Lightweight intelligent detection method and system for road landslide

Abstract

The application provides a lightweight intelligent detection method and system for road landslide, and relates to the technical field of intelligent monitoring of geological disasters. Through dynamic parameter configuration, multi-source data fusion to define the gradual change boundary, and optimization of feature extraction and fusion mechanism, the problems of insufficient boundary definition, weakened feature capture and lack of dynamic adjustment in the prior art are solved, and the application has high detection accuracy, fast response speed, strong environmental adaptability and user customization ability.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring technology for geological disasters, and in particular to a lightweight intelligent detection method and system for road landslides. Background Technology

[0002] Landslides, as geological disasters characterized by their sudden onset, wide impact, and destructive power, require precise and efficient segmentation for effective emergency response, loss assessment, and subsequent prevention and control. With the rapid development of remote sensing technology and artificial intelligence, automatic landslide segmentation methods based on satellite imagery are gradually replacing traditional manual annotation and becoming a core tool for disaster monitoring. However, existing technologies still face multiple challenges and struggle to meet the demands of practical applications.

[0003] Traditional image processing methods primarily rely on the grayscale and texture features of images to extract landslide areas through threshold segmentation or edge detection algorithms. While these methods are effective under simple terrain conditions, they are severely inadequate in defining the gradual transition boundaries of landslide areas when faced with complex terrain, vegetation cover, or weather disturbances. Their segmentation results are easily affected by factors such as lighting changes and cloud cover, leading to blurred boundaries, high false detection rates, and an inability to adapt to changing geological environments. Their robustness and accuracy fall short of practical standards for disaster monitoring. Deep learning semantic segmentation methods, while improving feature learning capabilities through encoder-decoder structures, have revealed significant shortcomings in practical deployments. To adapt to mobile devices or real-time monitoring scenarios, existing solutions often employ lightweight backbone networks. However, this network simplification directly weakens the feature extraction depth, reducing the model's ability to capture landslide features at different scales, especially performing poorly in identifying small-scale or concealed landslides. Furthermore, existing methods generally use fixed parameter configurations and uniform processing strategies, failing to dynamically adjust based on the terrain complexity of the target area, weather conditions, and user preferences. For example, users' emphasis on detection accuracy or speed cannot be effectively addressed, making it difficult to balance efficiency and accuracy in emergency disaster scenarios.

[0004] Significant shortcomings exist in the multi-source data fusion process. The integration of real-time meteorological data and high-resolution satellite imagery lacks spatiotemporal correlation guarantees, and time differences are not properly quantified and weighted, resulting in distorted fused data. In the feature fusion stage, background noise suppression mechanisms are weak, the weighting allocation logic for multi-scale features is overly simplistic, and historical landslide case experience is not utilized, reducing the distinction between landslide areas and the background in complex scenes. Furthermore, insufficient supplementation of detailed features during upsampling, ineffective fusion of shallow texture information, and a lack of targeted optimization for scene differences in post-processing strategies further limit the practicality of the segmentation results. These issues collectively prevent existing methods from providing high-precision data support for landslide disaster area calculation, exhibiting significant limitations in emergency response to sudden disasters.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] This application provides a lightweight intelligent detection method and system for road landslides, which has high detection accuracy, fast response speed, strong environmental adaptability and user customization capabilities.

[0007] Firstly, the lightweight intelligent detection method for road landslides provided in this application adopts the following technical solution: A lightweight intelligent detection method for road landslides includes: The detection model's parameter configuration is determined based on the geographic metadata of the target area and user role profile information. Simultaneously, scene determination is completed and the scene determination result is output. The user role profile information includes the weight settings of the user's preference for detection accuracy or detection speed. High-resolution optical satellite images of the target area are acquired and fused with real-time meteorological data and image data from historical image databases. The gradual transition boundary of the landslide area is defined through multi-temporal comparative analysis. Pixel-level annotation is completed based on the gradual transition boundary, and then an enhanced landslide segmentation dataset is constructed. The image data of the enhanced landslide segmentation dataset is preprocessed, and the preprocessed image data is input into the encoder. The encoder uses a lightweight MobileNetV4 network as the backbone network, performs efficient feature extraction and outputs high-order feature maps through the general inverted bottleneck module of the network, and dynamically adjusts the parameters of the encoder convolutional layer based on the scene determination result. The high-order feature map is input into the attention-diffuse pyramid module, and features of different scales are captured through the parallel branches of the module. Each parallel branch uses a 1×1 convolution and a depthwise separable dilated convolution structure to extract scale features. The scale features captured by each parallel branch are dynamically weighted and fused through a channel attention mechanism. At the same time, the feature weights of similar historical landslide cases are queried based on the integrated historical record module, and the attention allocation logic is optimized based on the feature weights. The feature map fused by the attention-free hollow pyramid module is input into the decoder. In the decoder, a large upsampling operation is performed through a lightweight dynamic upsampling unit. The upsampling result is then spliced ​​and fused with the shallow features output by the encoder. Based on the scene determination result, the corresponding post-processing strategy is selected, and the spliced ​​and fused feature map is post-processed to output a landslide segmentation map for calculating the landslide disaster area.

[0008] Optionally, the process of generating the user role profile information is as follows: first, collect the user's historical operation records, extract the user's historical detection task types, parameter adjustment trajectories and task result feedback data from the historical operation records, then analyze the extracted data through a feature clustering algorithm to quantify the user's preference for detection accuracy and speed, and finally generate user role profile information containing the preference weight settings.

[0009] Optionally, during the construction of the dataset, the fusion processing of real-time meteorological data and high-resolution optical satellite images adopts a weighted fusion algorithm based on data temporal consistency. This algorithm first aligns the temporal dimensions of meteorological data acquisition time and satellite image capture time, and then dynamically assigns weights according to the time difference between the two—the smaller the time difference, the higher the weight of the meteorological data, so as to ensure the spatiotemporal correlation of the fused data.

[0010] Optionally, after the encoder extracts features through the general inverted bottleneck module, it immediately performs batch normalization processing on the output feature map to improve the stability of feature extraction. The dynamically adjusted encoder convolutional layer parameters include the number of convolutional kernels, the dilation rate, and the stride. The parameter adjustment range is determined according to the terrain complexity level corresponding to the scene judgment result. The more complex the terrain, the greater the adjustment range to adapt to the feature capture requirements.

[0011] Optionally, in the multi-scale feature fusion process, the depth-separable dilated convolutional structure adopts a differential dilation setting with different dilation values ​​to capture features at different scales; the channel attention mechanism performs feature squeezing and activation operations on the channel dimensions of feature maps at each scale to strengthen the weight of effective feature channels and weaken ineffective channels, thereby achieving precise dynamic allocation of feature weights.

