Landslide hazard remote sensing detection method and system integrating spectral and terrain information

Abstract

The present application discloses a landslide hazard remote sensing detection method and system integrating spectral and terrain information, belonging to the field of remote sensing technology; the method comprises: establishing a landslide hazard sample list; establishing a spectral and terrain constrained semantic segmentation network model according to the landslide hazard sample list; establishing a weighted loss function for multiple scale spaces, and optimizing the spectral and terrain constrained semantic segmentation network model according to the weighted loss function; iteratively training the spectral and terrain constrained semantic segmentation network model according to the landslide hazard sample list and the weighted loss function; predicting the landslide hazard segmentation and verifying the accuracy of the terrain constrained semantic segmentation network model after iterative training according to the landslide hazard sample list; the spectral and terrain constrained semantic segmentation network model constructed by deep learning technology breaks through the limitations of the traditional single data source and network architecture, and improves the detection efficiency and accuracy.

Description

Technical Field

[0001] The present application belongs to the field of remote sensing technology, and specifically relates to a landslide hazard remote sensing detection method and system that integrates spectrum and terrain information. Background Art

[0002] Landslide disasters seriously endanger people's lives and property due to their extremely destructive power, suddenness, and wide range. Rapid and accurate detection of landslide hazards can timely reflect their spatial distribution and spatiotemporal dynamic processes, and has important practical significance in ensuring the safety of national and regional geological environments. Compared with high-cost and difficult-to-implement field survey methods, the rapid development of earth observation and remote sensing technology has provided major opportunities for large-scale landslide hazard detection, but its easily confused characteristics such as high position, high concealment, and uncertainty have put forward higher requirements for the intelligence and accuracy of remote sensing detection of landslide hazards.

[0003] At present, the semantic segmentation framework based on deep convolutional neural network (CNN) has broken through the bottleneck of traditional machine learning methods in the field of landslide hazard detection due to its powerful ability to characterize nonlinear features, and has achieved relatively excellent results; however, the local connection and weight sharing properties of the convolutional filter determine the inductive bias of the CNN structure, namely locality and translation invariance, making it difficult to represent global information in intensive prediction-type landslide hazard detection tasks; and the Transformer architecture has gradually been applied to landslide hazard detection due to its outstanding contributions in long-range dependency sequence modeling and transduction, and has shown better performance than CNN, but the time and space complexity of the multi-head self-attention mechanism on which the Transformer structure relies is quadratic in the input length, has low generalization ability and large computational requirements, resulting in prediction. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide a landslide hazard remote sensing detection method and system that integrates spectral and terrain information, which can solve the technical problem of low prediction accuracy caused by single data source and network architecture in landslide hazard prediction in the prior art.

[0005] In order to solve the above technical problems, this application is implemented as follows:

[0006] In a first aspect, an embodiment of the present application provides a landslide hazard remote sensing detection method that integrates spectrum and terrain information, the method comprising:

[0007] Acquiring multispectral remote sensing image data and terrain data to establish a landslide hazard sample list based on the multispectral remote sensing image data and terrain data;

[0008] According to the landslide hazard sample list, a spectral and terrain constrained semantic segmentation network model including a spectral feature encoding module, a terrain feature encoding module and a heterogeneous feature fusion decoding module is established;

[0009] Establishing a weighted loss function for multiple scale spaces, and optimizing the spectral and terrain constrained semantic segmentation network model according to the weighted loss function;

[0010] Iteratively training the spectral and terrain constrained semantic segmentation network model according to the landslide hazard sample list and the weighted loss function;

[0011] According to the landslide hazard sample list, the terrain-constrained semantic segmentation network model after iterative training is used to predict the landslide hazard segmentation and verify its accuracy.

[0012] As an optional implementation of the first aspect of the present application, the landslide hazard sample list is established based on the multispectral remote sensing image data and terrain data, specifically:

[0013] Performing a first preprocessing on the acquired multispectral remote sensing image data, and then processing the multispectral remote sensing image data after the first preprocessing using a conventional conversion algorithm to obtain multispectral remote sensing image feature data including a normalized vegetation index, a normalized water index, a brightness index, and a bare soil index; the first preprocessing includes geometric correction, atmospheric correction, and image mosaicking;

[0014] Performing a second preprocessing on the acquired terrain data, and then applying the conventional conversion algorithm to the terrain data after the second preprocessing to obtain terrain feature data including slope, aspect, curvature, shadow and terrain moisture index; the second preprocessing includes geo-registration and resampling;

[0015] Performing visual interpretation processing on the geographical facts generated by the multispectral remote sensing image feature data and the terrain feature data to obtain ground truth label data of landslide hazards;

[0016] The multispectral remote sensing image feature data, terrain feature data and ground truth label data are batch-cropped using a non-overlapping sliding window, and are divided into training samples and test samples according to a preset ratio.

[0017] As an optional implementation of the first aspect of the present application, the process of establishing the spectral and terrain constrained semantic segmentation network model is specifically as follows:

[0018] The spectral feature encoding module composed of efficient channel relationship enhancement blocks is established according to a deep convolutional neural network to obtain the local enhanced landslide hazard spectral features with multiple spatial scales; the spectral feature encoding module includes a residual convolution structure and an improved channel self-attention structure;

[0019] Establishing the terrain feature encoding module consisting of improved Swin-transformer blocks according to the Transformer architecture to obtain global enhanced terrain features with multiple spatial scales;

[0020] The heterogeneous feature fusion decoding module is established according to the local enhanced landslide hazard spectral characteristics and the global enhanced terrain characteristics to obtain landslide hazard segmentation prediction results with multiple spatial scales; the heterogeneous feature fusion decoding module includes a relationship weighted fusion structure and an expanded spatial pyramid pooling structure.

[0021] As an optional implementation of the first aspect of the present application, the acquisition of the spectral characteristics of the local enhanced landslide hazard with multiple spatial scales is specifically:

[0022] Establishing the residual convolution structure according to the bottleneck residual network, so as to obtain the deep spectral characteristics of the landslide hazard according to the residual convolution structure;

[0023] The improved channel self-attention structure is established according to the attention mechanism to enhance the deep-level spectral characteristics of the landslide hazard according to the improved channel self-attention structure, so as to obtain the locally enhanced landslide hazard spectral characteristics with multiple spatial scales.

