Landslide identification method and system based on staged feature adaptive transfer learning
By employing a phased feature adaptive transfer learning method and utilizing general visual pre-trained weights and covariance alignment mechanisms, the problem of sample scarcity and domain gap in landslide identification from UAV remote sensing images was solved, achieving efficient and accurate landslide identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2025-05-09
- Publication Date
- 2026-06-09
AI Technical Summary
Existing deep learning models face challenges in landslide identification from UAV remote sensing images, including scarce samples, high annotation costs, domain gaps, cross-source domain feature mismatches, and insufficient model processing capabilities for complex features, resulting in low identification accuracy and inefficiency.
A phased feature adaptive transfer learning method is adopted. By transferring general visual pre-trained weights, fine-tuning the source domain remote sensing image dataset, covariance alignment mechanism, and a small amount of target domain labeled data, a landslide identification model is constructed to achieve domain adaptation and feature alignment.
It improves the accuracy and efficiency of landslide identification, the model converges quickly under small sample conditions, saves training time and computing resources, and is adaptable to data from different fields.
Smart Images
Figure CN120673281B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of artificial intelligence and geological disaster monitoring and early warning, specifically to a landslide identification method and system based on phased feature adaptive transfer learning. Background Technology
[0002] Landslides are highly destructive geological hazards, and their accurate and rapid identification is crucial for disaster prevention and mitigation. Currently, UAV remote sensing has become an important means of acquiring imagery of landslide areas due to its high resolution, high timeliness, and flexibility. Although deep learning models based on convolutional neural networks or Transformer architectures have made significant progress in the field of computer vision, directly processing UAV imagery using deep learning models for landslide identification still faces the following serious challenges:
[0003] 1) Scarce samples and high labeling costs: Acquiring and accurately labeling a large number of drone landslide image samples is costly and time-consuming, resulting in a severe shortage of training data (target domain data). Deep learning models are difficult to train fully, are prone to overfitting, have poor generalization ability, and their recognition accuracy drops significantly under small sample conditions.
[0004] 2) Domain gap: Although transfer learning has been introduced to reduce dependence on target domain data, commonly used transfer strategies (such as using ImageNet pre-trained models) still have a significant domain gap problem due to the fundamental differences between the source domain (natural images) and the target domain (landslide remote sensing images) in terms of data distribution and feature space.
[0005] 3) Cross-source domain feature mismatch: Even if source domain data (such as satellite / aerial remote sensing images) that are more relevant to the task are introduced for migration, there are still significant differences between these data and the UAV images of the target domain in terms of resolution, illumination, and terrain details. This makes it difficult to align the feature spaces of the source domain and the target domain, and the model is prone to negative migration when migrating across domains.
[0006] 4) Insufficient ability of the model to handle complex features: Landslides have diverse morphologies, varying scales, and blurred boundaries, and are easily affected by vegetation, shadows, etc. Existing models (such as U-Net) use single-scale convolutional kernels or fixed receptive fields, making it difficult to capture multi-scale features of landslides (such as overall morphology and local cracks) and accurately delineate boundaries. Furthermore, they are not robust enough to noise interference (such as vegetation occlusion), and 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 finite target domain (UAV) labeled samples, make full use of relevant source domain (such as remote sensing) knowledge, accurately align cross-domain features (especially features that distinguish landslides from the background), and possess powerful multi-scale feature extraction capabilities, in order to solve the problems of low identification accuracy and low efficiency under small sample conditions. Summary of the Invention
[0008] To address the issues of low accuracy and efficiency in landslide identification under small sample conditions, this invention aims to provide a landslide identification method and system based on staged feature adaptive transfer learning. The specific technical solution adopted is as follows:
[0009] Firstly, this application discloses a landslide identification method based on staged feature adaptive transfer learning, the method comprising:
[0010] S1. Construct an initial landslide identification model F1 with an encoder-decoder;
[0011] S2, by transferring general visual pre-trained weights and utilizing the source domain remote sensing image dataset D S Fine-tuning the encoder parameters of model F1 yields a domain-adapted optimized landslide identification model F2.
