CN-121982519-A - Landslide intelligent recognition method based on multi-source remote sensing image fusion

CN121982519ACN 121982519 ACN121982519 ACN 121982519ACN-121982519-A

Abstract

The invention relates to the technical field of image semantic segmentation, in particular to a landslide intelligent recognition method based on multi-source remote sensing image fusion. The method comprises the steps of calculating a gradient matrix and a water flow convergence index based on a digital elevation model, determining a physical priori interval of cohesive force and an internal friction angle based on geological lithology data retrieval, constructing an input tensor, extracting a ground surface semantic feature map of an optical image by utilizing a visual perception branch through a double-branch coupling network, mapping the branch into a limited geotechnical parameter map by utilizing a physical parameter based on the physical priori interval, calling a micro infinite slope stability layer, performing forward guidance calculation based on a limit balance equation, generating a slope stability coefficient field, constructing physical attention gating, executing Hadamard product operation, and generating a semantic feature tensor. The invention ensures the true effectiveness of the obtained position data through the deep coupling integration and collaborative optimization of the visual perception branch and the physical parameter mapping branch.

Inventors

CHEN XIAOBO
Jiang Xiaori
FANG SHAOYAN
WANG YONGRUI
YANG LINLIN
YANG LEI
JI WENPENG

Assignees

山东省煤田地质局第四勘探队

Dates

Publication Date: 20260505
Application Date: 20251230

Claims (8)

