Landslide segmentation method, device and equipment based on mixed mamba and frequency domain calibration, and medium

Abstract

The application discloses a landslide segmentation method and device based on hybrid Mamba and frequency domain calibration, equipment and medium, relates to the landslide disaster detection technical field, and includes: through fusing optical satellite image and slope, lithology terrain constraint data, after preprocessing, inputting the model containing hybrid Mamba encoder, multi-scale adaptive gate module and progressive frequency domain calibration decoder for processing.Through multi-stage hybrid coding, context aggregation modulation and step-by-step resolution recovery, accurate landslide segmentation is realized, the global feature modeling capability is effectively improved, the multi-scale and boundary fuzzy landslide recognition accuracy is enhanced, the lightness and robustness are considered, and the actual monitoring demand of highway landslide is more suitable.

Description

Technical Field

[0001] This invention relates to the field of landslide disaster detection technology, and in particular to a landslide segmentation method, apparatus, equipment and medium based on hybrid Mamba and frequency domain calibration. Background Technology

[0002] Current landslide segmentation methods are mostly based on convolutional neural networks or Transformer architectures, using a single optical satellite image as input and employing an encoder-decoder structure to complete feature extraction and mask generation. Some methods introduce multi-scale feature fusion, attention mechanisms, or lightweight structures to optimize performance, while a few methods attempt to improve robustness by overlaying terrain data such as digital elevation models.

[0003] Pure convolutional neural networks have limited receptive fields, making it difficult to establish long-distance dependencies between landslide areas and the global geological background. Transformer-type models have computational complexity that increases quadratically with resolution, making it difficult to meet the requirements of lightweight design and real-time performance. Relying solely on optical images is susceptible to interference from lighting, vegetation, and surface exposure, and they do not make sufficient use of key landslide control factors such as slope and lithology, resulting in missed detections of small-scale landslides, blurred boundaries, and missegmentation of complex backgrounds. In highway landslide scenarios, they cannot achieve a coordinated improvement in accuracy and efficiency.

[0004] Therefore, how to perform high-precision and lightweight segmentation of highway landslides in optical satellite imagery, while improving the ability to identify multi-scale landslides with blurred boundaries, has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a landslide segmentation method, apparatus, device, and medium based on hybrid Mamba and frequency domain calibration, aiming to solve the technical problem of how to achieve high-precision and lightweight segmentation of highway landslides in optical satellite imagery.

[0006] To achieve the above objectives, this application proposes a landslide segmentation method based on hybrid Mamba and frequency domain calibration, comprising: Acquire optical satellite images to be segmented, and acquire terrain constraint data corresponding to the optical satellite images, wherein the terrain constraint data includes digital elevation slope data and lithological data; The optical satellite images are preprocessed to obtain standardized satellite images; The standardized satellite imagery and terrain constraint data are input into a preset satellite imagery landslide segmentation model for processing to obtain a landslide segmentation mask. The preset satellite imagery landslide segmentation model includes a hybrid Mamba multi-scale feature encoder, a multi-scale adaptive gating module, and a progressive frequency domain calibration fusion decoder. The step of inputting the standardized satellite imagery and terrain constraint data into a preset satellite imagery landslide segmentation model for processing to obtain a landslide segmentation mask includes: The standardized satellite imagery and terrain-constrained data are multi-stage hybrid encoded by a hybrid Mamba multi-scale feature encoder to obtain a multi-scale feature set, wherein the multi-stage hybrid encoding includes local multi-scale feature extraction and global long sequence dependency modeling. By using a multi-scale adaptive gating module, the hybrid encoded features in the multi-scale feature set are subjected to multi-scale context aggregation and adaptive semantic modulation to obtain enhanced deep semantic features. The enhanced deep semantic features are restored stepwise by a progressive frequency domain calibration fusion decoder, and the landslide segmentation mask is obtained by combining the initial shallow features in the multi-scale feature set with cross-layer fusion and boundary calibration.

[0007] In one embodiment, the standardized satellite imagery and terrain-constrained data are subjected to multi-stage hybrid encoding using a hybrid Mamba multi-scale feature encoder to obtain a multi-scale feature set. The multi-stage hybrid encoding includes steps of local multi-scale feature extraction and global long-sequence dependency modeling, comprising: The standardized satellite image is initially downsampled by convolution with a preset stride, and basic edge and color features are extracted by lightweight convolution to obtain initial shallow features; The initial shallow features are downsampled and channel-upgraded by convolution with a preset stride, and local multi-scale detail features are extracted by a dynamic scale mixing module to obtain the first encoded features. The first encoded feature is downsampled by convolution with a preset stride, and the second encoded feature is obtained by extracting local multi-scale features through the dynamic scale mixing module. The second encoded feature is downsampled by convolution with a preset stride, and local multi-scale features are extracted by the dynamic scale mixing module to obtain local encoded features; The local encoded features are input into the global feature extraction module based on the state space model to perform long-distance dependency modeling, thereby obtaining the global encoded features; The terrain constraint data is concatenated with the global coding features as an additional channel, and the concatenated features are weighted by terrain constraints through a gated fusion mechanism to obtain hybrid coding features; A multi-scale feature set is constructed based on the initial shallow features, the first encoded features, the second encoded features, and the hybrid encoded features.

[0008] In one embodiment, the step of downsampling and upscaling the initial shallow features through convolution with a preset stride, and extracting local multi-scale detail features through a dynamic scale mixing module to obtain the first encoded features includes: Multiple branch multi-scale features are obtained by extracting the receptive field of the initial shallow features through multiple parallel dilated depth convolutions, wherein the multiple parallel dilated depth convolutions have different preset dilation rates. The aggregated features are obtained by adding the multiple branch multi-scale features element by element. The aggregated features are spatially compressed using global adaptive average pooling to obtain global channel statistics. The global channel statistics are compressed and mapped to higher dimensions using a multilayer perceptron to generate dynamic fusion weights. Based on the dynamic fusion weights, channel-level soft attention weights are applied to the multiple branch multi-scale features, and element-wise summation is performed to obtain the fused features; The fused features are added to the initial shallow features by residual addition to obtain the first encoded features.

[0009] In one embodiment, the step of inputting the local encoded features into a global feature extraction module based on a state-space model for long-distance dependency modeling to obtain global encoded features includes: The local encoded features are spatially enhanced by performing local convolution to obtain spatially enhanced features; Perform dimensional transpose and flatten operations on the spatial enhancement features to obtain one-dimensional serialized features; The one-dimensional serialized features are input into the state space model for selective scanning and global context modeling to obtain global sequence features; The global sequence features are reverse-engineered and their dimensions restored to obtain the global encoded features.

[0010] In one embodiment, the step of concatenating the terrain constraint data as an additional channel with the global encoded features, and applying terrain constraint weighting to the concatenated features through a gating fusion mechanism to obtain hybrid encoded features, further includes: The digital elevation and slope data and lithological data of the terrain constraint data are spatially aligned to obtain a first terrain constraint feature and a second terrain constraint feature that match the size of the global coding feature. The first terrain constraint feature, the second terrain constraint feature, and the global encoding feature are concatenated along the channel dimension to obtain the concatenated feature; The spliced ​​features are fused by channel fusion using a preset convolution to obtain fused channel features; The gating fusion mechanism generates terrain constraint weights for the fusion channel features, wherein the terrain constraint weights are used to characterize the degree of influence of slope and lithology on the probability of landslide occurrence. The local coding features are weighted and modulated according to the terrain constraint weights to obtain hybrid coding features.

[0011] In one embodiment, the step of performing multi-scale context aggregation and adaptive semantic modulation on the hybrid encoded features in the multi-scale feature set through a multi-scale adaptive gating module to obtain enhanced deep semantic features includes: The hybrid encoded features are subjected to context information extraction at different ranges by multiple parallel dilated convolutions to obtain multiple scale context features, wherein the multiple parallel dilated convolutions have different preset dilation rates. The multiple scale context features are concatenated by channels, and a context weight map is generated by dimensionality reduction through preset convolution. Based on the hybrid coding features, multiple gated signals are generated through multiple parallel preset convolutions; Based on the first and second gate signals in the multi-gated signals, soft selection fusion is performed on the original features and context enhancement features to obtain intermediate modulation features; Based on the third gating signal in the multi-gating signals, the intermediate modulation features are filtered by channel-level importance to obtain the enhanced deep semantic features.

