A lightweight landslide image screening method, device, equipment, and medium with integrated gating.

Abstract

This application discloses a lightweight, gated landslide image screening method, apparatus, device, and medium, relating to the field of image data processing technology. The method includes: acquiring and preprocessing satellite image slices, then processing them through a pre-defined gated lightweight model containing a backbone network, a neck network, and a detection head; performing multi-scale feature extraction, high- and low-order feature fusion, and target detection and judgment; and outputting landslide-containing slices. This scheme employs enhanced spatial pyramid pooling, high- and low-order gated fusion, and an improved fast dual-branch module, effectively preserving detailed landslide features, reducing computational load and parameter count, and quickly and accurately screening satellite image slices containing landslide targets, meeting the needs of rapid screening of massive images and emergency monitoring.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a lightweight landslide image screening method, apparatus, device, and medium with fusion gating. Background Technology

[0002] Currently, landslide image screening mainly adopts two technical approaches: one is a global discrimination method based on image classification, and the other is a landslide target recognition method based on a general lightweight target detection network. Some solutions use lightweight models such as the YOLO series, which extract features through the backbone network, fuse features through the neck network, and output prediction results through the detection head to achieve preliminary screening of landslide images.

[0003] Traditional image classification models suffer from significant spatial information loss due to global pooling and multiple downsampling, making it easy for small-scale landslide features to be submerged by the background, resulting in a high false negative rate. General lightweight target detection models suffer from the problem of losing detailed features in their spatial pyramid pooling module, and conventional feature fusion methods cannot dynamically adapt to high and low order feature weights. Furthermore, the feature extraction module has computational redundancy, leading to insufficient recall and difficulty in balancing inference efficiency and accuracy in landslide screening tasks. Consequently, they cannot meet the needs of rapid and accurate screening of sparse landslides along highways.

[0004] Therefore, how to quickly and accurately select satellite image slices containing landslide targets from a large number of satellite image slices has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a lightweight landslide image screening method, apparatus, equipment, and medium with integrated gating, aiming to solve the technical problem of how to quickly and accurately screen satellite image slices containing landslide targets.

[0006] To achieve the above objectives, this application proposes a lightweight landslide image screening method with fusion gating, comprising: Obtain optical satellite image slices to be filtered; The optical satellite image slices are preprocessed to obtain standardized satellite images; The standardized satellite images are input into a pre-gated lightweight model for processing to obtain landslide screening results. The pre-gated lightweight model includes a backbone network, a neck network, and a detection head. The backbone network includes a convolutional downsampling layer, a continuous deep feature extraction layer, and an enhanced spatial pyramid pooling layer. The neck network includes a high- and low-order feature gating fusion module and an improved fast dual-branch feature extraction module. The detection head includes a category prediction branch and a location regression branch. Output a set of satellite image slices containing landslide targets based on the landslide screening results; The step of inputting the standardized satellite image into a preset gated lightweight model for processing to obtain landslide screening results includes: Multi-scale feature extraction is performed on the standardized satellite images using a backbone network to obtain multi-level feature maps. The multi-level feature map is fused with high and low-order features and rapidly reorganized by the neck network to obtain an enhanced fused feature map. The enhanced fusion feature map is analyzed by a detection head to detect landslide targets and determine their confidence levels, thus obtaining landslide screening results.

[0007] In one embodiment, the step of extracting multi-scale features from the standardized satellite image through a backbone network to obtain a multi-level feature map includes: The standardized satellite image is subjected to shallow feature extraction and resolution reduction through the convolutional downsampling layer to obtain a shallow feature map. The shallow feature map is subjected to layer-by-layer deep semantic feature extraction through the continuous deep feature extraction layer to obtain a deep semantic feature map; The enhanced spatial pyramid pooling layer is used to extract local salient features and compensate for global background information in the deep semantic feature map to obtain an enhanced backbone feature map. The shallow feature map, the deep semantic feature map, and the enhanced backbone feature map are output as a multi-level feature map.

[0008] In one embodiment, the step of extracting local salient features and compensating for global background information from the deep semantic feature map through the enhanced spatial pyramid pooling layer to obtain an enhanced backbone feature map includes: The deep semantic feature map is reduced in dimensionality by channel reduction through the first convolutional layer to obtain a dimensionality-reduced feature map. Local salient features are extracted from the reduced feature map using a max pooling layer to obtain a local salient feature map; Global background information is extracted from the reduced-dimensional feature map using an average pooling layer to obtain a global background feature map. The reduced feature map, the local salient feature map, and the global background feature map are concatenated along the channel dimension, and channel recovery is performed through a second convolutional layer to obtain an enhanced backbone feature map.

[0009] In one embodiment, the step of fusing high- and low-order features and rapidly recombining features from the multi-level feature map using a neck network to obtain an enhanced fused feature map includes: The high-order semantic features and low-order spatial features in the multi-level feature map are dynamically weighted and fused and texture enhanced by the high-order and low-order feature gating fusion module to obtain the gating fusion feature map. By improving the fast dual-branch feature extraction module, partial channel spatial convolution and cross-channel information recombination are performed on the gated fusion feature map to obtain a fast recombined feature map. The fast reconstructed feature map is aggregated by multi-scale feature aggregation through upsampling and lateral connection to obtain an aggregated feature map; The aggregated feature map is used as the enhanced fusion feature map.

[0010] In one embodiment, the step of dynamically weighting and fusing high-order semantic features and low-order spatial features in the multi-level feature map and enhancing texture through a high-low order feature gating fusion module to obtain a gated fused feature map further includes: The low-order spatial features are channel-upgraded by using a dimension-up convolutional layer, and then transformed low-order features are obtained by batch normalization and activation function processing. Global average pooling is performed on the higher-order semantic features and the transformed lower-order features respectively to obtain the first channel description vector and the second channel description vector. The first channel description vector and the second channel description vector are concatenated and then normalized to obtain the competitive gating weights; Based on the competitive gating weights, the higher-order semantic features and the transformed lower-order features are fused channel by channel to obtain a preliminary fused feature map; Channel normalization is performed on the preliminary fused feature map to obtain a normalized feature map; The low-frequency approximation components are extracted from the low-order spatial features by using a max pooling layer. The high-frequency residual component is obtained by subtracting the low-frequency approximation component from the low-order spatial features. The high-frequency residual components are adaptively weighted by depthwise convolution and activation functions to obtain weighted high-frequency components. The weighted high-frequency components and low-order spatial features are residually connected to obtain texture enhancement features; The normalized feature map and texture enhancement feature are concatenated along the channel dimension to obtain a gated fusion feature map.

