A method for rapid landslide area segmentation based on high-resolution satellite remote sensing data and deep learning algorithms

Abstract

This invention provides a rapid landslide area segmentation method based on high-resolution satellite remote sensing data and deep learning algorithms, belonging to the field of ground scene technology. The invention downsamples the original satellite remote sensing image to obtain a first image, extracts reflectance data from the red and near-infrared channels, calculates the reflectance difference and sum of reflectance values, and normalizes them to obtain a vegetation index to construct a vegetation layer. The first image and the vegetation layer are superimposed and stitched together, and feature enhancement is performed to obtain a vegetation feature map. This map is input into a convolutional neural network to output a landslide probability map. A two-dimensional coordinate set is constructed and cropped to obtain high-resolution test patches. The spatial gradient magnitude of the local vegetation index is calculated and weighted by a preset enhancement coefficient to generate edge-enhanced patches. The edge-enhanced patches are input into a semantic segmentation network to output a local landslide map, which is then backfilled into the initial image to obtain a landslide area segmentation map. This invention constructs a cascaded segmentation system to achieve efficient and accurate segmentation of landslide areas.

Description

Technical Field

[0001] This invention relates to the field of ground scene technology, specifically to a method for rapid segmentation of landslide areas based on high-resolution satellite remote sensing data and deep learning algorithms. Background Technology

[0002] As a highly destructive geological hazard, landslides necessitate the rapid and accurate acquisition of affected area information within the critical post-disaster rescue time, which is a core foundation for emergency rescue command. With the rapid development of high-resolution satellite remote sensing technology, massive amounts of remote sensing data have effectively overcome the shortcomings of traditional manual geological surveys, such as low efficiency and high risk. Against this backdrop, computer vision technology is deeply integrated with geological disaster monitoring, and the industry is rapidly transitioning from traditional object-oriented or manual feature extraction to end-to-end fully automated deep learning models.

[0003] In existing technologies, various deep convolutional neural networks have been widely used for landslide information extraction. A typical technique involves directly inputting images into a segmentation model, introducing complex structures such as hollow spatial pyramid pooling and attention mechanisms to extract high-level semantic features. To address the difficulty of annotation, some techniques also introduce weakly supervised frameworks based on category heatmaps and conditional random fields to approximate edge optimization. These techniques primarily rely on continuously deepening the network layers or stacking modules at a single resolution, allowing the network to implicitly learn the texture and spectral features of the landslide to complete the segmentation.

[0004] However, the aforementioned technologies have significant limitations in practical emergency response. The models' overemphasis on improving absolute accuracy leads to bloated architectures, easily creating computational bottlenecks when processing large-scale, high-resolution imagery, resulting in slow inference speeds and failing to meet the extremely high timeliness requirements of post-disaster detection. Existing networks are detached from the optical physics of remote sensing, failing to effectively utilize the significant physical prior of instantaneous destruction of surface vegetation, leading to degraded feature extraction capabilities in complex backgrounds. Furthermore, the lack of decoupling and edge-targeting enhancement for multi-scale features makes them prone to boundary fragmentation and false positives / missed detections when processing irregular edges. In summary, existing technologies fail to resolve the contradiction between fine feature extraction and efficient computational inference, necessitating a rapid segmentation scheme that deeply integrates multispectral physical priors with multi-scale network architectures.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method for rapid segmentation of landslide areas based on high-resolution satellite remote sensing data and deep learning algorithms, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for rapid landslide area segmentation based on high-resolution satellite remote sensing data and deep learning algorithms, comprising the following steps:

[0009] Step 1: Obtain the original satellite remote sensing image of the region to be segmented, downsample the original satellite remote sensing image to obtain the first image, and extract the reflectance data of each pixel in the red light channel and near-infrared channel from the first image;

[0010] Step 2: Obtain the vegetation index of each pixel based on the red light reflectance and near-infrared reflectance of each pixel in the first image. Based on the vegetation index and according to the position of the corresponding pixel, construct a vegetation layer and perform channel-level stitching of the first image and the vegetation layer to form a vegetation feature map.

[0011] Step 3: Input the vegetation feature map into a pre-trained convolutional neural network to generate a landslide probability map, extract the coordinates of pixels in the landslide probability map whose values ​​are greater than a preset probability threshold, construct a two-dimensional coordinate set, map the two-dimensional coordinate set to the corresponding position in the original satellite remote sensing image according to the downsampling ratio, and crop out a high-resolution image patch to be tested around the corresponding position.

[0012] Step 4: Determine the local vegetation index based on the reflectance data of each pixel in the red light channel and near-infrared channel within the high-definition test patch, and obtain the spatial gradient magnitude of the local vegetation index. Combine the spatial gradient magnitude with the preset enhancement coefficient to obtain the boundary weight. Use the enhancement weight to enhance the corresponding pixel to obtain the edge-enhanced patch.

[0013] Step 5: Input the edge enhancement patch into the pre-trained semantic segmentation network, extract the spatial features of the edge enhancement patch, determine the landslide state of each pixel of the edge enhancement patch based on the spatial features, summarize the landslide state of all pixels, and obtain the local landslide map of the edge enhancement patch.

[0014] Step 6: Construct an initial map with the same spatial dimensions as the original satellite remote sensing image. Based on the original satellite remote sensing image mapping coordinates corresponding to each high-resolution landslide patch during cropping, backfill all local landslide maps into the initial map. Then, stitch together adjacent local landslide maps to obtain a landslide area segmentation map.