[0012] In one embodiment, the step of progressively restoring the enhanced deep semantic features through a progressively frequency-domain calibrated fusion decoder, and then performing cross-layer fusion and boundary calibration in conjunction with the initial shallow features in the multi-scale feature set to obtain a landslide segmentation mask includes: The initial shallow features are enhanced with high-frequency edge information by using a texture compensation operator to obtain edge-enhanced initial shallow features. The resolution of the enhanced deep semantic features is improved by a lightweight dynamic upsampling module, and the number of channels is projected to the same level as the initial shallow features of the edge enhancement by a preset convolution to obtain the projected deep features. The projected deep features are concatenated with the edge-enhanced initial shallow features in the channel dimension to generate a feature calibration gating map. Based on the feature calibration gating map, the initial shallow features of edge enhancement are selectively preserved and noise suppressed to obtain the calibrated initial shallow features; The projected deep features are concatenated with the calibrated initial shallow features and fused using lightweight convolution to obtain the output features of the current decoding level. The output features of the current decoding level are progressively restored through multi-level cascade until the original input resolution is restored, thus obtaining the landslide segmentation mask.

[0013] Furthermore, to achieve the above objectives, this application also proposes a landslide segmentation device based on hybrid Mamba and frequency domain calibration, wherein the landslide segmentation device based on hybrid Mamba and frequency domain calibration includes: The acquisition module is used to acquire the optical satellite image to be segmented and to acquire the terrain constraint data corresponding to the optical satellite image, wherein the terrain constraint data includes digital elevation slope data and lithological data. The data preprocessing module is used to preprocess the optical satellite imagery to obtain standardized satellite imagery; The result module is used to input the standardized satellite imagery and terrain-constrained data into a preset satellite imagery landslide segmentation model for processing to obtain a landslide segmentation mask. The preset satellite imagery landslide segmentation model includes a hybrid Mamba multi-scale feature encoder, a multi-scale adaptive gating module, and a progressive frequency domain calibration fusion decoder. It is also used to perform multi-stage hybrid encoding on the standardized satellite imagery and terrain-constrained data using the hybrid Mamba multi-scale feature encoder to obtain a multi-scale feature set. The multi-stage hybrid encoding includes local multi-scale feature extraction and global long-sequence dependency modeling. The multi-scale adaptive gating module performs multi-scale context aggregation and adaptive semantic modulation on the hybrid encoded features in the multi-scale feature set to obtain enhanced deep semantic features. The progressive frequency domain calibration fusion decoder performs stepwise resolution recovery on the enhanced deep semantic features and combines them with the initial shallow features in the multi-scale feature set for cross-layer fusion and boundary calibration to obtain the landslide segmentation mask.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored, which, when executed by a processor, implements the steps of the landslide segmentation method based on hybrid Mamba and frequency domain calibration as described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the landslide segmentation method based on hybrid Mamba and frequency domain calibration as described above.

[0016] This application integrates optical satellite imagery with slope and lithological topographic constraints. After preprocessing, the data is input into a model containing a hybrid Mamba encoder, a multi-scale adaptive gating module, and a progressive frequency domain calibration decoder. Through multi-stage hybrid coding, contextual aggregation modulation, and progressive resolution recovery, accurate landslide segmentation is achieved, effectively improving global feature modeling capabilities and enhancing the identification accuracy of multi-scale and boundary-ambiguous landslides. It balances lightweight design and robustness, making it more suitable for the actual monitoring needs of highway landslides. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the first embodiment of the landslide segmentation method based on hybrid Mamba and frequency domain calibration in this application; Figure 2 This is a flowchart illustrating the second embodiment of the landslide segmentation method based on hybrid Mamba and frequency domain calibration in this application; Figure 3 This is a schematic diagram of the module structure of the landslide segmentation device based on hybrid Mamba and frequency domain calibration in this application; Figure 4 This is a schematic diagram of the device structure of the hardware operating environment involved in the landslide segmentation method based on hybrid Mamba and frequency domain calibration in the embodiments of this application.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] Current landslide segmentation methods are mostly based on convolutional neural networks or Transformer architectures, using a single optical satellite image as input and employing an encoder-decoder structure to complete feature extraction and mask generation. Some methods introduce multi-scale feature fusion, attention mechanisms, or lightweight structures to optimize performance, while a few methods attempt to improve robustness by overlaying terrain data such as digital elevation models.

[0023] Pure convolutional neural networks have limited receptive fields, making it difficult to establish long-distance dependencies between landslide areas and the global geological background. Transformer-type models have computational complexity that increases quadratically with resolution, making it difficult to meet the requirements of lightweight design and real-time performance. Relying solely on optical images is susceptible to interference from lighting, vegetation, and surface exposure, and they do not make sufficient use of key landslide control factors such as slope and lithology, resulting in missed detections of small-scale landslides, blurred boundaries, and missegmentation of complex backgrounds. In highway landslide scenarios, they cannot achieve a coordinated improvement in accuracy and efficiency.

[0024] Therefore, how to perform high-precision and lightweight segmentation of highway landslides in optical satellite imagery, while improving the ability to identify multi-scale landslides with blurred boundaries, has become an urgent problem to be solved.

[0025] Based on the above, this application also provides a landslide segmentation method based on hybrid Mamba and frequency domain calibration, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the landslide segmentation method based on hybrid Mamba and frequency domain calibration in this application.

[0026] In this embodiment, the landslide segmentation method based on hybrid Mamba and frequency domain calibration includes steps S10~S40: Step S10: Obtain the optical satellite image to be segmented and obtain the terrain constraint data corresponding to the optical satellite image.

[0027] Specifically, based on the latitude and longitude range or administrative boundaries of the target area, multispectral or panchromatic optical satellite imagery is acquired from Gaofen series satellites, resource satellites, or publicly available remote sensing data sources abroad. Simultaneously, a digital elevation model (DEM) for the corresponding area is downloaded from a geospatial data cloud platform to obtain corresponding terrain constraint data, including DEM slope data and lithological data. Slope data is generated using slope calculation tools, and lithological data is extracted from a geological map database. Then, spatial registration is performed on the optical satellite imagery, slope data, and lithological data to ensure a one-to-one spatial correspondence at the pixel level; that is, the optical image value, slope value, and lithological value at the same pixel location correspond to the same geographical location on the Earth's surface. The reason for this is that landslides have clear physical laws. The slope determines the stability threshold of the soil and rock mass under gravity, and the lithology determines the shear strength and permeability of the soil and rock mass. Introducing these two physical factors as constraints into the model can make the segmentation results not only based on visual features, but also more in line with the mechanical mechanism of landslides. This can effectively distinguish areas with similar spectral features but different physical conditions, such as bare land and real landslide bodies, road excavation faces and landslide deposits, which are easily confused.

[0028] Step S20: Preprocess the optical satellite imagery to obtain standardized satellite imagery.

[0029] Specifically, firstly, radiometric calibration is performed on the optical satellite imagery, converting the raw digital quantization values ​​recorded by the sensors into physical quantities (such as radiance or reflectance) to eliminate radiometric differences caused by different sensors and different imaging times. Then, atmospheric correction is performed to remove the influence of atmospheric scattering and absorption on the true reflectance of the surface, ensuring the comparability of images acquired at different times and in different regions. Next, geometric correction is performed, using ground control points or the satellite's own attitude parameters to correct the imagery to a standard map projection coordinate system, eliminating geometric distortions caused by terrain undulations and satellite attitude deviations. Afterward, the imagery is cropped or resampled to a preset input size (e.g., 512×512), and pixel values ​​are normalized, scaling the pixel values ​​from the original range to [0,1] to unify the numerical range of different images. Finally, data augmentation processing is performed on the imagery, including random scaling, random cropping, horizontal flipping, and photometric distortion operations, expanding the diversity of the training samples. The reason for this is that deep learning models are sensitive to the numerical range and distribution of input data. Unstandardized images may lead to unstable model training and slow convergence. Data augmentation can simulate image changes under different imaging conditions, improve the model's robustness to changes in lighting and shooting angles, and prevent the model from over-memorizing the specific appearance of training samples and losing its generalization ability.