[0012] Optionally, the lightweight dynamic upsampler adopts a structure that combines transposed convolution and pixel recombination—transposed convolution enlarges the feature map size, pixel recombination optimizes the feature resolution, and an adaptive threshold is set based on the grayscale difference of feature pixels during the upsampling process to filter out low-value redundant pixels; the shallow features concatenated with the upsampling results are low-level texture features output by the encoder in the early stage, which are used to supplement detailed information.

[0013] Optionally, the landslide disaster feature database integrated by the historical record module stores geographical, meteorological, and feature weight data of historical landslide cases. When querying similar cases, the geographical metadata and meteorological data of the target area are used as core search keywords to filter out cases whose feature similarity meets a preset threshold. Then, the feature weights of the cases are fused through a weighted average algorithm to optimize the current attention allocation logic.

[0014] Secondly, this application provides a lightweight intelligent detection system for road landslides, the lightweight intelligent detection system for road landslides comprising: The data acquisition module is used to determine the parameter configuration of the detection model based on the geographic metadata of the target area and the user role profile information, and at the same time complete the scene determination and output the scene determination result; the user role profile information includes the weight settings of the user's preference for detection accuracy or detection speed. The dataset construction module is used to acquire high-resolution optical satellite images of the target area, fuse them with real-time meteorological data and image data in the historical image database, define the gradual transition boundary of the landslide area through multi-temporal comparison analysis, complete pixel-level annotation based on the gradual transition boundary, and then construct an enhanced landslide segmentation dataset. The preprocessing module is used to preprocess the image data of the enhanced landslide segmentation dataset and input the preprocessed image data into the encoder. The encoder uses a lightweight MobileNetV4 network as the backbone network, performs efficient feature extraction and outputs high-order feature maps through the network's general inverted bottleneck module, and dynamically adjusts the parameters of the encoder's convolutional layers based on the scene determination results. The input module is used to input the high-order feature map into the attention-diffuse pyramid module. The parallel branches of this module capture features at different scales. Each parallel branch uses a 1×1 convolution and a depthwise separable dilated convolution structure to extract scale features. The scale features captured by each parallel branch are dynamically weighted and fused through a channel attention mechanism. At the same time, the integrated historical record module is used to query the feature weights of similar historical landslide cases, and the attention allocation logic is optimized based on the feature weights. The output module is used to input the feature map fused by the attention-drained pyramid module into the decoder. In the decoder, a large upsampling operation is performed through a lightweight dynamic upsampling unit. The upsampling result is then spliced ​​and fused with the shallow features output by the encoder. Based on the scene determination result, the corresponding post-processing strategy is selected, and the post-processing operation is performed on the spliced ​​and fused feature map to output a landslide segmentation map for calculating the landslide disaster area.

[0015] Thirdly, this application provides a computer device, the device comprising: a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method described above.

[0016] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the method described above.

[0017] In summary, this application includes the following beneficial technical effects: This application solves the problems of insufficient boundary definition, weakened feature capture, and lack of dynamic adjustment in the prior art by using dynamic parameter configuration, multi-source data fusion to define gradual boundaries, and optimizing feature extraction and fusion mechanisms. It has high detection accuracy, fast response speed, strong environmental adaptability, and user customization capabilities. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiments of this application.

[0019] Figure 2 This is a flowchart illustrating the first embodiment of the lightweight intelligent detection method for road landslides in this application.

[0020] Figure 3 This is a structural block diagram of the first embodiment of the lightweight intelligent detection system for road landslides in this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] Reference Figure 1 , Figure 1 This is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiments of this application.

[0023] like Figure 1 As shown, the computer device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0024] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0025] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a lightweight intelligent detection program for road landslides.

[0026] exist Figure 1 In the computer device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in this application can be set in the computer device, and the computer device calls the lightweight intelligent detection program for road landslides stored in the memory 1005 through the processor 1001, and executes the lightweight intelligent detection method for road landslides provided in the embodiments of this application.

[0027] This application provides a lightweight intelligent detection method for road landslides, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the lightweight intelligent detection method for road landslides in this application.

[0028] In this embodiment, the lightweight intelligent detection method for road landslides includes the following steps: Step S10: Determine the parameter configuration of the detection model based on the geographic metadata of the target area and the user role profile information, and at the same time complete the scene determination and output the scene determination result; the user role profile information includes the weight settings of the user's preference for detection accuracy or detection speed; Step S20: Acquire high-resolution optical satellite images of the target area, fuse them with real-time meteorological data and image data in the historical image database, define the gradual transition boundary of the landslide area through multi-time comparison analysis, complete pixel-level annotation based on the gradual transition boundary, and then construct an enhanced landslide segmentation dataset. Step S30: Preprocess the image data of the enhanced landslide segmentation dataset, and input the preprocessed image data into the encoder; the encoder uses a lightweight MobileNetV4 network as the backbone network, performs efficient feature extraction and outputs high-order feature maps through the general inverted bottleneck module of the network, and dynamically adjusts the parameters of the encoder convolutional layer based on the scene determination result. Step S40: Input the high-order feature map into the attention-diffuse pyramid module. The parallel branches of this module capture features at different scales. Each parallel branch uses a 1×1 convolution and a depthwise separable dilated convolution structure to extract scale features. The scale features captured by each parallel branch are dynamically weighted and fused through a channel attention mechanism. At the same time, the feature weights of similar historical landslide cases are queried based on the integrated historical record module. The attention allocation logic is optimized based on these feature weights. Step S50: Input the feature map fused by the attention-free hollow pyramid module into the decoder. In the decoder, perform a large-scale upsampling operation through a lightweight dynamic upsampler. The upsampling result is then spliced ​​and fused with the shallow features output by the encoder. Based on the scene determination result, select the corresponding post-processing strategy and perform post-processing operations on the spliced ​​and fused feature map to output a landslide segmentation map for calculating the landslide disaster area.

[0029] For ease of understanding, the following explains some key terms in this embodiment: Geographic metadata refers to data that describes the geographic attributes of a target area, such as topography, landforms, slope, altitude, and geological type, and is used to provide background information about the area.

[0030] User profile information refers to a set of data describing a user's preferences for detection tasks, such as the weighting of detection accuracy or detection speed, which is used to guide the personalized configuration of model parameters.

[0031] Scene determination refers to assessing and classifying the complexity, landslide type, and environmental conditions of a target area based on information such as geographic metadata and meteorological data, in order to determine a model processing strategy suitable for the current scene.

