Improved YOLOv8 active landslide detection method based on InSAR deformation

Abstract

The invention relates to the technical field of image processing, in particular to an improved YOLOv8 active landslide detection method based on InSAR deformation, and the method comprises the steps: obtaining a Sentinel-1A orbit drop SAR image, and calculating a deformation rate; processing the SAR image, and generating an earth surface deformation rate map and accumulated displacement of the target area; marking a landslide mass, cutting a deformation rate diagram, generating a data set after data enhancement, and dividing a training set, a verification set and a test set; a variable kernel convolution module is introduced to dynamically adjust a sampling position and a heavy parameter shuffling one-time aggregation module is introduced to adapt to small target and irregular deformation features; performing spatial distribution analysis on landslide hidden danger points in the research area, and paying attention to a high-risk area with severe deformation; a typical area is selected to carry out field investigation, false positive deformation and real landslide caused by engineering activities are distinguished, the accuracy of a model detection result is verified, and the method can be used for carrying out multi-dimensional verification and classification on a suspected landslide body in a severe deformation area.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an improved YOLOv8 active landslide detection method based on InSAR deformation. Background Technology

[0003] InSAR technology boasts all-day, all-weather advantages, and SBAS-InSAR can extract millimeter-level deformation data, providing data for landslide hazard identification. However, it suffers from poor real-time performance, low automation, low efficiency in processing wide-area data, and the need for manual interpretation, impacting large-scale monitoring. Furthermore, there is insufficient research on deep learning for InSAR deformation rate maps, and existing methods have limited accuracy in detecting small-target landslides. Landslide targets are small-scale and irregular, making it difficult for traditional convolutional neural networks to capture subtle features. Existing models rely on optical data and have poor adaptability to InSAR deformation patterns. In addition, standardized datasets are lacking, geological validation of detection results is insufficient, existing models primarily focus on large landslides, and their ability to identify small or hidden landslides is weak. The YOLO series models also face feature extraction and representation problems when processing small targets.

[0004] While traditional SBAS-InSAR technology can acquire millimeter-level surface deformation data, its complexity in data processing limits its ability to achieve real-time monitoring and automated analysis over a wide area. This results in low efficiency in identifying large-scale landslide hazards and fails to meet the timeliness requirements of disaster early warning. Secondly, existing deep learning methods are mostly designed for optical remote sensing data and lack specific detection models for InSAR deformation rate maps. They are particularly inadequate in extracting subtle deformation features of landslide targets (such as small-scale, irregular shapes), limiting detection accuracy. Finally, landslide datasets often contain a high proportion of small targets (<32×32 pixels). Traditional YOLO models suffer from feature loss due to downsampling and poor adaptability to multi-scale targets. They also exhibit weak generalization ability when applied to complex terrain or different regions, leading to frequent false negatives and missed detections. Furthermore, existing research lacks sufficient integration of InSAR results with high-resolution remote sensing, UAV data, and field exploration, resulting in a lack of a systematic disaster analysis framework. This makes it difficult to accurately distinguish between deformation caused by engineering activities and actual landslides, leading to weak geological interpretation of the detection results.

[0005] Therefore, there is an urgent need to integrate the advantages of InSAR deformation monitoring with the automation capabilities of deep learning to build an accurate, fast, and efficient small landslide target detection model suitable for InSAR data. As a result, the improved YOLOv8 active landslide detection method based on InSAR deformation has emerged. Summary of the Invention

[0006] The purpose of this invention is to provide an improved YOLOv8 active landslide detection method based on InSAR deformation, aiming to solve the problems of insufficient integration of InSAR results with high-resolution remote sensing, UAV data and field exploration, lack of a systematic disaster analysis framework, difficulty in accurately distinguishing between deformation caused by engineering activities and actual landslides, resulting in weak geological interpretation of detection results.

[0007] To achieve the above objectives, this invention provides an improved YOLOv8 active landslide detection method based on InSAR deformation, comprising the following steps;

[0008] Acquire Sentinel-1A descent SAR images, calculate deformation rate, and obtain basic data;

[0009] Based on SBAS-InSAR technology, SAR images are processed through data preprocessing, differential interferometry, orbit refinement, inversion and geocoding to generate surface deformation rate maps and cumulative displacement of the target area.

[0010] The InSAR results were visually interpreted, landslide bodies were labeled, deformation rate maps were cropped, and after data augmentation, a dataset was generated and divided into training, validation and test sets.

[0011] A variable kernel convolution module is introduced to dynamically adjust the sampling position, and a reparameter shuffling and one-time aggregation module is introduced to adapt to small targets and irregular deformation features;

[0012] By combining the identification results with Gaofen-2 remote sensing images and UAV oblique photography data, spatial distribution analysis of landslide hazard points in the study area was conducted, focusing on high-risk areas with severe deformation.