[0030] Step S30: Input standardized satellite imagery and terrain constraint data into a preset satellite imagery landslide segmentation model for processing to obtain a landslide segmentation mask.

[0031] It should be noted that the preset satellite image landslide segmentation model includes a hybrid Mamba multi-scale feature encoder, a multi-scale adaptive gating module, and a progressive frequency domain calibration fusion decoder. The hybrid Mamba multi-scale feature encoder includes a dynamic scale mixing module, a state-space model-based global feature extraction module, and a local-global gating fusion mechanism. The dynamic scale mixing module includes multiple parallel dilated deep convolutional branches, element-wise addition units, a global adaptive average pooling layer, a multilayer perceptron, and a channel-level soft attention weighting unit. The multiple parallel dilated deep convolutional branches have different preset dilation rates for extracting multi-scale receptive field features. The state-space model-based global feature extraction module includes deep convolutional layers, dimension transpose units, layer normalization layers, a state-space model core layer, and an inverse reshaping unit. The state-space model core layer includes a selective scanning mechanism and a hardware-aware perceptron algorithm. The local-global gating fusion mechanism includes a channel stitching unit, preset convolutional layers, and an adaptive gating weight generation unit. The multi-scale adaptive gating module includes a multi-scale context extraction branch and an adaptive gating modulation branch. The multi-scale context extraction branch includes multiple parallel dilated convolutional layers, channel stitching units, and pre-defined convolutional dimensionality reduction layers, where the parallel dilated convolutional layers have different pre-defined dilation rates. The adaptive gated modulation branch includes multiple parallel pre-defined convolutional layers, soft-selection fusion units, and channel-level importance filtering units. The progressive frequency domain calibration fusion decoder includes multiple cascaded frequency domain calibration fusion modules, each of which includes a texture compensation operator, a lightweight dynamic upsampling module, a feature calibration gating mechanism, and a lightweight convolutional fusion unit. The texture compensation operator includes a max-pooling layer and an interpolation unit. The lightweight dynamic upsampling module includes a sampling point generator, an offset calculation unit, and a resampling unit. The feature calibration gating mechanism includes a channel stitching unit, pre-defined convolutional layers, and a gated map generation unit. The lightweight convolutional fusion unit includes a lightweight convolutional layer, a batch normalization layer, and an activation function layer.

[0032] Furthermore, prior to step S30, the process includes: First, acquiring a training sample set, which includes optical satellite image samples labeled with real landslide masks and corresponding terrain constraint data samples. Specifically, optical satellite images are selected from publicly available landslide datasets and self-built datasets. Using a multi-temporal image comparison analysis method, professionals use annotation tools to delineate the landslide body at the pixel level based on historical images before and after the landslide, generating real landslide masks. Simultaneously, a digital elevation model corresponding to the space of each image is acquired, slope data is calculated, and lithological classification data is extracted from a geological map database as terrain constraint data samples. Then, the samples are divided into training, validation, and test sets according to a preset ratio, and data augmentation operations such as random scaling, random cropping, horizontal flipping with preset probabilities, and photometric distortion are performed on the training set samples. The reason for this is that the performance of deep learning models is highly dependent on the scale, quality and diversity of training data. Landslide pixels usually account for a very small percentage of the entire image, and landslides in different regions differ in terms of geomorphic conditions, imaging time and spectral characteristics. Only by providing sufficient, balanced and high-quality paired data can the model learn robust representations of landslide features.

[0033] Secondly, optical satellite image samples and corresponding terrain-constrained data samples are input into the initial satellite image landslide segmentation model for forward propagation to obtain a predicted landslide segmentation mask. Specifically, the optical satellite image samples undergo a four-stage hierarchical processing using a hybrid Mamba multi-scale feature encoder to progressively extract feature representations from local details to global semantics. In the fourth stage, the terrain-constrained data samples incorporate physical guidance information and are enhanced with deep semantic features through a multi-scale adaptive gating module. Finally, a progressively frequency-domain calibrated fusion decoder restores the resolution step by step, outputting a predicted landslide segmentation mask with the same size as the input image. This is done because forward propagation is a fundamental step in model training. Only by fully simulating the inference process during training can the loss be calculated based on the difference between the predicted output and the ground truth annotations, thereby guiding the optimization direction of the network parameters.

[0034] Then, based on the predicted landslide segmentation mask and the actual landslide mask, the total training loss is calculated using a joint loss function. Specifically, the focus loss adds a modulation coefficient to the standard cross-entropy loss, and reduces the weight of easily distinguishable samples by pre-setting a balance factor and focus parameters, shifting the optimization focus to difficult-to-distinguish samples such as landslides and bare land; the region overlap loss directly calculates the spatial overlap between the predicted mask and the actual mask, focusing on set similarity rather than pixel-by-pixel classification, and is insensitive to the size of the foreground region; the boundary loss is calculated based on the symbolic distance field, generating boundary penalty weights through the distance transformation of the actual mask, so that predicted pixels that deviate from the actual boundary are penalized more, and a dynamic introduction mechanism is used to superimpose the weights after the model has completed basic semantic learning. The joint loss function includes focus loss, region overlap loss, and boundary loss, and the specific formula is as follows: in Indicates total training loss. Indicates the focal loss value. Indicates the loss of regional overlap. Indicates the boundary loss value. The weighting coefficients representing the focus loss. The weighting coefficients represent the regional overlap loss. The weighting coefficients representing the boundary loss; This represents the total number of pixels, that is, the total number of pixels in the image; Indicates pixel index, Indicates the first Preset balance factor for each pixel, This represents the probability that the model predicts a pixel belongs to the true category, and its value ranges from 0 to 1. This indicates the preset focusing parameters, used to smoothly reduce the weight of easily distinguishable samples; Represents a logarithmic function; Indicates the first The true label value of each pixel is 1 for the landslide area and 0 for the background area; This indicates a preset smoothing term to prevent calculation errors caused by a denominator of 0. Indicates the first The signed distance field value corresponding to each pixel represents the signed distance from that pixel to the real boundary. It is positive inside the landslide area, negative outside, and zero at the boundary. The absolute value of the symbol distance field represents the distance from the pixel to the true boundary; the greater the distance, the greater the penalty weight. This is done because landslide segmentation faces three major challenges: extreme class imbalance, high requirements for morphological integrity, and blurred boundaries. A single loss function cannot address all of these simultaneously. Focus loss solves the problem of positive and negative sample imbalance, region overlap loss optimizes global morphological integrity, and boundary loss refines boundary localization. The three are used together to form complementary supervision.

[0035] Then, a deep supervision strategy is introduced into multiple stages of the decoder of the pre-set satellite image landslide segmentation model. The intermediate feature maps at different levels are upsampled to the original image size, and the corresponding total training loss is calculated for each. The total training loss is then weighted and summed according to the pre-set layer weights to obtain the deep supervision joint loss, which is expressed by the following formula: in This indicates that the deep supervision team suffered joint losses. This indicates the number of output levels involved in deep supervision. Indicates the output hierarchical index. Indicates the first Hierarchical scale supervision weights.

[0036] Finally, backpropagation is performed based on the deep supervised joint loss to update the network parameters of the preset satellite image landslide segmentation model, resulting in an optimized model. This process continues until a preset training stopping condition is met, at which point the optimized model is adopted as the preset satellite image landslide segmentation model. Specifically, an adaptive moment estimation optimizer is used, with preset momentum parameters and preset weight decay coefficients. The initial learning rate is set to a preset value, and a cosine annealing strategy is used to dynamically adjust the learning rate. In each iteration, the gradient is calculated, and the convolutional kernel weights, bias terms, and learnable parameters in the gating mechanism are updated to obtain the optimized model. A preset total number of training epochs is then set. After each training epoch, the performance of the optimized model is evaluated using a validation set. When the validation set metric no longer improves after several consecutive epochs, an early stopping strategy is triggered to terminate the training early, and the optimized model with the best validation set performance is saved as the final model. The reason for this is that the model needs a sufficient number of iterations to converge to a better solution, but overtraining can cause the model to memorize the noise and specificity of the training data and lose its generalization ability. The early stopping strategy and learning rate decay can prevent overfitting and ensure that the trained model maintains good segmentation accuracy and generalization performance on unseen test data.