[0024] As an optional implementation of the first aspect of the present application, the heterogeneous feature fusion decoding module is established to obtain a landslide hazard segmentation prediction result with multiple spatial scales, specifically:

[0025] Establishing a lightweight relational weighted fusion structure to fuse the local enhanced landslide hazard spectral features and the global enhanced terrain features according to the relational weighted fusion structure to obtain multi-scale fusion features;

[0026] The expanded spatial pyramid pooling structure is established to perform landslide hazard segmentation prediction on the multi-scale fusion features according to the expanded spatial pyramid pooling structure, so as to obtain the landslide hazard segmentation prediction result with multiple spatial scales.

[0027] As an optional implementation of the first aspect of the present application, the establishment of a weighted loss function for multiple scale spaces is specifically:

[0028] Establishing a class-balanced weighted cross entropy loss function according to the landslide hazard segmentation prediction result with multiple spatial scales and the ground truth label data;

[0029] Designing an adaptive weight value and a class-balanced weighted cross entropy loss function according to the loss value, and establishing the weighted loss function for multiple scale spaces;

[0030] The class-balanced weighted cross entropy loss function expression is:

[0031] ;

[0032] The weighted loss function expression for multiple scale spaces is:

[0033] ;

[0034] in, represents the class-balanced weighted cross entropy loss function, Indicates spatial scale, Indicates pixels, h represents the height of the pixel, w represents the width of the pixel, Indicates category labels, Indicates The pixels are labeled The ground truth value of Indicates The spatial scale The pixel label is The landslide hazard prediction score is represents the number of samples of the th category label, represents the hyperparameter, c represents the total number of categories, represents the weighted loss function for multiple scale spaces, and S represents the total number of spatial scales.

[0035] As an optional implementation of the first aspect of the present application, the iterative training of the spectral and terrain constrained semantic segmentation network model is specifically:

[0036] Performing data expansion on the training samples by using a geometric transformation enhancement method, wherein the geometric transformation enhancement method includes rotation, flipping, scaling and translation;

[0037] According to the training sample volume after the data expansion, setting training parameters for iterative training of the spectral and terrain constrained semantic segmentation network model;

[0038] The spectral and terrain constrained semantic segmentation network model is optimized and iteratively trained by using a stochastic gradient descent algorithm and an adaptive momentum estimation algorithm;

[0039] The spectral and terrain constrained semantic segmentation network model after the iterative training is saved.

[0040] In a second aspect, an embodiment of the present application provides a landslide hazard remote sensing detection system that integrates spectral and terrain information, the system comprising:

[0041] The first acquisition module is used to acquire multispectral remote sensing image data and terrain data;

[0042] A first processing module, used for processing the multispectral remote sensing image data and terrain data acquired by the first acquisition module to obtain a corresponding landslide hazard sample list;

[0043] A first establishing module is used to establish a spectral and terrain constrained semantic segmentation network model according to the slope hidden danger sample list obtained by the first processing module;

[0044] A second establishment module is used to establish a weighted loss function for multiple scale spaces, and optimize the spectral and terrain constrained semantic segmentation network model according to the weighted loss function for multiple scale spaces;

[0045] A first training module is used to iteratively train the optimized spectral and terrain constrained semantic segmentation network model;

[0046] The prediction verification module performs landslide hazard segmentation prediction and accuracy verification on the terrain-constrained semantic segmentation network model after iterative training according to the landslide hazard sample list.

[0047] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the method described in the first aspect.

[0048] In a fourth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented.

[0049] In the embodiments of the present application, by fusing multispectral remote sensing image data and terrain data, the spectral characteristics of surface materials and terrain morphological information are fully utilized to improve the accuracy of landslide hazard detection; the weighted loss function to multiple scale spaces can comprehensively consider the prediction errors at different scales, so that the model can show good performance at different scales, enhancing the generalization ability of the model; the spectral and terrain constrained semantic segmentation network model constructed using deep learning technology breaks through the limitations of traditional single data source and network architecture, and can realize automated and intelligent landslide hazard detection, improving detection efficiency and accuracy; through landslide hazard detection, a scientific basis can be provided for the early warning and prevention of landslide disasters, which helps to reduce the impact of landslide disasters on human society and the natural environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of a landslide hazard remote sensing detection method that integrates spectrum and terrain information provided by some embodiments of the present application;

[0051] Figure 2 It is a logic diagram of a landslide hazard remote sensing detection method that integrates spectrum and terrain information provided by some embodiments of the present application;

[0052] Figure 3 It is a schematic diagram of a spectral and terrain constrained semantic segmentation network of a landslide hazard remote sensing detection method that integrates spectral and terrain information provided by some embodiments of the present application;

[0053] Figure 4 It is a schematic diagram of the structure of a spectral feature encoding module of a landslide hazard remote sensing detection method that integrates spectral and terrain information provided by some embodiments of the present application;

[0054] Figure 5 It is a schematic diagram of the structure of a terrain feature encoding module of a landslide hazard remote sensing detection method that integrates spectrum and terrain information provided by some embodiments of the present application;

[0055] Figure 6 It is a schematic diagram of the structure of a heterogeneous feature fusion decoding module of a landslide hazard remote sensing detection method that fuses spectrum and terrain information provided by some embodiments of the present application. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0057] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here. In addition, the "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally represents that the objects associated with each other are in an "or" relationship.

[0058] In conjunction with the accompanying drawings, a landslide hazard remote sensing detection method and system that integrates spectrum and terrain information provided by the embodiment of the present application will be described in detail through specific embodiments and their application scenarios.