[0032] The enhanced landslide segmentation dataset refers to the collection of images and labeled data used to train and validate landslide segmentation models after multi-source data fusion, multi-time comparison analysis and pixel-level annotation processing. Its characteristics are more refined boundary annotation and richer data dimensions.

[0033] The encoder is a component of a deep learning model, responsible for extracting multi-level, multi-scale feature information from the input image, typically implemented through a series of convolutional and pooling layers.

[0034] The lightweight MobileNetV4 network is a high-efficiency convolutional neural network backbone network designed to significantly reduce the number of model parameters and computational complexity while maintaining high performance, making it suitable for resource-constrained devices or real-time applications.

[0035] The general inverted bottleneck module is a core building block in the MobileNetV4 network. It achieves efficient feature transformation and information transfer by first expanding the channels, then performing depthwise separable convolution, and finally compressing the channels.

[0036] The Attention-Diffuse Pyramid module is a module used to capture multi-scale contextual information. It enhances the ability to identify target regions by setting dilated convolution branches with different dilation rates in parallel and combining them with an attention mechanism to perform weighted fusion of features at different scales.

[0037] Channel attention is a deep learning mechanism that learns the channel dimension of a feature map and dynamically assigns weights to different channels, thereby strengthening important feature channels, suppressing unimportant feature channels, and improving feature representation capabilities.

[0038] The history module is a database and query system for storing and managing historical landslide case information. It is used to retrieve similar cases in the current detection task and to optimize model decisions based on their experience.

[0039] The decoder is another component in a deep learning model. It is responsible for progressively restoring the high-order feature maps output by the encoder to the original image resolution and generating the final segmentation result. This is usually achieved through upsampling and convolution operations.

[0040] The Lightweight Dynamic Upsampler is a high-efficiency upsampling module designed to increase feature map size and optimize resolution with minimal computational overhead, and dynamically adjust the upsampling process based on feature content.

[0041] Shallow features are feature maps output by the encoder in the early stages. They have high resolution but relatively little semantic information and usually contain rich texture and edge detail information.

[0042] Post-processing strategies refer to methods that correct and optimize the initial segmentation results after the model outputs them, based on specific rules or algorithms. Examples include removing small-area noise, smoothing boundaries, and filling holes, in order to improve the accuracy and usability of the segmentation map.

[0043] This embodiment provides a lightweight intelligent detection method for road landslides.

[0044] Specifically, during the parameter configuration phase of the detection model, the model parameters can be determined based on the geographic metadata of the target area and user role profile information. For example, geographic metadata can include information such as the slope, altitude, and geological type of the target area, while user role profile information can be simply set to the user's single preference for detection accuracy or detection speed. Based on this, scene determination can be performed, such as roughly dividing the scene into "gentle areas" or "steep areas" according to the average slope or vegetation coverage of the area, and outputting the corresponding scene determination results.

[0045] Furthermore, in terms of data acquisition and dataset construction, high-resolution optical satellite images of the target area can be obtained. These images can be simply overlaid and fused with real-time meteorological data (such as rainfall and temperature) and image data from historical image databases. Subsequently, through visual comparative analysis of images from different time points, the boundaries of the landslide area are manually or semi-automatically defined, and pixel-level annotations are completed based on these boundaries, thereby constructing a landslide segmentation dataset for model training.

[0046] Based on this, the image data in the constructed landslide segmentation dataset is preprocessed, such as by scaling and pixel value normalization. The preprocessed image data is then input into the encoder. This encoder can use a lightweight network as its backbone, such as MobileNetV1 or V2, whose inverted bottleneck module performs feature extraction and outputs high-order feature maps. The parameters of the encoder's convolutional layers can be set to preset fixed values ​​without dynamic adjustment.

[0047] Subsequently, the high-order feature map is input into the attention-based dilated pyramid module. This module can have multiple parallel branches, each employing a standard convolution or a dilated convolution structure with a fixed dilation rate to capture features at different scales. The captured scale features can be fused through simple concatenation or averaging, without using complex channel attention mechanisms. Simultaneously, a historical record module can be integrated, storing basic information about historical landslide cases for manual reference, but not directly involved in feature weight optimization.

[0048] Finally, the feature map fused by the attention-drilled pyramid module is input into the decoder. In the decoder, traditional bilinear interpolation or nearest-neighbor interpolation methods can be used to perform upsampling operations, restoring the feature map to near its original resolution. The upsampling result can be simply added element-wise to the earlier features output by the encoder. The post-processing stage can employ a single, pre-defined general strategy, such as performing morphological opening and closing operations on the segmentation results to smooth boundaries or remove small-area noise, ultimately outputting a landslide segmentation map for calculating the landslide disaster area.

[0049] This embodiment effectively balances detection accuracy and speed by dynamically adjusting model parameters to adapt to different scenarios and user needs. Through multi-source data fusion and multi-time comparison analysis, it can accurately define the gradual transition boundaries of landslide areas, improving segmentation accuracy in complex terrain. The combination of the encoder and the attention-based hollow pyramid module enables efficient multi-scale feature extraction and optimized fusion, enhancing the ability to identify landslide areas. The post-processing strategy combining the decoder and scene determination results further improves the accuracy and practicality of the landslide segmentation map, providing reliable data support for calculating landslide disaster area.

[0050] In some of the above embodiments of this example, although it is proposed to determine the parameter configuration of the detection model based on the geographic metadata of the target area and the user role profile information, and it is pointed out that the user role profile information includes the weight settings of the user's preference for detection accuracy or detection speed, if there is a lack of a systematic and data-driven way to accurately generate these preference weights, the model parameter configuration may not accurately reflect the user's actual needs, thereby affecting the intelligence and personalization level of the detection method.

[0051] In this regard, this embodiment further proposes that the process of generating user role profile information in the above-mentioned landslide segmentation method is as follows: first, collect the user's historical operation records, extract the user's historical detection task types, parameter adjustment trajectories and task result feedback data from the historical operation records, and then analyze the extracted data through feature clustering algorithm to quantify the user's preference for detection accuracy and speed, and finally generate user role profile information containing the preference weight settings.

[0052] Specifically, firstly, the system continuously collects historical operation records of users during road landslide detection tasks. These records are raw data of user interactions with the detection system, covering user behavior patterns in different task scenarios. For example, this may include the time the user starts the detection task, the selected area, the initial parameters set, the duration of viewing the detection results, whether parameter fine-tuning was performed, and whether the final result was accepted or rejected. These historical operation records are the foundation for understanding user preferences.