[0013] Field investigations were conducted in typical areas. By comparing high-resolution remote sensing images with UAV aerial survey data, the false positive deformation caused by engineering activities and the real landslides were distinguished, thus verifying the accuracy of the model detection results.

[0014] The input data required for calculating the deformation rate includes SAR image data, precise orbit data, external DEM data, GACOS data, and ground control GCPS points.

[0015] Specifically, the method for processing SAR images based on SBAS-InSAR technology, through data preprocessing, differential interferometry, orbit refinement, inversion, and geocoding, to generate a surface deformation rate map and cumulative displacement of the target area is as follows:

[0016] The study area is cropped, the target range is extracted from the SAR main and secondary images, the spatial baselines of the main and secondary images are estimated, the interferometric quality is assessed, the main and secondary images are selected, and the interferometric pair is constructed.

[0017] The main and secondary images are registered using DEM calibration, the true phase difference is recovered by unwrapping using minimum cost flow, the interferogram is enhanced by Goldstein filtering, the differential interferogram is generated, the connection diagram is edited, the correlation of image pairs is sorted out, and a small baseline set processing framework is built.

[0018] Select GCP points and generate control point data, correct satellite orbit parameters, and then perform terrain re-flattening, estimate and remove constant phase and residuals, remove unwrapped phase slopes, continuously purify phase information, and refine orbits.

[0019] Based on the orbit refinement results, orbital errors are removed, deformation rate, terrain residuals and average strength are estimated, the orbit is refined a second time and the phase is unwrapped, and further processing is performed.

[0020] Optimize the re-flattening process, use GCP calibration to more thoroughly eliminate the influence of topographic phase, remove atmospheric phase, extract key features using singular value decomposition, calculate the surface deformation rate, and output reliable monitoring results;

[0021] Set quality control thresholds for height and rate accuracy, then geocode deformation values, associate them with real geographical locations, generate a time deformation sequence, display the deformation process, and output vector results.

[0022] Specifically, the methods for introducing a variable kernel convolution module to dynamically adjust the sampling position and a reparameter shuffling and one-time aggregation module to adapt to small targets and irregular deformation features are as follows:

[0023] Input image data and enter the Conv module to complete the first round of basic feature extraction;

[0024] The backbone network extracts features sequentially through Conv, VkConv, C2f, and SPPF layers;

[0025] Multi-scale features are output from the Backbone. First, the low-resolution, high-semantic feature map is upscaled through sampling. Then, it is concatenated with the feature map of the same resolution level to merge the channel dimension features.

[0026] The channel dimension features are input into the RVS-OSA module for deep fusion and optimization, enhancing the expression of subtle features;

[0027] The subtle features processed by RVS-OSA are partly sent directly to the backend Detect for target detection, and partly participate in the multi-scale feature loop of the Neck layer, outputting optimized features at different scales to the corresponding detection head, thus helping the model to perform multi-scale target detection.

[0028] The input image data enters the Conv module to complete the first round of basic feature extraction. During the process of initially capturing the low-level spatial features of the image through conventional convolution operations, if there are differences in size or channels in the original image, resize and normalization operations will be performed in advance to ensure that the input format meets the model requirements.

[0029] This invention presents an improved YOLOv8 active landslide detection method based on InSAR deformation. It acquires Sentinel-1A down-orbit SAR imagery, calculates the deformation rate, and obtains basic data. The SAR imagery is processed using SBAS-InSAR technology, including data preprocessing, differential interferometry, orbit refinement, inversion, and geocoding, to generate a surface deformation rate map and cumulative displacement of the target area. The InSAR results are visually interpreted, landslide bodies are labeled, the deformation rate map is cropped, and after data augmentation, a dataset is generated, divided into training, validation, and test sets. A variable kernel convolution module is introduced to dynamically adjust the sampling position, adapting to small targets and irregular deformation features. The recognition results are then compared with... This study combines Gaofen-2 remote sensing imagery and UAV oblique photography data to analyze the spatial distribution of landslide hazard points in the research area, focusing on high-risk areas with severe deformation. Field investigations were conducted in typical areas, and by comparing high-resolution remote sensing imagery and UAV aerial survey data, false positive deformations caused by engineering activities and actual landslides were distinguished, verifying the accuracy of the model's detection results. The method first generates deformation rate maps based on SBAS-InSAR technology to reveal the evolution trend of surface deformation. Then, VRO_YOLO is proposed to achieve efficient identification of active landslides. Finally, combining high-resolution remote sensing imagery, UAV oblique photography, and field exploration, multi-dimensional verification analysis is performed on suspected landslide bodies in areas with severe deformation. The core idea of ​​the SBAS-InSAR technology for generating deformation rate maps is to extract millimeter-level surface deformation rates and cumulative deformation from multi-temporal SAR images through short-baseline interferometry combination, time-series analysis, and singular value decomposition (SVD). The VRO_YOLO described above is a composite network architecture in the YOLO v8 network architecture that employs a variable kernel convolution VkConv module and RVS-OSA dual-path enhancement features to more effectively capture and process targets of different scales, especially densely distributed small targets. This addresses the problems of insufficient integration of InSAR results with high-resolution remote sensing, UAV data, and field exploration; the lack of a systematic disaster analysis framework; the difficulty in accurately distinguishing between deformation caused by engineering activities and actual landslides; and the resulting weak geological interpretation of detection results.