[0059] Example

[0060] See also Figure 1 and Figure 2 As shown, a landslide hazard remote sensing detection method integrating spectrum and terrain information includes the following steps:

[0061] S100: Acquire multispectral remote sensing image data and terrain data, and establish a landslide hazard sample list according to the multispectral remote sensing image data and terrain data;

[0062] It is important to understand that multispectral image data of the target area are obtained through remote sensing technology. These data contain spectral information of different bands and can reflect the different characteristics of surface materials. At the same time, obtaining terrain data helps to understand the surface morphology and geological structure. Based on the acquired multispectral remote sensing image data and terrain data, landslide risk areas are identified through manual annotation or existing data, and a list of landslide risk samples is established. This list contains information such as the location, size, shape, and corresponding spectral and terrain characteristics of the landslide risk areas.

[0063] Furthermore, in S100, a sample list of landslide hazard samples is established based on the spectral remote sensing image data and the terrain data, specifically:

[0064] S110: performing a first preprocessing on the acquired multispectral remote sensing image data, and then processing the multispectral remote sensing image data after the first preprocessing using a conventional conversion algorithm to obtain multispectral remote sensing image feature data including a normalized vegetation index, a normalized water index, a brightness index, and a bare soil index;

[0065] S120: performing a second preprocessing on the acquired terrain data, and then processing the terrain data after the second preprocessing using a conventional conversion algorithm to obtain terrain feature data including slope, slope aspect, curvature, shadow, and terrain humidity index;

[0066] S130: Performing visual interpretation processing on the data geographic facts generated by the multispectral remote sensing image feature data and the terrain feature data to obtain ground truth label data of the landslide hazard;

[0067] S140: using a non-overlapping sliding window to batch-crop the multispectral remote sensing image feature data, terrain feature data, and ground truth label data, and divide them into training samples and test samples according to a preset ratio;

[0068] The first preprocessing includes geometric correction, atmospheric correction and image mosaicking; the second preprocessing includes geographic registration and resampling.

[0069] It should be understood that, based on the above-mentioned feature limitations, the acquired multispectral remote sensing image data is first subjected to geometric correction (ensuring the spatial position accuracy of the image), atmospheric correction (eliminating the influence of the atmosphere on the image) and image mosaicking (joining multiple images into a complete image); the multispectral remote sensing image data after the first preprocessing is processed using a conventional conversion algorithm to extract multispectral remote sensing image feature data such as the normalized vegetation index (reflecting vegetation coverage), the normalized water index (reflecting water distribution), the brightness index (reflecting surface reflectivity) and the bare soil index (reflecting the bare soil situation); the acquired terrain data is geo-referenced (ensuring spatial consistency between the terrain data and the multispectral remote sensing image data) and resampled (adjusting the terrain data to the same resolution as the multispectral remote sensing image data); the second preprocessing is processed using a conventional conversion algorithm The processed terrain data are used to extract terrain feature data such as slope (reflecting the degree of inclination of the surface), slope aspect (reflecting the direction of the inclination of the surface), curvature (reflecting the degree of curvature of the surface), shadow (reflecting the blocking of sunlight by the terrain) and terrain moisture index (reflecting the degree of wetness of the terrain); based on the data geographical facts generated by multispectral remote sensing image feature data and terrain feature data, the landslide hazard area is identified through visual interpretation method, and the ground truth label data of the landslide hazard is generated; the multispectral remote sensing image feature data, terrain feature data and ground truth label data are batch cropped using a non-overlapping sliding window to generate a series of small-sized sample data; the cropped sample data are divided into training samples and test samples according to the preset ratio (such as 70% training samples and 30% test samples) for subsequent model training and verification.

[0070] It should be noted that the quality of multispectral remote sensing image data and terrain data has been improved through preprocessing steps such as geometric correction, atmospheric correction, image mosaicking, geo-registration and resampling, providing a reliable data basis for subsequent feature extraction and model training; a variety of multispectral remote sensing image feature data and terrain feature data have been extracted, such as normalized vegetation index, normalized water index, slope, slope aspect, etc. These feature information can comprehensively reflect the characteristics of surface materials and terrain, and help to improve the accuracy of landslide hazard detection; batch cropping is performed through a non-overlapping sliding window, and the cropped sample data are divided into training samples and test samples, which helps the model learn more feature information during the training process and improves the generalization ability of the model; based on the preprocessed multispectral remote sensing image data and terrain data, the spectral and terrain constrained semantic segmentation network model constructed in combination with deep learning technology can realize efficient and accurate landslide hazard detection, and provide a scientific basis for early warning and prevention of landslide disasters.

[0071] S200: Based on the list of landslide hazard samples, a spectral and terrain constrained semantic segmentation network model including a spectral feature encoding module, a terrain feature encoding module and a heterogeneous feature fusion decoding module is established;

[0072] It should be understood that the spectral feature encoding module is responsible for extracting spectral features from multispectral remote sensing images, encoding spectral information using deep learning technology, and generating spectral feature vectors; the terrain feature encoding module is responsible for extracting terrain features from terrain data, such as elevation, slope, aspect, etc., and also uses deep learning technology to encode and generate terrain feature vectors; the heterogeneous feature fusion decoding module is responsible for fusing spectral feature vectors and terrain feature vectors, and generating fused feature vectors through specific fusion strategies (such as splicing, weighting, etc.); then, the decoder is used to decode the fused feature vector into the prediction result of landslide hazards.

[0073] Furthermore, the specific establishment process of the spectral and terrain constrained semantic segmentation network model in S200 is as follows:

[0074] S210: A spectral feature encoding module consisting of efficient channel relationship enhancement blocks is established based on a deep convolutional neural network to obtain the spectral features of locally enhanced landslide hazards with multiple spatial scales;

[0075] S220: Establish a terrain feature encoding module consisting of improved Swin-transformer blocks based on the Transformer architecture to obtain global enhanced terrain features with multiple spatial scales;

[0076] S230: A heterogeneous feature fusion decoding module is established according to the local enhanced landslide hazard spectral features and the global enhanced terrain features to obtain landslide hazard segmentation prediction results with multiple spatial scales.

[0077] Among them, the spectral feature encoding module in S210 includes a residual convolution structure and an improved channel self-attention structure; the heterogeneous feature fusion decoding module in S230 includes a relational weighted fusion structure and an expanded spatial pyramid pooling structure.

