Lightweight-based satellite landslide image intelligent detection method, device and equipment and medium

Abstract

The application discloses a kind of light-based satellite landslide image intelligent detection method, device, equipment and medium, it is related to disaster detection technical field, including: by obtaining the whole scene optical satellite image and topographic auxiliary data containing landslide and non-landslide area, based on topographic data, using dynamic segmentation strategy to image is differentially segmented and standardized preprocessing, obtain standardization image. Through double-layer machine learning model screening target landslide image, the target image is input to improved light convolutional neural network and is processed, and initial detection result is obtained, finally the result is edge optimized and coordinate calibration, and the accurate landslide area detection result is output. Through dynamic segmentation and double-layer screening, the recognition accuracy of landslide image is effectively improved, and the improved light convolutional neural network realizes efficient detection in a low-resource environment, further improves the accuracy of landslide detection result.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of disaster detection, and in particular to a satellite landslide image intelligent detection method and device based on lightness, equipment and a medium. BACKGROUND

[0002] In the current satellite image-based landslide detection technology, traditional methods mostly rely on manual visual interpretation or simple image feature threshold segmentation, which is low in efficiency and strong in subjectivity; the method based on deep learning mainly learns landslide deep semantic features from remote sensing images through a convolutional neural network to realize intelligent identification. In order to overcome the problem that the local receptive field of the traditional convolutional neural network is limited and it is difficult to capture the long-distance dependence relationship between the landslide body and the surrounding environment.

[0003] Although the prior art has improved the landslide detection accuracy, there are still significant defects: first, the uniform segmentation method with fixed overlapping pixel values is mostly used for whole scene image segmentation, without considering the difference of potential landslide areas, which is easy to cause the loss of landslide edge information or low segmentation efficiency; second, the preliminary screening link mostly relies on a single feature (such as only color feature) or a single machine learning model, which is insufficient in screening accuracy, and a large number of non-landslide images will still enter the subsequent detection process, occupying computing resources and affecting the overall detection efficiency.

[0004] Therefore, how to realize real-time and accurate detection of satellite landslide images while balancing the light-weight deployment requirements and detection accuracy has become a problem to be solved. SUMMARY

[0005] The main purpose of the present application is to provide a satellite landslide image intelligent detection method and device based on lightness, equipment and a medium, which aims to solve the technical problem of how to realize efficient and accurate detection of satellite landslide images on edge devices with limited computing resources.

[0006] To achieve the above purpose, the present application provides a satellite landslide image intelligent detection method based on lightness, which comprises the following steps:

[0007] Obtaining optical satellite images of a whole scene area and corresponding terrain auxiliary data, wherein the whole scene area comprises a landslide area and a non-landslide area;

[0008] Based on the terrain auxiliary data, a dynamic segmentation strategy is used to differentially segment and standardize the preprocessing of the optical satellite images, to obtain a standardized segmented image;

[0009] A double-layer machine learning model is used to double-screen the standardized segmented image, to obtain a target landslide image, wherein the double-layer machine learning model is constructed based on a support vector machine model and a random forest classifier;

[0010] Input the target landslide image into an improved lightweight convolutional neural network for processing to obtain an initial landslide detection result, wherein the improved lightweight convolutional neural network comprises a dynamic depth convolution module, a cross-level skip connection module, a multi-directional convolution shift enhancement module, an improved joint attention module, and an adaptive upsampling module.

[0011] Edge optimization and coordinate calibration are performed on the initial landslide detection result to output a target landslide region detection result.

[0012] In an embodiment, the improved lightweight convolutional neural network comprises a dynamic depth convolution module, a selectable residual connection module, a cross-level skip connection module, a multi-directional convolution shift enhancement module, an improved joint attention module, and an adaptive upsampling module, and further comprises:

[0013] The dynamic depth convolution module adopts a depth separable convolution structure in a MobileNetV3Small backbone network, comprising a 3x3 depth convolution layer, a batch normalization layer, and an activation function layer connected in sequence, and further comprises a selectable residual connection submodule; the cross-level skip connection module comprises an encoder and a decoder; the multi-directional convolution shift enhancement module comprises a channel grouping unit, a four-way shift unit, and a channel superposition unit; the improved joint attention module comprises a channel attention submodule and a spatial attention submodule, the channel attention submodule is composed of a global average pooling layer, two 1x1 convolution layers, and a Sigmoid activation layer, and the spatial attention submodule is composed of a channel average value map generation unit, a channel maximum value map generation unit, a channel splicing unit, a 7x7 convolution layer, and a Sigmoid activation layer; the adaptive upsampling module comprises a nearest neighbor interpolation unit, a 3x3 depth convolution alignment unit, and a 1x1 pointwise convolution unit.

[0014] In an embodiment, the step of adopting a dynamic segmentation strategy based on the terrain auxiliary data to perform differential segmentation and standardized preprocessing on the optical satellite image to obtain a standardized segmented image comprises:

[0015] The slope data and the terrain roughness data in the terrain auxiliary data are preprocessed to obtain a slope normalized value and a terrain roughness normalized value.

[0016] When the slope normalized value is greater than a preset first threshold value and the terrain roughness normalized value is greater than a preset second threshold value, the region corresponding to the slope normalized value and the terrain roughness is determined as a high-risk landslide region.

[0017] when either the slope normalization value is greater than a preset first threshold value or the terrain roughness normalization value is greater than a preset second threshold value, determining the region corresponding to the slope normalization value and terrain roughness as a medium-risk landslide region;

[0018] when neither the slope normalization value is greater than a preset first threshold value nor the terrain roughness normalization value is greater than a preset second threshold value, determining the region corresponding to the slope normalization value and terrain roughness as a low-risk landslide region;

[0019] performing pixel segmentation on a portion of the optical satellite image corresponding to the high-risk landslide region according to a preset first overlap pixel value to obtain a first segmented image;

[0020] performing pixel segmentation on a portion of the optical satellite image corresponding to the medium-risk landslide region according to a preset second overlap pixel value to obtain a second segmented image;

[0021] performing pixel segmentation on a portion of the optical satellite image corresponding to the low-risk region according to a preset third overlap pixel value to obtain a third segmented image, wherein the preset first overlap pixel value is greater than the preset second overlap pixel value, and the preset second overlap pixel value is greater than the preset third overlap pixel value;

[0022] performing numbering on the first segmented image, the second segmented image and the third segmented image according to a preset rule and storing original geographic coordinate information corresponding to each image to obtain a standardized segmented image associated with original geographic coordinates.

[0023] In an embodiment, the step of performing double-layer screening on the standardized segmented image by a double-layer machine learning model to obtain a target landslide image comprises:

[0024] performing color distribution statistics on three color channels of the standardized segmented image respectively, calculating a color histogram of each channel using a preset quantization interval to obtain a three-channel histogram;

[0025] splicing the three-channel histogram according to the channel dimension to generate a color histogram feature vector;

[0026] converting the standardized segmented image into a grayscale image and obtaining a texture feature vector using a gray level co-occurrence matrix algorithm;

[0027] performing standardization processing on the color histogram feature vector and the texture feature vector respectively by an L2 normalization method to obtain a standardized color feature vector and a standardized texture feature vector;

[0028] splicing the standardized color feature vector and the standardized texture feature vector along the feature dimension to obtain a fusion feature vector;

[0029] inputting the fusion feature vector into a random forest classifier of a double-layer machine learning model for processing and screening to obtain a first screened image, wherein the random forest classifier comprises a preset number of decision trees, and the first screened image comprises landslide features;

[0030] inputting a fusion feature vector corresponding to the first screened image into a support vector machine model in the double-layer machine learning model for processing to obtain a target landslide image, wherein the support vector machine model adopts a radial basis kernel function.

[0031] In an embodiment, the target landslide image is inputted into an improved lightweight convolutional neural network for processing to obtain an initial landslide detection result, wherein the improved lightweight convolutional neural network comprises the steps of a dynamic depth convolution module, a cross-layer jump connection module, a multi-directional convolution shift enhancement module, an improved joint attention module, and an adaptive upsampling module, comprising:

[0032] normalizing and calibrating the target landslide image to obtain a standardized input image;

[0033] inputting the standardized input image into the dynamic depth convolution module, passing through a 3x3 depth convolution layer, a batch normalization, and an activation function layer of MobileNetV3Small, and then passing through a selectable residual connection submodule for processing according to the matching relationship of input channel number, output channel number, and step length to obtain an initial feature map;

[0034] inputting the initial feature map into an encoder and a decoder of the cross-layer jump connection module in sequence to obtain a corresponding shallow encoder feature map and a deep decoder feature map;

[0035] cross-layer associating the shallow encoder feature map and the deep decoder feature map to obtain a cross-layer associated feature map;

[0036] inputting the cross-layer associated feature map into the multi-directional convolution shift enhancement module, grouping channels, shifting four-way pixels, and reorganizing 1x1 convolution to obtain a multi-directional enhanced feature map;

[0037] inputting the multi-directional enhanced feature map into a channel attention submodule and a spatial attention submodule of the improved joint attention module for element-by-element multiplication to obtain an attention weighted feature map;

[0038] inputting the attention weighted feature map into the adaptive upsampling module to obtain an initial landslide detection result.