[0011] In one embodiment, the step of performing partial channel spatial convolution and cross-channel information reconstruction on the gated fusion feature map using an improved fast dual-branch feature extraction module to obtain a fast reconstructed feature map includes: Spatial information is extracted from the first channel of the gated fusion feature map by using partial convolutional layers to obtain a partial spatial feature map; The second channel of the gated fusion feature map is used as a bypass feature map, wherein the first channel and the second channel are all channels of the gated fusion feature map; By performing cross-channel information fusion and dimensional transformation on the partial spatial feature map through a first bottleneck convolutional layer and a second bottleneck convolutional layer in series, a bottleneck fusion feature map is obtained. The partial spatial feature map and the bottleneck fusion feature map are added together to obtain a fast reconstructed feature map.

[0012] In one embodiment, the step of performing landslide target detection and confidence determination on the enhanced fusion feature map using a detection head to obtain landslide screening results includes: The enhanced fusion feature map is subjected to landslide category probability prediction through the category prediction branch to obtain the category prediction result; The enhanced fused feature map is regressed by the location regression branch to obtain the target bounding box coordinate regression and the location prediction result. Based on the category prediction results and the location prediction results, a confidence level is determined, and prediction results with a confidence level greater than a preset confidence threshold are identified as valid landslide detection results. A landslide screening result is generated based on the effective landslide detection results, wherein the landslide screening result includes slice identification information containing landslide targets.

[0013] Furthermore, to achieve the above objectives, this application also proposes a lightweight landslide image screening device with fusion gating, the lightweight landslide image screening device comprising: The acquisition module is used to acquire optical satellite image slices to be screened; The data preprocessing module is used to preprocess the optical satellite image slices to obtain standardized satellite images; The result module is used to input the standardized satellite image into a preset gated lightweight model for processing to obtain landslide screening results. The preset gated lightweight model includes a backbone network, a neck network, and a detection head. The backbone network includes convolutional downsampling layers, continuous deep feature extraction layers, and enhanced spatial pyramid pooling layers. The neck network includes a high- and low-order feature gating fusion module and an improved fast dual-branch feature extraction module. The detection head includes a category prediction branch and a location regression branch. The module is also used to perform multi-scale feature extraction on the standardized satellite image through the backbone network to obtain multi-level feature maps; to perform high- and low-order feature fusion and fast feature recombination on the multi-level feature maps through the neck network to obtain enhanced fused feature maps; and to perform landslide target detection and confidence determination on the enhanced fused feature maps through the detection head to obtain landslide screening results. The filtering module is used to output a set of satellite image slices containing landslide targets based on the landslide filtering results.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the fusion-gated lightweight landslide image screening method described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the fusion-gated lightweight landslide image screening method described above.

[0016] This application acquires and preprocesses satellite image slices, then processes them using a pre-gated lightweight model containing a backbone network, a neck network, and a detection head. After multi-scale feature extraction, high- and low-order feature fusion, and target detection, it outputs slices containing landslide targets. This scheme employs enhanced spatial pyramid pooling, high- and low-order gating fusion, and an improved fast dual-branch module, effectively preserving detailed landslide features, reducing computational load and parameter count, and quickly and accurately selecting satellite image slices containing landslide targets, meeting the needs of rapid screening of massive amounts of imagery and emergency monitoring. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the first embodiment of the lightweight landslide image screening method with integrated gating in this application. Figure 2 This is a flowchart illustrating the second embodiment of the landslide image screening method with integrated gating and lightweight design of this application. Figure 3 This is a schematic diagram of the modular structure of the lightweight landslide image screening device with integrated gating in this application. Figure 4 This is a schematic diagram of the hardware operating environment involved in the landslide image screening method with integrated gating and lightweight design in this application embodiment.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] Currently, landslide image screening mainly adopts two technical approaches: one is a global discrimination method based on image classification, and the other is a landslide target recognition method based on a general lightweight target detection network. Some solutions use lightweight models such as the YOLO series, which extract features through the backbone network, fuse features through the neck network, and output prediction results through the detection head to achieve preliminary screening of landslide images.

[0023] Traditional image classification models suffer from significant spatial information loss due to global pooling and multiple downsampling, making it easy for small-scale landslide features to be obscured by the background, resulting in a high false negative rate. General lightweight target detection models suffer from detail loss in their spatial pyramid pooling modules, conventional feature fusion methods cannot dynamically adapt to high and low-order feature weights, and feature extraction modules have computational redundancy. This leads to insufficient recall and an inability to balance inference efficiency and accuracy in landslide screening tasks, making them unsuitable for the rapid and accurate screening requirements of sparse landslides along highways. Therefore, how to quickly and accurately screen satellite image slices containing landslide targets from a large number of satellite image slices has become an urgent problem to be solved.

[0024] Based on the above, this application also provides a lightweight landslide image screening method with fusion gating, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the landslide image screening method with integrated gating and lightweight design, as described in this application.

[0025] In this embodiment, the fusion-gated lightweight landslide image screening method includes steps S10~S40: Step S10: Obtain the optical satellite image slices to be screened.

[0026] It should be noted that the optical satellite image slices to be screened are fixed-size image units segmented from high-resolution optical satellite imagery along the highway. These slices are used for the initial identification of landslide targets and are the basic data units for carrying out the image screening task. These slices originate from the regular segmentation of the original satellite imagery, retaining only localized ground feature information.

[0027] Specifically, the process begins by acquiring raw image data from optical satellite remote sensing data sources. This raw image data is typically high-resolution, multi-band satellite imagery containing red, green, and blue visible light bands. Next, the raw image data is gridded according to a preset tile size (e.g., 512 pixels × 512 pixels) and a preset overlap pixel value (e.g., 50 pixels), resulting in several optical satellite image tiles to be screened. The gridding employs nearest-neighbor interpolation resampling to maintain pixel value stability. Finally, a unique identifier is assigned to each optical satellite image tile to be screened, and a tile index table is established. This allows subsequent processing steps to track and backtrack to the corresponding tile based on the unique identifier. This approach is necessary because single-scene optical satellite images cover a large area and generate massive amounts of data. Directly processing the entire image would lead to memory overflow and low computational efficiency. By slicing the large image into fixed-size small image patches, the requirements of deep learning models for input size are met, parallel batch processing is facilitated, and preserving overlapping areas prevents landslide targets from being truncated at tile boundaries and thus missed.

[0028] Step S20: Preprocess the optical satellite image slices to obtain standardized satellite images.