[0078] It should be understood that, based on the above-mentioned feature limitations, the spectral feature encoding module is built on the basis of a deep convolutional neural network (CNN), and uses the convolutional layer, pooling layer and fully connected layer structures of CNN to automatically extract the spatial and spectral features in the image; this module contains efficient channel relationship enhancement blocks, which enhance the feature relationship between channels through an improved channel self-attention structure; the channel self-attention structure weights the features of different channels by learning channel weights, thereby improving the response of important channels and enhancing the expression ability of spectral features; the residual convolution structure is used to increase the network depth while avoiding the problem of gradient disappearance or explosion; by introducing residual connections, the network can learn deeper feature representations; through the above structure, the spectral feature encoding module can obtain the spectral features of locally enhanced landslide hazards with multiple spatial scales, which are crucial for the subsequent segmentation and prediction of landslide hazards; the terrain feature encoding module is built on the Transformer architecture, and uses the Transformer's self-attention mechanism and position encoding characteristics to capture global features; this module contains an improved Sw in-transformer blocks, which capture the global relationship of terrain features through the self-attention mechanism, and the Swin-transformer blocks realize the capture of features of different spatial scales through window division and shift operations; through the improved Swin-transformer blocks, the terrain feature encoding module can obtain global enhanced terrain features with multiple spatial scales; these features provide global context information for the segmentation prediction of landslide hazards; the heterogeneous feature fusion decoding module contains a relational weighted fusion structure, which performs weighted fusion of different features by learning the weight relationship between features. This fusion method can make full use of the advantages of spectral and terrain features and improve the accuracy of segmentation prediction; the expanded spatial pyramid pooling structure captures feature information of different spatial scales through pooling operations of different scales. This structure helps to enhance the model's ability to identify landslide hazards of different scales; through the joint action of the relational weighted fusion structure and the expanded spatial pyramid pooling structure, the heterogeneous feature fusion decoding module can output landslide hazard segmentation prediction results with multiple spatial scales.

[0079] Furthermore, in S210, the spectral characteristics of the local enhanced landslide hazard with multiple spatial scales are obtained, specifically:

[0080] S211: Establish a residual convolution structure based on the bottleneck residual network to obtain the deep spectral characteristics of landslide hazards according to the residual convolution structure;

[0081] S212: establishing an improved channel self-attention structure according to the attention mechanism, so as to enhance the deep-level spectral characteristics of landslide hazards according to the improved channel self-attention structure, and obtain the locally enhanced spectral characteristics of landslide hazards with multiple spatial scales;

[0082] Further, see Figure 4 As shown, S211 is specifically: using convolutional layers with kernel sizes of {1×1, 3×3, 1×1} to sequentially perform spatial resolution downsampling and local feature encoding, and then using residual identity mapping to perform deep feature refinement to obtain deep spectral features of landslide hazards; S212 is specifically: using a channel reshaping layer to reshape and group the deep spectral features of landslide hazards obtained in step S211 into four sub-features, and using a series of cross operations such as X-direction average pooling, Y-direction average pooling, 3×3 convolutional layers and softmax layers on three of the sub-features to generate keys and values ​​of channel relationships, and then generating local enhanced landslide hazard spectral features with multiple spatial scales based on matrix addition operations and dot product operations with the query represented by another sub-feature.

[0083] It should be understood that based on the above feature restrictions, in S211, a bottleneck residual network is used to construct a residual convolution structure. The bottleneck residual network reduces the computational complexity by reducing the number of channels in the middle layer while maintaining the performance of the network; the kernel sizes used are {1×1, 3×3, 1×1} convolutional layer is used to downsample the spatial resolution and encode local features; 1×1 convolutional layer is used to adjust the number of channels, 3×3 convolutional layer is used to capture local features, and 1×1 convolutional layer is used again to adjust the number of channels; in the bottleneck residual network, residual connection (also called identity mapping) is introduced to add the input directly to the output, which helps to solve the gradient vanishing problem in the deep network and promotes the refinement of deep features; through the above structure, the deep spectral features of landslide hazards can be obtained, which contain rich spectral information and spatial structure information; in S212, the channel reshaping layer is used to reshape and group the deep spectral features into 4 sub-features, which is for the subsequent attention calculation between channels; X square is used on the 3 sub-features respectively. A series of cross operations such as average pooling in the horizontal direction, average pooling in the Y direction, 3×3 convolutional layer and softmax layer are performed to generate the key and value of the channel relationship. These operations help capture the dependencies between channels. Another sub-feature is used as a query, and attention is calculated based on matrix addition and dot product operations with the generated key and value. This process is similar to the self-attention mechanism in Transformer, but here it focuses on the attention between channels. Through attention calculation, local enhanced landslide hazard spectral features with multiple spatial scales are obtained. These features not only contain the original spectral information, but also integrate the dependencies between channels, thereby enhancing the feature expression ability and the ability to identify landslide hazards.

[0084] Further, referring to Figure 5, S220 is specifically: using image block embedding layers with different overlaps to convert the terrain data obtained by S100 into a vector sequence, and then inputting the vector sequence into two consecutive multi-head attention network structures with similar structures. Both structures include a normalization layer, a multi-head attention block, a normalization layer and a feature feed-forward layer. The difference is that the latter's multi-head attention block is based on a hybrid shift operation, and the relationship between the remote landslide terrain features is modeled based on the generated shifted horizontal stripe window and the vertical stripe window.

[0085] It should be understood that, based on the above-mentioned feature definitions, the role of the image block embedding layer is to convert the terrain data (such as slope, aspect, curvature, etc.) obtained by S100 into a vector sequence, which will serve as the input of the subsequent multi-head attention network structure; to achieve this, this embodiment adopts an image block embedding method with different degrees of overlap. By setting different degrees of overlap, feature information of different scales in the terrain data can be captured. Overlapping image blocks help the model better capture the continuity between terrain features; two consecutive multi-head attention network structures are used to process the vector sequence. Each structure includes a normalization layer, a multi-head attention block, another normalization layer, and a feature feedforward layer; the normalization layer (such as LayerNorm) is used to stabilize the training process and prevent gradient explosion or disappearance; the multi-head attention block is the core part of the Transformer architecture, which allows the model to focus on different parts of the input sequence at the same time. By splitting the attention to multiple heads, the model can process different feature subsets in parallel; the feature feedforward layer is a simple two-layer fully connected network used to further process the output of the multi-head attention block; the main difference between the two consecutive multi-head attention network structures is that the multi-head attention block in the second structure adopts a hybrid shift operation; the hybrid shift operation is a special window partitioning method that combines horizontal stripe windows and vertical stripe windows. This method allows the model to capture local features while also establishing the relationship between remote landslide terrain features; through the shift operation, the model can gradually expand its field of view to capture a wider range of terrain features; using the hybrid shift operation, the second multi-head attention block is able to better model the remote relationship between landslide terrain features; this is crucial for understanding the potential mechanism of landslide occurrence and improving the accuracy of landslide early warning systems.