[0053] Based on this, the system extracts key user behavior feature data from the collected user historical operation records. This includes the user's historical detection task types, such as whether it is a quick overview or a detailed analysis; the user's parameter adjustment trajectory, i.e., how the user modifies or optimizes various parameters of the detection model in different tasks, such as adjusting the confidence threshold, changing processing speed options, etc.; and task result feedback data, which can be explicit evaluations given by the user (such as satisfaction ratings, text comments) or implicit behaviors (such as the frequency of repeatedly running tasks and manually correcting results). This extracted data provides structured input for subsequent preference quantification.

[0054] Subsequently, the extracted data is analyzed in depth using feature clustering algorithms. Feature clustering algorithms can identify user groups or task instances with similar behavioral patterns and categorize them into different types. For example, algorithms such as K-means, DBSCAN, or hierarchical clustering can classify users into different role profiles, such as "speed-oriented," "precision-oriented," or "balanced," based on the similarity of users in task type, parameter adjustment, and result feedback. This clustering analysis helps to discover the inherent patterns and potential preferences of user behavior.

[0055] Based on the clustering analysis results, the system further quantifies users' preferences for detection accuracy and speed. For each user or user group, a numerical preference weight is calculated according to the characteristics of their respective cluster. For example, if a user group frequently chooses the fast processing mode and accepts a lower confidence threshold in historical operations, their speed preference weight will be higher; conversely, if users tend to perform refined analysis and spend more time adjusting parameters to pursue higher accuracy, their accuracy preference weight will be higher. This quantification process transforms abstract user behavior into numerical indicators that can be directly used by the model.

[0056] Ultimately, the system generates user profile information containing the aforementioned preference weight settings. This profile is a structured dataset that associates the quantified precision and speed preference weights with specific users or user groups. This profile information can then be used to dynamically adjust the parameter configuration of the detection model, ensuring that the model can better adapt to the user's personalized needs when performing tasks.

[0057] Through the above technical solution, this embodiment provides a data-driven and highly personalized user profile information generation mechanism. By systematically collecting and analyzing users' historical operation records, including their task types, parameter adjustment behaviors, and feedback on results, it is possible to objectively and accurately quantify users' actual preferences for detection accuracy and speed. This preference weighting based on real behavioral data rather than subjectively set values ​​allows the subsequent configuration of detection model parameters to more accurately match users' personalized needs, thereby significantly improving the intelligence level of the detection method and user satisfaction, reducing the frequency of users manually adjusting parameters, and ensuring that the model can provide performance that meets user expectations in different application scenarios.

[0058] When constructing an enhanced landslide segmentation dataset, real-time meteorological data needs to be fused with high-resolution optical satellite imagery. However, simply fusing data collected at different times may lead to temporal inconsistencies between the data, affecting the accuracy of the fusion results and the reliability of subsequent landslide area delineation. This temporal deviation may prevent the fused data from truly reflecting the actual situation of the target area, thereby reducing the effectiveness of the entire detection method.

[0059] To address this, this embodiment further proposes a weighted fusion algorithm based on data temporal consistency for the fusion processing of real-time meteorological data and high-resolution optical satellite images during the dataset construction process. This algorithm first aligns the temporal dimensions of meteorological data acquisition time and satellite image capture time, and then dynamically assigns weights based on the time difference between the two. Specifically, the smaller the time difference, the higher the weight of the meteorological data, to ensure the spatiotemporal correlation of the fused data.

[0060] The weighted fusion algorithm based on data temporal consistency is a method designed to optimize the fusion effect of multi-source heterogeneous data. Its core lies in identifying and quantifying the differences between different data sources in the time dimension, and adjusting the contribution of each data source in the fusion process accordingly. Through precise time alignment and dynamic weight allocation mechanisms, this algorithm ensures that the fused data more accurately reflects the true state of the target region at a specific point in time.

[0061] Aligning the time-series dimension of meteorological data acquisition time with satellite image capture time refers to the system acquiring precise timestamp information for each data point before data fusion. For example, for high-resolution optical satellite images, the capture time is typically recorded precisely to the second; for real-time meteorological data, the acquisition frequency may be higher, such as once per hour or per minute. The alignment process involves unifying these timestamps to the same format and time base (such as UTC time), and then calculating the time difference between the meteorological data acquisition time and the satellite image capture time. For example, the meteorological data point closest to the satellite image capture time can be found, and its time difference recorded.

[0062] The system dynamically assigns weights based on the time difference between the meteorological data and the satellite image. The smaller the time difference, the higher the weight of the meteorological data. This means that after calculating the time difference between the meteorological data and the satellite image, the system assigns a corresponding weight value to the meteorological data according to a preset weighting strategy. For example, an exponential decay function can be used, such as weight = exp(-k The weight is calculated using the time difference (k), where k is a preset decay coefficient representing the sensitivity of the time difference to the weight's influence. When the time difference is zero, the weight reaches its maximum value (e.g., 1); as the time difference increases, the weight decreases rapidly. A linear decay function can also be used, such as weight = max(0, 1 - k). The weight of a data point is determined by a time difference (k), where k is the attenuation coefficient. When the time difference exceeds a certain threshold, the weight may drop to zero. This dynamic allocation mechanism ensures that meteorological data that is closer in time to the moment the satellite image was captured is given higher importance during the fusion process, thus more accurately reflecting the actual meteorological conditions at that time.

[0063] By employing the aforementioned technical solution, this embodiment effectively addresses the potential inconsistency between real-time meteorological data and high-resolution optical satellite imagery in terms of acquisition time. Through precise alignment and dynamic weight allocation based on time differences, the fused data achieves a high degree of correlation in both spatiotemporal dimensions. This means the fused dataset more accurately reflects the actual meteorological conditions and surface conditions of the target area at a specific point in time, providing a more reliable and accurate input for subsequent multi-time-contrast comparative analysis to define the gradual transition boundaries of landslide areas. This highly spatiotemporally correlated data significantly improves the quality of the enhanced landslide segmentation dataset, thereby enhancing the accuracy and robustness of the entire landslide detection method and avoiding misjudgments or missed detections caused by data temporal mismatches.

[0064] In some of the above implementations, the encoder uses a lightweight MobileNetV4 network as its backbone to perform efficient feature extraction. However, in practical applications, facing complex and ever-changing road landslide scenarios, relying solely on the backbone network for feature extraction may not guarantee the stability of the features or their adaptability to different terrain complexities, thus affecting the accuracy of subsequent landslide segmentation.

[0065] To address this, this embodiment further proposes that after the encoder extracts features through the general inverted bottleneck module, it immediately performs batch normalization processing on the output feature map to improve the stability of feature extraction. The dynamically adjusted encoder convolutional layer parameters include the number of convolutional kernels, the dilation rate, and the stride. The parameter adjustment range is determined according to the terrain complexity level corresponding to the scene judgment result. The more complex the terrain, the greater the adjustment range to adapt to the feature capture requirements.

