Intelligent landslide identification method based on multi-source remote sensing image deep learning

Abstract

The invention relates to the technical field of remote sensing image processing, in particular to an intelligent landslide identification method based on deep learning of a multi-source remote sensing image, which comprises the following steps of: firstly, acquiring high-resolution optical and radar data of the multi-source remote sensing image by adopting a spatio-temporal data expansion method; the problem that an existing landslide identification method is insufficient in data type and data quantity is effectively solved, more comprehensive and accurate landslide data are provided, the generalization ability of the model is enhanced, then a deep learning model with semi-supervised learning and adaptive convolution multi-scale feature fusion is adopted, the size of a convolution kernel can be dynamically adjusted, and the robustness of the landslide identification method is improved. Landslide features of different scales are accurately captured, understanding of landslide morphology and structure is enhanced, robustness of the model in a complex terrain is improved, finally, a landslide boundary and morphology are corrected through a positioning correction method, inaccurate or discontinuous boundaries caused by prediction errors are eliminated, and stability of the model in a complex environment is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and particularly relates to a deep learning landslide intelligent recognition method based on multi-source remote sensing images. Background Art

[0002] As one of the most common geological disasters in China, landslides pose a serious threat to the lives and property safety of the people and the construction of infrastructure. With the global climate change, there are countless landslide disasters caused by earthquakes and rainfall, and the losses caused by the disasters are extremely heavy. Therefore, accurately and timely carrying out rapid intelligent interpretation of landslide disasters has important scientific significance for scientific disaster prevention and mitigation.

[0003] With the rapid development of satellite earth observation technology and deep learning technology, optical and radar remote sensing satellites with high spatio-temporal resolution provide rich data sources for landslide disaster identification, making it possible to identify landslides from daily to hourly levels. Traditional technical means use 3D visualization software to achieve landslide identification with the help of expert visual interpretation, which requires a large amount of manpower and financial costs in a wide area, and the accuracy and reliability of the identification depend on the professional experience of experts. However, remote sensing image data often has the characteristics of high dimension, complexity, and large noise, and due to the diversity of landslides and the differences in geographical environments, traditional image processing methods face huge challenges.

[0004] In recent years, deep learning, as a powerful artificial intelligence technology, has developed rapidly in the field of landslide identification. At present, landslide identification methods represented by convolutional neural networks emerge in an endless stream. This method can automatically identify the landslide area by training a large number of landslide samples, greatly improving the efficiency and accuracy of landslide identification. However, the existing landslide identification methods first have problems in terms of data sources. They use single remote sensing data, and since single remote sensing data sources often cannot comprehensively and accurately identify all landslides, the identification effect and reliability are reduced. Secondly, in terms of models, the existing deep learning methods only improve and innovate the structure of the deep learning model, ignoring the morphological feature differences of the landslide body itself. The morphological feature differences of the landslide body, such as the type, scale, shape, and development stage of the landslide, directly affect its performance in the terrain and its features in the image. If the model fails to effectively capture and distinguish these subtle differences, it will lead to the inability to uniformly identify different types of landslides, thus unable to specifically improve the identification accuracy of different landslide morphologies, and further reducing the overall identification effect and reliability. Therefore, it is urgent to develop an adaptive convolutional multi-scale feature fusion deep learning intelligent recognition model that takes into account the morphological differences of the landslide body to provide more scientific and effective technical support for accurate identification and scientific prevention and control of landslide disasters. Summary of the Invention

[0005] In view of this, the object of the present invention is to provide a deep learning-based intelligent landslide recognition method for multi-source remote sensing images, so as to achieve the technical effect of making the intelligent landslide recognition more accurate and comprehensive and being able to handle various complex situations.

[0006] The first aspect of the embodiment of the present invention discloses a deep learning-based intelligent landslide recognition method for multi-source remote sensing images, including:

[0007] S1. Construct a multi-sensor spatial scale feature difference sample based on optical remote sensing images, radar remote sensing images and terrain data within the landslide area, and construct an intelligent landslide recognition data set by using a spatio-temporal data expansion method based on the multi-sensor spatial scale feature difference sample;

[0008] S2. Construct an adaptive convolutional multi-scale feature fusion landslide intelligent recognition model considering the morphological difference features of the landslide body based on a semi-supervised learning framework, and design a gradient weighted objective function to optimize the landslide intelligent recognition model;

[0009] S3. Use a positioning correction method to correct the landslide intelligent recognition result output by the landslide intelligent recognition model, and generate a landslide prediction result.