[0015] Further, acquire the original satellite remote sensing image of the region to be segmented;

[0016] The mapping coordinates are determined according to the preset downsampling ratio. In each spectral channel of the original satellite remote sensing image, the four nearest neighboring pixels to the mapping coordinates are extracted respectively, and the spatial distance between the mapping coordinates and the four neighboring pixels in the horizontal and vertical directions is calculated.

[0017] Distance weights are generated based on spatial distance. The pixel values ​​of four adjacent pixels are then weighted and averaged using these distance weights to obtain the first image.

[0018] For the first image, the reflectance data corresponding to the red light channel and near-infrared channel are extracted pixel by pixel.

[0019] Furthermore, the difference between the reflectance data of each pixel in the red light channel and the near-infrared channel is calculated as the reflectance difference value, and the sum of the reflectance data of each pixel in the red light channel and the near-infrared channel is calculated as the total reflectance value.

[0020] The vegetation difference value of each pixel is obtained by calculating the ratio of the reflectance difference value to the total reflectance value.

[0021] The vegetation difference values ​​of all pixels in the first image are traversed and normalized to obtain the vegetation index of each pixel.

[0022] The vegetation index of each pixel is repositioned according to the positions of all pixels in the first image to obtain a vegetation layer.

[0023] By using the pixel values ​​of each pixel in the vegetation layer, feature enhancement is performed on the pixel features of each spectral channel at the corresponding position in the first image to obtain a vegetation feature map.

[0024] Furthermore, the vegetation feature map is input into a pre-trained convolutional neural network to obtain a landslide probability map;

[0025] Traverse the pixels in the landslide probability map, mark the pixels with probability values ​​exceeding the preset probability threshold as potential landslide points, record the coordinates of the potential landslide points, and generate a two-dimensional coordinate set.

[0026] Using the reciprocal of the downsampling ratio as a scaling factor, the coordinate values ​​in the two-dimensional coordinate set are mapped to the original satellite remote sensing image to obtain the original two-dimensional coordinates in the original satellite remote sensing image.

[0027] Centered on the original two-dimensional coordinates, the corresponding local image is cropped according to the preset size to obtain a high-definition image block to be tested.

[0028] Furthermore, the pre-trained convolutional neural network is obtained by training and converging using a sample set containing historical downsampled first images, corresponding constructed vegetation layers, and real landslide annotation data, by minimizing the classification cross-entropy loss between the predicted probability and the real landslide annotation data.

[0029] Furthermore, for each high-definition image block to be tested, the reflectance data of each pixel in the red light channel and near-infrared channel are extracted, the vegetation difference value of each pixel is calculated, and normalization is performed to obtain the local vegetation index of the corresponding pixel location.

[0030] The spatial rate of change of local vegetation index in the horizontal and vertical directions is calculated by using the discrete central difference operator, and vector magnitude is synthesized to obtain the spatial gradient magnitude of each pixel.

[0031] Multiply the spatial gradient magnitude by the preset enhancement coefficient to obtain the boundary weight;

[0032] By using boundary weights, the original pixel features of the high-definition test patch are weighted pixel by pixel to obtain the edge-enhanced patch.

[0033] Furthermore, the local vegetation indices within the current high-definition image patch to be tested are traversed, and the local maximum and local minimum values ​​are extracted respectively. The local vegetation range between the local maximum and local minimum values ​​is then calculated.

[0034] The regional mean of the spatial gradient magnitude is statistically analyzed to extract the local average gradient value of the high-resolution image patch to be tested;

[0035] The enhancement coefficient is obtained by evaluating the relative proportion between the local vegetation range and the local average gradient value.

[0036] Furthermore, each edge enhancement patch is input into a pre-trained semantic segmentation network, and convolution is performed on the edge enhancement patches to obtain global semantic information. The global semantic information is then upsampled to obtain the spatial features of the edge enhancement patches.

[0037] Convolve the spatial features to obtain the classification score of each pixel. Normalize the classification score to obtain the probability value of each pixel. Compare the probability value with the preset classification threshold to obtain the landslide state of each pixel.

[0038] Based on the coordinates of each pixel in the edge enhancement patch, the landslide state is spatially arranged to obtain a local landslide map.

[0039] Furthermore, an all-zero matrix with the same spatial dimensions as the original satellite remote sensing image is constructed as the initialization map;

[0040] Based on the original satellite remote sensing image mapping coordinates corresponding to each high-definition map block during cropping, the corresponding local landslide map is backfilled in situ into the initial map;

[0041] During the process of backfilling the corresponding local landslide map into the initial map in situ, the overlapping area of ​​adjacent local landslide maps is extracted;

[0042] Extract multiple landslide states from pixels in different local landslide images. If at least one of the multiple landslide states is identified as a landslide, then the pixel is identified as a landslide pixel.

[0043] If multiple landslide states do not have a landslide category, the pixel will be classified as a background pixel.