[0039] In an embodiment, the step of inputting the attention weighted feature map into the adaptive upsampling module to obtain an initial landslide detection result comprises:

[0040] input the attention-weighted feature map into a nearest neighbor interpolation unit to obtain an up-sampling feature map;

[0041] input the up-sampling feature map into a 3x3 deep convolution alignment unit to independently perform a 3x3 deep convolution operation on each channel to obtain a feature map after deep convolution alignment;

[0042] input the feature map after deep convolution alignment into a 1x1 point-wise convolution unit to adjust the number of channels to a preset output dimension through a 1x1 convolution kernel to obtain an adjusted feature map;

[0043] input the adjusted feature map into a 3x3 standard convolution layer, a batch normalization layer and a ReLU activation function in sequence to perform feature smoothing and nonlinear enhancement to obtain an enhanced feature map;

[0044] input the enhanced feature map into a single-channel 1x1 convolution layer to map to a class prediction space to obtain a mapping result;

[0045] apply a Sigmoid activation function to the mapping result to output a binary classification probability map;

[0046] when a probability value corresponding to the binary classification probability map is greater than a preset threshold, determine a region corresponding to the probability value as a landslide region;

[0047] take the binary classification probability map corresponding to the region as the initial landslide detection result.

[0048] In an embodiment, the step of performing edge optimization and coordinate calibration on the initial landslide detection result to output a target landslide region detection result includes:

[0049] obtain a binary classification probability map corresponding to the initial landslide detection result;

[0050] process the binary classification probability map using morphological closing operation to obtain a smoothed probability map;

[0051] perform threshold segmentation on the smoothed probability map to mark a region with a probability value greater than a preset determination threshold in the smoothed probability map as a candidate landslide region;

[0052] extract an edge contour of the candidate landslide region to obtain an initial landslide edge;

[0053] call original geographic coordinate information associated with the standardized segmented image, and calibrate the initial landslide edge to obtain a calibrated landslide edge;

[0054] perform terrain adaptive adjustment on the calibrated landslide edge in combination with slope direction information in the terrain auxiliary data to obtain a precise landslide edge;

[0055] According to the accurate landslide edge, the boundary coordinates, the center point coordinates and the area of the target landslide region are calculated, and the landslide region description information is generated, wherein the landslide region description information includes geographic coordinates and range size;

[0056] The landslide region description information is packaged in a preset data format, and a target landslide region detection result is output.

[0057] In addition, in order to achieve the above-mentioned purpose, the application also provides a satellite landslide image intelligent detection device based on lightness, which comprises:

[0058] An acquisition module is configured to acquire optical satellite images of an entire scene region and corresponding terrain auxiliary data, wherein the entire scene region comprises a landslide region and a non-landslide region;

[0059] A processing module is configured to perform differential segmentation and standardized preprocessing on the optical satellite images based on the terrain auxiliary data by using a dynamic segmentation strategy, so as to obtain a standardized segmented image;

[0060] A screening module is configured to perform double-layer screening on the standardized segmented image by using a double-layer machine learning model, so as to obtain a target landslide image, wherein the double-layer machine learning model is constructed based on a support vector machine model and a random forest classifier;

[0061] A result module is configured to input the target landslide image into an improved light-weight convolutional neural network for processing, so as to obtain an initial landslide detection result, wherein the improved light-weight convolutional neural network comprises a dynamic depth convolution module, a cross-layer jump connection module, a multi-directional convolution shift enhancement module, an improved joint attention module and an adaptive up-sampling module;

[0062] An optimization module is configured to perform edge optimization and coordinate calibration on the initial landslide detection result, so as to output a target landslide region detection result.

[0063] In addition, in order to achieve the above-mentioned purpose, the application also provides a storage medium, which is a computer readable medium, and a computer program is stored on the storage medium, wherein the computer program is executed by a processor to realize the steps of the satellite landslide image intelligent detection method based on lightness as described above.

[0064] In addition, in order to achieve the above-mentioned purpose, the application also provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to realize the steps of the satellite landslide image intelligent detection method based on lightness as described above.

[0065] The application obtains whole-scene optical satellite images containing landslide and non-landslide areas and terrain auxiliary data, adopts a dynamic segmentation strategy based on terrain data to differentially segment and standardize the images, and obtains standardized images. Target landslide images are screened through a double-layer machine learning model, the target images are input into an improved lightweight convolutional neural network for processing to obtain initial detection results, and finally the results are edge-optimized and coordinate-calibrated to output accurate landslide area detection results. The dynamic segmentation and double-layer screening effectively improve the recognition accuracy of landslide images, the improved lightweight convolutional neural network realizes efficient detection in a low-resource environment, and further improves the accuracy of the landslide detection results. BRIEF DESCRIPTION OF DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor.

[0067] Figure 1 The flowchart of the first embodiment of the satellite landslide image intelligent detection method based on lightweight of the present application is shown.

[0068] Figure 2 The flowchart of the second embodiment of the satellite landslide image intelligent detection method based on lightweight of the present application is shown.

[0069] Figure 3 The flowchart of the third embodiment of the satellite landslide image intelligent detection method based on lightweight of the present application is shown.

[0070] Figure 4 The module structure diagram of the satellite landslide image intelligent detection device based on lightweight of the first embodiment of the satellite landslide image intelligent detection method based on lightweight of the present application is shown.

[0071] Figure 5 The device structure diagram of the hardware running environment involved in the satellite landslide image intelligent detection method based on lightweight in the embodiments of the present application is shown.

[0072] The purpose of the present application, the functional characteristics and the advantages will be further explained in combination with the embodiments and with reference to the drawings. DETAILED DESCRIPTION

[0073] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application, and are not used to limit the present application.

[0074] In order to better understand the technical solutions of the present application, the following will be described in detail in combination with the drawings in the specification and the specific embodiments.

[0075] In current satellite image-based landslide detection techniques, traditional methods mostly rely on manual visual interpretation or simple image feature threshold segmentation, which are inefficient and subjective. Deep learning-based methods mainly use convolutional neural networks to learn deep semantic features of landslides from remote sensing images for intelligent recognition. To overcome the limitations of traditional convolutional neural networks in capturing long-range dependencies between landslide bodies and their surroundings, a lightweight satellite landslide image intelligent detection method is proposed to achieve efficient and accurate detection of satellite landslide images on edge devices with limited computing resources.

[0076] Based on the above, the embodiment of the present application provides a lightweight satellite landslide image intelligent detection method, which is described with reference to Figure 1 , Figure 1 The flowchart of the first embodiment of the lightweight satellite landslide image intelligent detection method of the present application is shown. In this embodiment, the lightweight satellite landslide image intelligent detection method comprises steps S10-S50:

[0077] Step S10, obtaining optical satellite images of the whole scene area and corresponding terrain auxiliary data.

[0078] It should be noted that the whole scene area refers to the complete geographical range covered by a single imaging of an optical satellite. This range is not a scattered local plot, but a spatial area with continuous geographical boundaries. The whole scene area includes landslide areas and non-landslide areas. Optical satellite images refer to image data generated by optical remote sensing satellites through receiving visible light band electromagnetic wave signals reflected by ground objects, and then through photoelectric conversion and data processing. Terrain auxiliary data refers to terrain-related data used to assist in analyzing the characteristics of landslide areas, including slope data (i.e., the inclination of the ground unit), aspect data (i.e., the azimuth of the ground unit), and terrain roughness data (i.e., the degree of fluctuation of the ground surface).

[0079] Step S20, based on the terrain auxiliary data, using a dynamic segmentation strategy to differentially segment and standardize the preprocessing of the optical satellite images, to obtain standardized segmented images.

[0080] It should be noted that the dynamic segmentation strategy refers to a processing method that uses non-uniform segmentation rules for optical satellite images based on the different characteristics of terrain auxiliary data. Differential segmentation refers to setting different segmentation parameters for image division for different terrain risk levels in the optical satellite images. The core of this operation is to distinguish between high-risk, medium-risk, and low-risk landslide areas. Standardized segmented images refer to image data with uniform size specifications and geographical coordinate associations after dynamic segmentation strategy processing.

[0081] Further, the step S20 comprises: firstly, preprocessing the slope data and the terrain roughness data in the terrain auxiliary data to obtain a slope normalized value and a terrain roughness normalized value. It should be noted that the slope data refers to the terrain data representing the steepness of the ground unit, reflecting the angle between the inclination angle of a point on the ground and the horizontal direction. The terrain roughness data refers to the terrain data representing the complexity of the ground surface, reflecting the dispersion degree of the elevation value within the ground unit.