[0037] Furthermore, to comprehensively verify the effectiveness and generalization ability of the pre-defined satellite image landslide segmentation model (hereinafter referred to as HMC-Net) proposed in this embodiment, systematic experiments were conducted on three landslide datasets of different sizes and from different sources. The three datasets include: a subset of the large-scale optical satellite landslide dataset developed by the Institute of Mountain Hazards and Environment, Chinese Academy of Sciences, hereinafter referred to as the CAS-SAT dataset, which contains 7422 pixel-level annotated images, covering nine landslide-prone areas worldwide, with spatial resolutions ranging from 0.5 meters to 5 meters; a widely used landslide dataset from a certain city, containing 770 images, mainly targeting single landslide scenarios; and a high-resolution optical satellite landslide dataset built for highway landslide identification tasks, hereinafter referred to as the HOS-LSD dataset, which contains 1218 images, focusing on landslide samples collected in the southwestern mountainous areas and along major transportation routes, and adding scenarios where landslides coexist with highways. Each dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model parameter learning, the validation set is used for model selection and hyperparameter tuning, and the test set is used only for final performance evaluation.

[0038] The experiment constructed comparative experiments on two metrics: accuracy and lightweightness. A total of 11 models were used, including classic CNN-based architectures, attention-enhanced networks, emerging models based on Transformers and Mamba, and hybrid methods. In comparison with classic CNN architectures, models such as FCN, U-Net, U-Net++, and Deeplabv3+ were compared. In addition, other segmentation models that have achieved state-of-the-art results on various datasets in recent years were also included, such as CResU-Net, ResUBFA, CA-Unet, MobileUnet, SCSegamba, and SlideNet. The experimental data from the three datasets are analyzed below. The results are summarized in Table 1 (model comparison experiment data based on the CAS-SAT dataset), Table 2 (model comparison experiment data based on a city landslide dataset), and Table 3 (model comparison experiment data based on the HOS-LSD dataset).

[0039] Table 1. Summary of experimental data for model comparison based on the CAS-SAT dataset. Table 2 Summary of experimental data for model comparison based on a landslide dataset from a certain city Table 3 Summary of experimental data for model comparison based on the HOS-LSD dataset Experimental results on a subset of the CAS-SAT dataset show that the preset satellite imagery landslide segmentation model achieves an intersection-over-union (IoU) ratio of 88.92%, an accuracy of 95.43%, an F1 score of 94.14%, and an overall accuracy of 99.13%, all ranking first among all comparative models. With 6.97M parameters, a computational cost of 20.88G, and a weight file size of 26.75MB, it achieves an excellent balance between accuracy and lightweight design. Experimental results on a landslide dataset from a specific city show that the preset satellite imagery landslide segmentation model achieves an IoU ratio of 82.24%, an F1 score of 90.26%, and an overall accuracy of 97.93%, all the highest values ​​among all comparative models. The parameter count remains at 6.97M, significantly lower than other high-precision models with parameter counts of 18.65M and 117.44M. Experimental results on the HOS-LSD dataset show that the crossover ratio of the preset satellite image landslide segmentation model reaches 72.87%, the recall rate reaches 85.40%, and the F1 score reaches 84.31%, all of which are the highest levels among all comparison models. At the same time, the number of parameters is only about 70% of the 9.96M of the preset comparison model, and the computational cost is only about 69% of the 30.08G of the preset comparison model.

[0040] This embodiment integrates optical satellite imagery with slope and lithological topographic constraints. After preprocessing, the data is input into a model containing a hybrid Mamba encoder, a multi-scale adaptive gating module, and a progressive frequency domain calibration decoder. Through multi-stage hybrid coding, contextual aggregation modulation, and progressive resolution recovery, accurate landslide segmentation is achieved, effectively improving global feature modeling capabilities and enhancing the identification accuracy of multi-scale and boundary-ambiguous landslides. It balances lightweight design and robustness, making it more suitable for actual monitoring needs of highway landslides.

[0041] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The landslide segmentation method based on hybrid Mamba and frequency domain calibration, step S30, further includes steps S201 to S203: Step S201: The standardized satellite imagery and terrain-constrained data are encoded in multiple stages using a hybrid Mamba multi-scale feature encoder to obtain a multi-scale feature set.

[0042] It should be noted that multi-stage hybrid coding includes local multi-scale feature extraction and global long-sequence dependency modeling. Local multi-scale feature extraction is a multi-receptive-field feature acquisition operation performed on a local space. This operation achieves parallel extraction and adaptive fusion of features at different scales through depthwise convolutions with different dilation rates. This operation can effectively handle scenarios with diverse landslide morphologies and large scale spans. Global long-sequence dependency modeling is a long-distance association construction operation performed on a global space. This operation transforms two-dimensional features into one-dimensional sequences and completes global information modeling through a state-space model. This operation can overcome the limitations of the limited receptive field of traditional convolutions and capture large-scale geomorphological association information.

[0043] Further, step S201 includes: firstly, performing initial downsampling on the standardized satellite imagery through a pre-set stride convolution, and then extracting basic edge and color features through lightweight convolution to obtain initial shallow features; Secondly, the initial shallow features are downsampled and the channels are upsized using a pre-set stride convolution, and local multi-scale detail features are extracted using a dynamic scale mixing module to obtain the first encoded features. Specifically, the initial shallow features are first extracted using multiple parallel dilated depth convolutions to obtain multi-scale features from multiple branches. It should be noted that three parallel branches are set, each using dilated depth convolution as the core operation. These parallel dilated depth convolutions have different pre-set dilation rates, configured as 1, 2, and 4 respectively, with a kernel size of 3×3 and the number of groups set to the number of input channels. Dilated depth convolution expands the receptive field by inserting holes between standard convolution kernel elements. For example, a dilation rate of 1 results in a 3×3 receptive field, a dilation rate of 2 is equivalent to a 5×5 receptive field, and a dilation rate of 4 is equivalent to a 9×9 receptive field. Simultaneously, the specific... The characteristics of each branch are represented as follows: in Indicates the first Output features of each branch ( ), Indicates the void ratio The depthwise separable convolution is used, with SiLU as the activation function and BN as the normalization function. Then, the multi-branch, multi-scale features are element-wise summed to obtain the aggregated features. Specifically, it is expressed as This step adds the feature maps output by the three parallel branches at their corresponding spatial and channel positions, rather than performing a traditional channel concatenation operation.

[0044] Then, global adaptive average pooling is used to compress the spatial dimension of the aggregated features, yielding global channel statistics. It's important to understand that this step compresses the spatial dimension of the aggregated features into a single numerical value, extracting the global statistical characteristics of each channel. Specifically, it calculates the average of the pixel values ​​at all spatial locations for each channel in the aggregated features, outputting a one-dimensional vector with dimensions equal to the number of channels. The specific formula is as follows: in Indicates the first Global channel statistics for each channel, where each element represents the global average response strength of the corresponding channel; These represent the height and width of the feature map, respectively. These represent the indices in the height and width directions, respectively. Indicates the first Each channel, height position Width position Aggregation features at the location.

[0045] Next, a multilayer perceptron is used to perform channel compression and dimensionality upscaling on the global channel statistics to generate dynamic fusion weights. Specifically, the first fully connected layer compresses the number of channels C according to the formula d=max(C / 4,16) to reduce the number of parameters; the second fully connected layer upscals the compressed features to 3C and outputs the dynamic fusion weights, expressed by the following formula: in Indicates the first Dynamic fusion weights of each branch Indicates the first Intermediate variables of each branch Indicates a branch index; This represents the summation index, traversing all branches from 1 to 3; This represents the low-dimensional features after channel compression; This indicates the first layer of 1×1 convolution, which takes the number of input channels as an example. Compressed to low dimensions This enables channel compression. This displays global channel statistics. This represents the number of low-dimensional channels after compression; This indicates the second 1×1 convolution layer, which reduces the low-dimensional dimension. The dimensionality was restored to 3 times the original number of channels.