[0044] The determined pixels are updated to their corresponding positions in the initial image, and the updated initial image is used as the landslide area segmentation image.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] This invention uses the original satellite remote sensing image of the region to be segmented to downsample and obtain a first image, extracting its reflectance data. It calculates the reflectance difference and sum of reflectance values, performs normalization mapping to obtain a vegetation index, and constructs a vegetation layer. The first image and the vegetation layer are then stitched together to obtain a vegetation feature map, which is input into a pre-trained convolutional neural network to output a landslide probability map. Pixel coordinates greater than a preset probability threshold are extracted to construct a two-dimensional coordinate set, which is then mapped onto the original satellite remote sensing image to crop out high-resolution patches for testing. This invention integrates the underlying physical mechanisms of remote sensing with multi-scale spatial structures, highlighting the physical characteristics of vegetation damage during landslides through the vegetation layer. This guides the model to generate a strong targeted response in the early stages of feature extraction, overcoming spectral confusion of ground features in complex backgrounds. Simultaneously, downsampling features are used for coarse screening of the entire map region, only high-resolution cropping is performed on potential areas that meet the probability threshold. This completely avoids redundant calculations on massive non-disaster background areas, significantly reduces memory usage, breaks through the computational bottleneck when processing large-scale high-resolution data, and significantly improves the overall inference speed, meeting the timeliness requirements of rapid post-disaster response.

[0047] This invention also obtains the local vegetation index of each high-resolution test patch and calculates its spatial gradient magnitude. The spatial gradient magnitude, combined with a preset enhancement coefficient, is used as a weight to perform a weighted operation on the original pixels of the high-resolution test patch, generating an edge-enhanced patch. Subsequently, the edge-enhanced patch is input into a pre-trained semantic segmentation network to output a local landslide map. Finally, based on the location information of the two-dimensional coordinate set, all local landslide maps are backfilled into an initial map of the same size as the original satellite remote sensing image, ultimately outputting a landslide area segmentation map. This design abandons the conventional approach of relying on time-consuming post-processing algorithms to repair image edges, innovatively prioritizing edge optimization. By using spatial gradient magnitude for weighting, irregular boundaries such as landslide mudflows are sharpened at the bottom layer before features enter the semantic segmentation network. This weighting mechanism significantly reduces the feature convergence burden of the semantic segmentation network, resulting in a more continuous, smoother, and more realistic landslide area segmentation map boundary. While ensuring extremely high computational efficiency, it achieves refined extraction and accurate segmentation of features in large-scale landslide areas. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall method flow of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0050] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0051] Example:

[0052] Please see Figure 1 The present invention provides a technical solution:

[0053] A method for rapid landslide area segmentation based on high-resolution satellite remote sensing data and deep learning algorithms, comprising the following steps:

[0054] Step 1: Obtain the original satellite remote sensing image of the region to be segmented, downsample the original satellite remote sensing image to obtain the first image, and extract the reflectance data of each pixel in the red light channel and near-infrared channel from the first image.

[0055] In this embodiment, the original satellite remote sensing image of the region to be segmented is obtained;

[0056] The mapping coordinates are determined according to the preset downsampling ratio. In each spectral channel of the original satellite remote sensing image, the four nearest neighboring pixels to the mapping coordinates are extracted respectively, and the spatial distance between the mapping coordinates and the four neighboring pixels in the horizontal and vertical directions is calculated.

[0057] Distance weights are generated based on spatial distance. The pixel values ​​of four adjacent pixels are then weighted and averaged using these distance weights to obtain the first image.

[0058] For the first image, the reflectance data corresponding to the red light channel and near-infrared channel are extracted pixel by pixel.

[0059] Downsampling of raw satellite remote sensing images addresses the computational bottleneck caused by the massive data volume of high-resolution imagery. High-resolution satellite remote sensing data contains rich spatial details; directly inputting raw satellite remote sensing images into deep learning networks results in extremely high memory consumption and computational overhead. Downsampling compression significantly reduces the amount of feature data while better preserving the spatial geometric relative positions of ground features. A foundational data layer for macroscopic coarse screening is constructed, enabling subsequent networks to quickly locate potential disaster areas across the entire map with minimal computational cost, avoiding ineffective convolutional calculations over massive normal background areas.

[0060] In remote sensing optical physics, healthy vegetation exhibits a strong absorption band in the red light band due to chlorophyll absorption, and a very strong reflection peak in the near-infrared band due to the scattering effect of the internal cell structure of the leaves. In contrast, the reflectance characteristics of exposed landslide soil or rock in these two bands are drastically different from those of vegetation. The extracted reflectance data from the red and near-infrared channels refer to the average spectral value of a pixel within its corresponding spectral band. True physical reflectance manifests as a continuous spectral curve within a fixed band, accurately capturing the sensitive spectral characteristics of abrupt changes in vegetation cover. This alleviates the problem of spectral confusion of ground features against complex surface backgrounds and improves the model's sensitivity in distinguishing between landslide and non-landslide areas.

[0061] Extracting reflectivity data eliminates external environmental interference and preserves characteristic data. Raw pixel values ​​acquired by satellite sensors are easily affected by factors such as the solar altitude angle, atmospheric scattering and absorption conditions, and terrain shadows at the time of capture, leading to drastic fluctuations in values ​​for the same ground feature. Reflectivity data characterizes the true reflectivity of a ground surface to electromagnetic waves, eliminating the influence of external environment and lighting. Extracting reflectivity data ensures that the subsequently calculated vegetation index has strict numerical stability, improving the model's generalization ability when faced with different regions, weather conditions, and satellite data sources.

[0062] Step 2: Obtain the vegetation index of each pixel based on the red light reflectance and near-infrared reflectance of each pixel in the first image. Construct a vegetation layer based on the vegetation index and the position of the corresponding pixel. Then, perform channel-level stitching of the first image and the vegetation layer to form a vegetation feature map.