[0082] When the slope normalized value is greater than a preset first threshold value and the terrain roughness normalized value is greater than a preset second threshold value, the region corresponding to the slope normalized value and the terrain roughness is determined as a high-risk landslide region; when either the slope normalized value is greater than the preset first threshold value or the terrain roughness normalized value is greater than the preset second threshold value, the region corresponding to the slope normalized value and the terrain roughness is determined as a medium-risk landslide region; when neither the slope normalized value is greater than the preset first threshold value nor the terrain roughness normalized value is greater than the preset second threshold value, the region corresponding to the slope normalized value and the terrain roughness is determined as a low-risk landslide region. It should be noted that the preset first threshold value refers to a critical value for determining the landslide risk set for the slope normalized value, which is an empirical value obtained based on a large amount of landslide sample data, and is used to distinguish the high and low of the landslide risk in the slope dimension. The preset second threshold value refers to a critical value for determining the landslide risk set for the terrain roughness normalized value, which is also determined by statistical analysis of a large amount of landslide sample data, and is used to distinguish the high and low of the landslide risk in the terrain roughness dimension.

[0083] The part corresponding to the high-risk landslide region in the optical satellite image is pixel segmented according to a preset first overlap pixel value to obtain a first segmented image; the part corresponding to the medium-risk landslide region in the optical satellite image is pixel segmented according to a preset second overlap pixel value to obtain a second segmented image; and the part corresponding to the low-risk region in the optical satellite image is pixel segmented according to a preset third overlap pixel value to obtain a third segmented image. It should be noted that the preset first overlap pixel value is greater than the preset second overlap pixel value, and the preset second overlap pixel value is greater than the preset third overlap pixel value. Since the high-risk region has a greater possibility of landslide, a larger overlap pixel value can ensure that the segmented image has sufficient overlap region at the boundary, avoiding the omission of landslide features in the segmentation process, thereby ensuring the accuracy of subsequent detection, while the overlap pixel value of the low-risk region is the smallest, because the landslide risk of these regions is low, a smaller overlap pixel value can ensure the basic demand of detection while further improving the processing efficiency and reducing unnecessary computational burden.

[0084] Finally, the first, second, and third segmented images are numbered according to a preset rule and the original geographic coordinate information corresponding to each image is stored to obtain standardized segmented images associated with original geographic coordinates. It should be noted that the preset rule refers to a standard specification prepared in advance for numbering different types of segmented images. This rule usually sets the numbering content in combination with the risk level and spatial position information corresponding to the image, so that the numbering can intuitively reflect the attribute characteristics of the image. The original geographic coordinate information refers to the real geographic spatial coordinate data corresponding to each pixel point in the optical satellite image, which can accurately locate the position and range of the image in the real geographic space. Specifically, the first, second, and third segmented images obtained are numbered according to the preset rule to ensure that each segmented image has a unique identifier, facilitating subsequent processing and management. The numbering rule can be developed according to factors such as the source area (high, medium, and low risk) of the segmented image and the segmentation order. Then, the original geographic coordinate information corresponding to each segmented image is stored. This geographic coordinate information is extracted from the original optical satellite image and is used to record the position of each segmented image in the actual geographic space. Finally, the numbered segmented images are associated with the stored original geographic coordinate information to form standardized segmented images. This association operation enables each segmented image to carry an explicit geographic identifier, facilitating further analysis and application in a geographic information system (GIS) or other related software, and also provides accurate geographic reference for the final landslide detection results.

[0085] In step S30, the standardized segmented images are double-layer filtered by a double-layer machine learning model to obtain target landslide images.

[0086] It should be noted that the double-layer machine learning model refers to a hierarchical screening model composed of two different machine learning classifiers, with the first-stage classifier responsible for coarse screening and the second-stage classifier responsible for fine screening. In this embodiment, the double-layer machine learning model is constructed based on a support vector machine model and a random forest classifier. The support vector machine model is a supervised learning classification model based on statistical learning theory, whose core idea is to find an optimal classification hyperplane to accurately divide sample data of different classes. The random forest classifier is a classification model based on the idea of ensemble learning, composed of multiple decision tree classifiers, and determines the final classification result through a voting mechanism. The target landslide image refers to an image data containing landslide region features obtained after filtering by the double-layer machine learning model. This type of image eliminates a large number of invalid image data that do not contain landslide features and only retains images with suspected landslide region features.

[0087] Further, the step S30 comprises: performing color distribution statistics on three color channels of the normalized segmented image respectively, calculating a color histogram of each channel by using a preset quantization interval, obtaining a three-channel histogram; splicing the three-channel histogram according to the channel dimension to generate a color histogram feature vector; converting the normalized segmented image into a gray image and obtaining a texture feature vector by using a gray level co-occurrence matrix algorithm; performing standardization processing on the color histogram feature vector and the texture feature vector respectively by using an L2 normalization method, obtaining a standardized color feature vector and a standardized texture feature vector; splicing the standardized color feature vector and the standardized texture feature vector along the feature dimension to obtain a fusion feature vector; inputting the feature vector into a random forest classifier of the double-layer machine learning model for processing and screening to obtain a first screening image, wherein the random forest classifier comprises a preset number of decision trees, and the first screening image comprises a landslide feature; inputting the fusion feature vector corresponding to the first screening image into a support vector machine model in the double-layer machine learning model for processing to obtain a target landslide image, wherein the support vector machine model uses a radial basis kernel function. It should be noted that the color distribution statistics refers to an operation of frequency statistics on the pixel brightness values in the three color channels, which can quantitatively reflect the proportion of pixels with different brightness values in each color channel. The preset quantization interval refers to a numerical interval preset for dividing the pixel brightness values. The color histogram feature vector refers to a one-dimensional numerical vector formed by sequentially arranging the statistical data of the three-channel histogram, which is a digital expression form of the image color feature. The gray level co-occurrence matrix algorithm refers to an algorithm for counting the frequency of occurrence of different position pixel gray value combinations based on the spatial position relationship of the pixels in the gray image. The L2 normalization method refers to a commonly used method for normalizing the feature vector, and the core is to normalize the L2 norm of the feature vector, i.e., the square root of the sum of squares of the elements of the vector, to 1. The preset number of decision trees refers to the total number of decision trees preset when constructing the random forest classifier, which can be 100 in this embodiment. It can be understood that first, the red, green and blue color channels of the normalized segmented image are subjected to color distribution statistics respectively, and the color histogram of each channel is calculated by using the preset quantization interval, thereby obtaining the three-channel histogram. This process can quantitatively reflect the distribution of different colors in the image, providing basic data for subsequent feature extraction. Then, the three-channel histogram is spliced according to the channel dimension to generate a comprehensive color histogram feature vector for describing the color feature of the image. Then, the normalized segmented image is converted into a gray image, and a texture feature vector is calculated by using a gray level co-occurrence matrix algorithm. The gray level co-occurrence matrix can effectively extract the texture information of the image and reflect the spatial distribution relationship of the pixels in the image. Subsequently, the standardized color feature vector and the standardized texture feature vector are spliced along the feature dimension to obtain a fusion feature vector, which comprehensively reflects the color and texture features and provides more abundant information for subsequent classification and screening.The fusion feature vector is input into a random forest classifier in the double-layer machine learning model for processing and screening. The random forest filters out the first screening image containing landslide features through the synergistic effect of multiple decision trees. This process preliminarily screens out image regions that may contain landslides. Finally, the fusion feature vector corresponding to the first screening image is extracted and input into a support vector machine model in the double-layer machine learning model for further processing, thereby obtaining the final target landslide image, i.e., the image containing landslide features. This double-layer screening mechanism ensures the accuracy and reliability of the screening results through the preliminary screening of the random forest and the accurate classification of the support vector machine.

[0088] In step S40, the target landslide image is input into the improved lightweight convolutional neural network for processing to obtain an initial landslide detection result.

[0089] It should be noted that the improved lightweight convolutional neural network includes a dynamic depth convolution module, a cross-level skip connection module, a multi-directional convolution shift enhancement module, an improved joint attention module, and an adaptive upsampling module. The dynamic depth convolution module adopts the depth separable convolution structure in the MobileNetV3Small backbone network, including a 3x3 depth convolution layer, a batch normalization layer, and an activation function layer connected in turn. The dynamic depth convolution module further includes a selectable residual connection submodule. The selectable residual connection submodule adaptively introduces a skip connection according to the matching relationship between the input channel number, the output channel number, and the step size. The construction method of the selectable residual connection submodule is that when the input channel number and the output channel number are the same and the step size is 1, an identity mapping is used as the skip connection, otherwise, a projection mapping composed of a 1x1 convolution layer and a batch normalization layer connected in turn is used as the skip connection. The cross-level skip connection module includes an encoder and a decoder. The multi-directional convolution shift enhancement module includes a channel grouping unit, a four-way shift unit, and a channel superposition unit. The improved joint attention module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule is composed of a global average pooling layer, two 1x1 convolution layers, and a Sigmoid activation layer. The spatial attention submodule is composed of a channel average value map generation unit, a channel maximum value map generation unit, a channel splicing unit, a 7x7 convolution layer, and a Sigmoid activation layer. The adaptive upsampling module includes a nearest neighbor interpolation unit, a 3x3 depth convolution alignment unit, and a 1x1 pointwise convolution unit.