[0012] S3. Transfer the remote sensing image dataset D from the source domain. S And the UAV imagery dataset D for the target domain T The input is fed into model F2, and based on the covariance alignment mechanism, the distribution difference between the source domain and the target domain in the feature space is reduced to obtain the domain adaptive model F3;
[0013] S4. Transfer a small amount of labeled target domain small sample data D′ T The input is fed into model F3, and the specified parameters in model F3 are fine-tuned through a supervised learning mechanism to obtain the target domain-adapted model F4;
[0014] S5. Based on the target domain adaptation model F4, perform a landslide identification task on the input UAV image to be detected.
[0015] Furthermore, in step S1, the encoder part of the initial landslide identification model F1 adopts a 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, to effectively expand the receptive field and obtain richer contextual information by using dilated convolutions with different dilation rates in parallel without increasing the amount of computation.
[0017] Furthermore, in step S2, the transfer of general visual pre-trained weights and the utilization of the source domain remote sensing image dataset D... S Fine-tuning the encoder parameters of model F1 yields a domain-adapted optimized landslide identification model F2, including:
[0018] S21. The pre-trained encoder weights on a large-scale general image dataset are loaded into the encoder part of model F1. By utilizing general visual knowledge to initialize the encoder parameters of model F1, the landslide feature learning process is accelerated.
[0019] S22. Utilize the remote sensing image dataset D from the source domain. S The encoder parameters of the corresponding landslide mask label fine-tuning model F1 are obtained by pre-training in the intermediate domain to narrow the semantic gap between the landslide semantic distribution in the target domain and the target domain, thus obtaining the domain-adapted landslide recognition optimization model F2.
[0020] Furthermore, in step S3, the remote sensing image dataset D from the source domain... S And the UAV imagery dataset D for the target domain T The input is fed into model F2, and based on the covariance alignment mechanism, the distribution difference between the source domain and the target domain in the feature space is reduced to obtain the domain adaptive model F3:
[0021] S31. Transfer the remote sensing image dataset D from the source domain. S And the UAV imagery dataset D for the target domain T The input is fed into model F2, and through the feature extraction module of model F2, the data is extracted from dataset D. S D T Extracting 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 the target domain feature f T The covariance matrix C in tc ;
[0023] S33, the covariance matrix C sc C tcSubstituting the values into the predefined category-conditional covariance alignment loss function for backpropagation calculation, the parameters of model F2 are updated to minimize the distribution difference between the source and target domains in semantic categories, thus obtaining the domain-adaptive model F3.
[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, and d represents the feature dimension. This represents the square of the Frobenius norm.
[0027] Furthermore, in step S4, the specified parameters include all parameters in model F3 or decoder parameters and the last multi-layer encoder parameters.
[0028] Furthermore, in step S4, during the 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 and 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, in order to balance convergence speed and model stability.
[0029] Secondly, this application discloses a landslide recognition system based on staged feature adaptive transfer learning. The system includes a model building module, a basic feature transfer module, a category-conditional domain feature alignment module, a few-sample fine-tuning module, and a landslide recognition 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 transfer module is used to transfer general visual pre-trained weights and utilize the remote sensing image dataset D from the source domain. S Fine-tuning the encoder parameters of model F1 yields a domain-adapted optimized landslide identification model F2.
[0032] The category-conditional domain feature alignment module is used to align the remote sensing image dataset D from the source domain. S And the UAV imagery dataset D for the target domain T The input is fed into model F2, and based on the covariance alignment mechanism, the distribution difference between the source domain and the target domain in the feature space is reduced to obtain the domain adaptive model F3;
[0033] The small sample fine-tuning module is used to adjust a small amount of labeled target domain small sample data D′.T The input is fed into model F3, and the specified parameters in model F3 are fine-tuned through a supervised learning mechanism to obtain the target domain-adapted model F4;
[0034] The landslide recognition application module is used to perform landslide recognition tasks on the input UAV images to be detected based on the target domain adaptation model F4.
[0035] Thirdly, this application discloses a computer storage medium for storing computer execution instructions, which are used to execute the landslide identification method based on phased feature adaptive transfer learning as described in any of the preceding claims.
[0036] The present invention has the following beneficial effects:
[0037] 1) By transferring general visual pre-trained weights, which contain a large amount of general visual knowledge, such as basic features like edges, textures, and shapes, the model has a certain feature extraction capability in the initial stage, avoiding training from scratch, greatly improving the model's performance in the initial stage, and helping the model to generalize better on different datasets.