[0066] Specifically, after extracting features through the general inverted bottleneck module, the encoder immediately performs batch normalization on the output feature map. Batch normalization (BN) is a commonly used technique in deep learning, which standardizes the input of each layer in a neural network to a mean of 0 and a variance of 1. For each mini-batch of data, the BN layer calculates the mean and variance of that batch and uses these statistics to normalize the data. Subsequently, by introducing two learnable parameters—a scaling factor and an offset—a linear transformation is performed on the normalized data to restore the network's expressive power. This process effectively alleviates the internal covariate shift problem during training, keeping the input distribution of each layer of the network stable, thereby accelerating model convergence and improving the model's generalization ability. By performing batch normalization immediately after feature extraction through the general inverted bottleneck module, it ensures that the high-order feature maps output from the backbone network MobileNetV4 remain numerically stable, avoiding gradient vanishing or gradient exploding problems caused by excessively large or small feature values. This allows the network to learn and update parameters more stably during training, reducing its sensitivity to initial weight settings and thus improving the robustness and reliability of the feature extraction process.

[0067] Furthermore, the parameters of the encoder's convolutional layers are designed to be dynamically adjusted. These parameters include the number of kernels, the dilation rate, and the stride. "Dynamically adjusted" means that these convolutional layer parameters are not fixed before model training or deployment, but are adaptively modified in real-time or near real-time based on external inputs or internal states (in this case, the terrain complexity level corresponding to the scene determination result). This adjustment mechanism allows the model to flexibly optimize its feature extraction capabilities according to the specific characteristics of the scene being processed, rather than using a one-size-fits-all fixed configuration. The number of kernels determines the depth of the output feature map of the convolutional layer, i.e., the number of feature channels that the layer can learn. Increasing the number of kernels allows the network to capture more diverse and richer feature patterns, such as edges in different directions, textures, or more complex local structures. The dilation rate is a key parameter in dilated convolution, defining the spacing between kernel elements. By increasing the dilation rate, the receptive field of the kernel can be effectively expanded without increasing the number of kernel parameters or computational cost, allowing the network to capture a wider range of contextual information. The stride defines the distance the convolutional kernel moves across the input feature map. When the stride is greater than 1, the convolution operation performs downsampling, thereby reducing the spatial size of the feature map, reducing computational cost, and extracting more abstract features. Smaller strides help retain more spatial detail.

[0068] The adjustment range of the above parameters is determined based on the terrain complexity level corresponding to the scene determination result. The more complex the terrain, the larger the adjustment range, to adapt to the feature capture requirements. The scene determination result is determined based on the geographic metadata of the target area and user role profile information, and it includes a comprehensive evaluation of the current detection scene. The terrain complexity level is an important component of this scene determination result, which quantifies the complexity of the geomorphic features of the target area, such as topographic relief, slope variation, vegetation cover density, and geological structure complexity. These levels can be predefined as low, medium, high, or more detailed levels. "Adjustment range" refers to the degree to which parameters such as the number of convolutional kernels, the hole rate, and the stride are increased or decreased based on the baseline parameter values. When the terrain complexity level is high, it means that the landslide area may have more irregular shapes, more subtle texture variations, more complex boundary features, or require broader contextual information for differentiation. In this case, increasing the number of convolutional kernels can capture richer local feature patterns; increasing the hole rate can expand the receptive field and obtain more comprehensive global contextual information; and decreasing the stride can retain more spatial details and avoid the loss of important information during downsampling. This adaptive parameter adjustment ensures that the encoder can dynamically optimize its feature extraction strategy based on the challenges of the actual scenario, thereby more effectively capturing key features related to landslides.

[0069] By employing the aforementioned technical solution, batch normalization is performed immediately after feature extraction via the encoder's general inverted bottleneck module. This effectively addresses the potential numerical instability of features during deep network propagation, ensuring the robustness of feature representation and providing high-quality, stable input for subsequent feature fusion and segmentation tasks. Simultaneously, the number of convolutional kernels, dilation rate, and stride of the encoder's convolutional layers are dynamically adjusted based on scene determination results. This allows the encoder to adaptively optimize its feature capture capabilities according to changes in the terrain complexity of the target area. Specifically, when facing complex landslide areas, by increasing the adjustment range—for example, increasing the number of convolutional kernels to capture more details, increasing the dilation rate to expand the receptive field and obtain broader contextual information, or adjusting the stride to retain more spatial resolution—the encoder can extract landslide features more accurately and comprehensively in complex environments. This dynamic adjustment mechanism significantly improves the model's adaptability to different landslide scenarios and the effectiveness of feature extraction, avoiding the performance degradation of fixed-parameter models in complex scenarios, thereby enhancing the overall accuracy and reliability of road landslide detection.

[0070] In some of the embodiments described above, the attention-diffused pyramid module captures features at different scales through parallel branches and uses 1×1 convolutions and depthwise separable dilated convolution structures to extract scale features. However, in practical applications, landslide areas vary greatly in shape and scale. If these multi-scale features are not captured and fused sufficiently and precisely, the model may lack accuracy in recognizing landslide areas of different sizes and shapes. Especially in the feature fusion stage, the lack of an effective mechanism to distinguish and enhance key features may affect the final segmentation results.

[0071] To address this, this embodiment further proposes that during the multi-scale feature fusion process, the depth-separable dilated convolutional structure adopts different dilation rates to capture features at different scales. The channel attention mechanism performs feature squeezing and activation operations on the channel dimensions of feature maps at each scale, thereby strengthening the weights of effective feature channels and weakening ineffective channels, achieving precise dynamic allocation of feature weights.

[0072] Depthwise separable dilated convolutional structures are efficient convolutional operations that combine the parametric efficiency of depthwise separable convolutions with the receptive field expansion capabilities of dilated convolutions. In multi-scale feature fusion, to effectively capture feature information of landslide areas at different scales, different dilation rates can be configured for the depthwise separable dilated convolutional structures in each parallel branch of the attention dilated pyramid module. For example, dilation rates of 1, 2, 4, and 8 can be set, allowing each branch to perceive the input feature map with different receptive field sizes. Smaller dilation rates help extract local detail features, while larger dilation rates capture broader contextual information, thus comprehensively covering multi-scale features from fine textures to macroscopic structures. This differentiated dilation rate setting allows the model to better adapt to the varying geometry and size of landslide areas.