[0063] In this embodiment, the difference between the reflectance data of each pixel in the red light channel and the near-infrared channel is calculated as the reflectance difference value, and the sum of the reflectance data of each pixel in the red light channel and the near-infrared channel is calculated as the total reflectance value.

[0064] The vegetation difference value of each pixel is obtained by calculating the ratio of the reflectance difference value to the total reflectance value.

[0065] The vegetation difference values ​​of all pixels in the first image are traversed and normalized to obtain the vegetation index of each pixel.

[0066] The vegetation index of each pixel is repositioned according to the positions of all pixels in the first image to obtain a vegetation layer.

[0067] By using the pixel values ​​of each pixel in the vegetation layer, feature enhancement is performed on the pixel features of each spectral channel at the corresponding position in the first image to obtain a vegetation feature map.

[0068] The ratio of reflectance difference to total reflectance is calculated to eliminate noise from topographical undulations, uneven illumination, and sensor calibration errors. Since the absolute reflectance in the red and near-infrared bands scales proportionally with external lighting conditions, a simple difference cannot provide a consistent evaluation standard across different lighting areas. By introducing the total reflectance as the denominator in the ratio calculation, the results are normalized to a fixed interval, highlighting the unique spectral characteristics of vegetation. This makes the vegetation difference value more resistant to interference from topographical shadows and changes in solar altitude angle. This improves the stability of vegetation cover extraction and ensures the reliability of the data source under different environments.

[0069] Isolated vegetation difference values ​​represent only the spectral attributes of a single pixel and lack spatial context. By rigorously arranging vegetation according to the spatial physical coordinates of the first image, the spatial geometric topology of the features in the real physical world is preserved.

[0070] The relative difference levels of each pixel are stretched and mapped, converting the extreme value-scaled floating-point numerical values ​​into data representations conforming to standard image channel formats. This stretching and mapping process enhances the visual contrast and data variance of the image, making the boundary gradient between the landslide-damaged vegetation area and the normal background steeper and more significant. It also reduces the convergence difficulty of deep learning networks in shallow feature extraction and highlights high-frequency edge signals.

[0071] The pixel values ​​in the vegetation layer serve as spatial weight coefficients for physical priors. Utilizing these values ​​to target and enhance all spectral channels of the first image forces the deep learning network to break free from the limitation of treating all regions equally. This guides the model to generate a strong targeted response early in feature extraction, focusing on areas of vegetation cover anomalies while significantly suppressing useless signals from large areas of non-hazardous background. This not only avoids redundant computation caused by massive amounts of invalid background but also overcomes the problem of spectral obfuscation of ground features against complex backgrounds.

[0072] Step 3: Input the vegetation feature map into a pre-trained convolutional neural network to generate a landslide probability map. Extract the coordinates of pixels in the landslide probability map whose values ​​are greater than a preset probability threshold, construct a two-dimensional coordinate set, map the two-dimensional coordinate set to the corresponding position in the original satellite remote sensing image according to the downsampling ratio, and crop out a high-resolution image patch to be tested around the corresponding position.

[0073] In this embodiment, the vegetation feature map is input into a pre-trained convolutional neural network to obtain a landslide probability map;

[0074] Traverse the pixels in the landslide probability map, mark the pixels with probability values ​​exceeding the preset probability threshold as potential landslide points, record the coordinates of the potential landslide points, and generate a two-dimensional coordinate set.

[0075] Using the reciprocal of the downsampling ratio as a scaling factor, the coordinate values ​​in the two-dimensional coordinate set are mapped to the original satellite remote sensing image to obtain the original two-dimensional coordinates in the original satellite remote sensing image.

[0076] Centered on the original two-dimensional coordinates, the corresponding local image is cropped according to the preset size to obtain a high-definition image block to be tested.

[0077] The vegetation feature map, as input, not only contains multispectral optical reflectance information but also directly incorporates characteristics of vegetation damage. The network extracts deep features through multiple convolutional kernels and maps the multidimensional feature tensor to floating-point values ​​in the [0,1] interval using an activation function in the output layer. This value quantifies the statistical confidence that each pixel belongs to a landslide terrain, generating a landslide probability map. This transforms abstract multidimensional image features into concrete, quantifiable predictive indicators, providing a reliable mathematical basis for subsequent spatial location selection.

[0078] Potential landslide points are marked for rapid coarse screening, quickly eliminating massive amounts of absolutely safe background areas. A preset probability threshold is the key decision boundary determining the rigor of this coarse screening. This probability threshold is typically set based on a comprehensive analysis of statistical evaluation results from a large historical sample set during the validation phase, as well as the actual operational needs of disaster emergency response. Threshold filtering can instantly narrow the focus from the complex full-map data to highly suspicious areas of anomaly, reducing the data throughput of fine-tuning.

[0079] Feature extraction and probability prediction are performed on the downsampled first image, where the recorded two-dimensional coordinate set consists only of logical coordinates on a low-resolution feature map. To accurately obtain the high-precision edge morphology of the landslide, refined analysis must be performed on the high-resolution original image containing extremely high spatial details. By using the reciprocal of the downsampling ratio as a scaling factor for inverse coordinate mapping, the coordinates of suspected points in the first image can be accurately restored to their true geographic or physical spatial locations in the original high-resolution satellite remote sensing image.