[0090] It can be understood that first, the target landslide image is input into the dynamic depth convolution module of the improved lightweight convolutional neural network, and the depth separable convolution structure in the MobileNetV3Small backbone network is used to sequentially perform 3x3 depth convolution, batch normalization and HardSwish activation, through the selectable residual connection module, according to the matching relationship of the input channel number, the output channel number and the step length, the jump connection is adaptively introduced, and the initial feature map of the image is extracted. Then, the initial feature map is transmitted to the cross-level jump connection module, which includes an encoder and a decoder. In the encoder, the feature map is down-sampled to extract multi-scale features; at the same time, the low-level features in the encoder are directly transmitted to the decoder through the jump connection, which is used to retain more detailed information and enhance the hierarchical expression ability of the features, providing a basis for subsequent feature fusion. Subsequently, the feature map output by the encoder is input into the multi-directional convolution shift enhancement module. The module first groups the feature map by channels, then performs pixel-level shift operations in four directions up, down, left and right for each group, and finally reorganizes the shifted feature map through a channel stacking unit. This process can effectively expand the receptive field, enhance the model's ability to capture landslide boundaries and multi-scale context information, and improve the expression richness of the features. Then, the enhanced feature map is input into the improved joint attention module. The module performs global average pooling on the feature map through the channel attention submodule, and then passes through two 1x1 convolution layers and a Sigmoid activation function to generate a channel attention weight map; at the same time, the channel average value map and the maximum value map are calculated through the spatial attention submodule, and then the two maps are spliced and passed through a 7x7 convolution and a Sigmoid activation function to generate a spatial attention weight map. The channel attention weight map and the spatial attention weight map are fused and applied to the feature map for weighted processing, further improving the accuracy of feature selection and the expression ability of important features. Finally, the weighted feature map is input into the adaptive up-sampling module. The module gradually restores the resolution of the feature map to the target resolution through the nearest neighbor interpolation unit, aligns the feature map through the 3x3 depth convolution alignment unit to ensure the consistency of the features, and then adjusts the channel number of the feature map through the 1x1 pointwise convolution unit to adapt to the output requirements of the subsequent detection task. Finally, the processed feature map is nonlinearly transformed through an activation function and mapped to the target class space to generate the initial landslide detection result.

[0091] In step S50, the initial landslide detection result is edge optimized and coordinate calibrated, and the target landslide region detection result is output.

[0092] It should be noted that the initial landslide detection result refers to the binary classification probability map output by the improved lightweight convolutional neural network after processing the target landslide image. Each pixel point in the probability map corresponds to a value in the interval of 0 to 1, which is used to represent the possibility of the pixel point belonging to the landslide area. The edge optimization refers to the image post-processing operation carried out to solve the problems of fuzzy edge, small holes or burrs in the initial landslide detection result. This operation smoothes and corrects the edge contour of the landslide area through morphological operation and other methods. The coordinate calibration refers to the operation of matching and correcting the pixel coordinates of the initial landslide detection result with the associated original geographic coordinate information. This operation is based on the geographic coordinate data stored in the standardized segmented image to eliminate the coordinate offset generated in the image segmentation and model detection process. The target landslide area detection result refers to the final landslide detection data obtained after edge optimization and coordinate calibration. This result not only contains clear and regular landslide area edge contour, but also contains accurate geographic coordinate information, area parameters and other parameters.

[0093] It can be understood that first, the initial landslide detection result is subjected to edge optimization processing. The boundary information of the landslide area is extracted through an edge detection algorithm (such as Canny edge detection), and the noise points and discontinuous parts on the boundary are removed to make the boundary of the landslide area smoother and more continuous. Then, according to the geographic coordinate information contained in the initial landslide detection result (which has been stored when the image is segmented), the detected landslide area is subjected to coordinate calibration. By converting the pixel coordinates in the detection result into actual geographic coordinates, it is ensured that the position information of the landslide area is consistent with the actual geographic space. Finally, the landslide area detection result after edge optimization and coordinate calibration is integrated and output to form the final target landslide area detection result.

[0094] The present embodiment effectively improves the recognition accuracy of landslide images through dynamic segmentation and double-layer screening. The improved lightweight convolutional neural network realizes efficient detection in a low-resource environment, further improving the accuracy of the landslide detection result.

[0095] Based on the first embodiment of the present application, the same or similar contents as the above embodiment one can be referred to the above introduction, and will not be repeated hereinafter. On this basis, please refer to Figure 2 , the satellite landslide image intelligent detection method based on lightweight steps S40 further includes steps S201-S207:

[0096] Step S201, normalizing and size calibrating the target landslide image to obtain a standardized input image.

[0097] Specifically, first, the target landslide image is normalized to adjust the pixel values of the image to a unified numerical range, usually [0, 1], to eliminate the problem of inconsistent pixel value ranges between different images due to factors such as lighting, sensor differences, etc., thereby improving the adaptability and processing efficiency of the model to the input image. Then, the normalized image is calibrated in size, which is cropped or padded to a predetermined standard size to ensure that all input images have consistent resolution so that subsequent processing modules can be uniformly processed, further improving the stability and accuracy of the detection model. Finally, the image after normalization and size calibration is used as a standardized input image for subsequent landslide detection processes.

[0098] In step S202, the standardized input image is input into the dynamic depth convolution module, which processes the image through a 3x3 depth convolution layer of MobileNetV3Small, batch normalization and activation function layer, and then through a selectable residual connection sub-module according to the matching relationship of input channel number, output channel number and step length to obtain an initial feature map.

[0099] It should be noted that the dynamic depth convolution module is the core feature extraction module of the improved lightweight convolutional neural network, which is constructed based on the depth separable convolution structure of the MobileNetV3Small backbone network. This module splits the standard convolution into depth convolution and point-wise convolution, which greatly reduces the parameter size and computational complexity while efficiently extracting local features of the image. MobileNetV3Small is a lightweight convolutional neural network architecture designed for resource-constrained edge devices, which integrates depth separable convolution, attention mechanism and new activation function technology.

[0100] It can be understood that first, the standardized input image is sent into the dynamic depth convolution module. In this module, the input image is convolved by using the 3x3 depth convolution layer in the MobileNetV3Small architecture to extract the local features of the image. The depth separable convolution separates the depth and spatial dimensions of the standard convolution, significantly reducing the computational complexity while maintaining the effectiveness of feature extraction. Then, the normalized feature map is processed by the batch normalization layer to stabilize the training process and speed up the convergence, while reducing the internal covariate shift. Then, the normalized feature map is subjected to a nonlinear transformation by the activation function layer (such as ReLU or HardSwish) to enhance the expression ability of the features. Subsequently, the feature map enters the optional residual connection sub-module. This sub-module determines whether to apply residual connection according to the matching relationship of input channel number, output channel number and step length. If the input channel number and the output channel number are the same and the step length is 1, the input feature map is directly added to the current feature map to form a residual connection, which helps to alleviate the gradient vanishing problem in deep networks and enhance the transmission of features. If the input channel number and the output channel number are different or the step length is not 1, a projection mapping composed of a 1x1 convolution layer and a batch normalization layer connected in turn is used as a skip connection. Finally, the initial feature map is obtained after the above processing, which not only contains rich local feature information, but also retains part of the original features of the input image through residual connection, laying a foundation for subsequent feature enhancement and landslide detection tasks.

[0101] In step S203, the initial feature map is sequentially passed through the encoder and the decoder of the cross-level skip connection module to obtain the corresponding shallow encoder feature map and deep decoder feature map.

[0102] It should be noted that the cross-level skip connection module is a core module for feature cross-layer fusion in the improved lightweight convolutional neural network, which consists of an encoder and a decoder. This module breaks the traditional pattern of one-way propagation of features in convolutional neural networks, achieving bidirectional association of shallow features and deep features, and avoiding the loss of shallow detail features in network propagation.

[0103] It can be understood that first, the initial feature map is input into the encoder part of the cross-level skip connection module. In the encoder, the spatial resolution of the feature map is gradually reduced and the number of channels is gradually increased through a series of downsampling operations (such as convolution layers combined with operations with a step of 2), so as to extract multi-scale semantic features. These semantic features can capture more abstract landslide feature information in the image, but some detail information may be lost. The output of the encoder is the shallow encoder feature map. Then, the shallow encoder feature map is transmitted to the decoder part. The decoder gradually recovers the spatial resolution of the feature map and reduces the number of channels through a series of upsampling operations (such as nearest neighbor interpolation), so as to reconstruct the detail information of the image. In the process of upsampling, the decoder fuses the shallow feature map from the encoder, and directly transmits the low-level features in the encoder to the corresponding level of the decoder through the skip connection, so as to retain more detail information and enhance the hierarchical expression ability of the features. The output of the decoder is the deep decoder feature map.

[0104] In step S204, the shallow encoder feature map and the deep decoder feature map are cross-layer associated to obtain a cross-layer associated feature map.

[0105] Specifically, first, the shallow encoder feature map and the deep decoder feature map are obtained from the cross-level skip connection module. The shallow encoder feature map contains rich detail information, and the deep decoder feature map carries more abstract semantic information. Then, the two types of feature maps are fused through the cross-layer association operation. Specifically, the shallow encoder feature map and the deep decoder feature map are spliced in the channel dimension, and then they are feature fused through convolution operation to obtain the cross-layer associated feature map to realize the complementation of information.

[0106] In step S205, the cross-layer associated feature map is input into the multi-directional convolution shift enhancement module, and a multi-directional enhanced feature map is obtained through channel grouping, four-direction pixel shift and 1x1 convolution reorganization.