[0073] Channel attention mechanisms aim to dynamically learn and adjust the importance of different feature channels. Their implementation typically involves two steps: feature compression and feature activation. Feature compression, through operations such as global average pooling, compresses the two-dimensional feature map of each channel into a single numerical value to obtain the global spatial information of that channel. For example, for a feature map of size H x W x C, global average pooling transforms it into a 1 x 1 x C vector, where each element represents the average activation intensity of the corresponding channel. Subsequently, the feature activation step utilizes one or more fully connected layers (or 1x1 convolutional layers) and activation functions (such as the sigmoid function) to learn and generate weights for each channel based on the compressed global information. These weights, ranging from 0 to 1, indicate the importance of each channel. Finally, these weights are element-wise multiplied with the corresponding channels of the original feature map, thereby dynamically weighting the features of different channels. In this way, the model can adaptively strengthen feature channels that are more important to the current task based on the input content, while suppressing less important channels, ensuring the effectiveness and accuracy of feature fusion.

[0074] Through the aforementioned technical solutions, the attention-based hollow pyramid module employs differentiated void ratio settings within a depth-separable hollow convolutional structure during multi-scale feature fusion. This allows each parallel branch to capture richer and more comprehensive multi-scale feature information, effectively avoiding information omissions that might occur with a single receptive field, thereby enhancing the model's ability to identify landslide areas of different sizes and shapes. Simultaneously, the channel attention mechanism performs feature squeezing and activation operations on the channel dimensions of feature maps at each scale, achieving precise dynamic allocation of feature channel weights. This intelligently strengthens effective feature channels beneficial for landslide identification while weakening interference from background noise or irrelevant feature channels. This combination enables the model to focus more on key information when fusing multi-scale features, significantly improving the accuracy and robustness of landslide segmentation, especially in handling complex and variable road landslide scenarios, providing more reliable and refined detection results.

[0075] In decoders, traditional methods often face a trade-off between computational efficiency and detail preservation when performing large-scale upsampling on the fused feature maps to restore the original image resolution. Simple large-scale upsampling can lead to blurred results or introduce artifacts, especially when accurate identification of landslide boundaries is required. How to efficiently and effectively restore the feature map resolution while effectively utilizing shallow features to supplement details is a key challenge.

[0076] To address this, this embodiment further proposes that in the aforementioned decoder, the lightweight dynamic upsampler employs a structure where transposed convolution and pixel recombination work in tandem. The transposed convolution enlarges the feature map size, while pixel recombination optimizes feature resolution. During the upsampling process, an adaptive threshold is set based on the grayscale difference of the feature pixels to filter out low-value redundant pixels. The shallow features concatenated with the upsampling results are low-level texture features from the encoder's earlier output, used to supplement detailed information.

[0077] Specifically, the lightweight dynamic upsampling mechanism combines transposed convolution and pixel rearrangement to achieve efficient and high-quality feature map resolution enhancement. Transposed convolution, also known as deconvolution, works by inserting zeros between pixels in the input feature map and applying learnable kernels to perform convolution operations, thereby enlarging the feature map's spatial size. This effectively maps low-resolution features to a high-resolution space, laying the foundation for subsequent finer processing. Pixel rearrangement, or sub-pixel convolution, is a more refined resolution optimization technique. It rearranges the channel information of a low-resolution feature map, transforming it into high-resolution spatial information. For example, a feature map with C channels, H rows, and W columns can be transformed into a feature map with (C / r^2) channels and (H... r) line, (W) The feature map is represented by column r, where r is the upsampling factor. This method can effectively reduce the checkerboard effect that may be introduced by transposed convolution, generating smoother upsampling results with better detail preservation.

[0078] To further optimize the upsampling process, this embodiment proposes setting an adaptive threshold based on the gray-level differences of feature pixels to filter out low-value redundant pixels. During feature map upsampling, not all pixels have equal importance to the final landslide segmentation result. By calculating the local gray-level differences of feature pixels (e.g., through gradient or variance analysis), regions with insignificant changes and low information content can be identified and classified as "low-value redundant pixels." The adaptive threshold is not a fixed value but is dynamically adjusted based on the features or scene determination results of the current processing area. For example, in areas with drastic feature changes, the threshold can be appropriately lowered to retain more details; while in areas with flat features, the threshold can be appropriately increased to filter out more redundant information. By setting the feature values ​​of these low-value redundant pixels to zero or reducing their weights, the subsequent computational load can be effectively reduced, allowing the model to focus more on the features of key regions.

[0079] Furthermore, the shallow features concatenated with the upsampled results specifically refer to the low-level texture features output in the early stages of the encoder. The encoder extracts features of varying levels of abstraction at different levels. Early levels typically retain rich original image details, such as edges, textures, and corners. While these low-level texture features are enhanced in semantics after deeper processing by the encoder, spatial resolution and detail may be lost. Therefore, concatenating and fusing these early low-level texture features with the upsampled high-level semantic features in the decoder effectively recovers the lost detail, resulting in a final landslide segmentation map that maintains high semantic accuracy while possessing finer boundary and texture details.

[0080] Through the above technical solution, the lightweight dynamic upsampler adopts a structure that combines transposed convolution and pixel recombination, enabling efficient upsampling at large magnitudes. Transposed convolution handles the initial size enlargement, while pixel recombination refines and optimizes feature resolution, effectively avoiding blurring or artifacts that may occur with traditional single upsampling methods. Furthermore, an adaptive threshold is set based on the grayscale difference of feature pixels, intelligently filtering out low-value redundant pixels, further reducing the computational burden and allowing the model to focus more on features in key regions. Simultaneously, the low-order texture features output from the encoder are stitched together, effectively compensating for the loss of spatial detail in high-order features. This ensures that the final landslide segmentation map maintains semantic accuracy while possessing high-precision boundary details, thus significantly improving the precision and reliability of landslide detection.

[0081] In some of the embodiments described above, a method is proposed to query the feature weights of similar historical landslide cases by integrating a historical record module, and to optimize the attention allocation logic based on these feature weights. However, in practical applications, how to efficiently and accurately store and retrieve massive amounts of historical landslide case data, and ensure that the selected historical experience can accurately and effectively guide the allocation of the current attention mechanism to cope with complex and ever-changing landslide scenarios, remains a technical challenge that needs to be concretized and improved.

[0082] To address this, this embodiment further proposes that the historical record module integrates a landslide disaster feature database, which is specifically used to store geographical, meteorological, and feature weight data for historical landslide cases. This landslide disaster feature database is a core component of the historical record module, designed to systematically store relevant information about past landslide events. It includes not only geographical location information of the landslide occurrence, such as latitude and longitude, altitude, slope, and geological type, but also meteorological conditions before and after the event, such as rainfall, temperature, and humidity, as well as feature weight data learned during model training or inference for that landslide case, or set by expert experience. These feature weight data can be a quantitative representation of which image features (such as texture, color, and shape) are more critical for landslide identification under specific geological and meteorological conditions. The database can be constructed using a relational database (such as PostgreSQL or MySQL) or a non-relational database (such as MongoDB), the specific choice depending on the data volume, query complexity, and system performance requirements.