[0038] 2) Since source domain remote sensing image datasets usually contain rich landslide-related features and annotation information, by fine-tuning the encoder parameters, the model can better adapt to the distribution characteristics of source domain data, learn a representation that is more consistent with the characteristics of landslides in the source domain, thereby improving the recognition accuracy on source domain data and making full use of the existing data resources in the source domain.
[0039] 3) Due to potential differences in acquisition methods, resolutions, and perspectives between the source and target domains, the data distributions may differ. The covariance alignment mechanism adjusts the feature distribution to make the features learned by the model more similar in both 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 its adaptability and recognition capabilities across different domains and achieving domain adaptation.
[0040] 4) By combining a small amount of labeled data in the target domain, we can make use of the large amount of data in the source domain and the knowledge learned in the early stage, while also taking into full account the special characteristics of the target domain. Through effective knowledge transfer and feature alignment, the model converges very quickly in the small sample fine-tuning stage. It only needs 5-10 training rounds 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. Attached Figure Description
[0041] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 The flowchart illustrates a landslide identification method based on phased feature adaptive transfer learning, as provided in one embodiment of the present invention.
[0043] Figure 2 This is a system architecture diagram of a landslide identification system based on phased feature adaptive transfer learning, provided as an embodiment of the present invention. Detailed Implementation
[0044] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a landslide identification method and system based on staged feature adaptive transfer learning proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0045] 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 invention pertains.
[0046] The following description, in conjunction with the accompanying drawings, details the specific scheme of the landslide identification method and system based on phased feature adaptive transfer learning provided by this invention.
[0047] Please see Figure 1 The diagram illustrates a method flowchart for landslide identification based on staged feature adaptive transfer learning, according to an embodiment of the present invention. The method includes:
[0048] Step S1: Construct an initial landslide identification model F1 with an encoder-decoder.
[0049] Specifically, this application chooses to construct a semantic segmentation network with an encoder-decoder structure as the basic model. The encoder uses a pre-trained convolutional neural network to extract multi-scale features, while the decoder fuses low-level and high-level features through skip connections to recover spatial details.
[0050] Step S2 involves transferring general visual pre-trained weights and utilizing the source domain remote sensing image dataset D.S Fine-tuning the encoder parameters of model F1 yields a domain-adapted optimized landslide identification model F2.
[0051] Specifically, this application loads pre-trained encoder weights from a large-scale general image dataset (such as ImageNet) into the encoder part of model F1, and freezes some low-level weights to retain general feature extraction capabilities (such as edges and textures). The aim is to utilize general visual knowledge for the transfer initialization of the encoder's low-level parameters. Subsequently, to adapt the model to the unique spatial resolution and spectral characteristics of remote sensing images, this application further selects a source domain remote sensing image dataset (such as a medium-resolution satellite or aerial remote sensing landslide image dataset) that is relevant to the landslide recognition task and has a relatively abundant amount of data. Using this dataset and its corresponding Mask landslide labels, the encoder parameters of model F1 are further pre-trained or fine-tuned to enhance the model's spatial semantic understanding of landslide areas.
[0052] Step S3, transfer the remote sensing image dataset D from the source domain. S And the UAV imagery dataset D for the target domain T The input is fed into model F2, and based on the covariance alignment mechanism, the distribution difference between the source domain and the target domain in the feature space is reduced to obtain the domain adaptive model F3.
[0053] Specifically, in order to address the domain differences between source domain remote sensing images and target domain UAV images, especially the differences in feature distributions that distinguish between landslide and non-landslide categories, this application employs category-conditional covariance alignment loss. By calculating the covariance matrix of landslide / non-landslide category features in the source and target domains and minimizing the weighted Frobenius norm distance between the two, the network parameters are optimized through backpropagation, making the second-order statistics (covariance) of the corresponding categories in the source and target domains as close as possible.
[0054] Step S4, process a small amount of labeled target domain small sample data D′ T The input is fed into model F3, and the specified parameters in model F3 are fine-tuned through a supervised learning mechanism to obtain the target domain-adapted model F4.