[0107] It should be noted that the channel grouping refers to the operation of dividing the feature channels of the cross-layer associated feature map into a plurality of channel subsets, and each subset contains an equal number of feature channels. The four-direction pixel shift refers to performing pixel shift operations in the up, down, left and right four directions on each grouped feature channel subset, and the shift distance in each direction is a preset pixel value. The 1x1 convolution reorganization refers to the operation of using a 1x1 size convolution kernel to convolve each channel subset after the four-direction pixel shift, and then splicing and fusing the operation results.

[0108] Specifically, the cross-layer associated feature map is first input into the multi-directional convolution shift enhancement module. In this module, the cross-layer associated feature map is subjected to a channel grouping operation, and the feature map is uniformly divided into several groups in the channel dimension, and each group will be independently processed subsequently. Then, a four-direction pixel shift operation is performed on the feature map of each group, i.e., a pixel-level shift in four directions, up, down, left and right, is performed on the feature map of each group. This shift operation can simulate a multi-directional receptive field, enhance the model's ability to capture landslide boundaries and multi-scale context information. Finally, the shifted feature map is reorganized through a 1x1 convolution to fuse the feature information in different directions and generate a multi-directional enhanced feature map. This process not only expands the receptive field, but also enhances the expression ability of the feature map, providing more rich feature information for subsequent detection tasks.

[0109] In step S206, the multi-directional enhanced feature map is input into the channel attention submodule and the spatial attention submodule of the improved joint attention module for element-by-element multiplication to obtain an attention weighted feature map.

[0110] It should be noted that the channel attention submodule is used to assign weights to the feature map from the channel dimension, and this submodule assigns higher weights to feature channels that are strongly related to landslide regions by calculating the importance scores of each feature channel. The spatial attention submodule is used to assign weights to the feature map from the spatial dimension, and this submodule strengthens the spatial position response of the landslide region by calculating the importance scores of each pixel point in the feature map.

[0111] Specifically, the multi-directional enhanced feature map is first input into the channel attention submodule of the improved joint attention module. In this submodule, a global average pooling operation is performed on the multi-directional enhanced feature map to compress the feature map of each channel into a scalar to obtain global channel information. Subsequently, these scalars are subjected to nonlinear transformation through two 1x1 convolution layers, and ReLU activation function is used in the middle, and finally Sigmoid activation function is used to generate a channel attention weight map. Each element of the weight map corresponds to a channel of the input feature map, indicating the importance of the channel. Then, the multi-directional enhanced feature map is input into the spatial attention submodule. In this submodule, the average value and the maximum value of the feature map are calculated along the channel dimension to obtain two spatial feature maps. These two feature maps are spliced along the channel dimension, and then feature fusion is performed through a 7x7 convolution layer, and finally a spatial attention weight map is generated through a Sigmoid activation function. Each element of the weight map corresponds to a spatial position of the input feature map, indicating the importance of the position. Finally, the channel attention weight map and the spatial attention weight map are multiplied element by element to obtain a comprehensive attention weight map. This comprehensive attention weight map is multiplied element by element with the multi-directional enhanced feature map to obtain an attention weighted feature map.

[0112] Step S207, the attention weighted feature map is input into an adaptive up-sampling module to obtain an initial landslide detection result.

[0113] It should be noted that the adaptive up-sampling module is a functional module for restoring the resolution of the feature map in the improved lightweight convolutional neural network, and includes a nearest neighbor interpolation unit, a 3x3 deep convolution alignment unit, and a 1x1 pointwise convolution unit.

[0114] Further, step S207 includes: inputting the attention weighted feature map into the nearest neighbor interpolation unit to obtain an up-sampled feature map; inputting the up-sampled feature map into the 3x3 deep convolution alignment unit to independently perform a 3x3 deep convolution operation on each channel to obtain a feature map after deep convolution alignment; inputting the feature map after deep convolution alignment into the 1x1 pointwise convolution unit to adjust the number of channels to a preset output dimension through a 1x1 convolution kernel to obtain an adjusted feature map; sequentially passing the adjusted feature map through a 3x3 standard convolution layer, a batch normalization layer, and a ReLU activation function for feature smoothing and nonlinear enhancement to obtain an enhanced feature map; inputting the enhanced feature map into a single-channel 1x1 convolution layer to map to a class prediction space to obtain a mapping result; applying a Sigmoid activation function to the mapping result to output a binary classification probability map; when the probability value corresponding to the binary classification probability map is greater than a preset threshold, determining the region corresponding to the probability value as a landslide region; and taking the binary classification probability map corresponding to the region as the initial landslide detection result. It should be noted that the core principle of the nearest neighbor interpolation unit is to select the nearest known pixel value in the feature map to the target pixel point as the filling value to realize the enlargement of the feature map size. The core operation of the 3x3 deep convolution alignment unit is to independently perform a 3x3 convolution operation on each channel of the feature map. The 1x1 pointwise convolution unit is to perform a convolution operation on the input feature map using a 1x1 convolution kernel to realize the dimensionality of the channel without changing the spatial size of the feature map. The single-channel 1x1 convolution layer is a functional layer for mapping a high-dimensional feature map to a class prediction space, which compresses a multi-channel feature map into a single-channel feature map through a 1x1 convolution kernel to realize the conversion from a feature space to a prediction space. The binary classification probability map is a single-channel feature map obtained after processing by the Sigmoid activation function, and the value of each pixel point is a probability value in the interval of 0 to 1. The closer the value is to 1, the higher the possibility that the position belongs to the landslide region, and the closer the value is to 0, the higher the possibility that the position belongs to the non-landslide region.

[0115] It can be understood that firstly, the attention weighted feature map is input into the nearest neighbor interpolation unit, the feature map is up-sampled by the nearest neighbor interpolation method, the spatial resolution of the feature map is improved to a preset higher resolution, and an up-sampled feature map is obtained. Then, the up-sampled feature map is input into the 3x3 depth convolution alignment unit, and a 3x3 depth convolution operation is independently performed on each channel. The depth separable convolution effectively reduces the computational complexity by separating the depth and spatial dimensions of the standard convolution, and at the same time, the feature map is aligned to ensure the consistency of the feature expression of the feature map in the spatial dimension, and a feature map aligned by depth convolution is obtained. Subsequently, the feature map aligned by depth convolution is input into the 1x1 point-by-point convolution unit, and the number of channels of the feature map is adjusted to a preset output dimension by a 1x1 convolution kernel. The 1x1 convolution can efficiently adjust the number of channels while reducing the amount of calculation, and an adjusted feature map is obtained. Then, the adjusted feature map is sequentially subjected to a 3x3 standard convolution layer, a batch normalization layer and a ReLU activation function. The standard convolution layer is used to further extract local features of the feature map, the batch normalization layer is used to stabilize the training process and speed up the convergence, and the ReLU activation function introduces nonlinearity to the feature map, enhances the expression ability of the feature, and finally obtains an enhanced feature map. Then, the enhanced feature map is input into a single-channel 1x1 convolution layer, and the feature map is mapped to a class prediction space by a 1x1 convolution operation to obtain a mapping result. Subsequently, the Sigmoid activation function is applied to the mapping result to compress the value of the mapping result to the interval [0, 1], and a binary classification probability map is output. The Sigmoid activation function can convert the value of each pixel point to the probability value of landslide or non-landslide. Finally, when the probability value of a certain pixel point in the binary classification probability map is greater than a preset threshold, the region corresponding to the pixel point is determined as a landslide region. The binary classification probability map determined as the landslide region is used as the initial landslide detection result for subsequent edge optimization and coordinate calibration processing.

[0116] In this embodiment, first, the target landslide image is normalized and size calibrated to obtain a standardized input image, and then input into the dynamic depth convolution module to extract an initial feature map. The deep and shallow layer features are obtained through the cross-layer jump connection module and cross-layer association, and the features are strengthened through the multi-directional convolution shift enhancement module and the improved joint attention module. Finally, the initial landslide detection result is output through the adaptive up-sampling module, which significantly improves the calculation efficiency while ensuring the detection accuracy, and is suitable for real-time landslide monitoring scenarios.

[0117] Based on the first embodiment of the present application, in the third embodiment of the present application, the same or similar contents as the above-mentioned first embodiment can be referred to the above introduction, and the subsequent will not be described in detail. On this basis, please refer to Figure 3 , the satellite landslide image intelligent detection method based on light weight further includes steps S301-S308:

[0118] Step S301, obtain the binary classification probability map corresponding to the initial landslide detection result.

[0119] Specifically, first, the binary classification probability map corresponding to the initial landslide detection result output from the adaptive upsampling module is extracted. This probability map is obtained by inputting the enhanced feature map into a single-channel 1x1 convolution layer to map to the class prediction space, and then processing it through a Sigmoid activation function. The Sigmoid function compresses the value of each pixel point to the interval [0, 1], representing the probability of the pixel point belonging to the landslide area. The closer the probability value is to 1, the higher the likelihood that the pixel point is in the landslide area; the closer the probability value is to 0, the higher the likelihood that the pixel point is not in the landslide area. In this way, the binary classification probability map intuitively reflects the landslide risk level of each pixel point in the image, providing a quantitative basis for subsequent landslide area determination and analysis.