[0046] Next, channel-level soft attention weights are applied to the multi-branch, multi-scale features based on dynamic fusion weights, and the summation is performed element-wise to obtain the fused features. Then, the residuals of the fused features and the initial shallow features are added together to obtain the first encoded features. The specific formula is expressed as follows: Here, ⊙ represents element-wise multiplication. Indicates fusion characteristics, This represents the initial shallow features.

[0047] The first encoded feature is downsampled using a convolution with a preset stride, and local multi-scale features are extracted using a dynamic scale mixing module to obtain the second encoded feature. The second encoded feature is then downsampled using a convolution with a preset stride, and local multi-scale features are extracted using the same module to obtain local encoded features. Specifically, a standard convolution with a preset kernel size of 3×3 and a preset stride of 2 is used to downsample the first encoded feature, halving both the height and width while doubling the number of channels. The downsampled feature is then input into the dynamic scale mixing module, where the dilatancy is adjusted to a preset combination (1,2,3) to obtain the second encoded feature, adapting to the needs of medium-scale feature extraction. Similarly, a standard convolution with a preset kernel size of 3×3 and a preset stride of 2 is used to downsample the second encoded feature, which is then input into the dynamic scale mixing module for multi-scale feature extraction to obtain local encoded features.

[0048] Next, the local encoded features are input into the global feature extraction module based on the state-space model for long-distance dependency modeling to obtain global encoded features. It's important to understand that the local encoded features are first spatially enhanced through deep convolution, resulting in spatially enhanced features. This step injects a spatial inductive bias into the feature map to preserve local topological structure before converting the two-dimensional feature map into a one-dimensional sequence. Specifically, a deep convolution with a preset kernel size is used to perform channel-by-channel spatial filtering on the initial shallow features. Each channel undergoes independent convolution operations without mixing information across channels, outputting spatially enhanced features. The specific formula is as follows: in, Indicates spatial enhancement features, This represents a 3×3 depthwise convolution operation. This represents the local encoded features. Then, the spatial enhancement features are transposed and flattened to obtain one-dimensional serialized features, as shown in the following formula: in Represents one-dimensional serialization features. This indicates the dimension transpose operation. Presentation layer normalization operation, This indicates a flattening operation.

[0049] One-dimensional serialized features are input into a state-space model for selective scanning and global context modeling to obtain global sequence features. This step leverages the linear complexity of the state-space model to achieve global long-distance dependency modeling. Specifically, the one-dimensional serialized features are input into the core layer of the state-space model, which includes three sub-components: input projection, a selective scanning mechanism, and output projection. Input projection maps the features to an expanded dimension. The selective scanning mechanism dynamically adjusts the state transition matrix and input correlation matrix based on the input content, enabling the model to selectively focus on long-distance contextual information related to landslides and ignore irrelevant background regions. Output projection maps the scanning results back to the original dimension. The state dimension of the state-space model is set to 16, the internal local convolution kernel size is set to 4, and the feature expansion factor is strictly controlled to 1 to ensure the model's lightweight characteristics. The entire scanning process is completed with computational complexity linear with the sequence length, as shown in the following formula: in Represents global sequence features. It represents the operation of the core module of the state-space model, including a selective scanning mechanism and a hardware-aware algorithm, used to achieve global long-distance dependency modeling with linear computational complexity.

[0050] By performing inverse reshaping and dimensionality restoration on the global sequence features, the global encoded features are obtained, as expressed by the following formula: in This represents the global encoding feature.

[0051] Next, the terrain constraint data is concatenated with the global coding features as an additional channel. A gating fusion mechanism is then used to weight the concatenated features according to terrain constraints, resulting in hybrid coding features. Specifically, the digital elevation, slope, and lithology data of the terrain constraint data are spatially aligned to obtain first and second terrain constraint features that match the size of the global coding features. These first, second, and global coding features are then concatenated along the channel dimension to obtain concatenated features. Channel fusion is performed on the concatenated features using a pre-defined convolution to obtain fused channel features. A gating fusion mechanism is used to generate terrain constraint weights on the fused channel features, where these weights characterize the influence of slope and lithology on the probability of landslide occurrence. Finally, the local coding features are weighted and modulated according to these terrain constraint weights to obtain hybrid coding features. Specifically, firstly, digital elevation slope data and lithology data are resampled to ensure their spatial resolution is strictly aligned with the height and width of the local coding features, guaranteeing that optical features, slope values, and lithology values ​​at the same pixel location correspond to the same geographical location on the surface. Then, the first topographic constraint feature, the second topographic constraint feature, and the local coding features are stacked along the channel dimension to generate a spliced ​​feature with the sum of the three channels. Next, a 1×1 convolution is used to perform channel dimensionality reduction and fusion on the spliced ​​feature, initially transforming heterogeneous data into a joint feature representation. Then, two parallel branches of a gated fusion mechanism generate visual feature retention weights and topographic constraint modulation weights, respectively. These weights are mapped to a 0-1 range using a Sigmoid function, where the topographic constraint weights reflect the physical probability of a landslide at the current location. For example, the weights for steep slopes and soft lithology tend to be close to 1, while the weights for gentle slopes and hard lithology tend to be close to 0. Finally, the topographic constraint weights are multiplied element-wise with the local coding features to achieve weighted modulation, and then added to the local coding feature residuals to output a hybrid coding feature. The reason for this is that landslides have clear physical laws. The slope determines the stability threshold of the soil and rock mass under gravity, and the lithology determines the shear strength and permeability. Pure visual feature segmentation is prone to misjudging areas such as bare land and road excavation surfaces that are similar to landslide spectra but have very different physical conditions. By using the physical guidance of terrain constraint weights, the model can distinguish between areas that "look like landslides" and areas that "are physically landslides", making the segmentation results more consistent with the mechanical mechanism and improving the discrimination accuracy and interpretability of the results in complex backgrounds.

[0052] Finally, a multi-scale feature set is constructed based on the initial shallow features, the first coding features, the second coding features, and the mixed coding features. Specifically, the initial shallow features are obtained by initially downsampling the standardized input image using a pre-stance convolution and then extracting them through lightweight convolution. This preserves high-resolution spatial texture and color edge information, with a spatial size half that of the input image and a channel number that is a pre-defined base value. The first encoded feature is obtained by the dynamic scale mixing module extracting multi-scale local details from the initial shallow features. The spatial size is further halved, the number of channels is doubled, and it contains multi-scale local texture information. The second encoded feature is obtained by the dynamic scale mixing module extracting multi-scale local features from the first encoded feature again. The spatial size is halved again, the number of channels is doubled again, and it contains medium-scale local semantic information. The hybrid encoded feature is obtained by concatenating terrain constraint data with local encoded features and then weighting and modulating them using a physical attention gating mechanism. The spatial size is the same as the global encoded feature, and it is a hybrid perception feature that integrates physical constraints. These four features are organized into a multi-scale feature set in order from shallow to deep and from high resolution to low resolution. The initial shallow features, the first encoded feature, and the second encoded feature serve as the source of shallow features for subsequent decoder skip connections, while the hybrid encoded feature serves as the input to the multi-scale adaptive gating module. The reason for this is that landslide segmentation tasks need to simultaneously consider global morphological perception and precise local boundary localization. Features at a single scale cannot meet this requirement. Multi-scale feature sets cover a complete feature hierarchy from high-resolution spatial texture to low-resolution global semantics, enabling the subsequent decoder to dynamically select the most suitable feature combination at different stages through a progressive frequency domain calibration fusion strategy. This achieves a virtuous cycle of deep semantic guidance for shallow screening and shallow detail supplementation for deep recovery, ultimately outputting a refined landslide segmentation mask.