[0010] Preferably, the step S1 includes:

[0011] S11. Establish an intelligent interpretation mark considering the morphological difference features of the landslide body according to the optical remote sensing image, radar remote sensing image and terrain data, and preliminarily construct a multi-sensor spatial scale feature difference sample based on the intelligent interpretation mark;

[0012] S12. Simulate and generate landslide samples by using a spatio-temporal data expansion method based on the multi-sensor spatial scale feature difference sample. The calculation formula of the spatio-temporal data expansion method is as follows:

[0013] D′(x′,t′)=[α·g(x)+β·h(x)+γ·interp x (interp y (D(x,t),t′),x′)]+Ν(0,δ 2 )

[0014] Where D′(x′,t′) represents the expanded data set, x′ and t′ respectively represent the new spatial position and time point, α, β, and γ represent weight coefficients, g(x) represents the spatial non-linear transformation, h(x) represents the temporal non-linear transformation, D(x,t) represents the original data set, interp x (D(x,t),x′) and interp y (D(x,t),t′) respectively represent the spatial and temporal interpolation operations, and Ν(0,δ 2 ) represents additive noise;

[0015] S13. Establish image slices using the multi-scale balanced target region method, and mix and reshuffle them with the sample data in the steps S11 and S12 to construct an intelligent landslide recognition dataset. The calculation formula of the multi-scale balanced target region method is as follows:

[0016] C i (x,t) = slice(I i (x), window_size, stride)

[0017]

[0018] Among them, C i (x,t) represents the slice data at the i-th scale, I i (x) represents the original image data at the i-th scale, window_size represents the window size, stride represents the step size, C balanced represents the slice after regional balance and fusion, S i represents the resolution of the multi-source image, w(S i ) represents the weight, which is adaptively determined according to the size of the landslide area.

[0019] Preferably, in the step S13, the window size is 16, 32, 64, 128, 256 or 512; the step size is 16, 32, 64, 128, 256 or 512.

[0020] Preferably, the step S2 includes:

[0021] S21. Extract the morphological difference features of the landslide body according to the intelligent landslide recognition dataset in the step S1, and design an adaptive convolution kernel according to the morphological difference features of the landslide body. The calculation formula of the adaptive convolution kernel is as follows:

[0022] K(x,y) = α(x,y)·f(shape,texture,boundary)

[0023] Among them, K(x,y) represents the adaptive convolution kernel, α(x,y) represents the adaptive weight obtained by network training, and f(shape,texture,boundary) represents the construction function of the morphological, texture and boundary features;

[0024] S22. Based on the semi-supervised framework and the adaptive convolution kernel in the step S21, construct a landslide intelligent recognition model using the multi-sensor high-level abstract semantic feature fusion mechanism and the deep convolutional neural network. The calculation formula of the multi-sensor high-level abstract semantic feature fusion mechanism is as follows:

[0025]

[0026] Among them, represents the fused feature map, and x m represents the m-th input data, and f m (·) represents high-level abstract feature extraction, and G m (·) represents the multi-sensor feature alignment and normalization mapping function, and α m represents the weighted coefficient of each sensor feature, H(·) represents the mapping function of cross-modal learning and semantic enhancement, and g(·) represents the prediction output function;

[0027] S23. Design the gradient weighted objective function to optimize the landslide intelligent recognition model in step S22. The calculation formula of the gradient weighted objective function is:

[0028]

[0029] Among them, L grad_weighted represents the gradient weighted loss, and L base represents the cross-entropy loss of the model, and λ represents the weight hyperparameter that controls the gradient weighting, represents the gradient of the input image Ι(x,y), represents the norm of the gradient amplitude of the feature map at the position (x,y), represents the gradient of the loss function L with respect to the input Ι(x,y).

[0030] Preferably, the morphological difference features of the landslide body include landslide morphological features, landslide texture features, and / or landslide boundary features.