[0030] Compared with existing technologies, the present invention has the following advantages:

[0031] Breaking through the limitations of traditional InSAR technology and single deep learning models, this paper deeply integrates the deformation monitoring capabilities of SBAS-InSAR with the target detection advantages of the improved YOLO model, and constructs a full-chain technical path of "remote sensing deformation data - intelligent feature extraction - active landslide identification". This provides a fusion solution across remote sensing and artificial intelligence fields for geological disaster monitoring, and promotes the technological leap of landslide detection from manual interpretation to automated intelligent analysis.

[0032] By optimizing the InSAR data processing workflow and deep learning model architecture, the project has significantly improved the detection efficiency of large-scale landslide hazards, realizing full automation from data acquisition to hazard identification. It has overcome the problem of insufficient real-time performance caused by high human involvement and long processing cycles in traditional technologies for wide-area monitoring, and provided efficient technical support for regional geological disaster emergency early warning.

[0033] By organically integrating InSAR deformation analysis, high-resolution remote sensing interpretation, UAV aerial surveys, and field exploration data, a disaster analysis framework with multi-source data verification across air, space, and ground has been formed. This framework overcomes the limitations of single technical means in complex terrain or vegetated areas, improves the reliability of landslide identification and the scientific nature of geological interpretation, and lays the foundation for accurate disaster risk assessment.

[0034] It not only achieves efficient detection of active landslides, but also constructs a full-chain application system from disaster identification to hazard analysis and prevention and control decision-making by integrating multi-dimensional factors such as topography, climate conditions and human engineering activities. It provides intelligent solutions for early warning, emergency response and long-term prevention and control of landslide disasters, which is of great significance for improving the scientific and operational level of geological disaster management. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating the overall process of the improved YOLOv8 active landslide detection method based on InSAR deformation provided by this invention.

[0037] Figure 2 This is the algorithm network diagram of the improved YOLOv8 active landslide detection method based on InSAR deformation provided by the present invention.

[0038] Figure 3 This is a result image of the identification of publicly available landslide datasets.

[0039] Figure 4 This is a flowchart of the improved YOLOv8 active landslide detection method based on InSAR deformation provided by the present invention. Detailed Implementation

[0040] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0041] Please see Figures 1 to 4 This invention provides an improved YOLOv8 active landslide detection method based on InSAR deformation, comprising the following steps;

[0042] S1 acquires Sentinel-1A descent SAR images, calculates deformation rate, and obtains basic data;

[0043] In this embodiment of the invention, the input data required for calculating the deformation rate includes SAR image data, precise orbit data, external DEM data, GACOS data, and ground control GCPS points.

[0044] S2 processes SAR images based on SBAS-InSAR technology, and generates surface deformation rate maps and cumulative displacement of the target area through data preprocessing, differential interferometry, orbit refinement, inversion and geocoding.

[0045] In this embodiment of the invention, landslide areas are manually interpreted visually based on InSAR deformation rate maps, and the LabelImg tool is used to label the landslide locations in PASCALVOC format. Data augmentation operations such as rotation and flipping are performed on the original samples to generate a sufficient number of training samples, which are then divided into training, validation, and test sets in a 6:1:3 ratio.

[0046] Specific methods:

[0047] S21 cropping study area, extracting target range from SAR main and sub-images, estimating spatial baselines of main and sub-images, assisting in judging interferometric quality, selecting main and sub-images, and constructing interferometric pairs;

[0048] In this embodiment of the invention, SBAS-InSAR processing begins with data preprocessing. First, the study area is cropped, and the target range is extracted from the main and secondary SAR images. Then, the spatial baselines of the main and secondary images are estimated to help determine the interferometric quality. Finally, the main and secondary images are selected, and an interferometric pair is constructed to prepare the raw data for subsequent processes.