[0055] Specifically, here we have a small amount of labeled target domain small sample data D′ T This can be considered as encompassing 100 precisely labeled UAV imagery and their landslide masks. "Small sample size" is not limited to an absolute number (e.g., 100 samples), but rather refers to the relative sparsity relative to the size of the source domain dataset (e.g., the source domain contains 10,000 remote sensing images). The target domain small sample data D′ TAfter inputting into model F3, by fine-tuning all or some of the parameters of model F3 (such as the decoder and the last few encoder layers), combining a loss function suitable for pixel-level binary classification segmentation tasks, and by setting group learning rates and using a learning rate decay mechanism to prevent overfitting, the target domain-adapted model F4 can be obtained.
[0056] Step S5: Based on the target domain adaptation model F4, perform a landslide recognition task on the input UAV image to be detected.
[0057] Specifically, the final model F4 obtained after the above three stages of training is applied to new UAV images to be detected. The model outputs a pixel-level probability map, and by setting an adaptive threshold (such as 0.5) or morphological post-processing, the final landslide area segmentation mask (i.e., the recognition result) can be obtained.
[0058] As can be seen from the above, the landslide recognition method based on phased feature adaptive transfer learning disclosed in this application transfers general visual pre-trained weights. These pre-trained weights contain a large amount of general visual knowledge, such as basic features like edges, textures, and shapes, enabling the model to have a certain feature extraction capability in the initial stage. This avoids training from scratch, greatly improves the model's performance in the initial stage, and helps the model generalize better on different datasets. 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 better matches the landslide characteristics of the source domain, 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 terms of acquisition methods, resolution, and viewing angle, the data distribution may be different. By adjusting the feature distribution through covariance alignment, the features learned by the model are made more similar in the source and target domains. This allows the model to better utilize the knowledge learned in the source domain when processing data in the target domain, improving the model's adaptability and recognition ability in different domains and achieving domain adaptation. By combining a small amount of labeled data in the target domain, the model utilizes a large amount of data in the source domain and the knowledge learned in the early stages, while also fully considering the special characteristics of the target domain. Through effective knowledge transfer and feature alignment, the model converges very quickly in the small sample fine-tuning stage, achieving stable performance in only 5-10 training epochs, which is much faster than the 40-50 epochs 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 identification model F1 uses a ResNet network as the backbone feature extraction network to leverage the powerful feature extraction capabilities of the ResNet network and alleviate the gradient vanishing problem.
[0060] Specifically, the ResNet network introduces a residual block structure, enabling it to learn the residual (i.e., the difference) between the input and output, rather than directly learning the input-to-output mapping. This design allows for deeper network layers to be stacked while maintaining good feature extraction capabilities. Furthermore, ResNet directly propagates features from shallow layers to deeper layers through residual connections, allowing gradients to propagate more effectively through these connections, thus mitigating the vanishing gradient problem and enabling the training of 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, to effectively expand the receptive field and obtain richer contextual information by using dilated convolutions with different dilation rates in parallel without increasing the amount of computation.
[0062] Specifically, in landslide recognition tasks, landslide identification often relies on surrounding contextual information (such as terrain, vegetation, and water bodies). The ASPP module expands the receptive field, enabling the model to simultaneously focus on local details and global contextual information, thereby improving the accuracy of landslide recognition. It's worth noting that traditional pyramid pooling methods (such as SPP) require downsampling and upsampling operations to capture multi-scale information, which increases computational cost. The ASPP module, however, operates directly on the original image size using dilated convolution, avoiding complex downsampling and upsampling processes, thus achieving multi-scale information capture while maintaining computational efficiency.
[0063] In one embodiment, in step S2, the transfer of general visual pre-trained weights and the utilization of the source domain remote sensing image dataset D... S Fine-tuning the encoder parameters of model F1 yields a domain-adapted optimized landslide identification model F2, including:
[0064] Step S21 involves loading the pre-trained encoder weights on a large-scale general image dataset into the encoder part of model F1, thereby accelerating the landslide feature learning process by utilizing general visual knowledge to initialize the encoder parameters of model F1.
[0065] Specifically, this application selects to load pre-trained encoder weights on a large-scale general image dataset (such as ImageNet) into the encoder part of model F1. By utilizing the rich low-level and mid-level visual features (such as edges, textures, simple shapes, etc.) learned in the pre-trained weights, the encoder parameters of model F1 are quickly initialized, enabling it to have good feature extraction capabilities in the early stages of training.
[0066] Step S22, using the remote sensing image dataset D from the source domainS The encoder parameters of the corresponding landslide mask label fine-tuning model F1 are obtained by pre-training in the intermediate domain to narrow the semantic gap between the landslide semantic distribution in the target domain and the target domain, thus obtaining the domain-adapted landslide recognition optimization model F2.