[0031] Preferably, the adaptive convolution kernel includes morphological convolution, horizontal texture convolution, vertical texture convolution, horizontal boundary convolution, vertical boundary convolution, or edge detection convolution.

[0032] Preferably, step S3 is specifically:

[0033] Adopt the positioning correction method to correct the boundary error in the landslide intelligent recognition result output by the landslide intelligent recognition model in step S2 to generate the landslide prediction result. The calculation formula of the positioning correction method is:

[0034]

[0035] Among them, y represents the finally corrected landslide prediction result, y′ represents the result output by the landslide intelligent recognition model, M(·) represents spatial morphological correction, F(·) represents image filtering, represents the gradient of the error with respect to the prediction boundary, γ represents the correction step size, represents the gradient correction.

[0036] Preferably, the positioning correction includes spatial correction, morphological operation, and / or smooth gradient correction.

[0037] Preferably, the correction step size is 0.1, 0.5, or 1.

[0038] The present invention provides a deep learning landslide intelligent recognition method based on multi-source remote sensing images. First, by obtaining high-resolution optical and radar data of multi-source remote sensing images and adopting a spatio-temporal data expansion method, the problems of insufficient data types and data quantities in existing landslide recognition methods are effectively overcome, providing more comprehensive and accurate landslide data and enhancing the generalization ability of the model. Secondly, a deep learning model that combines semi-supervised learning and adaptive convolutional multi-scale feature fusion is used, which can dynamically adjust the size of the convolutional kernel, accurately capture landslide features at different scales, enhance the understanding of landslide morphology and structure, and improve the robustness of the model under complex terrains. Finally, the landslide boundary and morphology are corrected through a positioning correction method to eliminate inaccurate or discontinuous boundaries caused by prediction errors, further improving the stability of the model in complex environments. Through the organic combination of the above steps, the accuracy of model recognition is improved, the reliability of the model in different environments is enhanced, the landslide intelligent recognition is made more accurate and comprehensive, and various complex situations can be handled. Description of the Drawings

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0040] Figure 1 is a flowchart of a deep learning landslide intelligent recognition method based on multi-source remote sensing images disclosed in the first embodiment of the present invention;

[0041] Figure 2 is a schematic diagram of an adaptive convolutional kernel in a deep learning landslide intelligent recognition method based on multi-source remote sensing images disclosed in the first embodiment of the present invention;

[0042] Figure 3 is a structure diagram of a landslide intelligent recognition model in a deep learning landslide intelligent recognition method based on multi-source remote sensing images disclosed in the first embodiment of the present invention;

[0043] Figure 4 is a schematic diagram of a landslide prediction result in a deep learning landslide intelligent recognition method based on multi-source remote sensing images disclosed in the first embodiment of the present invention.

[0044] Figure 5It is a heat map comparing the FAST-SCNN model, SegFormer model, Swin model, ELN model, SC model with the model of the present invention in a landslide intelligent recognition method based on deep learning of multi-source remote sensing images disclosed in the first embodiment of the present invention. Detailed implementation manners

[0045] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0046] In the present invention, the orientation or positional relationship indicated by terms such as "upper", "lower", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings. These terms are mainly used to better describe the present invention and its embodiments, and are not used to limit that the indicated device, element or component must have a specific orientation, or be constructed and operated in a specific orientation.

[0047] Moreover, in addition to being able to represent an orientation or positional relationship, some of the above terms may also be used to represent other meanings. For example, the term "upper" may also be used to represent a certain attachment relationship or connection relationship in some cases. For those of ordinary skill in the art, the specific meanings of these terms in the present invention can be understood according to specific circumstances.

[0048] In addition, the terms "installed", "set", "provided with", "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral structure; it can be a mechanical connection or an electrical connection; it can be a direct connection, or an indirect connection through an intermediate medium, or an internal connection between two devices, elements or components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0049] In addition, terms such as "first", "second", etc. are mainly used to distinguish different devices, elements or components (the specific types and structures may be the same or different), and are not used to indicate or imply the relative importance and quantity of the indicated devices, elements or components. Unless otherwise stated, the meaning of "a plurality" is two or more.