[0029] Specifically, firstly, radiometric calibration is performed on the optical satellite image slices, converting the original digital quantization values ​​into physical radiance values ​​to ensure comparability of image data acquired by different sensors and at different times. Secondly, geometric correction is performed on the radiometrically calibrated images, eliminating geometric distortions caused by terrain undulations through ground control points and digital elevation models, and uniformly mapping the image coordinates to a preset geographic coordinate system (in this embodiment, the World Geodetic System 1984 coordinate system + Universal Transverse Mercator Projection 6-degree zone) to obtain a geometrically corrected image. Then, pixel value normalization is performed on the geometrically corrected image, linearly mapping the pixel values ​​to [0, 1] or a preset standardized distribution (mean 0, variance 1) to obtain a normalized image. Finally, the normalized image is adjusted to a preset input size (640 pixels × 640 pixels) to obtain a standardized satellite image. This is done because optical satellite imagery is affected by factors such as sensor characteristics, atmospheric conditions, and solar altitude angle during the acquisition process, resulting in radiometric differences and geometric distortions in pixel values. Radiometric calibration eliminates sensor response differences, geometric fine correction ensures the spatial accuracy of landslide locations, normalization stabilizes the distribution of model input data to avoid gradient explosion or convergence difficulties, and size adjustment adapts to the input layer requirements of the pre-gated lightweight model to ensure that the subsequent feature extraction network can process efficiently.

[0030] Step S30: Input the standardized satellite image into the preset gated lightweight model for processing to obtain the landslide screening results.

[0031] It should be noted that the pre-gated lightweight gating model is a deep learning model for rapid screening of landslide images. This model is based on a lightweight target detection structure and incorporates a gated feature fusion mechanism to achieve efficient identification of landslide targets. The pre-gated lightweight gating model consists of a backbone network, a neck network, and a detection head. The backbone network is the network structure responsible for feature extraction. This network extracts semantic and spatial features of different scales from the input image through multiple convolutional and pooling operations. The backbone network includes convolutional downsampling layers, continuous deep feature extraction layers, and enhanced spatial pyramid pooling layers. The neck network is the feature fusion structure connecting the backbone network and the detection head. This network integrates feature information from different levels, balancing semantic information and spatial resolution. The neck network includes a high- and low-order feature gating fusion module and an improved fast dual-branch feature extraction module. The detection head is the execution structure responsible for outputting the prediction results. This structure determines the target category and location based on the fused features. The detection head includes a category prediction branch and a location regression branch. Specifically, the convolutional downsampling layer is a convolutional layer used to reduce the size of the feature map. This layer reduces the feature map resolution and expands the receptive field through convolution operations. It provides the foundation for deep feature extraction. The continuous deep feature extraction layer is a network layer used to continuously extract high-level semantic information. This layer enhances feature representation capabilities through stacked convolutions and activation operations. This layer improves the model's ability to distinguish complex features. The enhanced spatial pyramid pooling layer is a module for enhancing features across multiple receptive fields. This layer supplements detailed texture information while retaining fast computation characteristics. This layer alleviates the feature loss problem caused by traditional pooling. The high-low order feature gating fusion module is used to dynamically fuse high- and low-level features. This module adaptively allocates feature weights through a gating mechanism, highlighting effective information. This module improves the accuracy and adaptability of feature fusion. The improved fast dual-branch feature extraction module is used for lightweight feature extraction. This module uses partial convolution and bottleneck structures to reduce computational cost and parameter count. This module ensures model inference efficiency. The category prediction branch is used to determine whether a landslide target exists in the image. This branch outputs the probability value that the corresponding region contains a landslide. This branch provides a category basis for landslide determination. The location regression branch is used to predict the location of the landslide target. This branch outputs the coordinates of the target's bounding box. This branch provides the location basis for landslide localization.

[0032] Specifically, firstly, standardized satellite images are input into the backbone network of a pre-gated lightweight model. A convolutional downsampling layer performs shallow feature extraction and resolution reduction on the standardized satellite images, resulting in a shallow feature map that preserves landslide edge texture and local spatial details. Then, a series of deep feature extraction layers extract deep semantic features from the shallow feature map, resulting in a deep semantic feature map. This deep semantic feature map contains abstract category information about the landslide body and background features. Next, an enhanced spatial pyramid pooling layer extracts local salient features and compensates for global background information in the deep semantic feature map, resulting in an enhanced backbone feature map. This enhanced backbone feature map combines local texture sharpness with global context awareness. Finally, the shallow feature map, deep semantic feature map, and enhanced backbone feature map are used as multiple... The hierarchical feature maps are input into the neck network. A high-low order feature gating fusion module dynamically weights and fuses high-order semantic features and low-order spatial features in the multi-level feature maps, enhancing texture to obtain a gated fusion feature map. The dynamic weighting fusion adaptively assigns contribution weights to high- and low-order features based on channel semantic importance, avoiding information redundancy caused by indiscriminate superposition. Then, an improved fast dual-branch feature extraction module performs partial channel spatial convolution and cross-channel information recombination on the gated fusion feature map, resulting in a fast reconstructed feature map. Partial channel spatial convolution only performs spatial operations on some channels, while other channels are directly bypassed, reducing redundant computation. Finally, a detection head performs category prediction and location regression on the fast reconstructed feature map to obtain landslide screening results. These results include information determining whether a slice contains a landslide target. This approach is taken because landslide disasters exhibit spatial sparseness in satellite imagery, characterized by "wide distribution across roads and scattered landslide bodies." Furthermore, the area of ​​landslide bodies is relatively small compared to the overall slice size. The global pooling operation of traditional image classification models leads to irreversible loss of spatial information. After multiple downsampling, weak landslide feature signals are easily submerged by backgrounds such as large areas of forest, bare land, and farmland, resulting in numerous missed detections. In contrast, the pre-gated gated lightweight model retains spatial location information using a target detection paradigm and directly regresses the target bounding box through a local receptive field mechanism. This effectively captures local texture abrupt changes and morphological features of landslide bodies. Simultaneously, it dynamically filters effective feature channels through a gated adaptive fusion mechanism and reduces the number of parameters and computational overhead through lightweight module design. This allows for rapid inference while maintaining high recall, meeting the engineering requirements for real-time screening of massive satellite image slices.

[0033] Furthermore, to verify the effectiveness of the pre-gated lightweight YOLO (GL-YOLO) model proposed in this embodiment, a dedicated experimental environment was constructed and a landslide image screening dataset was prepared. The landslide image screening model dataset consists of two dataset components: an open-source dataset of a certain city and a self-built High-resolution Optical Satellite Landslide Dataset (HOS-LSD). The city dataset contains 2773 images, of which 770 contain landslides and 2003 do not. The HOS-LSD dataset is composed of two satellite images with a spatial resolution of 0.5 meters, covering a total area of ​​175.69 square kilometers, and contains numerous landslide disasters triggered by heavy rain. The HOS-LSD dataset uses ArcGIS satellite imagery slices into 512×512 pixel images. Nearest neighbor interpolation resampling is employed, and images are compared with historical images from the same season before landslides occurred. The dataset focuses on landslide hazards in mountainous areas and along major transportation routes, resulting in 1218 images containing landslide data and 1667 images without landslide data. The two datasets are then combined and divided into training, validation, and test sets in a 7:2:1 ratio, with 3961 images in the training set, 1132 in the validation set, and 565 in the test set, totaling 5658 images. The model training hyperparameters are set as follows: input image size 640×640 pixels; 100 training epochs; initial learning rate 0.01; and 256 training batches. Ablation experiments aim to quantitatively evaluate the impact of each module on the overall algorithm performance, clarifying the contribution of each component. For the three modules proposed in this application's screening task, ablation experiments are used to quantitatively evaluate both model accuracy and inference efficiency. The experiments strictly followed the controlled variable method, using YOLOv8-n as the baseline model, and evaluated in detail the specific effects of each module in improving detection accuracy and inference speed. During the model training phase, the model utilized training set data to adjust model parameters through backpropagation. Simultaneously, to monitor the model's generalization ability and prevent overfitting, a validation set was introduced during training. The model training was validated on the validation set, and corresponding accuracy scores were calculated. After sufficient training, the final model was tested on the test set. The ablation experiment consisted of five groups, each with six metrics. Recall, precision, and F1 score were used to calculate the model's accuracy performance, while GFLOPs, parameter count, and images inferred per second were used to quantify the model's size and inference speed. Detailed results of the ablation experiment are shown in Table 1.