[0086] Furthermore, in S230, the landslide hazard segmentation prediction results with multiple spatial scales are obtained, specifically:

[0087] S231: Establish a lightweight relational weighted fusion structure to fuse the local enhanced landslide hazard spectral features and the global enhanced terrain features according to the relational weighted fusion structure to obtain multi-scale fusion features;

[0088] S232: Establishing an expanded spatial pyramid pooling structure to perform landslide hazard segmentation prediction on the multi-scale fusion features according to the expanded spatial pyramid pooling structure, and obtaining a landslide hazard segmentation prediction result with multiple spatial scales.

[0089] Further, see Figure 6As shown, S231 is specifically as follows: first, according to the channel splicing output of the spectrum and terrain features of different scales, a 3×3 convolution is used to generate an intermediate feature representation, and the correlation coefficient matrix is ​​calculated using continuous average pooling and 1×1 convolution layers. Then, the correlation coefficient matrix is ​​respectively subjected to dot product operations with the input spectrum features and terrain features, and then addition operations are performed to generate multi-scale fusion features; S232 is specifically as follows: a convolution layer with a kernel size of 3×3 is used for dimensionality reduction, and then a parallel global average pooling layer and four dilated convolution layers with void rates of {1, 2, 4, 6} are used to perform a landslide feature extraction process. After channel splicing, a convolution layer with a kernel size of 3×3 and a Softmax layer are used in sequence to generate a landslide hazard segmentation prediction result with multiple spatial scales.

[0090] It should be understood that, based on the above feature definitions, the relational weighted fusion structure guides the feature fusion process by calculating the correlation coefficients between different features. This lightweight design helps to reduce computational complexity while maintaining model performance. First, the spectra and topographic features of different scales are channel-joined, and then a 3×3 convolution is used to generate an intermediate feature representation. This step helps to integrate information from different features and provides a basis for subsequent correlation calculations. Next, the correlation coefficient matrix is ​​calculated using continuous average pooling and 1×1 convolution layers. This matrix reflects the degree of correlation between different features and provides a basis for subsequent feature fusion. Provide guidance; finally, perform dot product operations on the correlation coefficient matrix and the input spectral features and terrain features, and then perform addition operations to generate multi-scale fusion features. This fusion method fully considers the correlation between features and helps to improve the segmentation performance of the model; in S232, the multi-scale fusion features are used to predict the landslide hazard segmentation using the expanded spatial pyramid pooling structure, thereby obtaining landslide hazard segmentation prediction results with multiple spatial scales; first, a convolutional layer with a kernel size of 3×3 is used to reduce the dimension to reduce the complexity of subsequent calculations; then, a parallel global average pooling layer and a void rate of {1, 2, 4, 6} are used to extract landslide features. These operations help capture landslide features of different scales and improve the segmentation accuracy of the model. After extracting landslide features of different scales, channel splicing is performed, and 3×3 convolutional layers are used to integrate features to generate a more robust feature representation. Finally, the Softmax layer is used to perform segmentation prediction on the integrated features to generate landslide hazard segmentation prediction results with multiple spatial scales. These results reflect the distribution of landslide hazards at different spatial scales, providing an important basis for subsequent landslide warning and prevention work.

[0091] S300: Establish a weighted loss function for multiple scale spaces, and optimize the spectral and terrain constrained semantic segmentation network model based on the weighted loss function;

[0092] It should be understood that, considering that landslide hazards may exhibit different characteristics at different scales, a weighted loss function for multiple scale spaces is established; this function can comprehensively consider the prediction errors at different scales and balance the importance of different scales in a weighted manner.

[0093] Furthermore, in S300, a weighted loss function for multiple scale spaces is established, specifically:

[0094] S310: establishing a class-balanced weighted cross entropy loss function based on the landslide hazard segmentation prediction results with multiple spatial scales and the ground truth label data;

[0095] S320: Designing an adaptive weight value and a class-balanced weighted cross entropy loss function according to the loss value, and establishing the weighted loss function for multiple scale spaces;

[0096] The expression of the class-balanced weighted cross entropy loss function in S310 is:

[0097] ;

[0098] The weighted loss function expression for multiple scale spaces in S320 is:

[0099] ;

[0100] in, represents the class-balanced weighted cross entropy loss function, Indicates spatial scale, Indicates pixels, h represents the height of the pixel, w represents the width of the pixel, Indicates category labels, Indicates The pixel label is The ground truth value of Indicates The spatial scale The pixel label is The landslide hazard prediction score is represents the number of samples of the th category label, represents the hyperparameter, c represents the total number of categories, represents the weighted loss function for multiple scale spaces, and S represents the total number of spatial scales.

[0101] It should be understood that, based on the above-mentioned feature limitations, the class-balanced weighted cross entropy loss function is a method to measure the difference between the model prediction results and the ground truth label data. The role of class-balanced weighting is to solve the under-fitting problem of the model in this category caused by the small proportion of landslide samples; since the segmentation of landslide hazards involves multiple spatial scales, and the prediction results of different scales may have different contributions to the final performance, the above-mentioned weighted loss function is designed to adaptively balance the losses of different scales, thereby optimizing the spectral and terrain constrained semantic segmentation network model and making the model more accurate.

[0102] S400: Iteratively train the spectral and terrain constrained semantic segmentation network model based on the landslide hazard sample list and weighted loss function;

[0103] It is important to understand that the spectral and terrain-constrained semantic segmentation network model is optimized using the landslide hazard sample list and the weighted loss function. Through iterative training and continuous adjustment of model parameters, the performance of the model in the landslide hazard segmentation task is gradually improved.