1. A landslide intelligent identification method based on multi-source remote sensing image fusion is characterized by comprising the following steps: Calculating a gradient matrix and a water flow convergence index based on the digital elevation model, searching and determining a physical priori interval of cohesive force and an internal friction angle based on the geological lithology data, and constructing an input tensor through spatial alignment; Based on a physical prior interval, mapping the branches by using physical parameters, and nonlinear mapping a gradient matrix and a water flow convergence index into a limited geotechnical parameter map containing cohesive force and internal friction angle; the method comprises the steps of calling a micro infinite slope stability layer, receiving a limited geotechnical parameter diagram and a gradient matrix, executing forward direction guidance calculation based on a limit balance equation to generate a slope stability coefficient field, constructing physical attention gating based on the slope stability coefficient field, executing Hadamard product operation with a ground surface semantic feature diagram, inhibiting the feature response of a physical stability region, retaining the semantic features of a physical instability region and generating a semantic feature tensor; Based on the semantic feature tensor, outputting a landslide binarization recognition result through up-sampling and classification judgment.
2. The landslide intelligent identification method based on multi-source remote sensing image fusion is characterized in that the specific generation process of input tensor comprises the steps of obtaining an optical image of a monitoring area, a digital elevation model and geological lithology data, performing band separation on the optical image, extracting four spectral band matrixes of red light, green light, blue light and near infrared, performing hydrologic analysis calculation on the digital elevation model, generating a flow direction matrix through a single-flow algorithm, calculating a flow accumulation amount based on the flow direction matrix, performing logarithmic treatment on the flow accumulation amount to generate a water flow convergence index matrix, simultaneously, calculating a surface normal vector of the digital elevation model, generating a gradient feature matrix based on an included angle between the surface normal vector and a vertical coordinate axis, performing raster coding on the geological lithology data to generate a lithology index map, mapping each pixel point in the lithology index map into a four-channel physical boundary constraint matrix based on a preset lithology mechanical parameter database, wherein the four-channel is respectively a cohesive force lower boundary channel, a cohesive force upper boundary channel, an internal angle lower boundary channel and an internal angle upper boundary channel, performing linear interpolation on the four-channel, and performing cascade interpolation on the water flow convergence index matrix by using the spatial gradient feature matrix and the linear interpolation constraint matrix.
3. The landslide intelligent recognition method based on multi-source remote sensing image fusion is characterized in that the specific processing process of a visual sensing branch comprises the steps of arranging a multi-scale residual error encoder on the visual sensing branch, transmitting an input tensor into the multi-scale residual error encoder, performing feature extraction and downsampling operation through a cascade residual error convolution unit to generate a multi-scale feature sequence with gradually reduced spatial resolution and gradually increased semantic abstraction degree, wherein the multi-scale feature sequence comprises shallow feature tensors representing surface texture details and deep feature tensors representing environmental context information, performing transpose convolution upsampling on the deep feature tensors by using a feature aggregation decoder, performing channel splicing on the upsampled deep feature tensors and the shallow feature tensors with the same resolution in each upsampling process to generate a space-time multi-mode fusion tensor, performing channel dimension reduction and feature recombination by using convolution to check the space-time multi-mode fusion tensor, and outputting a surface semantic feature map.
4. The landslide intelligent identification method based on multi-source remote sensing image fusion is characterized in that the specific processing process of physical parameter mapping branches comprises the steps of extracting a water flow convergence index matrix, a gradient feature matrix and a physical boundary constraint matrix from an input tensor, wherein the physical boundary constraint matrix comprises a clay lower bound channel, a clay upper bound channel, an internal friction angle lower bound channel and an internal friction angle upper bound channel, inputting the water flow convergence index matrix and the gradient feature matrix into a lightweight parameter regression network, extracting environmental factor features through a convolution layer, restraining output values through an S-shaped activation function to generate a normalized parameter coefficient map, the normalized parameter coefficient map comprises a clay coefficient channel and an internal friction angle coefficient channel, constructing a micro-area mapping operator, receiving the normalized parameter coefficient map and the physical boundary constraint matrix, performing linear interpolation on each pixel position between the clay lower bound channel and the clay upper bound channel by using the clay coefficient channel as weights, generating an equivalent clay distribution matrix by using the clay coefficient channel as weights, performing linear interpolation between the clay lower bound channel and the clay upper bound channel, and performing linear interpolation coefficient distribution matrix interpolation coefficient, and linear interpolation coefficient distribution matrix interpolation coefficient distribution matrix between the clay inner friction coefficient and the lower bound channel as dimensions and output coefficient inner friction coefficient distribution channel limit interface friction coefficient, and linear friction coefficient distribution matrix.
5. The landslide intelligent identification method based on multi-source remote sensing image fusion is characterized in that the specific generation process of a slope stability coefficient field comprises the steps of splitting a limited geotechnical parameter map into an equivalent cohesive force distribution matrix and an equivalent internal friction angle distribution matrix, obtaining a slope characteristic matrix and a water flow convergence index matrix, setting soil physical constants, wherein the soil physical constants comprise soil natural volume weight and preset thickness of a sliding body, performing sine and cosine triangular transformation on the slope characteristic matrix to generate a slope sine matrix and a slope cosine matrix, constructing a sliding driving force calculation operator, multiplying the soil natural volume weight, the preset thickness of the sliding body, the slope sine matrix and the slope cosine matrix by Hadamard product operation to generate a sliding shearing stress tensor, constructing an anti-sliding resistance calculation operator, calculating a pore water pressure ratio matrix based on the water flow convergence index matrix, calculating total normal stress by using the slope matrix, subtracting the pore water pressure corresponding to the pore water pressure ratio to generate an effective directional stress matrix, multiplying the effective directional stress matrix by the equivalent internal friction angle distribution matrix, multiplying the effective directional stress matrix by the shear stress tangent value of the equivalent internal friction angle distribution matrix, and adding the shear stress tensor to the shear stress tensor, and generating a shearing resistance tensor, and adding the anti-sliding stability coefficient to a small shear stress value.
6. A landslide intelligent recognition method based on multi-source remote sensing image fusion is characterized in that the specific generation process of semantic feature tensor comprises the steps of calling a preset critical balance constant, subtracting the critical balance constant from each pixel value in a slope stability coefficient field, taking the opposite number, generating a stability difference response matrix, wherein a positive value region corresponds to a physical instability region in the stability difference response matrix, a non-positive value region corresponds to a physical stability region, performing nonlinear mapping on the stability difference response matrix by using an activation function, normalizing physical values to generate physical attention gating, obtaining the earth surface semantic feature map, performing broadcast Hadamard product operation with the physical attention gating, and using the physical attention gating as a space weight coefficient matrix, performing numerical value holding operation in the physical instability region, performing numerical value attenuation operation in the physical stability region, and outputting semantic feature tensor.
7. The landslide intelligent recognition method based on multi-source remote sensing image fusion is characterized in that the specific generation process of the landslide binarization recognition result comprises the steps of inputting the semantic feature tensor into a semantic segmentation decoder, performing up-sampling operation on the semantic feature tensor by utilizing a cascade convolution layer, gradually recovering the spatial resolution to be consistent with a monitoring area, mapping the up-sampled semantic feature tensor into a landslide posterior probability matrix by utilizing a Softmax classification layer, wherein the numerical value in the landslide posterior probability matrix represents the confidence degree of each pixel point belonging to a landslide category, setting a binarization judgment threshold value, performing pixel-by-pixel threshold segmentation on the landslide posterior probability matrix, marking pixels with the confidence degree higher than the binarization judgment threshold value as landslide objects, marking the rest pixels as backgrounds, and outputting the landslide binarization recognition result.
8. The landslide intelligent identification method based on multi-source remote sensing image fusion is characterized by further comprising the specific process of conducting parameter optimization through a dual-branch coupling network, wherein the specific process comprises the steps of obtaining a landslide truth value tag diagram corresponding to a monitoring area, calculating semantic segmentation supervision loss values through a binary cross entropy algorithm based on the landslide truth value tag diagram and a landslide posterior probability matrix, constructing a physical consistency constraint loss function, evaluating logic conflict between a slope stability coefficient field and the landslide truth value tag diagram, calculating forward deviation of the slope stability coefficient field exceeding a critical balance constant through a linear rectification function when the landslide truth value tag diagram indicates a landslide category, using the linear rectification function as a physical false negative penalty item, forcing the physical parameter mapping branch to converge towards a destabilization state, using the linear rectification function as a physical false positive penalty item, forcing the physical parameter mapping branch to converge towards a steady state, performing weighted summation on the semantic segmentation supervision loss values, the physical false negative penalty item and the physical false positive penalty item, generating a target function, and updating the visual perception target function based on the state computing target function and the visual perception parameters of the branch map through the reverse-direction synchronization algorithm.

