frobenius norm.

[0079] In one embodiment, in step S4, the specified parameters include all parameters or decoder parameters in model F3 and the last multi-layer encoder parameters.

[0080] Specifically, the selection of all parameters in model F3, or the decoder parameters, and the final multi-layer encoder parameters as designated parameters for fine-tuning is based on different considerations. Fine-tuning all parameters in model F3 has the advantage of enabling the model to fully leverage all knowledge learned during joint training of source and target domain data, comprehensively optimizing the model's capabilities for feature extraction, feature fusion, and classification prediction. This comprehensive fine-tuning approach is suitable for situations where the data distribution differences between the source and target domains are relatively small, or when the target domain data volume is relatively abundant. It allows the model to better adapt to the target domain task and improve overall performance in the target domain. Fine-tuning the decoder parameters and the final multi-layer encoder parameters is primarily based on the consideration that the first few layers of the encoder typically learn relatively common, low-level features such as edges and textures. These features have certain commonalities across data from different domains and do not require significant adjustments. The decoder is responsible for upsampling and classifying the features extracted by the encoder, while the final multi-layer encoder involves higher-level feature representation and semantic information extraction, which is more closely related to the target task (such as pixel-level segmentation in landslide identification). By fine-tuning only these parameters, the model can adapt to the characteristics of the target domain data more quickly while retaining the encoder's general feature extraction capabilities, reducing the risk of overfitting. This is especially suitable for scenarios where the amount of target domain data is limited or the data distribution between the source and target domains differs greatly.

[0081] In one embodiment, in step S4, during fine-tuning of specified parameters in model F3, the method further includes: performing pixel-level supervised optimization using a Logits binary cross entropy loss function suitable for pixel-level binary classification segmentation tasks, and dynamically setting the group learning rate based on the model structure using an Adam optimizer, or performing dynamic learning rate adjustment using a learning rate decay strategy to balance convergence speed and model stability.

[0082] Specifically, the Logits binary cross entropy loss function, suitable for pixel-level binary classification tasks, is used for pixel-level supervised optimization because in pixel-level binary classification tasks such as landslide identification, each pixel needs to be accurately classified as a landslide or non-landslide. The Logits binary cross entropy loss function can directly measure the difference between the pixel category probability distribution predicted by the model and the true label distribution. By minimizing this loss function, the model can be guided to continuously adjust parameters so that the prediction results are closer to the actual situation, thereby effectively improving the accuracy of pixel-level classification. Furthermore, the Adam optimizer is used in this application to update model parameters because the Adam optimizer combines the advantages of the momentum method and adaptive learning rate. It can automatically adjust the learning rate based on the historical gradient information of each parameter, assigning different learning rates to different parameters, so that the model can converge more efficiently during training. Furthermore, dynamically setting the group learning rate based on the model structure is a refined learning rate adjustment strategy. Because different parts of the model structure have different sensitivities to parameter updates during training, for example, the decoder part may be more sensitive to changes in the learning rate, while the encoder part may be relatively stable. By dividing model parameters into different groups and setting a different learning rate for each group, we can more precisely control the update speed of each parameter, making the model more stable during training while accelerating convergence. Finally, it should be noted that the use of a learning rate decay strategy for dynamic learning rate adjustment aims to balance convergence speed and model stability during training. In the early stages of training, using a larger learning rate can accelerate the update of model parameters and help the model quickly approach the optimal solution. Gradually reducing the learning rate as training progresses can prevent the model from oscillating around the optimal solution, improving the model's convergence accuracy and stability.

[0083] In summary, the core technical innovations of this application are:

[0084] 1) Systematic phased transfer learning strategy: The transfer learning process for landslide identification is innovatively decomposed into three organically integrated stages: basic feature transfer (general + intermediate domain) → category condition domain alignment → small sample fine-tuning, achieving gradual and refined adaptation of knowledge;

[0085] 2) Introducing Intermediate Domain Pre-training: Using remote sensing landslide data, which is more relevant to the target task but has a larger data volume, as the intermediate domain for pre-training effectively bridges the semantic gap between general data and target UAV data, providing better initialization for subsequent alignment and fine-tuning;

[0086] 3) Application of Class-Conditional Feature Alignment: We specifically use the Class-Wise CORAL Loss for domain adaptation, aligning landslide and non-landslide features at the category level. This effectively alleviates the problems of inconsistent category distribution and feature confusion in cross-domain data, significantly improving the model's discriminative ability in complex contexts.

[0087] 4) Synergy between architecture and strategy: The phased migration strategy is combined with the ResUNet+ASPP network architecture suitable for handling complex boundaries and multi-scale targets, giving full play to the synergistic effect of structural optimization and training strategy optimization.

[0088] 5) Efficient learning with small samples: Through staged knowledge injection and feature alignment, the model can converge quickly and achieve high accuracy even with only a small number of labeled samples in the target domain.

[0089] In summary, this application effectively solves the problems of low accuracy, poor efficiency, and blurred boundaries in small-sample landslide identification in drone images through an innovative phased transfer learning strategy, combined with network structure optimization and category condition feature alignment. It has the advantages of high accuracy, high efficiency, and strong practicality, and has important application value in the field of intelligent monitoring and early warning of geological disasters.