[0049] S22 uses DEM to calibrate and register primary and secondary images, unwraps them using minimum cost flow to recover the true phase difference, enhances the interferogram with Goldstein filtering, generates a differential interferogram, edits the connection diagram, sorts out the correlation of image pairs, and builds a small baseline set processing framework;

[0050] In this embodiment of the invention, the interferometric process is used to construct and optimize interferograms. The primary and secondary images are registered using DEM calibration, the true phase difference is recovered by unwrapping using minimum cost flow, the interferogram is enhanced by Goldstein filtering, a differential interferogram is generated, the connectivity graph is edited, image pair associations are sorted out, and a small baseline set processing framework is established.

[0051] S23 selects GCP points and generates control point data, corrects satellite orbit parameters, and sequentially performs terrain re-flattening, estimates and removes constant phase and residuals, removes unwrapped phase slopes, continuously purifies phase information, and refines orbits.

[0052] In this embodiment of the invention, orbit refinement aims to reduce the impact of orbital errors. First, GCP points are selected and control point data is generated to correct satellite orbital parameters. Then, terrain re-leveling, estimation and removal of constant phase and residuals, and removal of unwrapped phase slopes are performed sequentially to continuously purify phase information and lay the foundation for deformation inversion.

[0053] S24 Based on the orbit refinement results, remove orbit errors, estimate deformation rate, terrain residual and average strength, refine the orbit a second time and unwrap the phase, and perform in-depth processing;

[0054] In this embodiment of the invention, the first inversion initially extracts deformation information. Based on the orbit refinement results, orbital errors are first removed, and then the deformation rate, terrain residuals, and average intensity are estimated. The orbit is then refined a second time and the phase is unwrapped, providing high-quality data for subsequent accurate inversion.

[0055] S25 optimization and re-flattening, with the help of GCP calibration, more thoroughly eliminates the influence of terrain phase, removes atmospheric phase, extracts key features using singular value decomposition, calculates the surface deformation rate, and outputs reliable monitoring results;

[0056] In this embodiment of the invention, the second inversion focuses on accurately calculating the deformation rate. First, the surface is optimized and leveled, and the influence of topographic phase is eliminated more thoroughly by using GCP calibration; then, the atmospheric phase is removed, and key features are extracted using singular value decomposition, ultimately accurately calculating the surface deformation rate and outputting reliable monitoring results.

[0057] S26 sets height and rate accuracy thresholds for quality control, then geocodes deformation values, associates them with real geographical locations, generates a time deformation sequence, displays the deformation process, and outputs vector results.

[0058] In this embodiment of the invention, geocoding transforms deformation data into usable geographic information. First, quality control is achieved by setting accuracy thresholds for height and velocity; then, deformation values ​​are geocoded and associated with real geographic locations; next, a time-varying deformation sequence is generated to visualize the deformation process; finally, vector results are output for GIS analysis and mapping, serving practical applications.

[0059] S3 manual visual interpretation of InSAR results, landslide bodies are labeled, deformation rate maps are cropped, and after data augmentation, a dataset is generated and divided into training, validation and test sets;

[0060] In this embodiment of the invention, an improved YOLO network is constructed that integrates variable kernel convolution (VkConv) and the RVS-OSA module. VkConv adapts to irregular landslide shapes by dynamically adjusting the sampling position of the convolution kernel, while the RVS-OSA module combines RepVGG and ShuffleNet to enhance shallow feature interactions through channel shuffling and feature aggregation, reducing computational complexity by approximately 50% and optimizing small target detection capabilities.

[0061] S4 introduces a variable kernel convolution module to dynamically adjust the sampling position and a reparameter shuffling and one-time aggregation module to adapt to small targets and irregular deformation features;

[0062] Specific methods:

[0063] The S41 input image data enters the Conv module to complete the first round of basic feature extraction, and captures the low-level spatial features of the image through conventional convolution operations.

[0064] The S42 backbone network extracts features sequentially through Conv, VkConv, C2f, and SPPF layers. VkConv adjusts the sampling position based on the offset and resamples the feature map. The resampled result is then reshaped, reconvolved, and standardized before being activated by the SiLU activation function to obtain the output.

[0065] In this embodiment of the invention, the backbone network extracts features in layers through Conv, VkConv, C2f, and SPPF. The core is that the VkConv module replaces part of the conventional convolution: when faced with input features, VkConv dynamically adjusts the size of the convolution kernel with a variable kernel mechanism. This is different from the original model's method of extracting features with a single convolution kernel, which enhances the targeting of feature extraction. VkConv first determines the initial sampling position of the convolution kernel based on the target coordinates, then obtains image features through 2D convolution sampling, and calculates the sampling position offset based on the sampling results.

[0066] S43 outputs multi-scale features from the Backbone. First, it upsamples the low-resolution, high-semantic feature map to increase its resolution. Then, it concatenates the feature map with the same resolution level to merge the channel dimension features.