Description

Landslide intelligent recognition method based on multi-source remote sensing image fusion Technical Field The invention relates to the technical field of image semantic segmentation, in particular to a landslide intelligent recognition method based on multi-source remote sensing image fusion. Background The method is used for automatically identifying and positioning the wide-area landslide target in the complex geographical environment where geological disasters are easy to occur (such as mountain canyons and strong rainfall areas), and is an important technical application direction in the current disaster prevention and reduction field. In the prior art, a recognition method of combining a single optical remote sensing image with a deep learning semantic segmentation model is generally adopted, namely, by monitoring the spectral texture and morphological feature changes of a ground surface image, when the model detects pixel features similar to landslide samples, the model is automatically segmented in a region to be detected, and the positions and the ranges of all landslide targets are determined, so that the method is a main technical means for realizing large-scale geological disaster investigation at present. However, the prior art scheme has significant defects in recognition reliability and environmental adaptability, and the data source for recognizing landslide has the problems of great environmental restriction and incomplete information acquisition. The existing multi-source data fusion method is based on passive imaging, is extremely easy to be blocked by cloud and fog, uneven illumination and interference of mountain shadows, so that space-time faults appear on monitoring data, continuous tracking is difficult to achieve in a post-disaster golden rescue period, the existing multi-source data fusion method is based on simple input end splicing or shallow layer feature superposition, a fusion mechanism cannot be adaptively adjusted along with local quality changes of different mode data, feature differences among heterogeneous data are difficult to be accurately matched, deviation is caused in a positioning range under a complex scene, the problems of homogeneity and boundary blurring exist in recognition results, namely, indiscriminate segmentation results are provided for all areas, geological adaptation cannot be conducted according to special gradients or elevations of landslides, a large number of false reports occur, parameter data cannot be accurately obtained, and information lacks effectiveness. Therefore, a landslide intelligent recognition method based on multi-source remote sensing image fusion is provided. Disclosure of Invention The invention aims to provide a landslide intelligent recognition method based on multi-source remote sensing image fusion, which aims to solve the problems in the background technology. In order to achieve the purpose, the invention provides the technical scheme that the landslide intelligent identification method based on multi-source remote sensing image fusion comprises the following steps: Calculating a gradient matrix and a water flow convergence index based on the digital elevation model, searching and determining a physical priori interval of cohesive force and an internal friction angle based on the geological lithology data, and constructing an input tensor through spatial alignment; Based on a physical prior interval, mapping the branches by using physical parameters, and nonlinear mapping a gradient matrix and a water flow convergence index into a limited geotechnical parameter map containing cohesive force and internal friction angle; the method comprises the steps of calling a micro infinite slope stability layer, receiving a limited geotechnical parameter diagram and a gradient matrix, executing forward direction guidance calculation based on a limit balance equation to generate a slope stability coefficient field, constructing physical attention gating based on the slope stability coefficient field, executing Hadamard product operation with a ground surface semantic feature diagram, inhibiting the feature response of a physical stability region, retaining the semantic features of a physical instability region and generating a semantic feature tensor; Based on the semantic feature tensor, outputting a landslide binarization recognition result through up-sampling and classification judgment. The method comprises the steps of obtaining an optical image of a monitoring area, a digital elevation model and geological lithology data, performing band separation on the optical image, extracting four spectral band matrixes of red light, green light, blue light and near infrared light, performing hydrologic analysis calculation on the digital elevation model, generating a flow matrix through a single flow direction algorithm, calculating a flow accumulation amount based on the flow matrix, performing logarithmic processing on the flow accumulation a