[0034] Using YOLOv8 as the baseline model, YOLOv8+ESPPF is an improved model that introduces an enhanced spatial pyramid pooling layer on top of YOLOv8. YOLOv8+L2Hblock is an improved model that introduces a high- and low-order feature gating fusion module on top of YOLOv8. YOLOv8+Fc2f is an improved model that introduces an improved fast bi-branch feature extraction module on top of YOLOv8. Accuracy analysis focuses on the two core components: the backbone network and the neck network. In the backbone network, replacing SPPF in YOLOv8 with an enhanced spatial pyramid pooling layer reduces redundant feature extraction performance in the baseline model, improving accuracy by 4.52% while reducing the number of parameters. In the improvements to the neck network, replacing the feature stacking model in the baseline model with a high- and low-order feature gating fusion module fully leverages its unique gating mechanism, enabling more efficient fusion of features at different levels. This achieves a 1.64% improvement in recall and a 4.54% improvement in accuracy, with only a 3% increase in model parameters, resulting in better model performance at the cost of a relatively small increase in parameter count. To improve the inference speed of the model, the improved fast dual-branch feature extraction module was used to replace the c2f in the baseline model. This reduced the number of parameters by about 11%, decreased the recall by 1.65%, and increased the precision by 2.81%.

[0035] Step S40: Output a set of satellite image slices containing landslide targets based on the landslide screening results.

[0036] It should be noted that the landslide screening result is the judgment information output by the pre-gated lightweight model. The landslide screening result includes tile identification information containing landslide targets. This information specifically includes whether a landslide target exists in the image tile, as well as the location and confidence level data of the corresponding landslide. This information is the direct basis for distinguishing between valid and invalid tiles. The set of satellite image tiles containing landslide targets is a whole composed of all image tiles containing landslide targets. This set removes background tiles that do not contain landslides, retaining only the image data that needs to be further refined for segmentation. This set is used to reduce the amount of data in subsequent processing.

[0037] Specifically, firstly, the slice identifier information in the landslide screening results is parsed, and the corresponding optical satellite image slices are located from the slice index table to obtain the set of slices to be output. Secondly, the set of slices to be output is deduplicated to remove duplicate slices caused by preset overlapping areas, resulting in a deduplicated candidate slice set. Then, the deduplicated candidate slice set is sorted according to a preset confidence ranking rule (from high to low) to obtain an ordered candidate slice set. Finally, the ordered candidate slice set is packaged and output to the downstream fine segmentation module, so that the downstream fine segmentation module can perform pixel-level boundary extraction of landslide bodies in the ordered candidate slice set. This is done because there are overlapping areas between adjacent slices during the gridding process, and the same landslide target may be detected by multiple slices at the same time. If deduplication is not performed, the downstream module will repeatedly calculate the same target, resulting in wasted computing power and redundant results. Confidence ranking allows high-reliability slices to enter the subsequent processing stages first. When system resources are limited, the processing time of high-confidence slices can be prioritized, thereby achieving reasonable scheduling of computing resources and optimization of processing efficiency in the connection between massive image screening and fine segmentation.

[0038] This embodiment acquires and preprocesses satellite image slices, then processes them using a pre-gated lightweight model containing a backbone network, a neck network, and a detection head. After multi-scale feature extraction, high- and low-order feature fusion, and target detection, it outputs slices containing landslides. This scheme employs enhanced spatial pyramid pooling, high- and low-order gating fusion, and an improved fast dual-branch module, effectively preserving detailed landslide features, reducing computational load and parameter count, and quickly and accurately selecting satellite image slices containing landslide targets, meeting the needs of rapid screening of massive amounts of imagery and emergency monitoring.

[0039] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The fusion-gated lightweight landslide image screening method further includes steps S201-S203 in step S30: Step S201: Multi-scale feature extraction is performed on the standardized satellite image through the backbone network to obtain multi-level feature maps.

[0040] It should be noted that step S201 includes: First, shallow feature extraction and resolution reduction are performed on the standardized satellite image through a convolutional downsampling layer to obtain a shallow feature map. Specifically, the standardized satellite image is input into the convolutional downsampling layer, and low-level visual features such as image edges, textures, and colors are extracted through convolution operations with a preset kernel size (3 pixels × 3 pixels) to obtain an initial feature map; Second, a downsampling operation with a stride of 2 is performed on the initial feature map to reduce the spatial resolution of the feature map to half of the input size, resulting in a dimensionality-reduced feature map. The dimensionality-reduced feature map retains key spatial structure information while reducing computational load; Then, batch normalization is performed on the dimensionality-reduced feature map to stabilize the feature distribution and accelerate convergence, and then nonlinear expressive power is introduced through an activation function to obtain a nonlinear feature map; Finally, the nonlinear feature map is output as a shallow feature map to subsequent continuous deep feature extraction layers. This is done because landslides appear as localized areas with abrupt changes in texture and color in satellite imagery. Shallow feature extraction can sensitively capture these low-level visual cues, while resolution reduction reduces the computational burden of subsequent layers while preserving key spatial structures. This allows the model to efficiently process large-size satellite image inputs and provides regular feature map sizes for deep networks to support multi-scale feature fusion.