[0067] In this embodiment of the invention, multi-scale features (feature maps with different downsampling ratios) are output from the Backbone. First, the resolution of the low-resolution, high-semantic feature map is increased through the Upsampling operation. Then, it is concatted with the feature map of the same resolution level to merge the channel dimension features, thus preparing multi-source features for the RVS-OSA module.

[0068] The channel dimension features described in S44 are input into the RVS-OSA module for deep fusion and optimization, enhancing the expression of subtle features;

[0069] In this embodiment of the invention, the spliced ​​features are input into the RVS-OSA module. This module integrates RepVGG structure reparameterization, ShuffleNet channel interaction and One-ShotAggregation ideas to deeply fuse and optimize multi-scale and multi-channel features, enhance the expression of subtle features such as small targets, and make effective features more focused and more discriminative.

[0070] The subtle features described in S45, after being processed by RVS-OSA, are partly sent directly to the backend Detect for target detection, and partly continue to participate in the multi-scale feature loop of the Neck layer, outputting optimized features at different scales to the corresponding detection head, thus assisting the model in multi-scale target detection.

[0071] In this embodiment of the invention, part of the features processed by RVS-OSA are directly sent to the backend Detect for target detection; the other part continues to participate in the multi-scale feature loop of the Neck layer, outputting optimized features at different scales to the corresponding detection head, which helps the model improve detection accuracy and efficiency in multi-scale target detection, especially small and deformable targets, by leveraging the features enhanced by RVS-OSA, thus making up for the shortcomings of traditional Neck layer feature fusion.

[0072] S5 combines the identification results with Gaofen-2 remote sensing images and UAV oblique photography data to conduct spatial distribution analysis of landslide hazard points in the study area, focusing on high-risk areas with severe deformation.

[0073] S6 selected typical areas to conduct field investigations. By comparing high-resolution remote sensing images and UAV aerial survey data, it distinguished between false positive deformations caused by engineering activities and real landslides, and verified the accuracy of the model detection results.

[0074] To better understand this technical solution, the following embodiments are provided for further explanation:

[0075] The present invention will be further described in detail in the following steps:

[0076] Sentinel-1 SAR images of the target area were acquired using a down-orbiting method, and 30m DEM and precise orbit data were obtained simultaneously. The images were registered and processed using SBAS-InSAR technology to construct a set of small baseline interferograms. Phase unwrapping was performed using the minimum flow method, and the surface deformation rate was inverted using singular value decomposition to generate a rasterized deformation rate map.

[0077] Landslide areas were manually interpreted based on InSAR deformation rate maps, and the locations of landslide bodies were labeled in PASCALVOC format using the LabelImg tool. Data augmentation operations such as rotation and flipping were performed on the original samples to generate a sufficient number of training samples, which were then divided into training, validation, and test sets in a 6:1:3 ratio.

[0078] An improved YOLO network is constructed by fusing variable kernel convolution (VkConv) and the RVS-OSA module. VkConv adapts to irregular landslide shapes by dynamically adjusting the sampling position of the convolution kernel, while the RVS-OSA module combines RepVGG and ShuffleNet to enhance shallow feature interactions through channel shuffling and feature aggregation, reducing computational complexity by approximately 50% and optimizing small target detection capabilities.

[0079] The training environment was configured using the PyTorch framework, with 200 iteration cycles, a learning rate of 0.001, and a batch size of 8. A cosine annealing strategy was used to decay the learning rate. The model performance was periodically evaluated on the validation set using precision, recall, and mean AP50 as metrics, and the network parameters were adjusted to their optimal state.

[0080] The deformation rate map is cropped to a fixed size and then input into the trained VRO_YOLO model, which outputs the bounding box and confidence score of the landslide target. Detection results with a confidence score ≥0.5 are filtered, and the location, extent, and deformation features of active landslides are automatically extracted to generate a preliminary detection list.

[0081] By analyzing geomorphological features using Gaofen-2 remote sensing imagery and acquiring 0.1m resolution aerial survey data through UAV oblique photography, field exploration was conducted to record the soil and rock structure and the impact of engineering activities, verifying the accuracy of the detection results. Based on the detection results, a landslide hazard distribution map was drawn, and the risk level was assessed by combining topographic, rainfall, and human activity data, proposing emergency reinforcement and real-time monitoring recommendations.

[0082] The core idea of ​​the SBAS-InSAR technology is to extract millimeter-level surface deformation rate and cumulative deformation from multi-temporal SAR images by combining short baseline interferograms, time series analysis, and singular value decomposition (SVD).