[0120] Step S302, perform morphological closing operation on the binary classification probability map to obtain the smoothed probability map.

[0121] It should be noted that the morphological closing operation is a post-processing operation on images based on mathematical morphology, which is composed of dilation and erosion operations in sequence. The operation first performs dilation on the image to fill the small holes and gaps in the landslide area of the binary classification probability map, and then performs erosion to shrink the edge burrs of the landslide area, finally achieving the smoothing of the image.

[0122] Specifically, first, a suitable morphological structure element is selected, which is usually a circular or square kernel used to define the range of neighborhood operation. Then, dilation and erosion operations are performed on the binary classification probability map in sequence. The dilation operation expands the high-probability area (pixels close to 1) in the probability map to the surrounding low-probability area, filling small holes and disconnected parts, and enhancing connectivity. The erosion operation shrinks the high-probability area to remove small noise and burrs, making the boundary smoother. Through the closing operation of dilation followed by erosion, small holes and noise in the probability map can be effectively eliminated, and the boundary can be smoothed, obtaining the smoothed probability map.

[0123] Step S303, threshold segmentation is performed on the smoothed probability map, and the area in the smoothed probability map with a probability value greater than a preset decision threshold is marked as a candidate landslide area.

[0124] It should be noted that the preset decision threshold refers to a probability threshold value preset for distinguishing between landslide areas and non-landslide areas, which is determined based on statistical analysis and model verification of a large amount of landslide sample data. The candidate landslide area refers to the area in the smoothed probability map with a pixel value greater than the preset decision threshold, which is a suspected landslide area preliminarily selected after threshold segmentation.

[0125] Specifically, a preset determination threshold is first set, which is determined in advance according to the specific requirements of the landslide detection task and the data characteristics, and is used to distinguish the landslide area and the non-landslide area. Then, a threshold segmentation operation is performed on each pixel of the smoothed probability map. Specifically, each pixel in the smoothed probability map is traversed, and it is checked whether the probability value of the pixel is greater than the preset determination threshold. If the probability value of a certain pixel is greater than the threshold, the pixel is marked as part of the candidate landslide area; otherwise, if the probability value is less than or equal to the threshold, the pixel is marked as a non-landslide area. In this way, the smoothed probability map is divided into a candidate landslide area and a non-landslide area, thereby preliminarily determining the possible landslide position.

[0126] In step S304, the edge contour of the candidate landslide area is extracted to obtain an initial landslide edge.

[0127] It should be noted that the edge contour refers to the boundary line between the target region and the background region in the image, which reflects the geometric shape and spatial distribution characteristics of the target region. In the landslide detection scenario, the edge contour corresponds to the boundary line of the candidate landslide area, which can accurately depict the peripheral shape of the landslide area.

[0128] Specifically, first, edge detection processing is performed on the image marked as the candidate landslide area. Typically, a classic edge detection algorithm such as the Canny edge detection algorithm is used to scan each pixel in the candidate landslide area and calculate the gradient amplitude and direction of each pixel. By setting a gradient threshold, pixel points with large gradient amplitudes are selected, which are usually located at the boundaries of the region. Then, these boundary pixel points are connected to form a closed edge contour. This process can clearly outline the boundary shape of the candidate landslide area to obtain the initial landslide edge.

[0129] In step S305, the original geographic coordinate information associated with the standardized segmented image is called, and the initial landslide edge is calibrated to obtain a calibrated landslide edge.

[0130] Specifically, first, the original geographic coordinate information associated with the standardized segmented image is called from the storage system. These coordinate information has been bound to each segmented image during the image segmentation stage, recording the position of each image segment in the actual geographic space. Then, the initial landslide edge is matched and fused with these original geographic coordinate information. By converting the pixel coordinates of the initial landslide edge to corresponding geographic coordinates, the accurate expression of the landslide edge in geographic space is ensured. This process calibrates the initial landslide edge using geographic coordinate information to correct the deviations that may be introduced during the image processing process, thereby obtaining the calibrated landslide edge.

[0131] Step S306, the calibrated landslide edge is terrain-adapted and adjusted in combination with the slope direction information in the terrain auxiliary data, to obtain an accurate landslide edge.

[0132] Specifically, first, the slope direction information is extracted from the terrain auxiliary data, which reflects the inclination direction of the terrain surface and has important reference value for the adjustment of the landslide edge. Then, the calibrated landslide edge is compared and analyzed with the slope direction information. According to the slope direction information, the trend of the landslide edge is adapted and adjusted to ensure that the direction of the landslide edge is consistent with the inclination direction of the terrain. For example, if the slope direction information shows that the terrain in a certain area is inclined in the northeast direction, and the trend of the landslide edge deviates from it, the landslide edge is fine-tuned to match the slope direction information. This process further optimizes the accuracy of the landslide edge through terrain adaptation and adjustment, making it more consistent with the actual terrain characteristics. The final accurate landslide edge is not only accurate in geographical coordinates, but also highly consistent with the actual landslide phenomenon in terms of terrain characteristics, providing more accurate data support for landslide monitoring and disaster assessment.

[0133] Step S307, the boundary coordinates, center point coordinates and area of the target landslide region are calculated according to the accurate landslide edge, and landslide region description information is generated.

[0134] It should be noted that the boundary coordinates refer to the real geographical coordinate set of a series of discrete points on the accurate landslide edge, which can fully represent the peripheral spatial range of the landslide region. The center point coordinates refer to the geographical coordinates corresponding to the geometric center calculated based on the boundary coordinates of the landslide region, which are usually determined by averaging the boundary coordinates or using the barycentric algorithm. The area refers to the actual land area of the landslide region calculated based on the boundary coordinates of the accurate landslide edge through geographical space area calculation algorithm. The landslide region description information includes geographical coordinates and range size.

[0135] Specifically, first, the boundary coordinates of the target landslide region are determined based on the accurate landslide edge through geometric calculation method. The boundary coordinates are achieved by extracting the extreme points (such as the northernmost, southernmost, easternmost and westernmost points) of the landslide edge, which define the geographical range of the landslide region. Then, the center point coordinates of the landslide region are calculated, usually using the average value or geometric center of the boundary coordinate points. In addition, the range size of the landslide region is obtained by calculating the area enclosed by the landslide edge. This process uses geographical coordinate information and the geometric characteristics of the accurate landslide edge to generate landslide region description information. The landslide region description information not only includes the geographical coordinates of the landslide region, but also covers its range size, providing detailed and accurate data support for subsequent landslide monitoring, risk assessment and geographic information system analysis.

[0136] In step S308, the landslide area description information is packaged in a preset data format, and a target landslide area detection result is output.

[0137] It should be noted that the preset data format refers to a pre-defined structured data specification for standardizing the storage and transmission of landslide area description information, which usually includes uniform requirements such as field name, data type, field length, etc.

[0138] Specifically, first, the generated landslide area description information is sorted and packaged according to the requirements of the preset data format. The preset data format usually includes specific fields, such as the boundary coordinates, center point coordinates, and area of the landslide area, as well as possible additional information such as the risk level of the landslide area and the detection time. These information is organized into a structured data object according to the specified format, such as JSON format, XML format, or custom binary format, to ensure the standardization and consistency of the data. Then, the packaged landslide area description information is output as the target landslide area detection result, which can be directly used in subsequent landslide monitoring systems, geographic information system platforms, or disaster warning systems, providing accurate and timely landslide information to decision-makers to take appropriate measures.

[0139] The embodiment obtains a binary classification probability map of the initial detection result, smoothes it through morphological closing operation, threshold segmentation to obtain a candidate landslide area, extracts the initial edge, and then calibrates it in combination with the original geographic coordinates, and adjusts it to a precise landslide edge according to the terrain slope information. According to this, the boundary, center point coordinates and area are calculated, the description information is generated and packaged according to the preset format, and the target detection result is output, effectively improving the accuracy and reliability of landslide detection. Combined with the calibration and adaptation of geographic coordinates and terrain information, it ensures that the detection result is consistent with the actual terrain.

[0140] Based on the first embodiment of the present application, the present application also provides a lightweight satellite landslide image intelligent detection device, please refer to Figure 4 , the device comprises:

[0141] The acquisition module 10 is configured to acquire optical satellite images and corresponding terrain auxiliary data of an entire scene area, wherein the entire scene area includes a landslide area and a non-landslide area.

[0142] The processing module 20 is configured to perform differential segmentation and standardized preprocessing on the optical satellite images based on the terrain auxiliary data using a dynamic segmentation strategy, to obtain a standardized segmented image.

[0143] The screening module 30 is configured to perform double-layer screening on the standardized segmented image through a double-layer machine learning model to obtain a target landslide image, wherein the double-layer machine learning model is constructed based on a support vector machine model and a random forest classifier.

[0144] The result module 40 is configured to input the target landslide image into the improved lightweight convolutional neural network to obtain an initial landslide detection result, wherein the improved lightweight convolutional neural network comprises a dynamic depth convolution module, a cross-level skip connection module, a multi-directional convolution shift enhancement module, an improved joint attention module, and an adaptive up-sampling module.