[0041] Secondly, by performing layer-by-layer deep semantic feature extraction on the shallow feature map through continuous deep feature extraction layers, a deep semantic feature map is obtained. Specifically, firstly, the shallow feature map is input into the first deep feature extraction block. Multiple concatenated convolutional layers perform local feature aggregation and channel expansion on the shallow feature map to obtain the first deep feature map. This first deep feature map contains primary semantic distinction information between landslides and features such as bare land, farmland, and forests. Secondly, the first deep feature map is input into the second deep feature extraction block. Downsampling convolution further reduces the spatial dimension and expands the receptive field, allowing each feature point to cover a larger context area, resulting in the second deep feature map. This second deep feature map can perceive the topographic and geomorphological association information around the landslide. Then, the second deep feature map is input into the third deep feature extraction block. A bottleneck structure performs channel compression and re-expansion, reducing the number of parameters while enhancing the non-linear expressive power of the features, resulting in the third deep feature map. This third deep feature map possesses stronger category abstraction capabilities. Finally, the third deep feature map is output as a deep semantic feature map to the enhanced spatial pyramid pooling layer. This is done because landslide disaster identification not only relies on local texture features, but also requires understanding the semantic relationship between the landslide body and the surrounding environment. For example, landslides usually occur in specific terrain locations such as mountain slopes and riverbanks. Continuous deep feature extraction expands the receptive field and abstracts semantic information layer by layer, enabling the model to move from pixel-level texture to scene-level semantic understanding, thereby distinguishing real landslides from similar land features (such as exposed rocks and construction sites), reducing the false detection rate. At the same time, the layered and progressive structural design avoids the loss of small target information caused by single over-sampling.

[0042] Then, an enhanced spatial pyramid pooling layer is used to extract local salient features and compensate for global background information in the deep semantic feature map, resulting in an enhanced backbone feature map. Further, a first convolutional layer is used to perform channel dimensionality reduction on the deep semantic feature map, resulting in a dimensionality-reduced feature map; a max pooling layer is used to extract local salient features from the dimensionality-reduced feature map, resulting in a local salient feature map; an average pooling layer is used to extract global background information from the dimensionality-reduced feature map, resulting in a global background feature map; the dimensionality-reduced feature map, the local salient feature map, and the global background feature map are concatenated along the channel dimension, and a second convolutional layer is used to restore the channels, resulting in the enhanced backbone feature map. Specifically, firstly, the deep semantic feature map is input into the first convolutional layer, where a 1×1 convolution operation is used to compress the number of channels in the feature map to half of its original value, resulting in a dimensionality-reduced feature map. This dimensionality-reduced feature map reduces the computational overhead of subsequent pooling operations while preserving the original semantic information. Secondly, the dimensionality-reduced feature map is input into a max-pooling layer, where a max-pooling operation with a kernel size of 5 pixels × 5 pixels and a stride of 1 is used to extract the maximum value within each local region, resulting in a locally salient feature map. This locally salient feature map preserves the most significant feature responses, such as landslide edges and texture abrupt changes. Then, the dimensionality-reduced feature map is further processed... The feature map is simultaneously input into the average pooling layer. The mean value of all pixels in each local region is calculated through average pooling operations with the same kernel size and stride to obtain the global background feature map. The global background feature map supplements the overall background and texture information of the non-maximum region that was ignored during the max pooling process. Finally, the dimensionality-reduced feature map, the local salient feature map and the global background feature map are concatenated along the channel dimension to obtain the concatenated feature map. The concatenated feature map is then input into the second convolutional layer. The number of channels is restored to the original number of channels before input to the enhanced spatial pyramid pooling layer through a 1×1 convolution operation to obtain the enhanced backbone feature map. This approach is adopted because traditional spatial pyramid pooling modules employ three consecutive max pooling operations, each performed on the output of the previous layer. Multiple max pooling operations retain only local maxima, leading to the continuous loss of detailed information in non-maximum regions (such as subtle textures and edge gradations within the landslide body), resulting in feature coarsening. In contrast, this scheme simplifies the three max pooling operations into a parallel structure of one max pooling operation plus one average pooling operation. Max pooling is responsible for extracting salient features of the landslide body to ensure no target is missed, while average pooling is responsible for compensating for background texture details to maintain fine-grained perception capabilities. After extraction by both operations in parallel, the features are concatenated and fused with the original dimensionality-reduced features, and then information is integrated through channel recovery convolution. This approach retains the advantages of large receptive field feature extraction, alleviates the problem of feature coarsening, and avoids the additional computational load brought by the introduction of attention mechanisms, achieving a balance between lightweight design and accuracy.

[0043] Finally, the shallow feature map, deep semantic feature map, and enhanced backbone feature map are output as multi-level feature maps. Specifically, firstly, the shallow feature map is labeled as the first-level feature output. The shallow feature map retains the edge contour, texture details, and spatial location information of the landslide body, providing high-resolution spatial localization basis for the subsequent neck network. Secondly, the deep semantic feature map is labeled as the second-level feature output. The deep semantic feature map contains abstract category distinction information between the landslide body and background features, providing semantic discrimination basis for the subsequent neck network. Then, the enhanced backbone feature map is labeled as the third-level feature output. The enhanced backbone feature map has both local saliency features and global background compensation information, providing enhanced comprehensive representation for the subsequent neck network. Finally, the first-level feature output, second-level feature output, and third-level feature output are combined into a multi-level feature map. The multi-level feature map is transmitted from the backbone network to the neck network through lateral connection pathways, enabling the neck network to fuse and reorganize features from different levels. This is done because landslides exhibit multi-scale distribution characteristics in satellite imagery. Small-area landslides require high-resolution shallow features for precise localization, while large-area landslide clusters require deep semantic features for overall discrimination. Enhancing the backbone features provides a comprehensive representation optimized by the global context. Outputting features from all three levels simultaneously to the neck network enables the subsequent high- and low-order feature gating fusion module to adaptively select the optimal feature source based on the semantic importance of the channel, avoiding the information bottleneck of single-scale features, thereby improving the robustness of detecting landslide targets of different sizes.

[0044] Step S202: The neck network is used to fuse high and low-order features and quickly reorganize features in the multi-level feature maps to obtain an enhanced fused feature map.

[0045] It should be noted that step S202 includes: First, dynamically weighting and fusing high-order semantic features and low-order spatial features in the multi-level feature map and enhancing texture through a high-low order feature gating fusion module to obtain a gated fusion feature map. Further, channel-wise upscaling is performed on the low-order spatial features through an upscaling convolutional layer, and batch normalization and activation function processing are applied to obtain transformed low-order features. Global average pooling is performed on both high-order semantic features and low-order transformed features to obtain first-channel and second-channel description vectors. These vectors are then concatenated and normalized to obtain competitive gating weights. High-order semantic features and low-order transformed features are then weighted and fused channel-by-channel according to these competitive gating weights to obtain a preliminary fused feature map. This preliminary fused feature map is then normalized to obtain a normalized feature map. Low-frequency approximation components are extracted from the low-order spatial features using a max-pooling layer to obtain low-frequency approximation components. The difference between the low-frequency approximation components and the low-order spatial features is calculated to obtain high-frequency residual components. These high-frequency residual components are adaptively weighted using depthwise convolution and activation functions to obtain weighted high-frequency components. The weighted high-frequency components and low-order spatial features are then residually concatenated to obtain texture enhancement features. Finally, the normalized feature map and texture enhancement features are concatenated along the channel dimension to obtain a gated fused feature map.