[0083] Acquire N+1 SAR images (t0,...,t) of the same region in chronological order. NOne of the images is selected as the master image, and the other SAR images are registered onto the master image. M+1 multi-view differential interferograms are generated from N+1 SAR images. M satisfies the following inequality:

[0084]

[0085] For image t A and the t obtained from the main image B (t B >t A The interferometric phase of the j-th differential interferogram generated from a SAR image, with respect to the azimuth coordinate x and range coordinate r of the pixel, can be represented as follows:

[0086]

[0087] Wherein d(t) B ,x,r) and d(t A (x, r) represents time t B and t A Cumulative deformation relative to the radar line-of-sight direction of d(t0,x,r); It refers to the residual in the topographic phase difference interferogram. Describe atmospheric phase error. This represents the phase of random noise. Wherein, and It can be represented as:

[0088]

[0089] Where: λ is the wavelength of radar propagation, B ⊥j Δz is the vertical baseline j of the interferogram, Δz is the DEM error, R is the distance between the radar and the target point, and θ is the incident angle.

[0090] Regardless of the atmospheric delay phase, residual terrain phase, and noise phase, it can be simplified as follows:

[0091]

[0092] To obtain a physically meaningful deposition sequence, the phase is represented as the product of the average phase velocity and time between two acquisition times:

[0093]

[0094] The phase value of the j-th interferogram can be written as:

[0095]

[0096] The velocity for each time interval is the integral of the time interval between the master and slave images. It is written in matrix form as follows:

[0097] B v =δφ

[0098] Among them B v Corresponding to an M×N coefficient matrix, and V T It can be represented as follows

[0099]

[0100] Since the differential interferogram of the small baseline set uses a multi-principal image method, matrix B is prone to rank loss. Therefore, the singular value method (SVD) is used to find the generalized inverse matrix of matrix B, thereby obtaining the minimum norm solution of the velocity vector. Finally, by integrating the velocity over each time period, the deformation of each time period can be obtained.

[0101] The core idea of ​​VkConv is a variable kernel convolution mechanism that dynamically adjusts the sampling position of the convolution kernel to capture changes in the shape of the target (such as rotation and scaling), making it especially suitable for small targets and irregularly shaped objects.

[0102] VkConv defines a novel coordinate generation algorithm that provides the convolution kernel with an arbitrary number of parameters and sampling shapes, enabling irregular convolution operations with a linearly increasing number of parameters. VkConv first determines the initial sampling position of the convolution kernel based on the target coordinates, then obtains image features through 2D convolution sampling, and calculates the sampling position offset based on the sampling results.

[0103] VkConv adjusts the sampling position based on the offset and resamples the feature map. The resampled result is then reshaped, reconvolved, and standardized before being output through the SiLU activation function. In this paper, landslide targets are small in size and irregular in shape in the image. Standard convolution with a fixed sampling pattern is insufficient to accurately extract their features. VkConv, by flexibly adjusting the sampling shape, can better fit the contour of small targets, thereby improving the detection capability of landslides.

[0104] Convolutional neural networks define initial sampling positions based on convolution operations, and locate features at the corresponding positions using a regular sampling grid. Assuming R represents the sampling grid, then R is expressed as follows:

[0105] R={(-1,-1),(-1,0),...,(0,1),(1,1)}

[0106] To allow irregular convolutional kernels to have sampling grids, an algorithm for convolutions of arbitrary sizes was created that generates the initial sampling coordinates of the convolutional kernel. This method first generates a regular sampling grid, then generates an irregular sampling grid for the remaining sampling points, and finally stitches them together to form a unified sampling grid.

[0107] The initial sampling shape for these convolutional kernels with different numbers of parameters needs to be considered in two ways. First, an approximately square shape, given the same size, facilitates offset learning. Second, it facilitates the generation of convolutional kernels of arbitrary sizes. Therefore, the sampling shape takes these two points into account, generating the corresponding initial sampling shape for convolutions of arbitrary sizes. The sampling grid for regular convolutions uses the center as the origin (0, 0), while most irregular convolutions are not centered. To accommodate the size of the convolutions used, the top-left corner is set as the sampling origin (0, 0) in the algorithm, and the corresponding convolution operation at that position is defined as follows:

[0108]

[0109] This represents the generated sampling grid, convolution parameters, and the pixel at the corresponding position. However, irregular convolution operations are impossible because irregular sampling coordinates cannot be matched with convolution operations of the corresponding sizes, such as convolutions of sizes 5, 7, and 13.

[0110] During training, VkConv requires frequent memory access to calculate offsets and sample features, which significantly reduces detection speed and efficiency.

[0111] We propose the RVS-OSA module. It integrates RepVGG and ShuffleNet into the OSA module. The RVS module enhances shallow feature interaction and preserves landslide details by improving the information flow of grouped convolutions and utilizing structural reparameterization and channel shuffling. The OSA module simplifies feature concatenation, avoids feature dilution of small targets, and aggregates features with multiple receptive fields to enhance the network's feature extraction capabilities.