[0067] Specifically, this application selects a medium-resolution satellite or aerial remote sensing landslide image dataset as the source domain's remote sensing image dataset D. S Leveraging its large-scale data samples and rich feature information covering various terrains, landforms, and landslide types, the model is pre-trained in the intermediate domain to narrow the semantic gap between the model and the target domain's landslide semantic distribution. It should be noted that large-scale data samples provide more diverse landslide representations, while rich feature information helps the model learn more comprehensive and generalized landslide features. This allows the model to better adapt to landslide semantics in different scenarios during intermediate domain pre-training, thus bridging the gap with the target domain's semantic distribution.
[0068] In one embodiment, in step S3, the remote sensing image dataset D from the source domain is... S And the UAV imagery dataset D for the target domain T The input is fed into model F2, and based on the covariance alignment mechanism, the distribution difference between the source domain and the target domain in the feature space is reduced to obtain the domain adaptive model F3:
[0069] Step S31, transfer the remote sensing image dataset D from the source domain. S And the UAV imagery dataset D for the target domain T The input is fed into model F2, and through the feature extraction module of model F2, the data is extracted from dataset D. S D T Extracting source domain features f s and target domain features f T .
[0070] Specifically, during the feature extraction process, the feature extraction module of model F2 first processes the input source domain remote sensing image dataset D. S and target domain UAV imagery dataset D T Preliminary pixel-level processing is performed to convert the raw image data into a numerical form suitable for subsequent feature learning. Then, layer-by-layer convolutional operations are used to progressively extract features at different levels within the image. Finally, the data is processed from dataset D. S Extract source domain features f s From dataset D T Extract target domain features f T These images reflect the unique characteristics of landslides in the source and target domains, respectively.
[0071] Step S32: For each predefined semantic category, calculate its feature f in the source domain. s The covariance matrix C in sc and the target domain feature f T The covariance matrix C in tc .
[0072] Specifically, the source domain features f are calculated 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 differences in feature distributions between the source and target domains under a given semantic category. The covariance matrix reflects the correlation between different dimensions of the feature vectors. By comparing C... sc and C tc This allows us to understand whether the changing trends and correlations of features under the same semantic category in the source and target domains are consistent across different dimensions.
[0073] Step S33, the covariance matrix C sc C tc Substituting the values into the predefined category-conditional covariance alignment loss function for backpropagation calculation, the parameters of model F2 are updated to minimize the distribution difference between the source and target domains in semantic categories, thus obtaining the domain-adaptive model F3.
[0074] Specifically, the predefined category-conditional covariance alignment loss function is constructed based on a difference measure of the covariance matrix. Its core idea is to calculate the source domain covariance matrix C. sc Covariance matrix C of the target domain tc The distance between them is measured, and this is used to measure the difference in their semantic categories.
[0075] In the above embodiments, by inputting source domain remote sensing images and target domain UAV images into the model to extract features, and calculating the covariance matrix of each semantic category in the source and target domain features, and then using the category conditional covariance alignment loss function for backpropagation to update the model parameters, the feature distribution difference between the source and target domains in semantic categories is effectively reduced. This enables the model to learn more domain-invariant feature representations, thereby improving the model's generalization ability in the target domain, enhancing the model's accuracy and robustness in identifying targets such as landslides in image data from different sources (such as remote sensing images and UAV images), and reducing the performance degradation caused by domain differences.
[0076] In one embodiment, in step S33, the category-conditional covariance alignment loss function L C-TORAL As shown below:
[0077]
[0078] Where C represents the total number of predefined semantic categories, and d represents the feature dimension. This 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 multilayer encoder parameters.
[0080] Specifically, choosing all parameters in model F3, or the decoder parameters and the final multi-layer encoder parameters as the specified parameters for fine-tuning, is based on considerations at different levels. Choosing to fine-tune all parameters in model F3 has the advantage of allowing the model to fully utilize all the knowledge learned during joint training of source and target domain data, comprehensively optimizing the model's feature extraction, feature fusion, and classification prediction capabilities. This comprehensive fine-tuning approach is suitable when the differences in the distribution of source and target domain data are relatively small, or when the amount of target domain data is sufficient, enabling the model to better adapt to the target domain task and improve its overall performance in the target domain. Choosing to fine-tune the decoder parameters and the final multi-layer encoder parameters is mainly because the first few layers of the encoder typically learn relatively general low-level features, such as edges and textures. These features have certain commonalities across different domain data and do not require significant adjustments. The decoder is responsible for upsampling and classification prediction of the features extracted by the encoder, while the final multi-layer encoder involves higher-level feature representation and semantic information extraction, and is more closely related to the target task (such as pixel-level segmentation in landslide recognition). 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 where there are significant differences in the distribution of source and target domain data.