[0145] The optimization module 50 is configured to perform edge optimization and coordinate calibration on the initial landslide detection result to output a target landslide region detection result.

[0146] The satellite landslide image intelligent detection device based on lightweight provided in the application adopts the satellite landslide image intelligent detection method based on lightweight in the above embodiment, and can solve the technical problem of how to realize efficient and accurate detection of satellite landslide images on edge devices with limited computing resources. Compared with the prior art, the satellite landslide image intelligent detection device based on lightweight provided in the application has the same beneficial effects as the satellite landslide image intelligent detection method based on lightweight provided in the above embodiment, and other technical features in the satellite landslide image intelligent detection device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0147] The application provides a satellite landslide image intelligent detection device based on lightweight, which comprises at least one processor and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the satellite landslide image intelligent detection method based on lightweight in the above embodiment one.

[0148] Reference will now be made to Figure 5 , which shows a structural schematic diagram of a satellite landslide image intelligent detection device based on lightweight suitable for implementing the embodiments of the application. The satellite landslide image intelligent detection device based on lightweight in the embodiments of the application can include but is not limited to mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistant), PADs (Portable Application Description), PMPs (Portable Media Player), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. Figure 5 The satellite landslide image intelligent detection device based on lightweight shown is only an example, and should not impose any limitation on the functions and use range of the embodiments of the application.

[0149] As shown in Figure 5 The lightweight-based satellite landslide image intelligent detection device can include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. In the RAM 1004, various programs and data required for the lightweight-based satellite landslide image intelligent detection device to operate are also stored. The processing device 1001, the ROM 1002, and the RAM 1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; the storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the lightweight-based satellite landslide image intelligent detection device to communicate wirelessly or by wire with other devices to exchange data. Although the figure shows the lightweight-based satellite landslide image intelligent detection device with various components, it should be understood that not all of the shown components are required to be implemented or possessed. More or less can be alternatively implemented or possessed.

[0150] In particular, according to embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network through the communication device, or installed from the storage device 1003, or installed from the ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the methods of the embodiments of the present disclosure are performed.

[0151] The satellite landslide image intelligent detection device based on light weight provided in the application adopts the satellite landslide image intelligent detection method based on light weight in the above embodiment, and can solve the technical problem of how to realize efficient and accurate detection of satellite landslide images on edge devices with limited computing resources. Compared with the prior art, the satellite landslide image intelligent detection device based on light weight provided in the application has the same beneficial effects as the satellite landslide image intelligent detection method based on light weight provided in the above embodiment, and other technical features in the satellite landslide image intelligent detection device based on light weight are the same as the features disclosed in the previous embodiment method, which will not be repeated here.

[0152] It should be understood that parts of the present application can be realized by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0153] The above is merely specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0154] The present application provides a computer readable medium having computer readable program instructions (i.e. computer programs) stored thereon, the computer readable program instructions being used to execute the satellite landslide image intelligent detection method based on light weight in the above embodiment.

[0155] The computer readable medium provided in the present application may, for example, be a U disk, but is not limited to an electric, magnetic, optical, electromagnetic, infrared, or semiconductor, or device, or any combination of the above. More specific examples of the computer readable medium can include, but are not limited to, an electric connection with one or more conductive wires, a portable computer disk, a hard disk, a random access memory (RAM), a read only memory (ROM), an erasable programmable read only memory (EPROM or flash memory), an optical fiber, a portable compact disk read only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present embodiment, the computer readable medium can be any tangible medium containing or storing a program that can be executed by or used in conjunction with an instruction execution device. The program code contained on the computer readable medium can be transmitted in any suitable medium, including but not limited to an electric wire, an optical cable, an RF (Radio Frequency), and the like, or any suitable combination of the above.

[0156] The above computer readable medium can be contained in the lightweight satellite landslide image intelligent detection device, or can exist separately without being assembled into the lightweight satellite landslide image intelligent detection device.

[0157] The above computer readable medium carries one or more programs, which, when executed by the lightweight satellite landslide image intelligent detection device, enable the lightweight satellite landslide image intelligent detection device to be written in one or more programming languages or combinations thereof for executing computer program codes for performing the operations of the present application. The programming languages include object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" language or similar programming languages. The program code can be executed entirely on a user computer, partially on a user computer, as a separate software package, partially on a user computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, through the Internet using an Internet service provider).

[0158] The flowcharts and block diagrams in the drawings illustrate the architecture, functionality, and operation of possible implementations of various embodiments of the methods and computer program products according to the present application. In this regard, all block s in a flowchart or block diagram represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the block can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that all or some of the blocks, and combinations of blocks, can be implemented by special purpose hardware-based systems, which perform the specified functions or operations, in addition to, or instead of, a software-based system.

[0159] The modules involved in the embodiments of the present application can be implemented in the form of software or in the form of hardware. In some cases, the name of the module does not constitute a limitation on the unit itself.

[0160] The readable medium provided by the present application is a computer readable medium, which stores computer readable program instructions (i.e., computer programs) for executing the above-mentioned lightweight-based satellite landslide image intelligent detection method, and can solve the technical problem of how to realize efficient and accurate detection of satellite landslide images on edge devices with limited computing resources. Compared with the prior art, the computer readable medium provided by the present application has the same beneficial effects as the lightweight-based satellite landslide image intelligent detection method provided by the above-mentioned embodiments, and will not be described here.

[0161] The present application also provides a computer program product comprising a computer program, which, when executed by a processor, implements the steps of the lightweight-based satellite landslide image intelligent detection method as described above.

[0162] The computer program product provided by the present application can solve the technical problem of how to realize efficient and accurate detection of satellite landslide images on edge devices with limited computing resources. Compared with the prior art, the computer program product provided by the present application has the same beneficial effects as the lightweight-based satellite landslide image intelligent detection method provided by the above-mentioned embodiments, and will not be described here.

[0163] The above only describes some embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structural transformation made by using the content of the present application specification and drawings, or direct / indirect application in other related technical fields is included in the patent protection scope of the present application.

Claims

1. A lightweight intelligent detection method for satellite landslide images, characterized in that, The method includes: Acquire optical satellite imagery and corresponding terrain-aided data of the entire landscape area, wherein the entire landscape area includes landslide areas and non-landslide areas; Based on the terrain-aided data, a dynamic segmentation strategy is used to perform differential segmentation and standardized preprocessing on the optical satellite imagery to obtain a standardized segmented image. The standardized segmented image is filtered in two layers by a two-layer machine learning model to obtain the target landslide image. The two-layer machine learning model is constructed based on a support vector machine model and a random forest classifier. The target landslide image is input into an improved lightweight convolutional neural network for processing to obtain an initial landslide detection result. The improved lightweight convolutional neural network includes a dynamic depth convolution module, a cross-level skip connection module, a multi-directional convolution shift enhancement module, an improved joint attention module, and an adaptive upsampling module. The initial landslide detection results are then optimized for edges and calibrated for coordinates to output the detection results for the target landslide area. The step of performing differential segmentation and standardization preprocessing on the optical satellite imagery based on the terrain-aided data using a dynamic segmentation strategy to obtain a standardized segmented image includes: The slope data and terrain roughness data in the terrain auxiliary data are preprocessed to obtain normalized slope values ​​and normalized terrain roughness values. When the slope normalization value is greater than a preset first threshold and the terrain roughness normalization value is greater than a preset second threshold, the area corresponding to the slope normalization value and the terrain roughness is determined to be a high-risk landslide area. When either the slope normalization value is greater than a preset first threshold or the terrain roughness normalization value is greater than a preset second threshold, the area corresponding to the slope normalization value and the terrain roughness is determined as a medium-risk landslide area. When neither the slope normalization value nor the terrain roughness normalization value is greater than a preset first threshold is met, the area corresponding to the slope normalization value and the terrain roughness is determined to be a low-risk landslide area. The portion of the high-risk landslide area in the optical satellite image is segmented according to a preset first overlapping pixel value to obtain a first segmented image; The portion corresponding to the medium-risk landslide area in the optical satellite image is segmented into pixels according to a preset second overlapping pixel value to obtain a second segmented image; The portion corresponding to the low-risk area in the optical satellite image is segmented by pixels according to a preset third overlapping pixel value to obtain a third segmented image, wherein the preset first overlapping pixel value is greater than the preset second overlapping pixel value, and the preset second overlapping pixel value is greater than the preset third overlapping pixel value. The first segmented image, the second segmented image, and the third segmented image are numbered according to a preset rule, and the original geographic coordinate information corresponding to each image is stored to obtain a standardized segmented image with associated original geographic coordinates.