[0080] In this embodiment, the pre-trained convolutional neural network is obtained by training and converging using a sample set containing historical downsampled first images, corresponding constructed vegetation layers, and real landslide annotation data, by minimizing the classification cross-entropy loss between the predicted probability and the real landslide annotation data.

[0081] Real landslide annotation data refers to manually annotated pixel-level labels that completely correspond to the input image in terms of spatial resolution. Pixels in the landslide area are assigned positive sample labels, while pixels in the background area where no landslide has occurred are assigned negative sample labels.

[0082] A sample set of real landslide labeled data is used for model training to ensure that the feature distribution extracted by the convolutional neural network during the training phase remains consistent with that in the actual application inference phase, and to deeply integrate remote sensing physical mechanisms. Using historical downsampled first images enables the network to learn the macroscopic morphology of landslides in a low-resolution feature space; introducing vegetation layers as training input injects optical physical priors into the deep learning model. This allows the network to focus on the essential physical characteristics of abrupt changes in vegetation cover rather than blindly searching for patterns in complex visible light textures during parameter optimization. Real landslide labeled data provides supervisory signals for the mapping between these physical characteristics and actual disasters, guiding the network gradient in the correct descent direction, reducing the risk of overfitting in complex backgrounds, and improving the network's recognition accuracy when facing different geological landforms.

[0083] Minimizing the classification cross-entropy loss drives network training convergence, reflecting the underlying task attributes of probability evaluation and macroscopic coarse screening. The core task of convolutional neural networks is to perform pixel-level binary classification, i.e., determining whether the current region belongs to a potential landslide or a safe background. The classification cross-entropy loss function can accurately measure the difference between the probability distribution predicted by the network and the true label distribution, and penalizes high-confidence false predictions.

[0084] By continuously minimizing this loss value during training, the network can be forced to produce highly discriminative probability values ​​at the output, generating a landslide probability map with strong contrast and clear boundaries. Optimizing the network's classification decision boundary provides a solid and reliable data basis for efficient and accurate filtering of potential landslide points using preset probability thresholds, reducing the probability of missed and false detections in the macro-screening stage.

[0085] Step 4: Determine the local vegetation index based on the reflectance data of each pixel in the red light channel and near-infrared channel within the high-definition test patch, and obtain the spatial gradient magnitude of the local vegetation index. Combine the spatial gradient magnitude with the preset enhancement coefficient to obtain the boundary weight. Use the enhancement weight to enhance the corresponding pixel to obtain the edge-enhanced patch.

[0086] In this embodiment, for each high-definition image block to be tested, the reflectance data of each pixel point in the red light channel and the near-infrared channel are extracted respectively, the vegetation difference value of each pixel point is calculated, and normalization processing is performed to obtain the local vegetation index of the corresponding pixel position.

[0087] The spatial rate of change of local vegetation index in the horizontal and vertical directions is calculated by using the discrete central difference operator, and vector magnitude is synthesized to obtain the spatial gradient magnitude of each pixel.

[0088] Multiply the spatial gradient magnitude by the preset enhancement coefficient to obtain the boundary weight;

[0089] By using boundary weights, the original pixel features of the high-definition test patch are weighted pixel by pixel to obtain the edge-enhanced patch.

[0090] High-resolution local vegetation indices are extracted from corresponding pixel locations and introduced to characterize the vegetation damage caused by landslides. While low-resolution macroscopic vegetation indices can only determine approximate location, high-resolution local vegetation indices can depict the microscopic surface anomalies within and around the landslide body, providing the physical data basis for pixel-level edge sharpening and enhancement.

[0091] The essential manifestation of landslide boundaries in optical remote sensing imagery is the dramatic alternation between normal vegetation and exposed soil, reflected as a step-like abrupt change in vegetation index values. The discrete central difference operator (CDD) is a classic numerical differentiation method that approximates the derivative by calculating the difference between the current pixel and its adjacent pixels. The CDD operator offers better symmetry, effectively smoothing local isolated noise while accurately and sensitively capturing the spatial rate of change of the vegetation index in both horizontal and vertical directions. It explicitly transforms the implicit numerical abrupt changes between pixels into a rate of change index.

[0092] Spatial rate of change is a directional vector in both the horizontal and vertical directions, while landslide boundaries in real terrain are irregular and vary in orientation. To objectively evaluate whether a pixel is on a landslide edge without being influenced by the specific orientation of the boundary, vector magnitude synthesis is required. The synthesized spatial gradient magnitude is a scalar, stripped of its directional attribute, purely and intuitively quantifying the severity of vegetation mutation at the pixel's location. Pixels with larger gradient magnitudes are more likely to be located at the actual physical boundary of a landslide.

[0093] By combining the spatial gradient magnitude with a preset enhancement coefficient to generate boundary weights, a spatial edge attention mask is constructed. Using these weights, multiplicative or additive weighting operations are performed on the original pixels of the high-resolution test patch, which can target and amplify the pixel features of the landslide boundary while relatively suppressing the feature responses of the internal flat or background areas.

[0094] The weighted edge enhancement patches significantly increase the contrast of the landslide boundaries, reducing the feature convergence burden on the subsequent semantic segmentation network. This allows the network to quickly output segmentation patches with smooth, continuous boundaries that closely resemble the real mudslide morphology, perfectly balancing computational efficiency and refined extraction accuracy.