[0050] The inventive concept of the present invention is:

[0051] In the landslide recognition methods in the prior art, first, in terms of data sources, single remote sensing data is often used. Since a single remote sensing data source often fails to comprehensively and accurately identify all landslides, the recognition effect and reliability are reduced. Second, in terms of models, existing deep learning methods only improve and innovate the structure of the deep learning model, ignoring the morphological feature differences of the landslide body itself. Due to the morphological feature differences of the landslide body, such as the type, scale, shape, and development stage of the landslide, it directly affects its performance in the terrain and features in the image. If the model fails to effectively capture and distinguish these subtle differences, it will lead to the inability to uniformly identify different types of landslides, thus unable to specifically improve the recognition accuracy of different landslide morphologies, and further reducing the overall recognition effect and reliability. In summary, the existing landslide recognition has the technical problem of being unable to accurately identify landslides due to the deficiencies in both data sources and models. To solve the technical problem of low accuracy in landslide recognition in the prior art, the present invention provides a deep learning-based intelligent landslide recognition method for multi-source remote sensing images. First, by obtaining high-resolution optical and radar data of multi-source remote sensing images and adopting a spatio-temporal data expansion method, the problems of insufficient data types and data quantities in existing landslide recognition methods are effectively overcome, providing more comprehensive and accurate landslide data and enhancing the generalization ability of the model. Second, a deep learning model that combines semi-supervised learning and adaptive convolutional multi-scale feature fusion is used, which can dynamically adjust the size of the convolutional kernel, accurately capture landslide features at different scales, enhance the understanding of landslide morphology and structure, and improve the robustness of the model in complex terrains. Finally, a positioning correction method is used to correct the landslide boundary and morphology, eliminating inaccurate or discontinuous boundaries caused by prediction errors, and further improving the stability of the model in complex environments. Through the organic combination of the above steps, the accuracy of model recognition is improved, the reliability of the model in different environments is enhanced, the intelligent landslide recognition is made more accurate and comprehensive, and various complex situations can be handled.

[0052] Specifically:

[0053] Please refer to Figures 1-4 As shown, a deep learning-based intelligent landslide recognition method for multi-source remote sensing images proposed in the first embodiment of the present invention includes:

[0054] S1. Construct a multi-sensor spatial scale feature difference sample based on the optical remote sensing image, radar remote sensing image, and terrain data within the landslide area, and construct an intelligent landslide recognition data set by adopting a spatio-temporal data expansion method based on the multi-sensor spatial scale feature difference sample;

[0055] S2. Construct a landslide intelligent recognition model that combines semi-supervised learning and adaptive convolutional multi-scale feature fusion considering the morphological difference features of the landslide body, and design a gradient-weighted objective function to optimize the landslide intelligent recognition model;

[0056] S3. Use the positioning correction method to correct the landslide intelligent recognition result output by the landslide intelligent recognition model and generate a landslide prediction result.

[0057] Preferably, the step S1 specifically includes:

[0058] S11. Establish intelligent interpretation marks considering the morphological difference characteristics of the landslide body based on optical remote sensing images, radar remote sensing images, and terrain data. The intelligent interpretation marks refer to the marks or rules for automatically identifying and extracting ground object features from remote sensing images using artificial intelligence and machine learning technologies. First, optical remote sensing images can help identify key parts such as the boundary and landslide surface of the landslide. Radar remote sensing images (such as InSAR data) provide ground displacement and deformation information, revealing the dynamic changes of the landslide. Terrain data (such as digital elevation model DEM) can describe terrain features such as the slope and aspect of the landslide. Based on the intelligent interpretation marks, initially construct a multi-sensor spatial scale feature difference sample. Specifically, for each landslide area, by collecting data from different sensors (such as optical remote sensing images, radar remote sensing images, and terrain data) and fusing them, it can be achieved through multi-source data registration, feature extraction, and fusion strategies. The fused data can construct a sample map reflecting the differences in spatial scales of different sensors, thereby generating a multi-sensor spatial scale feature difference sample;

[0059] S12. Use the spatio-temporal data expansion method to simulate and generate landslide samples with multiple sensors, multiple morphologies, multiple perspectives, and multiple scales. The calculation formula of the spatio-temporal data expansion method is as follows:

[0060] D′(x′,t′) = [α·g(x) + β·h(x) + γ·interp x (interp y (D(x,t),t′),x′)] + Ν(0,δ 2 )

[0061] Among them, D′(x′,t′) represents the expanded data set, x′ and t′ respectively represent the new spatial position and time point, α, β, and γ represent weight coefficients, g(x) represents the non-linear transformation of space, h(x) represents the non-linear transformation of time, D(x,t) represents the original data set, interp x (D(x,t),x′) and interp y (D(x,t),t′) respectively represent the interpolation operations of space and time, Ν(0,δ 2 ) represents additive noise;

[0062] S13. Establish image slices using the multi-scale balanced target region method. Considering the problem of unbalanced sample categories, preferably select image slices with a landslide pixel ratio greater than 60%. The landslide pixel ratio refers to the ratio of the number of pixels in the landslide area to the total number of pixels in the entire study area in remote sensing images or geographic information systems, and mix and reshuffle it with the sample data in steps S11 and S12 to construct an intelligent landslide recognition dataset. The calculation formula of the multi-scale balanced target region method is as follows:

[0063] C i (x,t) = slice(I i (x), window_size, stride)

[0064]

[0065] Among them, C i (x,t) represents the slice data at the i-th scale, I i (x) represents the original image data at the i-th scale, window_size represents the window size, preferably set to 16, 32, 64, 128, 256 or 512, stride represents the step size, preferably set to 16, 32, 64, 128, 256 or 512, C balanced represents the slice after regional balance and fusion, S i represents the resolution of the multi-source image, w(S i ) represents the weight, which is adaptively determined according to the size of the landslide area.

[0066] Preferably, step S2 specifically includes:

[0067] S21. Extract the morphological difference features of the landslide body according to the intelligent landslide recognition dataset in step S1. Preferably, the morphological difference features of the landslide body include landslide morphological features, landslide texture features or landslide boundary features, and design an adaptive convolution kernel according to the morphological difference features of the landslide body. Preferably, as Figure 2 shown, the adaptive convolution kernel includes morphological convolution, horizontal texture convolution, vertical texture convolution, horizontal boundary convolution, vertical boundary convolution or edge detection convolution kernel. By designing multiple convolution kernels to meet landslide samples of different shapes and scales, the purpose of fully extracting the semantic information of multi-source remote sensing landslides is achieved. The calculation formula of the adaptive convolution kernel is as follows:

[0068] K(x,y) = α(x,y)·f(shape,texture,boundary)

[0069] Among them, K(x, y) represents the adaptive convolution kernel, α(x, y) represents the adaptive weight obtained by network training, and f(shape, texture, boundary) represents the constructor of morphological, texture, and boundary features;

[0070] S22. Based on the semi-supervised framework and the adaptive convolution kernel in the step S21, a landslide intelligent recognition model is constructed by using a multi-sensor high-level abstract semantic feature fusion mechanism and a deep convolutional neural network. Specifically, as Figure 3 shown, the deep convolutional neural network (Deep Convolutional Neural Networks, DCNNs or CNNs) is a deep learning model dedicated to processing data with a grid structure, such as images and videos. Through multi-layer convolution and pooling operations, the CNN can automatically extract hierarchical features of the data and is widely used in fields such as computer vision, image classification, and object detection. The calculation formula of the multi-sensor high-level abstract semantic feature fusion mechanism is as follows:

[0071]

[0072] Among them, represents the fused feature map, x m represents the m-th input data, f m (·) represents high-level abstract feature extraction, that is, the feature extraction function applied to x m data, G m (·) represents the multi-sensor feature alignment and normalization mapping function, enabling the feature representations of different sensors to be compared and fused in the same semantic space. α m represents the weighting coefficient of each sensor feature, H(·) represents the mapping function of cross-modal learning and semantic enhancement, and g(·) represents the prediction output function; that is, the Softmax layer, and pixels with a probability greater than 0.7 are used as the recognition result;

[0073] Among them,

[0074] F m = f m (x m )

[0075] F m ′ = G m (F m )

[0076]

[0077] F enhanced = H(F fusion )

[0078]