[0112] By employing channel segmentation and shuffling operators, the computational complexity is reduced by approximately half, and gradient vanishing is mitigated through single aggregation, improving landslide localization accuracy and simplifying the network structure and computational process. In the RVS-OSA module, the input is divided into two parts: one part passes directly, and the other part is processed through stacked RVS modules. The processed features and the directly passed features are merged after channel shuffling. During the training phase, a multi-branch topology is used to learn rich feature information, while the inference phase is simplified to a single-branch structure to reduce memory consumption and achieve fast inference. Simultaneously, channel shuffling enhances feature fusion, and stacking RVS modules reuses features and enhances information flow between channels.

[0113] RVS-OSA maintains the same number of input channels and a minimum number of output channels, thereby reducing memory access cost (MAC). It also employs different numbers of RCS stacks to achieve semantic information extraction at different stages of the backbone and neck network. A common metric for evaluating computational efficiency (or time complexity) is floating-point operations (FLOPs). FLOPs are only an indirect indicator of inference speed; however, object detectors with a DenseNet backbone exhibit considerably slow speed and low energy efficiency because the linearly increasing number of channels through dense connections leads to heavy memory access costs (MAC). Given M×M dimensional input features, K×K convolutional kernels, C1 input channels, and C2 output channels, then FLOPs and MAC can be calculated as follows:

[0114] FLOPs = M 2 K 2 C1C2

[0115] MAC = M 2 (C1+C2)+K 2 C1C2

[0116] The RVS-OSA module aggregates all processed and fused features in one go to obtain the final feature representation for subsequent object detection.

[0117] By integrating the VkConv and RVS-OSA modules, a powerful and efficient object detection framework is constructed. It can detect landslides in all feature maps at multiple scales and is well adapted to the problem of uneven sample distribution and a large proportion of small target samples in the study area. In addition, it learns rich feature information through a multi-branch structure during the training phase and reduces memory consumption by simplifying it to a single-branch structure during the inference phase, thereby achieving fast inference and comprehensively improving YOLO's feature extraction capability and detection accuracy.

[0118] In the field of object detection, mean accuracy (mAP), precision (P), and recall (R) are commonly used to evaluate the detection performance of a model. To more accurately and rigorously evaluate detector performance, we obtain corresponding mAP values ​​using different IoU (Intersection over Union) thresholds. mAP@0.5 refers to the mAP value calculated when the IoU threshold is 0.5. mAP@0.5:0.95 represents the average mAP at different IoU thresholds, where the IoU value ranges from 0.5 to 0.95, with a step size of 0.05. The calculation method for the evaluation metrics is as follows:

[0119]

[0120] Where TP represents the number of correctly predicted positive samples; FP represents the number of incorrectly predicted positive samples; and FN represents the number of incorrectly predicted negative samples.

[0121] In this invention, performance evaluations of multiple models were conducted for landslide target detection tasks, as shown in Table 1. The experimental results show that VRO_YOLO performs exceptionally well in terms of P (49.2%), F (51.6%), and Map50 (43.8%), outperforming most of the comparison models. Compared to classic models such as YOLOv5, the P-value is similar, but the F and Map50 are superior. Compared to Retinanet and others, it exhibits superior overall performance, indicating that VRO_YOLO may possess better accuracy and comprehensive detection capabilities in target detection tasks.

[0122] Table 1 Comparison of different object detection networks

[0123]

[0124] This invention integrates SBAS-InSAR technology with the improved YOLOv8 algorithm. First, it uses SBAS-InSAR to generate a surface deformation rate map, then uses VRO_YOLO to detect landslides from the deformation rate map, and finally combines high-resolution remote sensing, UAV oblique photography and field exploration to conduct disaster analysis.

[0125] This invention replaces the original convolutional module with a variable kernel convolution (VkConv) in the YOLOv8 network. By dynamically adjusting the sampling position of the convolutional kernel, it captures changes in the target shape. With the help of a new coordinate generation algorithm, it achieves irregular convolution with arbitrary number of parameters and sampling shape, which can further improve the detection capability of small targets on landslides.

[0126] This invention proposes a reparameterized shuffled one-time aggregation (RVS-OSA) in the neck network of YOLOv8. It integrates RepVGG and ShuffleNet into OSA, and enhances the cross-channel interaction of shallow features through structural reparameterization and channel shuffling. This preserves landslide details, avoids the dilution of small target features, reduces computational complexity, reduces gradient vanishing, and improves monitoring accuracy.