[0095] In this embodiment, the local vegetation index within the current high-definition image patch to be tested is traversed, the local maximum and local minimum values ​​are extracted respectively, and the local vegetation range between the local maximum and local minimum values ​​is calculated.

[0096] The regional mean of the spatial gradient magnitude is statistically analyzed to extract the local average gradient value of the high-resolution image patch to be tested;

[0097] The enhancement coefficient is obtained by evaluating the relative proportion between the local vegetation range and the local average gradient value.

[0098] Calculating local vegetation range quantifies the absolute contrast of vegetation cover within a current high-resolution map tile. Local maximum values ​​typically represent undisturbed, healthy forest within the tile, while local minimum values ​​represent complete exposure of soil or rock due to landslides. Local vegetation range reflects the most severe span of physical and ecological environmental damage within a local area. Establishing a macroscopic contrast benchmark can intuitively characterize the severity of landslide hazards and potential feature stretching space within the current map tile, providing reliable basic data support for adaptive weight adjustment.

[0099] Regional mean statistics are performed on spatial gradient magnitudes to assess the overall boundary complexity and high-frequency texture density of local patches. The gradient magnitude of a single pixel is affected by isolated noise points or minor surface debris in the remote sensing image. The local average gradient value, through smoothed statistical analysis in the spatial domain, reflects the overall basis strength of the edge signal within the patch. This filters out the accidental influence of local high-frequency noise, providing a stable and representative evaluation standard for regional texture features.

[0100] High-resolution image tiles with different geographical backgrounds and varying degrees of damage exhibit significant inherent differences in contrast and noise levels. Using a fixed enhancement factor can easily lead to over-sharpening of originally clear areas, resulting in false edges, or insufficient enhancement of blurred landslide boundaries.

[0101] The local vegetation range represents the contrast potential of the current image patch, while the average gradient value represents the existing edge strength. By calculating the relative ratio between the two, a customized enhancement coefficient can be tailored for each high-resolution image patch. Regions with high contrast but blurred boundaries receive a strong boost, while regions with already extremely dense textures are moderately suppressed. This improves the environmental adaptability of edge enhancement in the face of varied and complex terrain, ensuring that the feature maps fed into the semantic segmentation network always maintain optimal boundary resolution.

[0102] Step 5: Input the edge enhancement patch into the pre-trained semantic segmentation network, extract the spatial features of the edge enhancement patch, determine the landslide state of each pixel of the edge enhancement patch based on the spatial features, summarize the landslide states of all pixels, and obtain the local landslide map of the edge enhancement patch.

[0103] In this embodiment, each edge enhancement patch is input into a pre-trained semantic segmentation network, the edge enhancement patches are convolved to obtain global semantic information, and the global semantic information is upsampled to obtain the spatial features of the edge enhancement patches;

[0104] Convolve the spatial features to obtain the classification score of each pixel. Normalize the classification score to obtain the probability value of each pixel. Compare the probability value with the preset classification threshold to obtain the landslide state of each pixel.

[0105] Based on the coordinates of each pixel in the edge enhancement patch, the landslide state is spatially arranged to obtain a local landslide map.

[0106] Upon receiving edge-enhanced patches, the encoder extracts global semantic features from the patches layer by layer using multi-layer convolution and downsampling operations, such as the macroscopic texture and shape of landslides. Subsequently, the decoder performs upsampling to restore spatial resolution, introducing skip connections in this process to fuse low-level features from the encoder that retain rich physical boundary details with high-level semantic features. Finally, the network outputs a feature matrix rich in spatial detail information, i.e., high-dimensional spatial features.

[0107] Each physical pixel within the edge enhancement patch is scored independently, outputting a continuous classification score. The scores, ranging from 0 to 1, are then forcibly normalized to obtain the probability value of each pixel belonging to the landslide category. The probability values ​​of all pixels are then compared one by one with a preset classification threshold, which is set to 0.5.

[0108] Semantic segmentation networks can fully perceive the global contextual semantic relationships between different features within a map tile through downsampling, and restore lost spatial resolution layer by layer through upsampling. Since the input edge-enhanced map tiles have already undergone boundary sharpening by the underlying physical gradient, the semantic segmentation network can easily overcome the feature convergence difficulties of complex boundaries such as muddy water flow under guidance, and then make high-precision binary classification judgments for each physical pixel, outputting a local prediction map with smooth boundaries and a high degree of fit to the real terrain.

[0109] Step 6: Construct an initial map with the same spatial dimensions as the original satellite remote sensing image. Based on the original satellite remote sensing image mapping coordinates corresponding to each high-resolution landslide patch during cropping, backfill all local landslide maps into the initial map. Then, stitch together adjacent local landslide maps to obtain a landslide area segmentation map.

[0110] In this embodiment, an all-zero matrix with the same spatial size as the original satellite remote sensing image is constructed as the initialization map;

[0111] Based on the original satellite remote sensing image mapping coordinates corresponding to each high-definition map block during cropping, the corresponding local landslide map is backfilled in situ into the initial map;

[0112] During the process of backfilling the corresponding local landslide map into the initial map in situ, the overlapping area of ​​adjacent local landslide maps is extracted;

[0113] Extract multiple landslide states from pixels in different local landslide images. If at least one of the multiple landslide states is identified as a landslide, then the pixel is identified as a landslide pixel.

[0114] If multiple landslide states do not have a landslide category, the pixel will be classified as a background pixel.