2. The intelligent detection method for lightweight satellite landslide imagery as described in claim 1, characterized in that, The improved lightweight convolutional neural network includes the steps of a dynamic deep convolution module, a cross-layer skip connection module, a multi-directional convolution shift enhancement module, an improved joint attention module, and an adaptive upsampling module, and further includes: The dynamic deep convolution module adopts the depth-separable convolution structure in the MobileNetV3Small backbone network, including a 3×3 deep convolutional layer, a batch normalization layer, and an activation function layer connected in sequence. The dynamic deep convolution module also includes a selective residual connection submodule. The cross-layer skip connection module includes an encoder and a decoder. The multi-directional convolution shift enhancement module includes a channel grouping unit, a four-way shifting unit, and a channel stacking unit. The improved joint attention module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule consists of a global average pooling layer, two 1×1 convolutional layers, and a sigmoid activation layer. The spatial attention submodule consists of a channel average map generation unit, a channel maximum map generation unit, a channel splicing unit, a 7×7 convolutional layer, and a sigmoid activation layer. The adaptive upsampling module includes a nearest neighbor interpolation unit, a 3×3 deep convolutional alignment unit, and a 1×1 pointwise convolutional unit.

3. The intelligent detection method for lightweight satellite landslide imagery as described in claim 1, characterized in that, The step of performing a two-layer screening of the standardized segmented image using a two-layer machine learning model to obtain the target landslide image includes: The color distribution of the three color channels of the standardized segmented image is statistically analyzed, and the color histogram of each channel is calculated using a preset quantization interval to obtain the three-channel histogram. The three-channel histograms are concatenated along the channel dimension to generate a color histogram feature vector. The standardized segmented image is converted into a grayscale image and the texture feature vector is obtained using the gray-level co-occurrence matrix algorithm. The color histogram feature vector and the texture feature vector are standardized by L2 normalization to obtain standardized color feature vector and standardized texture feature vector, respectively. The standardized color feature vector and the standardized texture feature vector are concatenated along the feature dimension to obtain the fused feature vector; The fused feature vector is input into a random forest classifier of a two-layer machine learning model for processing and filtering to obtain a first filtered image, wherein the random forest classifier contains a preset number of decision trees and the first filtered image includes landslide features. The fusion feature vector corresponding to the first screened image is extracted and input into the support vector machine model in the two-layer machine learning model for processing to obtain the target landslide image, wherein the support vector machine model adopts the radial basis kernel function.

4. The intelligent detection method for lightweight satellite landslide imagery as described in claim 1, characterized in that, The step of inputting the target landslide image into an improved lightweight convolutional neural network for processing to obtain an initial landslide detection result, wherein the improved lightweight convolutional neural network includes a dynamic depthwise convolution module, a cross-level skip connection module, a multi-directional convolutional shift enhancement module, an improved joint attention module, and an adaptive upsampling module, includes: The target landslide image is normalized and sized to obtain a standardized input image; The standardized input image is input into the dynamic depth convolution module, and then processed through a 3×3 depth convolution layer, batch normalization and activation function layer of MobileNetV3Small. The initial feature map is obtained by processing the image through the selective residual connection submodule according to the matching relationship between the number of input channels, the number of output channels and the stride. The initial feature map is sequentially passed through the encoder and decoder of the cross-level skip connection module to obtain the corresponding shallow encoder feature map and deep decoder feature map. The shallow encoder feature map and the deep decoder feature map are correlated across layers to obtain a cross-layer correlated feature map. The cross-layer associated feature map is input into the multi-directional convolutional shift enhancement module, and after channel grouping, four-way pixel shifting and 1×1 convolutional recombination, a multi-directional enhanced feature map is obtained. The multi-directional enhanced feature map is input into the channel attention submodule and spatial attention submodule of the improved joint attention module and multiplied element-wise to obtain the attention-weighted feature map. The attention-weighted feature map is input into the adaptive upsampling module to obtain the initial landslide detection result.

5. The intelligent detection method for lightweight satellite landslide imagery as described in claim 4, characterized in that, The step of inputting the attention-weighted feature map into the adaptive upsampling module to obtain the initial landslide detection result includes: The attention-weighted feature map is input into the nearest neighbor interpolation unit to obtain the upsampled feature map; The upsampled feature map is input into a 3×3 depth convolution alignment unit to perform a 3×3 depth convolution operation independently on each channel, resulting in a feature map after depth convolution alignment. The feature map aligned by depth convolution is input into a 1×1 pointwise convolution unit, and the number of channels is adjusted to the preset output dimension through a 1×1 convolution kernel to obtain the adjusted feature map. The adjusted feature map is sequentially passed through a 3×3 standard convolutional layer, a batch normalization layer, and a ReLU activation function for feature smoothing and non-linear enhancement to obtain the enhanced feature map. The enhanced feature map is input into a single-channel 1×1 convolutional layer and mapped to the class prediction space to obtain the mapping result; Apply the Sigmoid activation function to the mapping result to output a binary classification probability map; When the probability value corresponding to the binary classification probability map is greater than a preset threshold, the area corresponding to the probability value is determined to be a landslide area. The binary probability map corresponding to the landslide area is used as the initial landslide detection result.

6. The intelligent detection method for lightweight satellite landslide imagery as described in claim 1, characterized in that, The step of performing edge optimization and coordinate calibration on the initial landslide detection results and outputting the target landslide area detection results includes: Obtain the binary classification probability map corresponding to the initial landslide detection results; The binary classification probability map is processed using morphological closing operations to obtain a smoothed probability map; The smoothed probability map is segmented by a threshold, and the regions in the smoothed probability map with probability values ​​greater than a preset judgment threshold are marked as candidate landslide regions; Extract the edge contour of the candidate landslide area to obtain the initial landslide edge; The original geographic coordinate information associated with the standardized segmented image is called and calibrated in combination with the initial landslide edge to obtain the calibrated landslide edge; By combining the slope aspect information in the terrain auxiliary data, the calibrated landslide edge is adjusted for terrain adaptation to obtain a precise landslide edge. Based on the precise landslide edge, calculate the boundary coordinates, center point coordinates, and area of ​​the target landslide region, and generate landslide region description information, wherein the landslide region description information includes geographic coordinates and range dimensions; The landslide area description information is encapsulated according to a preset data format, and the target landslide area detection result is output.

7. A lightweight intelligent detection device based on satellite landslide imagery, characterized in that, The device includes: The acquisition module is used to acquire optical satellite images of the entire scene area and corresponding terrain auxiliary data, wherein the entire scene area includes landslide areas and non-landslide areas; The processing module is used to perform differential segmentation and standardized preprocessing on the optical satellite imagery based on the terrain-aided data using a dynamic segmentation strategy, to obtain a standardized segmented image; it is also used to preprocess the slope data and terrain roughness data in the terrain-aided data to obtain normalized slope values ​​and normalized terrain roughness values; when the normalized slope value is greater than a preset first threshold and the normalized terrain roughness value is greater than a preset second threshold, the area corresponding to the normalized slope value and terrain roughness is identified as a high-risk landslide area; when either the normalized slope value is greater than the preset first threshold or the normalized terrain roughness value is greater than the preset second threshold, the area corresponding to the normalized slope value and terrain roughness is identified as a medium-risk landslide area; when neither the normalized slope value is greater than the preset first threshold nor the normalized terrain roughness value is greater than the preset second threshold, the area is identified as a medium-risk landslide area; when neither the normalized slope value is greater than the preset first threshold nor the normalized terrain roughness value is greater than the preset second threshold, the area is identified as a high-risk landslide area. In this process, the areas corresponding to the slope normalization value and terrain roughness are identified as low-risk landslide areas; the portions corresponding to high-risk landslide areas in the optical satellite image are segmented by pixels according to a preset first overlap pixel value to obtain a first segmented image; the portions corresponding to medium-risk landslide areas in the optical satellite image are segmented by pixels according to a preset second overlap pixel value to obtain a second segmented image; the portions corresponding to low-risk areas in the optical satellite image are segmented by pixels according to a preset third overlap pixel value to obtain a third segmented image, wherein the preset first overlap pixel value is greater than the preset second overlap pixel value, and the preset second overlap pixel value is greater than the preset third overlap pixel value; the first segmented image, the second segmented image, and the third segmented image are numbered according to preset rules, and the original geographic coordinate information corresponding to each image is stored to obtain a standardized segmented image associated with the original geographic coordinates; The filtering module is used to perform two-layer filtering on the standardized segmented image through a two-layer machine learning model to obtain the target landslide image, wherein the two-layer machine learning model is constructed based on a support vector machine model and a random forest classifier. The result module is used to input the target landslide image into an improved lightweight convolutional neural network for processing to obtain an initial landslide detection result. The improved lightweight convolutional neural network includes a dynamic depth convolution module, a cross-level skip connection module, a multi-directional convolution shift enhancement module, an improved joint attention module, and an adaptive upsampling module. The optimization module is used to perform edge optimization and coordinate calibration on the initial landslide detection results and output the detection results of the target landslide area.

8. A lightweight intelligent detection device based on satellite landslide imagery, characterized in that, The device includes: a memory, a processor, and a lightweight satellite landslide image-based intelligent detection program stored in the memory and running on the processor, the lightweight satellite landslide image-based intelligent detection program being configured to implement the steps of the lightweight satellite landslide image-based intelligent detection method as described in any one of claims 1-6.

9. A storage medium, characterized in that, The storage medium stores a lightweight satellite landslide image intelligent detection program, which, when executed by a processor, implements the steps of the lightweight satellite landslide image intelligent detection method as described in any one of claims 1-6.

Images