[0046] Specifically, firstly, the low-order spatial features Inputting a dimension-upgrading convolutional layer, the number of channels for low-order spatial features is increased to the level of high-order semantic features through 1×1 convolution operations. Using the same channel dimension, an upgraded feature map is obtained. This upgraded feature map is then subjected to distribution stabilization processing via a batch normalization layer. Finally, a nonlinear transformation is introduced through an activation function to obtain the transformed low-order features. The specific formula is as follows: in This represents a low-order feature of the transformation. This represents the batch normalization function. Represents a linear activation function. This represents a 1×1 convolution operation. Transforming low-order features provides semantic representation capabilities that match those of high-order semantic features. Next, global average pooling is performed on both the high-order semantic features and the transformed low-order features, compressing the spatial dimension into channel description vectors, resulting in the first and second channel description vectors. The specific formula is as follows: in This represents the first channel description vector, which characterizes the global semantic strength of each channel of the higher-order features; This represents the second-channel description vector, which characterizes the global semantic strength of each channel of the transformed low-order features. This represents the global average pooling function; then, the first channel description vector and the second channel description vector are concatenated along the spatial dimension to form a joint feature tensor. Then, normalization is performed on the channel dimension using a normalization function to generate two sets of competing gating weights. and The specific calculation formula is as follows: in, Indicates the first The higher-order gating weights of each channel. Indicates the first The low-order gating weights of each channel, Indicates the channel index number. This represents the number of output channels. Competitive gating weights are used to measure the importance ratio of high-order to low-order information in the current channel. Each channel can only choose its optimal source from two paths. Next, high-order semantic features and transformed low-order features are weighted and fused channel-by-channel according to the competitive gating weights. The two sets of gating weights are element-wise multiplied into their respective feature maps and then summed to obtain a preliminary fused feature map. The specific formula is as follows: Here, ⊙ represents element-wise multiplication, and the initial fusion of feature maps achieves dynamic adaptive fusion of high- and low-order features. Subsequently, the standard deviation of each channel of the initial fused feature map is calculated, and all pixel values ​​of each channel are divided by the sum of the standard deviation of its corresponding channel and a minimum constant to obtain the normalized feature map. The specific formula is as follows: in This represents a minimal constant. The normalized feature map channels are re-approximated to a standard normal distribution, avoiding increased optimization difficulty in subsequent layers due to distribution shifts. Simultaneously, low-order spatial features are input into a max-pooling layer. A max-pooling operation with a preset kernel size (e.g., 3 pixels × 3 pixels) extracts low-frequency approximation components, which represent the smoothing trend information within local regions. Then, the low-frequency approximation components are subtracted element-wise from the low-order spatial features to separate the high-frequency residual components. The specific formula is as follows: in Represents high-frequency residual components. Represents the low-frequency approximate component. This represents the max pooling function. The high-frequency residual components characterize the intensity of local texture abrupt changes, i.e., high-frequency detail information such as landslide boundaries. Subsequently, the high-frequency residual components are input into a depth-oriented convolutional layer. Through independent channel-by-channel convolution operations, batch normalization, and activation function processing, an adaptive weight map is generated. The adaptive weight map is then multiplied element-wise with the high-frequency residual components to obtain weighted high-frequency components. These weighted high-frequency components dynamically enhance high-frequency details based on the importance of channels and spatial locations. Finally, the weighted high-frequency components are residually concatenated with low-order spatial features, and element-wise summed to obtain the texture enhancement features. The specific formula is as follows: in Indicates texture enhancement features, This represents the Sigmoid function. This is a depth-oriented convolution. Texture enhancement features strengthen edge responses while preserving the original low-order spatial information. Finally, the normalized feature map and the texture enhancement features are concatenated along the channel dimension to obtain a gated fusion feature map. Gated fusion feature maps simultaneously contain dynamically weighted high- and low-order semantic information and explicitly enhanced texture details. This is because traditional feature fusion methods typically use simple channel splicing or direct addition, indiscriminately superimposing all information, resulting in redundant features interfering with the effective signal, and the landslide boundary details are severely degraded after multiple downsamplings in deep networks.

[0047] Secondly, by improving the fast dual-branch feature extraction module, partial channel spatial convolution and cross-channel information recombination are performed on the gated fusion feature map to obtain a fast reconstructed feature map. Further, spatial information is extracted from the first channel of the gated fusion feature map through partial convolutional layers to obtain a partial spatial feature map; the second channel of the gated fusion feature map is used as a bypass feature map, where the first and second channels are all channels of the gated fusion feature map; cross-channel information fusion and dimensionality transformation are performed on the partial spatial feature map through a concatenated first and second bottleneck convolutional layers to obtain a bottleneck fusion feature map; the partial spatial feature map and the bottleneck fusion feature map are added together to obtain the fast reconstructed feature map. Specifically, firstly, all channels of the gated fusion feature map are divided into a first channel set and a second channel set. The first channel set contains a preset proportion (preset channel proportion value (1 / 4)) of the number of channels, and the second channel set contains the remaining number of channels. Secondly, spatial convolution operations with a preset kernel size (3 pixels × 3 pixels) are performed on the first channel set through partial convolutional layers to extract spatial structure information and obtain partial spatial feature maps. These partial spatial feature maps retain the local morphology and edge layout information of the landslide body. At the same time, the second channel set is directly retained as a bypass feature map, which participates in subsequent fusion without undergoing spatial convolution operations. To reduce redundant computation, a portion of the spatial feature map is input into a concatenated first bottleneck convolutional layer. Channel compression is performed using 1×1 convolution, and nonlinearity is introduced through an activation function to obtain the first bottleneck feature map. This first bottleneck feature map is then input into a second bottleneck convolutional layer, where channel expansion is performed using 1×1 convolution to restore the original dimension, resulting in the second bottleneck feature map. The second bottleneck feature map achieves deep fusion and dimensional transformation of cross-channel information. Finally, the portion of the spatial feature map and the second bottleneck feature map are added element-wise to obtain a rapidly reconstructed feature map. This rapidly reconstructed feature map contains both directly extracted spatial structure information and cross-channel semantic information reconstructed through the bottleneck structure. This approach is used because conventional convolution performs spatial operations on all channels, resulting in significant redundant computation. Partial convolution, however, performs spatial convolution on only a small number of channels, with the remaining channels directly bypassed. This significantly reduces computation while maintaining feature interaction capabilities. Combined with bottleneck-style 1×1 convolution for cross-channel information reconstructing, it maintains feature expressive power while significantly reducing the number of model parameters and floating-point operations, thereby accelerating the inference process and lowering the deployment threshold for edge devices, meeting the timeliness requirements for real-time screening of massive satellite imagery.