[0079] Among them, F m is the high-level semantic feature map of each sensor, that is, the feature extraction function applied to the x m data, and F m ′ is the feature map after feature alignment and normalization of each sensor, enabling the feature representations of different sensors to be compared and fused in the same semantic space. F fusion is the fused multi-sensor high-level semantic feature, and F enhanced is the feature map after semantic enhancement;

[0080] S23. Design the landslide intelligent recognition model in step S22 by optimizing the gradient-weighted objective function. The calculation formula of the gradient-weighted objective function is:

[0081]

[0082] Among them, L grad_weighted represents the gradient-weighted loss, L base represents the cross-entropy loss of the model, λ represents the weight hyperparameter for controlling the gradient weighting, and is generally set to 0.1, 0.5, or 1 through cross-validation. represents the gradient of the input image Ι(x,y), represents the norm of the gradient magnitude of the feature map at the position (x,y), which is used to measure the change intensity in the local area of the image. represents the gradient of the loss function L with respect to the input Ι(x,y). By calculating the gradient of the loss with respect to the image pixels, the contribution of each pixel in the image to the model loss can be measured.

[0083] Preferably, step S3 is specifically:

[0084] Adopt a positioning correction method to correct the boundary error in the landslide intelligent recognition result output by the landslide intelligent recognition model in step S2 to generate a landslide prediction result. Preferably, the positioning correction includes spatial correction, morphological operation, or smoothing and gradient correction. The calculation formula of the positioning correction method is:

[0085]

[0086] Among them, y represents the finally corrected landslide prediction result, y′ represents the result output by the landslide intelligent recognition model, M(·) represents spatial morphological correction, F(·) represents image filtering, represents the gradient of the error with respect to the prediction boundary, γ represents the correction step size, Indicates gradient correction. Among them, gradient correction means using gradient information to guide the correction direction, and the irregular boundaries can be adjusted by calculating the local gradient; smoothing the prediction results, removing jagged and irregular boundaries, and ensuring the coherence and continuity of the boundaries, using convolution kernels and image filtering to smooth the boundaries; morphological correction is used to eliminate the noise or small-scale incorrect boundaries in the prediction results.

[0087] In this embodiment, first, optical remote sensing images and radar remote sensing images are obtained. The optical remote sensing images provide rich ground object features and can clearly display the boundaries and shapes of landslides. The radar remote sensing images are not restricted by light and weather, can penetrate vegetation and capture potential features of landslides. Combining the data of the two can make up for their respective limitations and provide more comprehensive and accurate landslide information. In addition, through the spatio-temporal data expansion method, more representative data can be generated to cover landslide features under different times, spaces, and environmental conditions, thereby increasing the diversity of data and enhancing the generalization ability of the subsequent landslide intelligent recognition model. This not only helps to balance the sample numbers of landslide and non-landslide areas but also enhances the recognition ability of the landslide intelligent recognition model for various landslide types and development stages. In this way, the problem of insufficient data types and data quantities in existing landslide recognition methods can be effectively overcome. Secondly, the semi-supervised learning framework effectively expands the training dataset by using a large amount of unlabeled data and labels, improving the generalization ability and recognition accuracy of the landslide intelligent recognition model, especially outstanding in the case of scarce labeled data. The adaptive convolution can dynamically adjust the convolution kernel size according to the specific features of the landslide, thereby accurately capturing landslide features at different scales and enhancing the recognition ability for landslide areas of various sizes. The multi-scale feature fusion further enhances the comprehensive understanding of the landslide intelligent recognition model for diverse features such as landslide morphology and structure, making the landslide intelligent recognition model more robust under complex terrains. At the same time, by introducing image gradient information through the gradient-weighted objective function, the performance of the landslide intelligent recognition model in landslide intelligent recognition is optimized. In an image, landslide areas usually have obvious feature changes, such as boundaries, cracks, or complex terrains, and the gradient values in these areas are usually large. By weighting the areas with larger gradients in the loss function, the landslide intelligent recognition model can pay more attention to these important areas and improve the learning of landslide area features. Specifically, the gradient-weighted objective function adjusts the training focus of the landslide intelligent recognition model according to the magnitude of the local gradient, making it focus on the boundaries of the landslide and other significant change parts. By adjusting the gradient-weighted objective function, while enhancing the attention to landslide features, the balance of the overall performance can be maintained. This method effectively avoids over-focusing on the background or smooth areas, thereby improving the accuracy and robustness of landslide recognition. Especially when facing complex terrains and limited labeled samples, it can significantly improve the recognition accuracy and generalization ability of the landslide intelligent recognition model. Finally, a positioning correction method is used to process the landslide intelligent recognition results output by the landslide intelligent recognition model, correcting boundary errors and morphological deviations, and being able to eliminate the problems of discontinuous or inaccurate boundaries caused by prediction errors. This method smooths the landslide boundary, eliminates the boundary irregularities caused by noise or landslide intelligent recognition model errors, and ensures that the boundary of the landslide area is more continuous and accurate.In addition, the positioning correction method combines terrain and prior knowledge to correct the morphology of landslides, avoiding morphological deviations caused by terrain complexity and landslide type diversity, and improving the overall recognition accuracy. This method not only improves the recognition effect of the landslide intelligent recognition model, but also enhances the stability and adaptability under complex terrain and different weather conditions, providing a complete technical solution for the accurate intelligent recognition of landslide disasters. Therefore, through the organic combination of the above steps, the accuracy of model recognition is improved, and the reliability of the model in different environments is enhanced, making the landslide intelligent recognition more accurate and comprehensive, and capable of handling various complex situations.