[0127] The above-disclosed embodiments are merely preferred embodiments of the improved YOLOv8 active landslide detection method based on InSAR deformation of the present invention. Of course, they should not be construed as limiting the scope of the present invention. Those skilled in the art can understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. An improved YOLOv8 active landslide detection method based on InSAR deformation, characterized in that, Includes the following steps; Acquire Sentinel-1A descent SAR images, calculate deformation rate, and obtain basic data; Based on SBAS-InSAR technology, SAR images are processed through data preprocessing, differential interferometry, orbit refinement, inversion and geocoding to generate surface deformation rate maps and cumulative displacement of the target area. The InSAR results were visually interpreted, landslide bodies were labeled, deformation rate maps were cropped, and after data augmentation, a dataset was generated and divided into training, validation and test sets. A variable kernel convolution module is introduced to dynamically adjust the sampling position, and a reparameter shuffling and one-time aggregation module is introduced to adapt to small targets and irregular deformation features; By combining the identification results with Gaofen-2 remote sensing images and UAV oblique photography data, spatial distribution analysis of landslide hazard points in the study area was conducted, focusing on high-risk areas with severe deformation. Field investigations were conducted in typical areas. By comparing high-resolution remote sensing images with UAV aerial survey data, the false positive deformation caused by engineering activities and the real landslides were distinguished, thus verifying the accuracy of the model detection results.

2. The improved YOLOv8 active landslide detection method based on InSAR deformation as described in claim 1, characterized in that... ； The input data required for calculating the deformation rate includes SAR image data, precise orbit data, external DEM data, GACOS data, and ground control GCPS points.

3. The improved YOLOv8 active landslide detection method based on InSAR deformation as described in claim 1, Its characteristics are: The specific method for processing SAR images based on SBAS-InSAR technology, through data preprocessing, differential interferometry, orbit refinement, inversion, and geocoding, to generate a surface deformation rate map and cumulative displacement of the target area is as follows: The study area is cropped, the target range is extracted from the SAR main and secondary images, the spatial baselines of the main and secondary images are estimated, the interferometric quality is assessed, the main and secondary images are selected, and the interferometric pair is constructed. The main and secondary images are registered using DEM calibration, the true phase difference is recovered by unwrapping using minimum cost flow, the interferogram is enhanced by Goldstein filtering, the differential interferogram is generated, the connection diagram is edited, the correlation of image pairs is sorted out, and a small baseline set processing framework is built. Select GCP points and generate control point data, correct satellite orbit parameters, and then perform terrain re-flattening, estimate and remove constant phase and residuals, remove unwrapped phase slopes, continuously purify phase information, and refine orbits. Based on the orbit refinement results, orbital errors are removed, deformation rate, terrain residuals and average strength are estimated, the orbit is refined a second time and the phase is unwrapped, and further processing is performed. Optimize the re-flattening process, use GCP calibration to more thoroughly eliminate the influence of topographic phase, remove atmospheric phase, extract key features using singular value decomposition, calculate the surface deformation rate, and output reliable monitoring results; Set quality control thresholds for height and rate accuracy, then geocode deformation values, associate them with real geographical locations, generate a time deformation sequence, display the deformation process, and output vector results.

4. The improved YOLOv8 active landslide detection method based on InSAR deformation as described in claim 1, characterized in that... ； The specific methods for introducing a variable kernel convolution module to dynamically adjust the sampling position and a reparameter shuffling and one-time aggregation module to adapt to small targets and irregular deformation features are as follows: Input image data and enter the Conv module to complete the first round of basic feature extraction. The low-level spatial features of the image are initially captured through conventional convolution operations. The backbone network extracts features sequentially through Conv, VkConv, C2f, and SPPF layers, and initially captures the low-level spatial features of the image through conventional convolution operations. Multi-scale features are output from the Backbone. First, the low-resolution, high-semantic feature map is upsampled to increase its resolution. Then, it is concatenated with the feature map of the same resolution level to merge the channel dimension features. The channel dimension feature input reparameter mixing and one-time aggregation module performs deep fusion and optimization, enhancing the expression of subtle features and improving detection accuracy and efficiency; The subtle features processed by RVS-OSA are partly sent directly to the backend Detect for target detection, and partly continue to participate in the multi-scale feature loop of the Neck layer, so that all processed and fused features are aggregated at once. Output optimized features at different scales to the corresponding detection head to help the model detect targets at multiple scales.

5. The improved YOLOv8 active landslide detection method based on InSAR deformation as described in claim 4, characterized in that... ； The input image data enters the Conv module to complete the first round of basic feature extraction. During the process of initially capturing the low-level spatial features of the image through conventional convolution operations, if there are differences in size or channels in the original image, resize and normalization operations will be performed in advance to ensure that the input format meets the model requirements.

Images