[0048] Then, multi-scale feature aggregation is performed on the fast reconstructed feature map through upsampling and lateral connections to obtain an aggregated feature map. Finally, the aggregated feature map is used as an enhanced fusion feature map. Specifically, first, the fast reconstructed feature map is input into the upsampling layer, and the spatial resolution of the feature map is enlarged to a preset target size (the same size as the shallow feature map) through nearest neighbor interpolation or bilinear interpolation to obtain an upsampled feature map. The upsampled feature map restores the spatial granularity of deep features to adapt to subsequent fusion. Second, the upsampled feature map is concatenated with the shallow feature map from the same level of the backbone network in the channel dimension through the lateral connection path to obtain a concatenated fusion feature map. The lateral connection realizes the direct fusion of deep semantic information and shallow spatial details. Then, the concatenated fusion feature map is subjected to convolutional fusion processing. The aliasing effect caused by channel concatenation is eliminated by convolution operation with a preset convolution kernel size (3 pixels × 3 pixels) to obtain a smooth fusion feature map. Finally, the smooth fusion feature map is output as an aggregated feature map, and then the aggregated feature map is passed to the detection head as an enhanced fusion feature map. This is done because deep features have low spatial resolution after multiple downsampling layers. Although they have strong semantic abstraction capabilities, they lack precise localization information. On the other hand, shallow features have high spatial resolution but weak semantic discriminative power. By upsampling to restore the spatial size of deep features and connecting them laterally with shallow features at the same level, the model can simultaneously "accurately identify the location" and "distinguish the category". This multi-scale feature aggregation mechanism is the core design of the neck structure of the target detection network, ensuring that the detection head can obtain rich and aligned feature representations at multiple scales, thereby improving the robustness of detecting landslide targets of different sizes.

[0049] Step S203: The detection head performs landslide target detection and confidence determination on the enhanced fusion feature map to obtain landslide screening results.

[0050] It should be noted that step S203 includes: performing landslide category probability prediction on the enhanced fusion feature map through the category prediction branch to obtain the category prediction result; performing target bounding box coordinate regression on the enhanced fusion feature map through the location regression branch to obtain the location prediction result; determining the confidence level based on the category prediction result and the location prediction result, and determining the prediction result with a confidence level greater than a preset confidence threshold as a valid landslide detection result; generating landslide screening results based on the valid landslide detection results, wherein the landslide screening results include slice identification information containing landslide targets.

[0051] Specifically, firstly, the enhanced fusion feature map is input into the category prediction branch. Multiple cascaded convolutional layers perform channel compression and feature abstraction on the enhanced fusion feature map. Then, a fully connected layer maps it to a preset number of categories (two categories: landslide and non-landslide). Finally, a normalized exponential function converts the output into a probability distribution, yielding the category prediction result, which represents the probability that the current slice contains a landslide target. Secondly, the enhanced fusion feature map is simultaneously input into the location regression branch. Multiple cascaded convolutional layers extract target localization-related features. A fully connected layer outputs a preset number of regression parameters (four parameters in this embodiment: the x-coordinate and y-coordinate of the bounding box center point, width, and height), yielding the location prediction result. The location prediction result indicates the spatial location and size information of the landslide target in the slice. Then, the landslide category probability value in the category prediction result is multiplied by the bounding box quality score of the location prediction result to obtain the comprehensive confidence score. The comprehensive confidence score is compared with a preset confidence threshold (e.g., 0.02). Prediction results greater than the preset confidence threshold are selected as valid landslide detection results, while prediction results less than or equal to the preset confidence threshold are removed as background noise. Finally, landslide screening results are generated based on the valid landslide detection results. The landslide screening results include slice identification information containing landslide targets (specifically, slice number, slice coordinate location, and confidence score), so that the subsequent fine segmentation module can locate and process the corresponding slices based on the slice identification information. This is done because the detection head needs to answer two questions simultaneously: "Is there a landslide?" and "Where is the landslide?". The category prediction branch is responsible for determining the semantic category, and the location regression branch is responsible for locating the spatial range. The two work together to output a complete detection result. The confidence determination stage ensures a high recall rate by setting a low threshold to avoid missing potential landslide slices. The slice identification information in the landslide screening results provides accurate data indexes for downstream modules, achieving seamless connection between the screening stage and the segmentation stage.

[0052] This embodiment extracts multi-scale features from standardized satellite images through a backbone network, resulting in a multi-level feature map composed of a shallow feature map, a deep semantic feature map, and an enhanced backbone feature map. A neck network then dynamically weights and fuses high- and low-order features in the multi-level feature map, followed by rapid bi-branch feature reorganization to obtain an enhanced fused feature map. A detection head performs category prediction, location regression, and confidence assessment on the enhanced fused feature map to obtain landslide screening results. A gated adaptive fusion mechanism dynamically filters effective feature channels, while enhanced spatial pyramid pooling balances local saliency with global background compensation. An improved rapid bi-branch module reduces redundant computation, significantly improving landslide target recall and reducing false negatives while maintaining a lightweight model, thus meeting the real-time screening requirements of massive satellite imagery.

[0053] Based on the first embodiment of this application, this application also provides a lightweight landslide image screening device with fusion gating, please refer to... Figure 3 The device includes: The acquisition module 10 is used to acquire optical satellite image slices to be screened.

[0054] The data preprocessing module 20 is used to preprocess optical satellite image slices to obtain standardized satellite images.

[0055] The result module 30 is used to input standardized satellite images into a pre-gated lightweight model for processing to obtain landslide screening results. The pre-gated lightweight model includes a backbone network, a neck network, and a detection head. The backbone network includes convolutional downsampling layers, continuous deep feature extraction layers, and enhanced spatial pyramid pooling layers. The neck network includes a high- and low-order feature gating fusion module and an improved fast dual-branch feature extraction module. The detection head includes a category prediction branch and a location regression branch. The module is also used to perform multi-scale feature extraction on the standardized satellite images through the backbone network to obtain multi-level feature maps. The neck network performs high- and low-order feature fusion and fast feature recombination on the multi-level feature maps to obtain enhanced fused feature maps. The detection head performs landslide target detection and confidence determination on the enhanced fused feature maps to obtain landslide screening results.

[0056] The filtering module 40 is used to output a set of satellite image slices containing landslide targets based on the landslide filtering results.

[0057] The lightweight landslide image screening device with fusion gating provided in this application, employing the lightweight landslide image screening method with fusion gating in the above embodiments, can solve the technical problem of how to quickly and accurately screen satellite image slices containing landslide targets. Compared with the prior art, the beneficial effects of the lightweight landslide image screening device with fusion gating provided in this application are the same as those of the lightweight landslide image screening method with fusion gating provided in the above embodiments, and other technical features in the lightweight landslide image screening device with fusion gating are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0058] This application provides a lightweight landslide image screening device with fusion gating. The lightweight landslide image screening device with fusion gating includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the lightweight landslide image screening method with fusion gating in the first embodiment described above.