[0104] Furthermore, in S400, the spectral and terrain constrained semantic segmentation network model is iteratively trained, specifically:

[0105] S410: Performing data expansion on the training samples by using a geometric transformation enhancement method, where the geometric transformation enhancement method includes rotation, flipping, scaling and translation;

[0106] S420: setting training parameters for iterative training of the spectral and terrain constrained semantic segmentation network model according to the training sample volume after data expansion;

[0107] S430: Using the stochastic gradient descent algorithm and the adaptive momentum estimation algorithm, the spectral and terrain constrained semantic segmentation network model is optimized and iteratively trained;

[0108] S440: Save the spectral and terrain constrained semantic segmentation network model after iterative training.

[0109] It should be understood that, based on the above-mentioned feature limitations, the training samples are augmented with geometric transformation enhancement. These geometric transformations include rotation, flipping (horizontally or vertically), scaling (zooming in or out), and translation (moving within the image plane); these transformations can be used alone or in combination to generate more training samples; thereby increasing the diversity of training samples, reducing the risk of overfitting, and improving the generalization ability of the model; according to the volume of expanded training samples, the training parameters (including batch size, learning rate, number of epochs, and weight decay) are reasonably configured to ensure the effectiveness and efficiency of model training; according to the number and complexity of training samples, as well as the limitations of computing resources, appropriate training parameters are selected; for example, larger training sample sets allow for larger batch sizes, while complex tasks require more iterations and smaller learning rates; through optimization algorithms (including Stochastic Gradient Descent, SGD) and adaptive momentum estimation algorithms (such as Adam) continuously adjust model parameters to minimize the loss function and improve model performance; SGD updates model parameters by randomly selecting a portion of training samples to accelerate the training process and reduce the amount of calculation; the Adam algorithm combines the advantages of momentum and adaptive learning rate adjustment (RMSprop), and can usually converge to a better solution faster; in each iteration, SGD or Adam algorithm is used to update model parameters, and the loss function value is calculated to evaluate model performance. According to the changes in the loss function, the learning rate and other training parameters are adjusted in time; finally, the trained model is saved for subsequent use in landslide hazard segmentation tasks.

[0110] S500: According to the landslide hazard sample list, the terrain-constrained semantic segmentation network model after iterative training is used to predict the landslide hazard segmentation and verify its accuracy;

[0111] It is important to understand that the trained spectral and terrain constrained semantic segmentation network model is used to perform landslide hazard segmentation prediction on new multispectral remote sensing image data and terrain data; the prediction results are compared with the real labels in the landslide hazard sample list, and the prediction performance of the model is evaluated by calculating indicators such as accuracy, recall rate, and F1 score; if the prediction performance meets the requirements, the model can be applied to actual landslide hazard detection tasks.

[0112] Furthermore, the prediction and accuracy verification in S500 are specifically as follows:

[0113] S510: using the test sample data obtained in step S140 as input data of the trained spectral and terrain constraint semantic segmentation network model saved in step S440 to perform inference prediction to generate a landslide hazard prediction result;

[0114] S520: Construct a confusion matrix based on the landslide hazard prediction result generated in S510 and the ground truth label obtained in step S130, and establish producer accuracy and user accuracy for accuracy verification.

[0115] It should be understood that, based on the above-mentioned feature definition, the test sample data (including spectral information and terrain data) obtained in step S140 is used as input and input into the trained spectral and terrain constrained semantic segmentation network model saved in step S440; the model will process the input data and output the prediction result of the landslide hazard; the test sample data is preprocessed (such as normalization, cropping, etc.) according to the requirements of the model input, and then input into the model; the model will output the probability or label of each pixel belonging to the landslide hazard category; according to the set threshold (such as 0.5), the probability is converted into a binary label (landslide hazard / non-landslide hazard), thereby generating a landslide hazard prediction result map; according to S The landslide hazard prediction result generated in step S10 and the ground truth label (i.e., the distribution of actual landslide hazards) obtained in step S130 are used to construct a confusion matrix; a confusion matrix is ​​a table used to show the relationship between actual categories and predicted categories; through the confusion matrix, the producer precision (Precision) and user precision (Recall, also known as sensitivity or true positive rate) can be calculated; the producer precision represents the proportion of pixels predicted as landslide hazards that are actually landslide hazards, which reflects the accuracy of the model in predicting landslide hazards; the user precision represents the proportion of pixels actually being landslide hazards that are correctly predicted as landslide hazards, which reflects the ability of the model to identify landslide hazards.

[0116] According to a landslide hazard remote sensing detection method that integrates spectrum and terrain information of this embodiment, first, multispectral remote sensing image data is obtained through remote sensing technology. These data contain reflection information of surface objects in different spectral bands, which can reflect different characteristics of ground objects; at the same time, terrain data such as elevation, slope, slope direction, etc. are obtained, which are crucial for analyzing landslide hazards; multispectral remote sensing image data and terrain data are combined, and a list of landslide hazard samples is established through professional analysis and processing. This list contains known landslide hazard areas and their corresponding multispectral and terrain features; based on the list of landslide hazard samples, a spectral and terrain constrained semantic segmentation network model is constructed. The model consists of a spectral feature encoding module, a terrain feature encoding module and a heterogeneous feature fusion decoding module; the spectral feature encoding module is responsible for extracting spectral features from multispectral remote sensing images; the terrain feature encoding module is responsible for extracting terrain features from terrain data; the heterogeneous feature fusion decoding module is responsible for fusing these two features and decoding them to generate the segmentation result of landslide hazards; in order to improve the segmentation accuracy of the model, a weighted loss function for multiple scale spaces is established. This function takes into account the segmentation errors at different scales and performs weighted processing on them. By minimizing this weighted loss function, the spectral and terrain constrained semantic segmentation network model is optimized to better adapt to the landslide hazard segmentation tasks at different scales. The model is iteratively trained using the landslide hazard sample list and the weighted loss function. During the training process, the model will continuously learn and adjust parameters to improve the segmentation capability of landslide hazards. After the model training is completed, the landslide hazard sample list is used to predict the landslide hazard segmentation of the model. The segmentation accuracy and reliability of the model are verified by comparing it with the actual landslide hazard area.