[0090] Please refer to Figure 2 The present application discloses a landslide identification system based on staged feature adaptive transfer learning, which includes a model construction module, a basic feature transfer module, a category condition domain feature alignment module, a small sample fine-tuning module, and a landslide identification application module, wherein:

[0091] The model building module is used to build an initial landslide identification model F1 with an encoder-decoder.

[0092] The basic feature migration module is used to migrate the general visual pre-training weights and utilize the remote sensing image dataset D of the source domain S Fine-tune the encoder parameters of model F1 to obtain the domain-adapted landslide identification optimization model F2.

[0093] The category condition domain feature alignment module is used to align the remote sensing image dataset D S , and the target domain drone image dataset D T The input is input into model F2, and the domain adaptation model F3 is obtained by reducing the distribution difference between the source domain and the target domain in the feature space based on the covariance alignment mechanism.

[0094] The small sample fine-tuning module is used to convert a small amount of labeled target domain small sample data D′ TThe input is sent to model F3, and the specified parameters in model F3 are fine-tuned through the supervised learning mechanism to obtain the target domain adaptation model F4.

[0095] The landslide identification application module is used to perform a landslide identification task on the input UAV image to be detected based on the target domain adaptation model F4.

[0096] In one embodiment, the above modules are also used to implement a landslide identification method based on staged feature adaptive transfer learning as described in any one of the above method embodiments, which is not limited in this application.

[0097] As can be seen from the above, the present application discloses a landslide recognition system based on staged feature adaptive transfer learning. By transferring general visual pre-training weights, these pre-training weights contain a large amount of general visual knowledge, such as basic features such as edges, textures, and shapes, so that the model has a certain feature extraction capability in the initial stage, avoiding training from scratch, greatly improving the performance of the model in the initial stage, and helping the model to better generalize on different data sets; since the source domain remote sensing image data set usually contains rich landslide-related features and annotation information, by fine-tuning the encoder parameters, the model can better adapt to the distribution characteristics of the source domain data and learn a representation that is more consistent with the source domain landslide characteristics, thereby improving the recognition accuracy on the source domain data and making full use of the existing data resources in the source domain; since the source domain and target domain data may differ in acquisition method, resolution, viewing angle, etc., resulting in different data distribution. Based on the covariance alignment mechanism, by adjusting the feature distribution, the features learned by the model are made more similar in the source domain and the target domain, so that the model can better utilize the knowledge learned in the source domain when processing the target domain data, improving the model's adaptability and recognition ability on data in different domains, and realizing domain adaptation; combined with a small amount of labeled data in the target domain, it not only utilizes a large amount of data and previously learned knowledge in the source domain, but also fully considers the particularity of the target domain. Through effective knowledge transfer and feature alignment, the model converges extremely quickly in the small sample fine-tuning stage, and only 5-10 training rounds are required to achieve stable performance, which is much faster than the 40-50 rounds required for traditional training from scratch, saving a lot of training time and computing resources.

[0098] Furthermore, the present application discloses a computer storage medium, which is used to store computer-executable instructions, and the computer-executable instructions are used to execute the landslide identification method based on staged feature adaptive transfer learning described in any of the aforementioned embodiments.

[0099] As can be seen from the above, a computer storage medium disclosed in the present application, by migrating general visual pre-training weights, these pre-training weights contain a large amount of general visual knowledge, such as basic features such as edges, textures, and shapes, so that the model has a certain feature extraction capability in the initial stage, avoiding training from scratch, greatly improving the performance of the model in the initial stage, and helping the model to better generalize on different data sets; since the source domain remote sensing image data set usually contains rich landslide-related features and annotation information, by fine-tuning the encoder parameters, the model can better adapt to the distribution characteristics of the source domain data and learn a representation that is more consistent with the source domain landslide characteristics, thereby improving the recognition accuracy on the source domain data and making full use of the existing data resources in the source domain; since the source domain and target domain data may differ in acquisition methods, resolution, viewing angle, etc., resulting in different data distributions. Based on the covariance alignment mechanism, by adjusting the feature distribution, the features learned by the model are made more similar in the source domain and the target domain, so that the model can better utilize the knowledge learned in the source domain when processing the target domain data, improving the model's adaptability and recognition ability on data in different domains, and realizing domain adaptation; combined with a small amount of labeled data in the target domain, it not only utilizes a large amount of data and previously learned knowledge in the source domain, but also fully considers the particularity of the target domain. Through effective knowledge transfer and feature alignment, the model converges extremely quickly in the small sample fine-tuning stage, and only 5-10 training rounds are required to achieve stable performance, which is much faster than the 40-50 rounds required for traditional training from scratch, saving a lot of training time and computing resources.