[0053] Step S202: Multi-scale context aggregation and adaptive semantic modulation are performed on the hybrid encoded features in the multi-scale feature set through the multi-scale adaptive gating module to obtain the enhanced deep semantic features.

[0054] It should be noted that step S202 includes: extracting context information of different ranges from the hybrid coding features through multiple parallel dilated convolutions to obtain context features of multiple scales, wherein the multiple parallel dilated convolutions have different preset dilation rates; concatenating the multiple scale context features by channels and generating a context weight map by dimensionality reduction through preset convolutions; generating multi-channel gating signals based on the hybrid coding features through multiple parallel preset convolutions; performing soft selection fusion on the original features and context enhancement features according to the first and second gating signals in the multi-channel gating signals to obtain intermediate modulation features; and performing channel-level importance filtering on the intermediate modulation features according to the third gating signal in the multi-channel gating signals to obtain enhanced deep semantic features.

[0055] Specifically, three parallel dilated convolutions are first used to extract contextual information of different ranges from the hybrid encoded features. The dilation rates are configured as 1, 2, and 4, respectively, resulting in 3×3, 5×5, and 9×9 features. These focus on local details, medium-range context, and large-scale geological background, respectively, outputting contextual features at multiple scales. The specific formula is as follows: in Indicates hybrid coding features, The local scale context features are extracted by dilated convolution with a dilation rate of 1. This represents a 3×3 dilated convolution with a dilation rate of 1, and an effective receptive field of 3×3, used to extract contextual information within a local area. The medium-scale contextual features are extracted by dilated convolution with a dilation rate of 2. This represents a 3×3 dilated convolution with a dilation rate of 2, and an effective receptive field of 5×5, used to extract contextual information within a moderate range. The medium-scale contextual features are extracted by dilated convolution with a dilation rate of 4. This represents a 3×3 dilated convolution with a dilation rate of 4, and an effective receptive field of 9×9, used to extract a large range of contextual information.

[0056] Then, the multiple scale context features are concatenated along the channel dimension, and dimensionality is reduced by a 1×1 convolution. A context weight map is then generated using the Sigmoid activation function, with the specific formula as follows: in It is the Sigmoid activation function. The context weight map represents the importance of each spatial location and each channel within the context. Then, multiple gated signals are generated through three parallel convolutions with a pre-defined kernel size of 1×1: a first gated signal, a second gated signal, and a third gated signal. The specific formula is as follows: in Indicates the first gating signal. This represents a 1×1 convolution that generates the first gate signal; the superscript k indicates that the convolution is specifically used to generate the first gate signal. Indicates the second gating signal. This represents a 1×1 convolution that generates the second gate signal; the superscript g indicates that the convolution is specifically used to generate the second gate signal. Indicates the third gating signal. This represents a 1×1 convolution that generates the third gating signal; the superscript 'q' indicates that this convolution is specifically used to generate the third gating signal. Then, soft-selective fusion is performed on the original features and context-enhanced features based on the first and second gating signals. The first gating signal determines how much original detail is retained, and the second gating signal determines how much contextual information is introduced. The two are adaptively fused through element-wise multiplication and addition, outputting intermediate modulation features. The specific formula is as follows: Where ⊙ denotes element-wise multiplication. The intermediate modulation features are represented; finally, the intermediate modulation features are filtered by channel-level importance based on the third gating signal. The third gating signal introduces stronger nonlinearity through element-wise multiplication, suppressing background noise and highlighting the landslide region. The output is an enhanced deep semantic feature that maintains the same dimension as the input feature map. The specific formula is as follows: in The enhanced deep semantic features are represented by the hybrid encoded features, which contain both high-level semantic information and compressed spatial details. These features are key nodes connecting contextual reasoning and accurate segmentation. Multi-scale context extraction enables the model to perceive landslide-related environmental information at different ranges simultaneously. Adaptive gating modulation allows the network to dynamically determine the fusion ratio of original details and contextual information based on local features, avoiding a one-size-fits-all approach to feature processing. Channel-level importance filtering further enhances the non-linear expressive power of the features, making the landslide area stand out in a complex background and providing a high-quality deep semantic foundation for accurate recovery by the subsequent decoder.

[0057] Step S203: The enhanced deep semantic features are restored step by step through progressive frequency domain calibration fusion decoder, and cross-layer fusion and boundary calibration are performed by combining the initial shallow features in the multi-scale feature set to obtain the landslide segmentation mask.

[0058] It should be noted that step S203 includes: enhancing the high-frequency edge information of the initial shallow features using a texture compensation operator to obtain edge-enhanced initial shallow features; improving the resolution of the enhanced deep semantic features using a lightweight dynamic upsampling module, and projecting the number of channels to the same level as the edge-enhanced initial shallow features using a preset convolution to obtain projected deep features; concatenating the projected deep features and the edge-enhanced initial shallow features in the channel dimension to generate a feature calibration gating map; selectively preserving and suppressing noise in the edge-enhanced initial shallow features according to the feature calibration gating map to obtain calibrated initial shallow features; concatenating the projected deep features and the calibrated initial shallow features, and fusing them using a lightweight convolution to obtain the current decoding level output features; and performing multi-level cascaded progressive restoration on the current decoding level output features until the original input resolution is restored to obtain a landslide segmentation mask.

[0059] Specifically, the initial shallow features transmitted by skip connections are first enhanced with high-frequency edge information using a texture compensation operator. This operator uses max pooling to extract low-frequency smooth components and obtains high-frequency edge information by the difference between the original image and the low-frequency components. The high-frequency edge information is then injected into the initial shallow features to enhance the model's ability to perceive landslide boundary textures and output edge-enhanced initial shallow features. At the same time, the resolution of the enhanced deep semantic features is improved by a lightweight dynamic upsampling module. This module generates dynamic offsets through a sampling point generator and combines them with the original sampling grid to form a sampling set for resampling, achieving efficient upsampling operations. Then, the number of channels is projected to the same level as the edge-enhanced initial shallow features through a convolution with a preset kernel size (1×1) to obtain projected deep features. Next, the projected deep features and the edge-enhanced initial shallow features are concatenated in the channel dimension. A feature calibration gating map is generated through a convolution with a preset kernel size (1×1). Regions with gating values ​​close to the preset high value (1) indicate that the deep and shallow features are consistent, preserving shallow details. Regions with gating values ​​close to the preset low value (0) The domain suppresses potential noise responses in the shallow layers; then, based on the feature calibration gating map, the initial shallow features for edge enhancement are selectively preserved and noise suppressed, outputting calibrated initial shallow features; then, the projected deep features are concatenated with the calibrated initial shallow features and fused through lightweight convolution. This lightweight convolution reduces redundant computation by introducing a preset module (Ghost module) and promotes cross-channel information interaction using a preset operation (Shuffle operation). The fused features are batch normalized and activated by a preset activation function (ReLU) to output the current decoding level output features; finally, the current decoding level output features are progressively restored through multi-level cascade. The first level receives the deep features processed by the multi-scale adaptive gating module and fuses them with the third layer features of the encoder, outputting the first-level features with half the number of channels and doubled resolution; the second level receives the first-level features and fuses them with the second layer features of the encoder, again halving the number of channels and doubling the resolution; the third level receives the second-level features and fuses them with the first layer features of the encoder, finally restoring the original input resolution and outputting the landslide segmentation mask. The reason for this approach is that landslide disasters in satellite imagery typically appear as irregularly shaped, blurred-boundary areas with spectral characteristics similar to the surrounding environment. Traditional decoders' simple upsampling and skip-connection stitching are insufficient to effectively address the semantic gap between deep semantics and shallow textures, leading to blurry or mis-segmented landslide boundaries. The progressive frequency domain calibration fusion strategy uses a semantically guided feature calibration gating mechanism to selectively enhance shallow features with deep semantics, avoiding shallow noise interference in the decoding process. Simultaneously, the injection of high-frequency edge information supplements the spatial details lost during decoding. The multi-level cascaded progressive recovery avoids information loss caused by excessively large single upsampling factors, ensuring precise positioning of landslide boundaries.