[0117] Technical effect: By integrating multi-spectral remote sensing image data and terrain data, the spectral characteristics of surface materials and terrain morphology information are fully utilized to improve the accuracy of landslide hazard detection; the weighted loss function to multiple scale spaces can comprehensively consider the prediction errors at different scales, so that the model can show good performance at different scales and enhance the generalization ability of the model; the spectral and terrain constrained semantic segmentation network model constructed using deep learning technology breaks through the limitations of traditional single data source and network architecture, and can realize automated and intelligent landslide hazard detection, improving detection efficiency and accuracy; through landslide hazard detection, it can provide a scientific basis for the early warning and prevention of landslide disasters, which will help reduce the impact of landslide disasters on human society and the natural environment.

[0118] It should be noted that the execution subject of the landslide hazard remote sensing detection method that integrates spectrum and terrain information provided in the embodiment of the present application may be a landslide hazard remote sensing detection system that integrates spectrum and terrain information, or, or a control module in the landslide hazard remote sensing detection system that integrates spectrum and terrain information for executing and loading a landslide hazard remote sensing detection method that integrates spectrum and terrain information. In the embodiment of the present application, a landslide hazard remote sensing detection system that integrates spectrum and terrain information executes a landslide hazard remote sensing detection method that integrates spectrum and terrain information as an example to illustrate a landslide hazard remote sensing detection method that integrates spectrum and terrain information provided in the embodiment of the present application.

[0119] A landslide hazard remote sensing detection system integrating spectral and terrain information, comprising:

[0120] The first acquisition module is used to acquire multispectral remote sensing image data and terrain data;

[0121] A first processing module is used to process the multispectral remote sensing image data and terrain data acquired by the first acquisition module to obtain a corresponding landslide hazard sample list;

[0122] The first establishment module is used to establish a spectral and terrain constraint semantic segmentation network model according to the slope hazard sample list obtained by the first processing module;

[0123] The second building module is used to establish a weighted loss function for multiple scale spaces, and optimize the spectral and terrain constrained semantic segmentation network model according to the weighted loss function for multiple scale spaces;

[0124] The first training module is used to iteratively train the optimized spectral and terrain constrained semantic segmentation network model;

[0125] The prediction and verification module performs landslide hazard segmentation prediction and accuracy verification on the terrain-constrained semantic segmentation network model after iterative training according to the landslide hazard sample list.

[0126] A landslide hazard remote sensing detection system that integrates spectrum and terrain information in the embodiment of the present application can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, a wearable device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), etc. The non-mobile electronic device can be a server, a network attached storage (NAS), a personal computer (PC), etc., which is not specifically limited in the embodiment of the present application.

[0127] In the embodiment of the present application, a landslide hazard remote sensing detection system integrating spectrum and terrain information can be a device with an operating system. The operating system can be an Android operating system, an iOS operating system, or other possible operating systems, which are not specifically limited in the embodiment of the present application.

[0128] The landslide hazard remote sensing detection system provided by the embodiment of the present application can realize Figures 1 to 6 In the method embodiment, each process of implementing a landslide hazard remote sensing detection method that integrates spectrum and terrain information is not described here to avoid repetition.

[0129] According to a landslide hazard remote sensing detection system that integrates spectral and terrain information in this embodiment, multispectral remote sensing image data and terrain data are acquired through a first acquisition module, and the system can comprehensively utilize the advantages of these two data sources; multispectral remote sensing images provide spectral information of surface objects, which helps to distinguish different types of surface coverage; and terrain data reveals the surface morphology and height changes, which are crucial for identifying landslide hazards; the combination of the two can significantly improve the accuracy and comprehensiveness of landslide hazard detection; the first processing module processes the acquired data, can automatically identify landslide hazard areas, and generate a corresponding landslide hazard sample list. This process has a high degree of automation, which reduces manual intervention and improves work efficiency. The system uses the first establishment module to establish a spectral and terrain-constrained semantic segmentation network model based on the landslide hazard sample list. This model can be customized for specific landslide hazard characteristics. The weighted loss function for multiple scale spaces established by the second establishment module helps to more comprehensively consider the landslide hazard characteristics of different scales during the model training process, thereby improving the accuracy and generalization ability of the model. The first training module iteratively trains the optimized spectral and terrain-constrained semantic segmentation network model. By continuously learning and adjusting the model parameters, the model's landslide hazard detection capability can be further improved. The prediction and verification module uses the landslide hazard sample list to predict and verify the landslide hazard segmentation of the iteratively trained model, which can ensure that the model's prediction results have high accuracy and reliability. This helps to more effectively identify landslide hazard areas in practical applications and provide strong support for disaster prevention and mitigation.

[0130] Optionally, an embodiment of the present application also provides an electronic device, including a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, each process of the above-mentioned embodiment of a remote sensing detection method for landslide hazards that integrates spectral and terrain information is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0131] An embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, each process of the above-mentioned embodiment of a remote sensing detection method for landslide hazards that integrates spectral and terrain information is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0132] The processor is the processor in the electronic device in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0133] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0134] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods of each embodiment of the present application.