[0100] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0101] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

Claims

1. A landslide identification method based on staged feature adaptive transfer learning, characterized in that: The method comprises: S1, constructing an initial landslide identification model F1 with an encoder-decoder; S2, by migrating general visual pre-training weights and utilizing the remote sensing image dataset D of the source domain S Fine-tune the encoder parameters of model F1 to obtain the domain-adapted landslide identification optimization model F2; S3, the remote sensing image dataset D of the source domain S , and the target domain drone image dataset D T The input is fed into model F2, and the domain adaptation model F3 is obtained by reducing the distribution difference between the source domain and the target domain in the feature space based on the covariance alignment mechanism. S4, a small amount of labeled target domain small sample data D' T Input into model F3, fine-tune the specified parameters in model F3 through the supervised learning mechanism, and obtain the target domain adaptation model F4; S5. Based on the target domain adaptation model F4, a landslide identification task is performed on the input UAV image to be detected.

2. The method according to claim 1, characterized in that In step S1, the encoder part of the initial landslide recognition model F1 adopts the ResNet network as the backbone feature extraction network to utilize the powerful feature extraction capability of the ResNet network to alleviate the gradient vanishing problem.

3. The method according to claim 1, characterized in that In step S1, an ASSP module is integrated between the encoder and decoder of the initial landslide identification model F1 or in the decoder of the initial landslide identification model F1 to effectively expand the receptive field and obtain richer context information by using dilated convolutions with different dilation rates in parallel without increasing the amount of computation.

4. The method according to claim 1, wherein In step S2, the method transfers the general visual pre-training weights and utilizes the remote sensing image dataset D of the source domain. S Fine-tune the encoder parameters of model F1 to obtain the domain-adapted landslide identification optimization model F2, including: S21, loading the encoder weights pre-trained on a large-scale general image dataset into the encoder part of model F1, and accelerating the landslide feature learning process by utilizing general visual knowledge to initialize the encoder parameters of model F1; S22. Using the remote sensing image dataset D in the source domain S , and its corresponding landslide mask labels are used to fine-tune the encoder parameters of the model F1. Through pre-training in the intermediate domain to narrow the semantic gap between the semantic distribution of landslides in the target domain, a domain-adapted landslide recognition optimization model F2 is obtained.

5. The method according to claim 1, characterized in that In step S3, the remote sensing image dataset D of the source domain is S , and the target domain drone image dataset D T The input is fed into model F2, and the domain adaptation model F3 is obtained by reducing the distribution difference between the source domain and the target domain in the feature space based on the covariance alignment mechanism: S31, the remote sensing image dataset D of the source domain S , and the target domain drone image dataset D T Input to model F2, through the feature extraction module of model F2, respectively from the dataset D S 、D T Extract the source domain features f s , and target domain features f T ; S32, for each predefined semantic category, calculate its feature f in the source domain s The covariance matrix C in sc , and in the target domain feature f T The covariance matrix C in tc ; S33, the covariance matrix C sc 、C tc Substitute it into the predefined category-conditional covariance alignment loss function for backpropagation calculation, and obtain the domain adaptation model F3 by updating the parameters of model F2 to minimize the distribution difference between the source domain and the target domain in semantic categories.

6. The method according to claim 5, characterized in that In step S33, the class conditional covariance alignment loss function L C-CORAL As shown below: Where C represents the total number of predefined semantic categories, d represents the feature dimension, represents the square of the Frobenius norm.

7. The method according to claim 1, characterized in that In step S4, the specified parameters include all parameters or decoder parameters in model F3 and the last multi-layer encoder parameters.

8. The method according to claim 1, characterized in that In step S4, during the process of fine-tuning the specified parameters in model F3, the method further includes: performing pixel-level supervised optimization using the Logits binary cross entropy loss function suitable for pixel-level binary classification segmentation tasks, and using the Adam optimizer to dynamically set the group learning rate based on the model structure, or using a learning rate decay strategy to dynamically adjust the learning rate to balance the convergence speed and model stability.

9. A landslide identification system based on staged feature adaptive transfer learning, characterized in that: The system includes a model building module, a basic feature migration module, a category condition domain feature alignment module, a small sample fine-tuning module, and a landslide identification application module, wherein: The model building module is used to build an initial landslide identification model F1 with an encoder-decoder; The basic feature migration module is used to migrate the general visual pre-training weights and utilize the remote sensing image dataset D of the source domain S Fine-tune the encoder parameters of model F1 to obtain the domain-adapted landslide identification optimization model F2; The category condition domain feature alignment module is used to align the remote sensing image dataset D S , and the target domain drone image dataset D T The input is fed into model F2, and the domain adaptation model F3 is obtained by reducing the distribution difference between the source domain and the target domain in the feature space based on the covariance alignment mechanism. The small sample fine-tuning module is used to convert a small amount of labeled target domain small sample data D′ T Input into model F3, fine-tune the specified parameters in model F3 through the supervised learning mechanism, and obtain the target domain adaptation model F4; The landslide identification application module is used to perform a landslide identification task on the input UAV image to be detected based on the target domain adaptation model F4.

10. A computer storage medium, characterized in that The computer storage medium is used to store computer-executable instructions, and the computer-executable instructions are used to execute the landslide identification method based on staged feature adaptive transfer learning according to any one of claims 1 to 8.

Images