[0060] This embodiment employs a hybrid Mamba multi-scale feature encoder to perform multi-stage hybrid encoding on standardized satellite imagery and terrain-constrained data, obtaining a multi-scale feature set. This set is then processed by a multi-scale adaptive gating module for context aggregation and semantic modulation. Finally, a progressive frequency domain calibration fusion decoder completes resolution restoration, cross-layer fusion, and boundary calibration, outputting a landslide segmentation mask. This approach efficiently captures global context and multi-scale features, improving the accuracy of landslide identification with blurred boundaries. Simultaneously, the integration of terrain constraints enhances robustness, balancing segmentation accuracy with model lightweighting, making it more suitable for intelligent monitoring needs of highway landslides.

[0061] Based on the first embodiment of this application, this application also provides a landslide segmentation device based on hybrid Mamba and frequency domain calibration. Please refer to... Figure 3 The device includes: The acquisition module 10 is used to acquire the optical satellite image to be segmented and to acquire the terrain constraint data corresponding to the optical satellite image, wherein the terrain constraint data includes digital elevation slope data and lithological data.

[0062] The data preprocessing module 20 is used to preprocess optical satellite images to obtain standardized satellite images.

[0063] The result module 30 is used to process standardized satellite imagery and terrain-constrained data into a preset satellite imagery landslide segmentation model to obtain a landslide segmentation mask. The preset satellite imagery landslide segmentation model includes a hybrid Mamba multi-scale feature encoder, a multi-scale adaptive gating module, and a progressive frequency domain calibration fusion decoder. It is also used to perform multi-stage hybrid encoding on the standardized satellite imagery and terrain-constrained data using the hybrid Mamba multi-scale feature encoder to obtain a multi-scale feature set. The multi-stage hybrid encoding includes local multi-scale feature extraction and global long sequence dependency modeling. The multi-scale adaptive gating module performs multi-scale context aggregation and adaptive semantic modulation on the hybrid encoded features in the multi-scale feature set to obtain enhanced deep semantic features. The progressive frequency domain calibration fusion decoder performs stepwise resolution recovery on the enhanced deep semantic features and combines them with the initial shallow features in the multi-scale feature set for cross-layer fusion and boundary calibration to obtain the landslide segmentation mask.

[0064] The landslide segmentation device based on hybrid Mamba and frequency domain calibration provided in this application, employing the landslide segmentation method based on hybrid Mamba and frequency domain calibration in the above embodiments, can solve the technical problem of how to achieve high-precision and lightweight segmentation of highway landslides in optical satellite imagery. Compared with the prior art, the beneficial effects of the landslide segmentation device based on hybrid Mamba and frequency domain calibration provided in this application are the same as those of the landslide segmentation method based on hybrid Mamba and frequency domain calibration provided in the above embodiments, and other technical features in the landslide segmentation device based on hybrid Mamba and frequency domain calibration are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0065] This application provides a landslide segmentation device based on hybrid Mamba and frequency domain calibration. The landslide segmentation device based on hybrid Mamba and frequency domain calibration includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the landslide segmentation method based on hybrid Mamba and frequency domain calibration in the above embodiment 1.

[0066] The following is for reference. Figure 4 This document illustrates a structural schematic diagram of a landslide segmentation device based on hybrid Mamba and frequency domain calibration, suitable for implementing embodiments of this application. The landslide segmentation device based on hybrid Mamba and frequency domain calibration in this application embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The landslide segmentation device based on hybrid Mamba and frequency domain calibration shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0067] like Figure 4As shown, the landslide segmentation device based on hybrid Mamba and frequency domain calibration may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the landslide segmentation device based on hybrid Mamba and frequency domain calibration. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the landslide segmentation device based on hybrid Mamba and frequency domain calibration to communicate wirelessly or wiredly with other devices to exchange data. Although various types of landslide segmentation devices based on hybrid Mamba and frequency domain calibration are shown in the figures, it should be understood that implementation or possession of all of them is not required. More or fewer of these devices may be implemented alternatively.

[0068] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0069] The landslide segmentation device based on hybrid Mamba and frequency domain calibration provided in this application, employing the landslide segmentation method based on hybrid Mamba and frequency domain calibration described in the above embodiments, can solve the technical problem of how to achieve high-precision and lightweight segmentation of highway landslides in optical satellite imagery. Compared with the prior art, the beneficial effects of the landslide segmentation device based on hybrid Mamba and frequency domain calibration provided in this application are the same as those of the landslide segmentation method based on hybrid Mamba and frequency domain calibration provided in the above embodiments, and other technical features of this landslide segmentation device based on hybrid Mamba and frequency domain calibration are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0070] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0071] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0072] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the landslide segmentation method based on hybrid Mamba and frequency domain calibration in the above embodiments.

[0073] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0074] The aforementioned computer-readable medium may be included in a landslide segmentation device based on hybrid Mamba and frequency domain calibration; or it may exist independently and not assembled into a landslide segmentation device based on hybrid Mamba and frequency domain calibration.

[0075] The aforementioned computer-readable medium carries one or more programs that, when executed by a landslide partitioning device based on hybrid Mamba and frequency domain calibration, enable the device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0076] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0077] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0078] The readable medium provided in this application is a computer-readable medium that stores computer-readable program instructions (i.e., a computer program) for executing the landslide segmentation method based on hybrid Mamba and frequency domain calibration described above. This solves the technical problem of how to achieve high-precision, lightweight segmentation of highway landslides in optical satellite imagery. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the landslide segmentation method based on hybrid Mamba and frequency domain calibration provided in the above embodiments, and will not be repeated here.

[0079] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the landslide segmentation method based on hybrid Mamba and frequency domain calibration as described above.

[0080] The computer program product provided in this application solves the technical problem of how to achieve high-precision and lightweight segmentation of highway landslides in optical satellite imagery. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the landslide segmentation method based on hybrid Mamba and frequency domain calibration provided in the above embodiments, and will not be repeated here.

[0081] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A landslide segmentation method based on hybrid Mamba and frequency domain calibration, characterized in that, The method includes: Acquire optical satellite images to be segmented, and acquire terrain constraint data corresponding to the optical satellite images, wherein the terrain constraint data includes digital elevation slope data and lithological data; The optical satellite images are preprocessed to obtain standardized satellite images; The standardized satellite imagery and terrain constraint data are input into a preset satellite imagery landslide segmentation model for processing to obtain a landslide segmentation mask. The preset satellite imagery landslide segmentation model includes a hybrid Mamba multi-scale feature encoder, a multi-scale adaptive gating module, and a progressive frequency domain calibration fusion decoder. The step of inputting the standardized satellite imagery and terrain constraint data into a preset satellite imagery landslide segmentation model for processing to obtain a landslide segmentation mask includes: The standardized satellite imagery and terrain-constrained data are multi-stage hybrid encoded using a hybrid Mamba multi-scale feature encoder to obtain a multi-scale feature set. This multi-stage hybrid encoding includes local multi-scale feature extraction and global long sequence dependency modeling. The hybrid Mamba multi-scale feature encoder comprises a dynamic scale mixing module, a state-space model-based global feature extraction module, and a local-global gating fusion mechanism. The dynamic scale mixing module includes multiple parallel dilated deep convolutional branches, element-wise addition units, a global adaptive average pooling layer, a multilayer perceptron, and a channel-level soft attention weighting unit. The multiple parallel dilated deep convolutional branches have different preset dilation rates for extracting multi-scale receptive field features. The state-space model-based global feature extraction module includes deep convolutional layers, dimension transpose units, layer normalization layers, a state-space model core layer, and an inverse reshaping unit. The state-space model core layer includes a selective scanning mechanism and a hardware-aware sensing algorithm. The local-global gating fusion mechanism includes a channel stitching unit, preset convolutional layers, and an adaptive gating weight generation unit. By using a multi-scale adaptive gating module, the hybrid encoded features in the multi-scale feature set are subjected to multi-scale context aggregation and adaptive semantic modulation to obtain enhanced deep semantic features. The enhanced deep semantic features are restored step by step through a progressive frequency domain calibration fusion decoder, and cross-layer fusion and boundary calibration are performed by combining the initial shallow features in the multi-scale feature set to obtain a landslide segmentation mask. The standardized satellite imagery and terrain-constrained data are subjected to multi-stage hybrid encoding using a hybrid Mamba multi-scale feature encoder to obtain a multi-scale feature set. The multi-stage hybrid encoding includes steps of local multi-scale feature extraction and global long-sequence dependency modeling, including: The standardized satellite image is initially downsampled by convolution with a preset stride, and basic edge and color features are extracted by lightweight convolution to obtain initial shallow features; The initial shallow features are downsampled and channel-upgraded by convolution with a preset stride, and local multi-scale detail features are extracted by a dynamic scale mixing module to obtain the first encoded features. The first encoded feature is downsampled by convolution with a preset stride, and the second encoded feature is obtained by extracting local multi-scale features through the dynamic scale mixing module. The second encoded feature is downsampled by convolution with a preset stride, and local multi-scale features are extracted by the dynamic scale mixing module to obtain local encoded features; The local encoded features are input into the global feature extraction module based on the state space model to perform long-distance dependency modeling, thereby obtaining the global encoded features; The terrain constraint data is concatenated with the global coding features as an additional channel, and the concatenated features are weighted by terrain constraints through a gated fusion mechanism to obtain hybrid coding features; A multi-scale feature set is constructed based on the initial shallow features, the first encoded features, the second encoded features, and the hybrid encoded features.