[0115] The determined pixels are updated to their corresponding positions in the initial image, and the updated initial image is used as the landslide area segmentation image.

[0116] A zero-matrix with the same spatial dimensions as the original satellite remote sensing image was constructed to create a panoramic physical canvas covering the complete macroscopic geographic coordinates of the disaster area. During the downsampling and coarse screening stage, the vast majority of absolutely safe, non-disaster background areas were actively stripped away. The zero values ​​in the zero-matrix represent the initial pixel background color, which in binary classification logic represents non-landslide areas. Areas that previously underwent massive invalid calculations were directly assigned as safe backgrounds by default, perfectly reconstructing the spatial dimensions of the original image while avoiding redundant feature reconstruction calculations for non-disaster areas.

[0117] When cropping and extracting suspected landslide points, in order to preserve sufficient local boundary context, physical overlap inevitably occurs between adjacent high-resolution map patches to be measured. At the overlapping boundaries, obvious splicing gaps or spatial faults are easily generated due to conflicts in the results of different local prediction maps.

[0118] Extracting overlapping areas and performing logical fusion of pixel states can effectively resolve prediction discrepancies at the boundary intersections of adjacent tiles. This smooths prediction differences from local perspectives, eliminates stitching seams, and ensures that large landslides spanning multiple tiles appear as a continuous, complete geological entity with coherent physical boundaries in the final landslide area segmentation map, thus improving the overall visual quality and interpretation accuracy of the final segmentation results.

[0119] Table 1: Accuracy evaluation results of different segmentation methods

[0120]

[0121] As shown in Table 1, landslide detection in high-resolution remote sensing images is a typical task with extremely imbalanced sample distribution, making conventional detection methods highly susceptible to interference when faced with massive and complex backgrounds. Regarding conventional methods, while existing methods such as Deeplabv3+ have a certain precision, their recall is low due to the difficulty in overcoming the loss of microscopic edge features and interference from complex terrain shadows, indicating that a large number of landslides with blurred or occluded edges are not effectively identified. In terms of recall, the method proposed in this application, through targeted enhancement of vegetation layers and spatial gradient edge sharpening, can more sensitively capture and identify more complex landslides compared to other methods, demonstrating that the model in this application improves the detection rate of landslides. In the field of landslide disaster emergency response and detection, to avoid serious hidden dangers caused by missed reports, more emphasis is often placed on the recall rate of landslide detection rather than just the precision rate. The F1 score, the harmonic value of precision and recall, is the most important parameter for evaluating the overall accuracy of the model, and the method in this application achieves the highest F1 score. Furthermore, the overall precision and mean intersection-union ratio (OCR) also significantly exceed those of other methods. Therefore, the landslide area rapid segmentation method proposed in this application based on high-resolution satellite remote sensing data and deep learning algorithms can significantly improve the accuracy of landslide area identification and segmentation.

[0122] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0123] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0124] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for rapid landslide area segmentation based on high-resolution satellite remote sensing data and deep learning algorithms, characterized in that, The specific steps include: Step 1: Obtain the original satellite remote sensing image of the region to be segmented, downsample the original satellite remote sensing image to obtain the first image, and extract the reflectance data of each pixel in the red light channel and near-infrared channel from the first image; The steps for obtaining the original satellite remote sensing image of the region to be segmented, downsampling the original satellite remote sensing image to obtain a first image, and extracting the reflectance data of each pixel in the red and near-infrared channels from the first image are as follows: Obtain the original satellite remote sensing image of the region to be segmented; The mapping coordinates are determined according to the preset downsampling ratio. In each spectral channel of the original satellite remote sensing image, the four nearest neighboring pixels to the mapping coordinates are extracted respectively, and the spatial distance between the mapping coordinates and the four neighboring pixels in the horizontal and vertical directions is calculated. Distance weights are generated based on spatial distance. The pixel values ​​of four adjacent pixels are then weighted and averaged using these distance weights to obtain the first image. For the first image, the reflectance data corresponding to the red light channel and the near-infrared channel are extracted pixel by pixel; Step 2: Obtain the vegetation index of each pixel based on the red light reflectance and near-infrared reflectance of each pixel in the first image. Based on the vegetation index and according to the position of the corresponding pixel, construct a vegetation layer and perform channel-level stitching of the first image and the vegetation layer to form a vegetation feature map. The vegetation index of each pixel is obtained based on the red light reflectance and near-infrared reflectance of each pixel in the first image. A vegetation layer is constructed based on the vegetation index and the position of the corresponding pixel. The first image and the vegetation layer are then stitched together at the channel level to form a vegetation feature map. The specific steps are as follows: The difference between the reflectance data of each pixel in the red light channel and the near-infrared channel is calculated as the reflectance difference value, and the sum of the reflectance data of each pixel in the red light channel and the near-infrared channel is calculated as the total reflectance value. The vegetation difference value of each pixel is obtained by calculating the ratio of the reflectance difference value to the total reflectance value. The vegetation difference values ​​of all pixels in the first image are traversed and normalized to obtain the vegetation index of each pixel. The vegetation index of each pixel is repositioned according to the positions of all pixels in the first image to obtain a vegetation layer. By using the pixel values ​​of each pixel in the vegetation layer, the pixel features of each spectral channel at the corresponding position in the first image are enhanced to obtain a vegetation feature map. Step 3: Input the vegetation feature map into a pre-trained convolutional neural network to generate a landslide probability map, extract the coordinates of pixels in the landslide probability map whose values ​​are greater than a preset probability threshold, construct a two-dimensional coordinate set, map the two-dimensional coordinate set to the corresponding position in the original satellite remote sensing image according to the downsampling ratio, and crop out a high-resolution image patch to be tested around the corresponding position. Step 4: Determine the local vegetation index based on the reflectance data of each pixel in the red light channel and near-infrared channel within the high-definition test patch, and obtain the spatial gradient magnitude of the local vegetation index. Combine the spatial gradient magnitude with the preset enhancement coefficient to obtain the boundary weight. Use the enhancement weight to enhance the corresponding pixel to obtain the edge-enhanced patch. Step 5: Input the edge enhancement patch into the pre-trained semantic segmentation network, extract the spatial features of the edge enhancement patch, determine the landslide state of each pixel of the edge enhancement patch based on the spatial features, summarize the landslide state of all pixels, and obtain the local landslide map of the edge enhancement patch. Step 6: Construct an initial map with the same spatial dimensions as the original satellite remote sensing image. Based on the original satellite remote sensing image mapping coordinates corresponding to each high-resolution landslide patch during cropping, backfill all local landslide maps into the initial map. Then, stitch together adjacent local landslide maps to obtain a landslide area segmentation map.