[0088] In summary, the landslide prediction results generated by the deep learning-based landslide intelligent recognition method using multi-source remote sensing images provided by the present invention are as Figure 4 and Figure 5 shown, and the accuracy comparison indicators are shown in Table 1. Figure 4 The original image is shown on the left, the true label data is shown in the middle, and the landslide prediction result generated by the present invention is shown on the right. Figure 5 It is a heat map comparing the FAST-SCNN, SegFormer, Swin, ELN, SC models with the model of the present invention. By comparing the original image, the true label data, the model prediction result and the heat map, it can be seen that this method has high accuracy and detail capture ability in the recognition of landslide areas, and at the same time shows good robustness in different scenarios such as rocky landslides, rock-soil landslides, and vegetation-covered landslides, significantly improving the automation degree of landslide recognition.

[0089] Table 1

[0090]

[0091] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other, and the key points of each embodiment are the differences from other embodiments. The above are only the embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the present invention.

Claims

1. A landslide intelligent identification method based on deep learning of multi-source remote sensing images, characterized in that: include: S1. Constructing multi-sensor spatial scale feature difference samples based on optical remote sensing images, radar remote sensing images and terrain data in the landslide area, and constructing an intelligent landslide identification dataset based on the multi-sensor spatial scale feature difference samples using a spatiotemporal data expansion method; S2. Based on a semi-supervised learning framework, an adaptive convolution multi-scale feature fusion landslide intelligent identification model is constructed that takes into account the morphological differences of landslide bodies, and a gradient weighted objective function is designed to optimize the landslide intelligent identification model; S3. Correcting the landslide intelligent identification result output by the landslide intelligent identification model using a positioning correction method, and generating a landslide prediction result.

2. According to claim 1, a landslide intelligent identification method based on multi-source remote sensing image deep learning is characterized in that: The step S1 comprises: S11, establishing intelligent interpretation marks that take into account the morphological differences of landslide bodies based on optical remote sensing images, radar remote sensing images and terrain data, and preliminarily constructing multi-sensor spatial scale feature difference samples based on the intelligent interpretation marks; S12. Based on the multi-sensor spatial scale characteristic difference samples, a spatiotemporal data expansion method is used to simulate and generate landslide samples. The calculation formula of the spatiotemporal data expansion method is as follows: D′(x′,t′)=[α·g(x)+β·h(x)+γ·interp x (interp y (D(x,t),t′),x′)]+Ν(0,δ 2 ) Among them, D′(x′,t′) represents the expanded data set, x′,t′ represent the new spatial position and time point respectively, α, β, γ represent weight coefficients, g(x) represents the nonlinear transformation of space, h(x) represents the nonlinear transformation of time, D(x,t) represents the original data set, interp x (D(x,t),x′) and interp y (D(x, t), t′) represent the interpolation operations in space and time respectively, Ν(0,δ 2 ) represents additive noise; S13, using a multi-scale balanced target region method to establish image slices, and mixing with the sample data in steps S11 and S12 to re-construct an intelligent landslide identification data set. The calculation formula of the multi-scale balanced target region method is as follows: C i (x,t)=slice(I i (x),window_size,stride) Among them, C i (x, t) represents the slice data at the i-th scale, I i (x) represents the original image data at the i-th scale, window_size represents the window size, stride represents the step size, C balanced represents the slice after regional balance and fusion, S i represents the resolution of the multi-source image, w(S i ) represents the weight, which is determined adaptively according to the size of the landslide area.