2. The method as described in claim 1, characterized in that, The step of downsampling and upscaling the initial shallow features through convolution with a preset stride, and extracting local multi-scale detail features through a dynamic scale mixing module to obtain the first encoded features includes: Multiple branch multi-scale features are obtained by extracting the receptive field of the initial shallow features through multiple parallel dilated depth convolutions, wherein the multiple parallel dilated depth convolutions have different preset dilation rates. The aggregated features are obtained by adding the multiple branch multi-scale features element by element. The aggregated features are spatially compressed using global adaptive average pooling to obtain global channel statistics. The global channel statistics are compressed and mapped to higher dimensions using a multilayer perceptron to generate dynamic fusion weights. Based on the dynamic fusion weights, channel-level soft attention weights are applied to the multiple branch multi-scale features, and element-wise summation is performed to obtain the fused features; The fused features are added to the initial shallow features by residual addition to obtain the first encoded features.

3. The method as described in claim 1, characterized in that, The step of inputting the local encoded features into the global feature extraction module based on the state-space model for long-distance dependency modeling to obtain global encoded features includes: The local encoded features are spatially enhanced by performing local convolution to obtain spatially enhanced features; Perform dimensional transpose and flatten operations on the spatial enhancement features to obtain one-dimensional serialized features; The one-dimensional serialized features are input into the state space model for selective scanning and global context modeling to obtain global sequence features; The global sequence features are reverse-engineered and their dimensions restored to obtain the global encoded features.

4. The method as described in claim 1, characterized in that, The step of concatenating the terrain constraint data as an additional channel with the global encoded features, and applying terrain constraint weighting to the concatenated features through a gated fusion mechanism to obtain hybrid encoded features, further includes: The digital elevation and slope data and lithological data of the terrain constraint data are spatially aligned to obtain a first terrain constraint feature and a second terrain constraint feature that match the size of the global coding feature. The first terrain constraint feature, the second terrain constraint feature, and the global encoding feature are concatenated along the channel dimension to obtain the concatenated feature; The spliced ​​features are fused by channel fusion using a preset convolution to obtain fused channel features; The gating fusion mechanism generates terrain constraint weights for the fusion channel features, wherein the terrain constraint weights are used to characterize the degree of influence of slope and lithology on the probability of landslide occurrence. The local coding features are weighted and modulated according to the terrain constraint weights to obtain hybrid coding features.

5. The method as described in claim 1, characterized in that, The step of performing multi-scale context aggregation and adaptive semantic modulation on the hybrid encoded features in the multi-scale feature set through a multi-scale adaptive gating module to obtain enhanced deep semantic features includes: The hybrid encoded features are subjected to context information extraction at different ranges by multiple parallel dilated convolutions to obtain multiple scale context features, wherein the multiple parallel dilated convolutions have different preset dilation rates. The multiple scale context features are concatenated by channels, and a context weight map is generated by dimensionality reduction through preset convolution. Based on the hybrid coding features, multiple gated signals are generated through multiple parallel preset convolutions; Based on the first and second gate signals in the multi-gated signals, soft selection fusion is performed on the original features and context enhancement features to obtain intermediate modulation features; Based on the third gating signal in the multi-gating signals, the intermediate modulation features are filtered by channel-level importance to obtain the enhanced deep semantic features.

6. The method as described in claim 1, characterized in that, The step of progressively restoring the enhanced deep semantic features through a progressively frequency-domain calibrated fusion decoder, and then performing cross-layer fusion and boundary calibration in conjunction with the initial shallow features in the multi-scale feature set to obtain the landslide segmentation mask includes: The initial shallow features are enhanced with high-frequency edge information by using a texture compensation operator to obtain edge-enhanced initial shallow features. The resolution of the enhanced deep semantic features is improved by a lightweight dynamic upsampling module, and the number of channels is projected to the same level as the initial shallow features of the edge enhancement by a preset convolution to obtain the projected deep features. The projected deep features are concatenated with the edge-enhanced initial shallow features in the channel dimension to generate a feature calibration gating map. Based on the feature calibration gating map, the initial shallow features of edge enhancement are selectively preserved and noise suppressed to obtain the calibrated initial shallow features; The projected deep features are concatenated with the calibrated initial shallow features and fused using lightweight convolution to obtain the output features of the current decoding level. The output features of the current decoding level are progressively restored through multi-level cascade until the original input resolution is restored, thus obtaining the landslide segmentation mask.

7. A landslide segmentation device based on hybrid Mamba and frequency domain calibration, characterized in that, The apparatus is applied to the landslide segmentation method based on hybrid Mamba and frequency domain calibration as described in any one of claims 1-6, and the apparatus comprises: The acquisition module is used to acquire the optical satellite image to be segmented and to acquire the terrain constraint data corresponding to the optical satellite image, wherein the terrain constraint data includes digital elevation slope data and lithological data. The data preprocessing module is used to preprocess the optical satellite imagery to obtain standardized satellite imagery; The result module is used to input the standardized satellite imagery and terrain-constrained data into a preset satellite imagery landslide segmentation model for processing to obtain a landslide segmentation mask. The preset satellite imagery landslide segmentation model includes a hybrid Mamba multi-scale feature encoder, a multi-scale adaptive gating module, and a progressive frequency domain calibration fusion decoder. It is also used to perform multi-stage hybrid encoding on the standardized satellite imagery and terrain-constrained data using the hybrid Mamba multi-scale feature encoder to obtain a multi-scale feature set. The multi-stage hybrid encoding includes local multi-scale feature extraction and global long-sequence dependency modeling. The multi-scale adaptive gating module performs multi-scale context aggregation and adaptive semantic modulation on the hybrid encoded features in the multi-scale feature set to obtain enhanced deep semantic features. The progressive frequency domain calibration fusion decoder performs stepwise resolution recovery on the enhanced deep semantic features and combines them with the initial shallow features in the multi-scale feature set for cross-layer fusion and boundary calibration to obtain the landslide segmentation mask.

8. A landslide segmentation device based on hybrid Mamba and frequency domain calibration, characterized in that, The device includes: a memory, a processor, and a landslide segmentation program based on hybrid Mamba and frequency domain calibration stored in the memory and running on the processor, the landslide segmentation program based on hybrid Mamba and frequency domain calibration being configured to implement the steps of the landslide segmentation method based on hybrid Mamba and frequency domain calibration as described in any one of claims 1-6.

9. A storage medium, characterized in that, The storage medium stores a landslide segmentation program based on hybrid Mamba and frequency domain calibration, which, when executed by a processor, implements the steps of the landslide segmentation method based on hybrid Mamba and frequency domain calibration as described in any one of claims 1-6.

Images