[0081] In one embodiment, during step S4, when 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, so as to balance convergence speed and model stability.
[0082] Specifically, the Logits binary cross-entropy loss function, suitable for pixel-level binary classification segmentation tasks, is used for pixel-level supervised optimization because in pixel-level binary classification segmentation tasks such as landslide recognition, each pixel needs to be accurately classified as either a landslide or a non-landslide. The Logits binary cross-entropy loss function can directly measure the difference between the pixel class probability distribution predicted by the model and the true label distribution. By minimizing this loss function, the model can be guided to continuously adjust its parameters, making the prediction results closer to the real situation, thereby effectively improving the accuracy of pixel-level classification. Furthermore, this application uses the Adam optimizer for model parameter updates because the Adam optimizer combines the advantages of momentum and adaptive learning rates. It can automatically adjust the learning rate based on the historical gradient information of each parameter, assigning different learning rates to different parameters, enabling the model to converge more efficiently during training. Furthermore, dynamically setting the group learning rate based on the model structure is a refined learning rate adjustment strategy. Since 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 the model parameters into different groups and setting different learning rates for each group, the update speed of each part of the parameters can be more precisely controlled, making the model more stable during training and accelerating convergence. Finally, it should be noted that using a learning rate decay strategy for dynamic learning rate adjustment is to balance convergence speed and model stability during training. In the early stages of training, using a larger learning rate can accelerate the update speed of model parameters, allowing the model to quickly approach the optimal solution. As training progresses, gradually decreasing the learning rate can prevent the model from oscillating around the optimal solution, improving the model's convergence accuracy and stability.
[0083] In summary, the core technological innovation of this application lies in:
[0084] 1) A systematic, phased transfer learning strategy: The transfer learning process for landslide identification is innovatively decomposed into three organically combined phases: “basic feature transfer (general + intermediate domain) → category condition domain alignment → small sample fine-tuning”, which realizes the gradual and refined adaptation of knowledge.
[0085] 2) Introducing intermediate domain pre-training: Using remote sensing landslide data that is more relevant to the target task but has a larger data volume as an intermediate domain for pre-training, effectively bridging the semantic gap between general data and target UAV data, and providing better initialization for subsequent alignment and fine-tuning;
[0086] 3) Application of Class Conditional Feature Alignment: Class-wise CORAL Loss is used for domain adaptation to align sliding and non-sliding features at the class level, which effectively alleviates the problem of inconsistent class distribution and feature confusion in cross-domain data and significantly improves the model's discrimination ability in complex backgrounds.
[0087] 4) Synergy between architecture and strategy: The phased transfer strategy is combined with the ResUNet+ASPP network architecture, which is suitable for handling complex boundaries and multi-scale targets, to give full play to the synergistic effect of structural optimization and training strategy optimization.
[0088] 5) Efficient learning with small samples: Through phased 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 using UAV images by adopting 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 This application discloses a landslide recognition system based on staged feature adaptive transfer learning. The system includes a model building module, a basic feature transfer module, a category conditional domain feature alignment module, a few-sample fine-tuning module, and a landslide recognition 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 transfer module is used to transfer general visual pre-trained weights and utilize the remote sensing image dataset D from the source domain. S Fine-tuning the encoder parameters of model F1 yields a domain-adapted optimized landslide identification model F2.
[0093] The category-conditional domain feature alignment module is used to align the remote sensing image dataset D from the source domain. S And the UAV imagery dataset D for the target domain T The input is fed into model F2, and based on the covariance alignment mechanism, the distribution difference between the source domain and the target domain in the feature space is reduced to obtain the domain adaptive model F3.
[0094] The small sample fine-tuning module is used to adjust a small amount of labeled target domain small sample data D′. TThe input is fed into model F3, and the specified parameters in model F3 are fine-tuned through a supervised learning mechanism to obtain the target domain-adapted model F4.