[0083] When searching for similar cases, the system uses geographic metadata and meteorological data of the target area as core search keywords to filter cases whose feature similarity meets a preset threshold. Then, a weighted average algorithm is used to fuse the feature weights of these cases, thereby optimizing the current attention allocation logic. Specifically, when landslide detection is needed for a new target area, the system extracts the geographic metadata (such as geographical location and topographic features) and real-time meteorological data of that target area. This data is used as the "key" to retrieve similar historical cases in the landslide hazard feature database. For example, a matching query can be performed in the database based on the target area's latitude and longitude, altitude range, geological type, and current rainfall and temperature. The retrieval algorithm can use distance-based similarity measures (such as Euclidean distance and cosine similarity) or rule-based matching. Not all retrieved historical cases have equal reference value; to ensure the effectiveness of the selected cases, further filtering of the search results is required. Feature similarity refers to the degree of matching between the target area and historical cases in key features such as geography and meteorology. The preset threshold is a configurable parameter used to define the degree of "similarity." For example, cases with a geographical feature similarity greater than 0.8 and a meteorological feature similarity greater than 0.7 can be considered similar cases. This screening mechanism ensures that only historical experience highly relevant to the current detection scenario is considered, avoiding interference from irrelevant data on the attention allocation logic. After screening, a set of historical landslide cases similar to the target area is obtained, each case with its corresponding feature weight data. To comprehensively utilize these historical experiences, this embodiment uses a weighted average algorithm to fuse these feature weights. The weight of the weighted average can be dynamically determined based on the similarity between each similar case and the target area; that is, the higher the similarity, the greater the proportion of its feature weight in the fusion. For example, if case A has a similarity of 0.9 with the target area and case B has a similarity of 0.7, then the contribution of case A's feature weight in the fusion result will be greater than that of case B. This approach can more finely integrate historical experience, forming a more representative feature weight set applicable to the current scenario. The fused feature weight set is used to adjust the allocation logic of the channel attention mechanism in the attention void pyramid module. Specifically, channel attention mechanisms typically learn to assign weights to different feature channels to highlight important features. By introducing fused historical feature weights, the initial weights or learning process of the channel attention mechanism can be guided and corrected. For example, the fused feature weights can be used as prior knowledge to weight the output of the channel attention mechanism, or they can be included as part of the loss function to encourage the attention mechanism to favor key feature channels indicated by historical experience.

[0084] Through the above technical solution, this embodiment clarifies the internal structure and working mechanism of the historical data module. Specifically, it systematically stores geographical, meteorological, and feature weight data of historical landslide cases through an integrated landslide disaster feature database. When querying similar cases, the geographic metadata and meteorological data of the target area are used as core search keywords, and cases with feature similarity meeting a preset threshold are rigorously selected, ensuring the relevance and effectiveness of the referenced historical data. Furthermore, by fusing the feature weights of these similar cases using a weighted average algorithm, historical experience can be integrated more precisely and reasonably, allowing historical cases with higher similarity to the current scenario to contribute more significantly to the optimization of the attention allocation logic. This mechanism effectively solves the problem of how to efficiently and accurately utilize historical data to optimize attention allocation, enabling the attention void pyramid module to more accurately focus on key features related to landslides, thereby significantly improving the accuracy and robustness of road landslide detection. Especially when facing landslide scenarios with diverse and complex terrain conditions, the model's adaptability and recognition capabilities are enhanced.

[0085] Furthermore, embodiments of this application also propose a computer-readable storage medium storing a lightweight intelligent detection program for road landslides. When the lightweight intelligent detection program for road landslides is executed by a processor, it implements the steps of the lightweight intelligent detection method for road landslides as described above.

[0086] Reference Figure 3 , Figure 3 This is a structural block diagram of the first embodiment of the lightweight intelligent detection system for road landslides in this application.

[0087] like Figure 3 As shown in the embodiments of this application, the lightweight intelligent detection system for road landslides includes: The data acquisition module 10 is used to determine the parameter configuration of the detection model based on the geographic metadata of the target area and the user role profile information, and at the same time complete the scene judgment and output the scene judgment result; the user role profile information includes the weight settings of the user's preference for detection accuracy or detection speed. The dataset construction module 20 is used to acquire high-resolution optical satellite images of the target area, fuse them with real-time meteorological data and image data in the historical image database, define the gradual transition boundary of the landslide area through multi-time comparison analysis, complete pixel-level annotation based on the gradual transition boundary, and then construct an enhanced landslide segmentation dataset. The preprocessing module 30 is used to preprocess the image data of the enhanced landslide segmentation dataset and input the preprocessed image data into the encoder. The encoder uses a lightweight MobileNetV4 network as the backbone network, performs efficient feature extraction and outputs high-order feature maps through the general inverted bottleneck module of the network, and dynamically adjusts the parameters of the encoder convolutional layer based on the scene determination result. The input module 40 is used to input the high-order feature map into the attention-diffuse pyramid module. The parallel branches of this module capture features at different scales. Each parallel branch uses a 1×1 convolution and a depthwise separable dilated convolution structure to extract scale features. The scale features captured by each parallel branch are dynamically weighted and fused through a channel attention mechanism. At the same time, the integrated historical record module is used to query the feature weights of similar historical landslide cases, and the attention allocation logic is optimized based on the feature weights. The output module 50 is used to input the feature map fused by the attention-drained pyramid module into the decoder. In the decoder, a large upsampling operation is performed by a lightweight dynamic upsampling device. The upsampling result is then spliced ​​and fused with the shallow features output by the encoder. Based on the scene determination result, the corresponding post-processing strategy is selected, and the post-processing operation is performed on the spliced ​​and fused feature map to output a landslide segmentation map for calculating the landslide disaster area.

[0088] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solution of this application. In specific applications, those skilled in the art can make settings as needed, and this application does not impose any restrictions on this.

[0089] This embodiment solves the problems of insufficient boundary definition, weakened feature capture, and lack of dynamic adjustment in the prior art by using dynamic parameter configuration, multi-source data fusion to define gradual boundaries, and optimizing feature extraction and fusion mechanisms. It has high detection accuracy, fast response speed, strong environmental adaptability, and user customization capabilities.