2. The method for rapid landslide area segmentation based on high-resolution satellite remote sensing data and deep learning algorithms according to claim 1, characterized in that: The vegetation feature map is input into a pre-trained convolutional neural network to obtain a landslide probability map; Traverse the pixels in the landslide probability map, mark the pixels with probability values ​​exceeding the preset probability threshold as potential landslide points, record the coordinates of the potential landslide points, and generate a two-dimensional coordinate set. Using the reciprocal of the downsampling ratio as a scaling factor, the coordinate values ​​in the two-dimensional coordinate set are mapped to the original satellite remote sensing image to obtain the original two-dimensional coordinates in the original satellite remote sensing image. Centered on the original two-dimensional coordinates, the corresponding local image is cropped according to the preset size to obtain a high-definition image block to be tested.

3. The method for rapid landslide area segmentation based on high-resolution satellite remote sensing data and deep learning algorithms according to claim 2, characterized in that: The pre-trained convolutional neural network is obtained by training and converging using a sample set containing historical downsampled first images, corresponding constructed vegetation layers, and real landslide annotation data, by minimizing the classification cross-entropy loss between the predicted probability and the real landslide annotation data.

4. The method for rapid landslide area segmentation based on high-resolution satellite remote sensing data and deep learning algorithms according to claim 1, characterized in that: For each high-definition image block to be tested, the reflectance data of each pixel in the red light channel and near-infrared channel are extracted, the vegetation difference value of each pixel is calculated, and normalization is performed to obtain the local vegetation index of the corresponding pixel location. The spatial rate of change of local vegetation index in the horizontal and vertical directions is calculated by using the discrete central difference operator, and vector magnitude is synthesized to obtain the spatial gradient magnitude of each pixel. Multiply the spatial gradient magnitude by the preset enhancement coefficient to obtain the boundary weight; By using boundary weights, the original pixel features of the high-definition test patch are weighted pixel by pixel to obtain the edge-enhanced patch.

5. The method for rapid landslide area segmentation based on high-resolution satellite remote sensing data and deep learning algorithms according to claim 4, characterized in that: Traverse the local vegetation indices within the current high-definition image patch to be tested, extract the local maximum and local minimum values ​​respectively, and calculate the local vegetation range between the local maximum and local minimum values; The regional mean of the spatial gradient magnitude is statistically analyzed to extract the local average gradient value of the high-resolution image patch to be tested; The enhancement coefficient is obtained by evaluating the relative proportion between the local vegetation range and the local average gradient value.

6. The method for rapid landslide area segmentation based on high-resolution satellite remote sensing data and deep learning algorithms according to claim 1, characterized in that: Each edge enhancement patch is input into a pre-trained semantic segmentation network. Convolution is performed on the edge enhancement patches to obtain global semantic information. The global semantic information is then upsampled to obtain the spatial features of the edge enhancement patches. Convolve the spatial features to obtain the classification score of each pixel. Normalize the classification score to obtain the probability value of each pixel. Compare the probability value with the preset classification threshold to obtain the landslide state of each pixel. Based on the coordinates of each pixel in the edge enhancement patch, the landslide state is spatially arranged to obtain a local landslide map.

7. The method for rapid landslide area segmentation based on high-resolution satellite remote sensing data and deep learning algorithms according to claim 6, characterized in that: Construct an all-zero matrix with the same spatial dimensions as the original satellite remote sensing image as the initialization map; Based on the original satellite remote sensing image mapping coordinates corresponding to each high-definition map block during cropping, the corresponding local landslide map is backfilled in situ into the initial map; During the process of backfilling the corresponding local landslide map into the initial map in situ, the overlapping area of ​​adjacent local landslide maps is extracted; Extract multiple landslide states from pixels in different local landslide images. If at least one of the multiple landslide states is identified as a landslide, then the pixel is identified as a landslide pixel. If multiple landslide states do not have a landslide category, the pixel will be classified as a background pixel. The determined pixels are updated to their corresponding positions in the initial image, and the updated initial image is used as the landslide area segmentation image.

Images