[0095] The landslide recognition application module is used to perform landslide recognition tasks on the input UAV images 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 phased feature adaptive transfer learning as described in any of the foregoing method embodiments, and this application does not limit this method.
[0097] As can be seen from the above, the landslide recognition system based on phased feature adaptive transfer learning disclosed in this application transfers general visual pre-trained weights. These pre-trained weights contain a large amount of general visual knowledge, such as basic features like edges, textures, and shapes, enabling the model to have a certain feature extraction capability in the initial stage. This avoids training from scratch, greatly improves the model's performance in the initial stage, and helps the model generalize better on different datasets. 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 better matches the landslide features of the source domain, 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 terms of acquisition methods, resolution, and viewing angle, the data distribution may be different. By adjusting the feature distribution through covariance alignment, the features learned by the model are made more similar in the source and target domains. This allows the model to better utilize the knowledge learned in the source domain when processing data in the target domain, improving the model's adaptability and recognition ability in different domains and achieving domain adaptation. By combining a small amount of labeled data in the target domain, the model utilizes a large amount of data in the source domain and the knowledge learned in the early stages, while also fully considering the special characteristics of the target domain. Through effective knowledge transfer and feature alignment, the model converges very quickly in the small sample fine-tuning stage, achieving stable performance in only 5-10 training epochs, which is much faster than the 40-50 epochs required for traditional training from scratch, saving a lot of training time and computing resources.
[0098] Furthermore, this application discloses a computer storage medium for storing computer execution instructions, which are used to execute the landslide identification method based on phased feature adaptive transfer learning as described in any of the foregoing embodiments.
[0099] As can be seen from the above, the computer storage medium disclosed in this application, by transferring general visual pre-trained weights, which contain a large amount of general visual knowledge, such as basic features like edges, textures, and shapes, enables the model to have a certain feature extraction capability in the initial stage, avoiding training from scratch, greatly improving the model's performance in the initial stage, and helping the model to generalize better on different datasets; 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, learn a representation that is more consistent with the landslide characteristics of the source domain, thereby improving the recognition accuracy on the source domain data and making full use of the existing data resources of the source domain; since the source domain and target domain data may differ in terms of acquisition methods, resolution, and viewing angle, resulting in different data distributions. By adjusting the feature distribution through covariance alignment, the features learned by the model are made more similar in the source and target domains. This allows the model to better utilize the knowledge learned in the source domain when processing data in the target domain, improving the model's adaptability and recognition ability in different domains and achieving domain adaptation. By combining a small amount of labeled data in the target domain, the model utilizes a large amount of data in the source domain and the knowledge learned in the early stages, while also fully considering the special characteristics of the target domain. Through effective knowledge transfer and feature alignment, the model converges very quickly in the small sample fine-tuning stage, achieving stable performance in only 5-10 training epochs, which is much faster than the 40-50 epochs required for traditional training from scratch, saving a lot of training time and computing resources.
[0100] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some 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. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
Claims
1. A landslide identification method based on phased feature adaptive transfer learning, characterized in that, The method includes: S1. Construct an initial landslide identification model with an encoder-decoder. F 1; S2, by transferring general visual pre-trained weights and utilizing remote sensing image datasets from the source domain. D S Fine-tuning model F The encoder parameters of 1 are used to obtain a domain-adapted landslide identification optimization model. F 2; S3. Transfer the remote sensing image dataset from the source domain. D S and the drone imagery dataset of the target domain. D T Input to model F In section 2, a domain-adaptive model is obtained by reducing the distribution differences between the source and target domains in the feature space based on the covariance alignment mechanism. F 3, including: S31. Transfer the remote sensing image dataset from the source domain. D S and the drone imagery dataset of the target domain. D T Input to model F In section 2, through the model F The feature extraction module of module 2 extracts features from the dataset. D S , D T Source domain features were extracted. f s and target domain features f T ; S32. For each predefined semantic category, calculate its features in the source domain. f s The covariance matrix in C sc and features in the target domain f T The covariance matrix in C tc ; S33, the covariance matrix C sc , C tc Substituting these values into the following category-conditional covariance alignment loss function for backpropagation calculation, the model is updated. F The parameters of 2 are used to minimize the difference in semantic category distribution between the source and target domains, resulting in a domain adaptive model. F 3: ； in, C This represents the total number of predefined semantic categories. d Representing feature dimension, Denotes the square of the Frobenius norm; S4. Use a small amount of labeled target domain sample data. Input to model F In section 3, a supervised learning mechanism is used to train the model. F Fine-tuning the specified parameters in step 3 yields the target domain adaptation model. F 4; S5. Based on the target domain adaptation model F 4. Perform landslide identification tasks on the input drone images to be detected.