[0090] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this application. In practical applications, those skilled in the art can select some or all of it to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0091] In addition, for technical details not described in detail in this embodiment, please refer to the lightweight intelligent detection method for road landslides provided in any embodiment of this application, which will not be repeated here.

[0092] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0093] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0094] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0095] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A lightweight intelligent detection method for road landslides, characterized in that, include: The parameter configuration of the detection model is determined based on the geographic metadata of the target area and the user role profile information. At the same time, the scene determination is completed and the scene determination result is output. The user role profile information includes weight settings for the user's preference for detection accuracy or detection speed; High-resolution optical satellite images of the target area are acquired and fused with real-time meteorological data and image data from historical image databases. The gradual transition boundary of the landslide area is defined through multi-temporal comparative analysis. Pixel-level annotation is completed based on the gradual transition boundary, and then an enhanced landslide segmentation dataset is constructed. The image data of the enhanced landslide segmentation dataset is preprocessed, and the preprocessed image data is input into the encoder. The encoder uses a lightweight MobileNetV4 network as the backbone network, performs efficient feature extraction and outputs high-order feature maps through the general inverted bottleneck module of the network, and dynamically adjusts the parameters of the encoder convolutional layer based on the scene determination result. The high-order feature map is input into the attention-diffuse pyramid module, and features of different scales are captured through the parallel branches of the module. Each parallel branch uses a 1×1 convolution and a depthwise separable dilated convolution structure to extract scale features. The scale features captured by each parallel branch are dynamically weighted and fused through a channel attention mechanism. At the same time, the feature weights of similar historical landslide cases are queried based on the integrated historical record module, and the attention allocation logic is optimized based on the feature weights. The feature map fused by the attention-drilled hollow pyramid module is input into the decoder. In the decoder, a large upsampling operation is performed through a lightweight dynamic upsampling unit. The upsampling result is then concatenated and fused with the shallow features output by the encoder. Based on the scene determination results, the corresponding post-processing strategy is selected, and the post-processing operation is performed on the spliced ​​and fused feature map to output the landslide segmentation map for the calculation of the landslide disaster area.

2. The landslide segmentation method according to claim 1, characterized in that, The process of generating the user role profile information is as follows: First, collect the user's historical operation records, extract the user's historical detection task types, parameter adjustment trajectories and task result feedback data from the historical operation records, and then analyze the extracted data through a feature clustering algorithm to quantify the user's preference for detection accuracy and speed, and finally generate user role profile information containing the preference weight settings.

3. The landslide segmentation method according to claim 1, characterized in that, During the construction of the dataset, the fusion processing of real-time meteorological data and high-resolution optical satellite images adopts a weighted fusion algorithm based on data temporal consistency. This algorithm first aligns the temporal dimensions of meteorological data acquisition time and satellite image capture time, and then dynamically assigns weights according to the time difference between the two—the smaller the time difference, the higher the weight of the meteorological data, so as to ensure the spatiotemporal correlation of the fused data.

4. The landslide segmentation method according to claim 1, characterized in that, After the encoder extracts features through the general inverted bottleneck module, it immediately performs batch normalization on the output feature map to improve the stability of feature extraction. The dynamically adjusted encoder convolutional layer parameters include the number of convolutional kernels, the dilation rate, and the stride. The parameter adjustment range is determined according to the terrain complexity level corresponding to the scene judgment result. The more complex the terrain, the greater the adjustment range to adapt to the feature capture requirements.

5. The landslide segmentation method according to claim 1, characterized in that, In the multi-scale feature fusion process, the depth-separable dilated convolutional structure adopts different dilation values ​​to capture features at different scales. The channel attention mechanism performs feature squeezing and activation operations on the channel dimensions of feature maps at various scales, thereby strengthening the weights of effective feature channels and weakening ineffective channels, achieving precise dynamic allocation of feature weights.

6. The landslide segmentation method according to claim 1, characterized in that, The lightweight dynamic upsampler employs a structure that combines transposed convolution with pixel recombination—transposed convolution enlarges the feature map size, pixel recombination optimizes the feature resolution, and an adaptive threshold is set based on the grayscale difference of feature pixels during the upsampling process to filter out low-value redundant pixels. The shallow features concatenated with the upsampling results are low-level texture features from the early output of the encoder, used to supplement detailed information.

7. The landslide segmentation method according to claim 1, characterized in that, The historical record module integrates a landslide disaster feature database, which stores geographical, meteorological, and feature weight data of historical landslide cases. When querying similar cases, the system uses the geographical metadata and meteorological data of the target area as core search keywords to filter out cases whose feature similarity meets a preset threshold. Then, the feature weights of the cases are fused through a weighted average algorithm to optimize the current attention allocation logic.

8. A lightweight intelligent detection system for road landslides, characterized in that, The lightweight intelligent detection system for road landslides includes: The data acquisition module is used to determine the parameter configuration of the detection model based on the geographic metadata of the target area and the user role profile information, and at the same time complete the scene determination and output the scene determination result; the user role profile information includes the weight settings of the user's preference for detection accuracy or detection speed. The dataset construction module is used to acquire high-resolution optical satellite images of the target area, fuse them with real-time meteorological data and image data in the historical image database, define the gradual transition boundary of the landslide area through multi-temporal comparison analysis, complete pixel-level annotation based on the gradual transition boundary, and then construct an enhanced landslide segmentation dataset. The preprocessing module is used to preprocess the image data of the enhanced landslide segmentation dataset and input the preprocessed image data into the encoder. The encoder uses a lightweight MobileNetV4 network as the backbone network, performs efficient feature extraction and outputs high-order feature maps through the network's general inverted bottleneck module, and dynamically adjusts the parameters of the encoder's convolutional layers based on the scene determination results. The input module is used to input the high-order feature map into the attention-diffuse pyramid module. The parallel branches of this module capture features at different scales. Each parallel branch uses a 1×1 convolution and a depthwise separable dilated convolution structure to extract scale features. The scale features captured by each parallel branch are dynamically weighted and fused through a channel attention mechanism. At the same time, the integrated historical record module is used to query the feature weights of similar historical landslide cases, and the attention allocation logic is optimized based on the feature weights. The output module is used to input the feature map fused by the attention-drained pyramid module into the decoder. In the decoder, a large upsampling operation is performed through a lightweight dynamic upsampling unit. The upsampling result is then spliced ​​and fused with the shallow features output by the encoder. Based on the scene determination result, the corresponding post-processing strategy is selected, and the post-processing operation is performed on the spliced ​​and fused feature map to output a landslide segmentation map for calculating the landslide disaster area.

9. A computer device, characterized in that, The device includes a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Includes instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.

Images