3. The method for intelligent landslide identification based on deep learning of multi-source remote sensing images according to claim 2 is characterized in that: In step S13, the window size is 16, 32, 64, 128, 256 or 512; the step size is 16, 32, 64, 128, 256 or 512.

4. The method for intelligent landslide identification based on deep learning of multi-source remote sensing images according to claim 1 is characterized in that: The step S2 comprises: S21, extracting landslide body morphological difference characteristics according to the intelligent landslide identification data set in step S1, and designing an adaptive convolution kernel according to the landslide body morphological difference characteristics, and the calculation formula of the adaptive convolution kernel is as follows: K(x,y)=α(x,y)·f(shape,texture,boundary) Among them, K(x,y) represents the adaptive convolution kernel, α(x,y) represents the adaptive weight obtained by network training, and f(shape, texture, boundary) represents the constructor of morphology, texture and boundary features; S22, based on the semi-supervised framework and the adaptive convolution kernel in step S21, a multi-sensor high-level abstract semantic feature fusion mechanism and a deep convolutional neural network are used to construct a landslide intelligent recognition model, and the calculation formula of the multi-sensor high-level abstract semantic feature fusion mechanism is as follows: in, Represents the fused feature map, x m represents the mth input data, f m (·) represents high-level abstract feature extraction, G m (·) represents the multi-sensor feature alignment and normalization mapping function, α m represents the weighting coefficient of each sensor feature, H(·) represents the mapping function for cross-modal learning and semantic enhancement, and g(·) represents the prediction output function; S23, designing a gradient weighted objective function to optimize the landslide intelligent identification model in step S22, wherein the calculation formula of the gradient weighted objective function is: Among them, L grad_weighted represents the gradient weighted loss, L base represents the cross entropy loss of the model, λ represents the weight hyperparameter that controls the gradient weighting, represents the gradient of the input image Ι(x,y), Represents the norm of the gradient amplitude of the feature map at position (x, y), Represents the gradient of the loss function L with respect to the input Ι(x,y).

5. The method for intelligent landslide identification based on deep learning of multi-source remote sensing images according to claim 4 is characterized in that: The landslide body morphological difference characteristics include landslide morphological characteristics, landslide texture characteristics and / or landslide boundary characteristics.

6. The method for intelligent landslide identification based on deep learning of multi-source remote sensing images according to claim 4 is characterized in that: The adaptive convolution kernel includes morphological convolution, horizontal texture convolution, vertical texture convolution, horizontal boundary convolution, vertical boundary convolution or edge detection convolution.

7. The method for intelligent landslide identification based on deep learning of multi-source remote sensing images according to claim 1 is characterized in that: The step S3 is specifically as follows: The positioning correction method is used to correct the boundary error in the landslide intelligent identification result output by the landslide intelligent identification model in step S2 to generate a landslide prediction result. The calculation formula of the positioning correction method is: Among them, y represents the final corrected landslide prediction result, y′ represents the output result of the landslide intelligent identification model, M(·) represents spatial morphological correction, F(·) represents image filtering, represents the gradient of the error to the prediction boundary, γ represents the correction step size, represents the gradient correction.

8. The method for intelligent landslide identification based on deep learning of multi-source remote sensing images according to claim 7 is characterized in that: The positioning correction includes space correction, morphological operation and / or smooth gradient correction.

9. The method for intelligent landslide identification based on deep learning of multi-source remote sensing images according to claim 7 is characterized in that: The correction step size is 0.1, 0.5 or 1.

Images