[0135] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A landslide hazard remote sensing detection method integrating spectrum and terrain information, characterized in that: The method comprises: Acquiring multispectral remote sensing image data and terrain data to establish a landslide hazard sample list based on the multispectral remote sensing image data and terrain data; According to the landslide hazard sample list, a spectral and terrain constrained semantic segmentation network model including a spectral feature encoding module, a terrain feature encoding module and a heterogeneous feature fusion decoding module is established; Establishing a weighted loss function for multiple scale spaces, and optimizing the spectral and terrain constrained semantic segmentation network model according to the weighted loss function; Iteratively training the spectral and terrain constrained semantic segmentation network model according to the landslide hazard sample list and the weighted loss function; According to the landslide hazard sample list, the terrain-constrained semantic segmentation network model after iterative training is used to predict the landslide hazard segmentation and verify its accuracy; The process of establishing the spectral and terrain constrained semantic segmentation network model is specifically as follows: The spectral feature encoding module composed of efficient channel relationship enhancement blocks is established according to a deep convolutional neural network to obtain the local enhanced landslide hazard spectral features with multiple spatial scales; the spectral feature encoding module includes a residual convolution structure and an improved channel self-attention structure; Establishing the terrain feature encoding module consisting of improved Swin-transformer blocks according to the Transformer architecture to obtain global enhanced terrain features with multiple spatial scales; The heterogeneous feature fusion decoding module is established according to the local enhanced landslide hazard spectral features and the global enhanced terrain features to obtain landslide hazard segmentation prediction results with multiple spatial scales; the heterogeneous feature fusion decoding module includes a relation weighted fusion structure and an expanded spatial pyramid pooling structure; The method of obtaining the spectral characteristics of local enhanced landslide hazard with multiple spatial scales is specifically as follows: Establishing the residual convolution structure according to the bottleneck residual network, so as to obtain the deep spectral characteristics of the landslide hazard according to the residual convolution structure; Establishing the improved channel self-attention structure according to the attention mechanism, so as to enhance the deep-level spectral characteristics of the landslide hazard according to the improved channel self-attention structure, and obtaining the locally enhanced spectral characteristics of the landslide hazard with multiple spatial scales; The heterogeneous feature fusion decoding module is established to obtain landslide hazard segmentation prediction results with multiple spatial scales, specifically: Establishing a lightweight relational weighted fusion structure to fuse the local enhanced landslide hazard spectral features and the global enhanced terrain features according to the relational weighted fusion structure to obtain multi-scale fusion features; Establishing the expanded spatial pyramid pooling structure, so as to perform landslide hazard segmentation prediction on the multi-scale fusion features according to the expanded spatial pyramid pooling structure, and obtaining the landslide hazard segmentation prediction result with multiple spatial scales; The weighted loss function for multiple scale spaces is established as follows: A class-balanced weighted cross entropy loss function is established based on the landslide hazard segmentation prediction results with multiple spatial scales and ground truth label data; Designing an adaptive weight value and a class-balanced weighted cross entropy loss function according to the loss value, and establishing the weighted loss function for multiple scale spaces; The class-balanced weighted cross entropy loss function expression is: ； The weighted loss function expression for multiple scale spaces is: ； in, represents the class-balanced weighted cross entropy loss function, Indicates spatial scale, Indicates pixels, h represents the height of the pixel, w represents the width of the pixel, Indicates category labels, Indicates The pixel label is The ground truth value of Indicates The spatial scale The pixel label is The landslide hazard prediction score is represents the number of samples of the th category label, represents the hyperparameter, c represents the total number of categories, represents the weighted loss function for multiple scale spaces, and S represents the total number of spatial scales.

2. The method for remote sensing detection of landslide hazards by integrating spectrum and terrain information according to claim 1, characterized in that: The landslide hazard sample list is established based on the multispectral remote sensing image data and terrain data, specifically: Performing a first preprocessing on the acquired multispectral remote sensing image data, and then processing the multispectral remote sensing image data after the first preprocessing using a conventional conversion algorithm to obtain multispectral remote sensing image feature data including a normalized vegetation index, a normalized water index, a brightness index, and a bare soil index; the first preprocessing includes geometric correction, atmospheric correction, and image mosaicking; Performing a second preprocessing on the acquired terrain data, and then applying the conventional conversion algorithm to the terrain data after the second preprocessing to obtain terrain feature data including slope, aspect, curvature, shadow and terrain moisture index; the second preprocessing includes geo-registration and resampling; Performing visual interpretation processing on the data geographic facts generated by the multispectral remote sensing image feature data and the terrain feature data to obtain ground truth label data of landslide hazards; The multispectral remote sensing image feature data, terrain feature data and ground truth label data are batch-cropped using a non-overlapping sliding window, and are divided into training samples and test samples according to a preset ratio.

3. The method for remote sensing detection of landslide hazards by integrating spectrum and terrain information according to claim 2, characterized in that: The iterative training of the spectral and terrain constrained semantic segmentation network model is specifically as follows: Performing data expansion on the training samples by using a geometric transformation enhancement method, wherein the geometric transformation enhancement method includes rotation, flipping, scaling and translation; According to the training sample volume after the data expansion, setting training parameters for iterative training of the spectral and terrain constrained semantic segmentation network model; The spectral and terrain constrained semantic segmentation network model is optimized and iteratively trained by using a stochastic gradient descent algorithm and an adaptive momentum estimation algorithm; The spectral and terrain constrained semantic segmentation network model after the iterative training is saved.

4. A landslide hazard remote sensing detection system integrating spectrum and terrain information, capable of implementing a landslide hazard remote sensing detection method integrating spectrum and terrain information as described in any one of claims 1 to 3, characterized in that: The system comprises: The first acquisition module is used to acquire multispectral remote sensing image data and terrain data; A first processing module, used for processing the multispectral remote sensing image data and terrain data acquired by the first acquisition module to obtain a corresponding landslide hazard sample list; A first establishing module is used to establish a spectral and terrain constrained semantic segmentation network model according to the slope hidden danger sample list obtained by the first processing module; A second establishment module is used to establish a weighted loss function for multiple scale spaces, and optimize the spectral and terrain constrained semantic segmentation network model according to the weighted loss function for multiple scale spaces; A first training module is used to iteratively train the optimized spectral and terrain constrained semantic segmentation network model; The prediction verification module performs landslide hazard segmentation prediction and accuracy verification on the terrain-constrained semantic segmentation network model after iterative training according to the landslide hazard sample list.

5. An electronic device, characterized in that: The method comprises a processor, a memory and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, the steps of a remote sensing detection method for landslide hazards integrating spectral and terrain information as described in any one of claims 1 to 3 are implemented.

6. A readable storage medium, characterized in that: The readable storage medium stores a program or instruction, and when the program or instruction is executed by the processor, the steps of the landslide hazard remote sensing detection method that integrates spectrum and terrain information as described in any one of claims 1 to 3 are implemented.

Images