2. The method according to claim 1, characterized in that, In step S1, the encoder part of the initial landslide identification model F1 uses a ResNet network as the backbone feature extraction network to leverage the powerful feature extraction capabilities of the ResNet network and 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, so as to effectively expand the receptive field and obtain richer contextual information by using dilated convolutions with different dilation rates in parallel without increasing the amount of computation.
4. The method according to claim 1, characterized in that, In step S2, the method involves transferring general visual pre-trained weights and utilizing remote sensing image datasets from the source domain. D S Fine-tuning model F The encoder parameters of 1 are used to obtain a domain-adapted landslide identification optimization model. F 2, including: S21. Load the encoder weights pre-trained on a large-scale general image dataset into the model. F The encoder part of 1 initializes the model by utilizing general visual knowledge. F The encoder parameters of 1 accelerate the landslide feature learning process; S22. Utilize the remote sensing image dataset from the source domain. D S and its corresponding landslide mask label fine-tuning model F The encoder parameters of 1 are pre-trained in the intermediate domain to narrow the semantic gap between the semantic distribution of landslides in the target domain and the intermediate domain, resulting in a domain-adapted optimized landslide recognition model. F 2.
5. The method according to claim 1, characterized in that, In step S4, the specified parameters include the model. F All parameters in section 3, or decoder parameters, and finally, multi-layer encoder parameters.
6. The method according to claim 1, characterized in that, In step S4, during the model... F During the fine-tuning of the specified parameters in 3, the method further includes: performing pixel-level supervised optimization using the Logits binary cross-entropy loss function suitable for pixel-level binary classification and 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, so as to balance convergence speed and model stability.
7. A landslide identification system based on phased feature adaptive transfer learning, characterized in that, The system includes a model building module, a basic feature transfer module, a category-conditional domain feature alignment module, a few-sample fine-tuning module, and a landslide recognition application module, wherein: The model building module is used to construct an initial landslide recognition model with an encoder-decoder architecture. F 1; The basic feature transfer module is used to transfer general visual pre-trained weights and utilize remote sensing image datasets from the source domain. D S Fine-tuning model F The encoder parameters of 1 are used to obtain a domain-adapted landslide identification optimization model. F 2; The category-conditional domain feature alignment module is used to align the remote sensing image dataset from the source domain. D S and the drone imagery dataset of the target domain. D T Input to model F In section 2, a domain-adaptive model is obtained by reducing the distribution differences between the source and target domains in the feature space based on the covariance alignment mechanism. F 3, which includes: the remote sensing image dataset of the source domain. D S and the drone imagery dataset of the target domain. D T Input to model F In section 2, through the model F The feature extraction module of module 2 extracts features from the dataset. D S , D T Source domain features were extracted. f s and target domain features f T For each predefined semantic category, its features in the source domain are calculated respectively. f s The covariance matrix in C sc and features in the target domain f T The covariance matrix in C tc ; The covariance matrix C sc , C tc Substituting these values into the following category-conditional covariance alignment loss function for backpropagation calculation, the model is updated. F The parameters of 2 are used to minimize the difference in semantic category distribution between the source and target domains, resulting in a domain adaptive model. F 3: ； in, C This represents the total number of predefined semantic categories. d Representing feature dimension, Denotes the square of the Frobenius norm; The small sample fine-tuning module is used to adjust a small amount of labeled target domain small sample data. Input to model F In section 3, a supervised learning mechanism is used to train the model. F Fine-tuning the specified parameters in step 3 yields the target domain adaptation model. F 4; The landslide identification application module is used to adapt the model based on the target domain. F 4. Perform landslide identification tasks on the input drone images to be detected.
8. A computer storage medium, characterized in that, The computer storage medium is used to store computer execution instructions, which are used to execute the landslide identification method based on phased feature adaptive transfer learning as described in any one of claims 1 to 7.