[0059] The following is for reference. Figure 4This document illustrates a structural schematic diagram of a landslide image screening device suitable for implementing the fusion-gated lightweight landslide image screening device in the embodiments of this application. The fusion-gated lightweight landslide image screening device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The illustrated lightweight landslide image screening device with fusion gating is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0060] like Figure 4 As shown, the fusion-gated lightweight landslide image screening device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the fusion-gated lightweight landslide image screening device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the fusion-gated lightweight landslide image screening device to wirelessly or wiredly communicate with other devices to exchange data. Although various fusion-gated lightweight landslide image screening devices are shown in the figures, it should be understood that implementation or possession of all of them is not required. More or fewer of these devices may be implemented alternatively.

[0061] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0062] The lightweight landslide image screening device with fusion gating provided in this application, employing the lightweight landslide image screening method with fusion gating in the above embodiments, can solve the technical problem of how to quickly and accurately screen satellite image slices containing landslide targets. Compared with the prior art, the beneficial effects of the lightweight landslide image screening device with fusion gating provided in this application are the same as those of the lightweight landslide image screening method with fusion gating provided in the above embodiments, and other technical features in this lightweight landslide image screening device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0063] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0064] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0065] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the fusion-gated lightweight landslide image screening method in the above embodiments.

[0066] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0067] The aforementioned computer-readable medium may be included in the fusion-gated lightweight landslide image screening device; or it may exist independently and not be assembled into the fusion-gated lightweight landslide image screening device.

[0068] The aforementioned computer-readable medium carries one or more programs that, when executed by the fusion-gated lightweight landslide image screening device, enable the fusion-gated lightweight landslide image screening device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0069] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0070] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0071] The readable medium provided in this application is a computer-readable medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described fusion-gated lightweight landslide image screening method, which can solve the technical problem of how to quickly and accurately screen satellite image slices containing landslide targets. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the fusion-gated lightweight landslide image screening method provided in the above embodiments, and will not be repeated here.

[0072] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the fusion-gated lightweight landslide image screening method described above.

[0073] The computer program product provided in this application can solve the technical problem of how to quickly and accurately screen satellite image slices containing landslide targets. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the fusion-gated lightweight landslide image screening method provided in the above embodiments, and will not be repeated here.

[0074] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A lightweight landslide image screening method with integrated gating, characterized in that, The method includes: Obtain optical satellite image slices to be filtered; The optical satellite image slices are preprocessed to obtain standardized satellite images; The standardized satellite images are input into a pre-gated lightweight model for processing to obtain landslide screening results. The pre-gated lightweight model includes a backbone network, a neck network, and a detection head. The backbone network includes a convolutional downsampling layer, a continuous deep feature extraction layer, and an enhanced spatial pyramid pooling layer. The neck network includes a high- and low-order feature gating fusion module and an improved fast dual-branch feature extraction module. The detection head includes a category prediction branch and a location regression branch. Output a set of satellite image slices containing landslide targets based on the landslide screening results; The step of inputting the standardized satellite image into a preset gated lightweight model for processing to obtain landslide screening results includes: Multi-scale feature extraction is performed on the standardized satellite images using a backbone network to obtain multi-level feature maps. The multi-level feature map is fused with high and low-order features and rapidly reorganized by the neck network to obtain an enhanced fused feature map. The enhanced fusion feature map is analyzed by a detection head to detect landslide targets and determine their confidence levels, thus obtaining landslide screening results.

2. The method as described in claim 1, characterized in that, The step of extracting multi-scale features from the standardized satellite image through a backbone network to obtain multi-level feature maps includes: The standardized satellite image is subjected to shallow feature extraction and resolution reduction through the convolutional downsampling layer to obtain a shallow feature map. The shallow feature map is subjected to layer-by-layer deep semantic feature extraction through the continuous deep feature extraction layer to obtain a deep semantic feature map; The enhanced spatial pyramid pooling layer is used to extract local salient features and compensate for global background information in the deep semantic feature map to obtain an enhanced backbone feature map. The shallow feature map, the deep semantic feature map, and the enhanced backbone feature map are output as a multi-level feature map.

3. The method as described in claim 2, characterized in that, The step of extracting local salient features and compensating for global background information from the deep semantic feature map through the enhanced spatial pyramid pooling layer to obtain an enhanced backbone feature map includes: The deep semantic feature map is reduced in dimensionality by channel reduction through the first convolutional layer to obtain a dimensionality-reduced feature map. Local salient features are extracted from the reduced feature map using a max pooling layer to obtain a local salient feature map; Global background information is extracted from the reduced-dimensional feature map using an average pooling layer to obtain a global background feature map. The reduced feature map, the local salient feature map, and the global background feature map are concatenated along the channel dimension, and channel recovery is performed through a second convolutional layer to obtain an enhanced backbone feature map.

4. The method as described in claim 1, characterized in that, The step of fusing high- and low-order features and rapidly recombining features from the multi-level feature maps through the neck network to obtain an enhanced fused feature map includes: The high-order semantic features and low-order spatial features in the multi-level feature map are dynamically weighted and fused and texture enhanced by the high-order and low-order feature gating fusion module to obtain the gating fusion feature map. By improving the fast dual-branch feature extraction module, partial channel spatial convolution and cross-channel information recombination are performed on the gated fusion feature map to obtain a fast recombined feature map. The fast reconstructed feature map is aggregated by multi-scale feature aggregation through upsampling and lateral connection to obtain an aggregated feature map; The aggregated feature map is used as the enhanced fusion feature map.

5. The method as described in claim 4, characterized in that, The step of dynamically weighting and fusing high-order semantic features and low-order spatial features in the multi-level feature map and enhancing texture through the high- and low-order feature gating fusion module to obtain a gated fused feature map further includes: The low-order spatial features are channel-upgraded by using a dimension-up convolutional layer, and then transformed low-order features are obtained by batch normalization and activation function processing. Global average pooling is performed on the higher-order semantic features and the transformed lower-order features respectively to obtain the first channel description vector and the second channel description vector. The first channel description vector and the second channel description vector are concatenated and then normalized to obtain the competitive gating weights; Based on the competitive gating weights, the higher-order semantic features and the transformed lower-order features are fused channel by channel to obtain a preliminary fused feature map; Channel normalization is performed on the preliminary fused feature map to obtain a normalized feature map; The low-frequency approximation components are extracted from the low-order spatial features by using a max pooling layer. The high-frequency residual component is obtained by subtracting the low-frequency approximation component from the low-order spatial features. The high-frequency residual components are adaptively weighted by depthwise convolution and activation functions to obtain weighted high-frequency components. The weighted high-frequency components and low-order spatial features are residually connected to obtain texture enhancement features; The normalized feature map and texture enhancement feature are concatenated along the channel dimension to obtain a gated fusion feature map.

6. The method as described in claim 4, characterized in that, The step of performing partial channel spatial convolution and cross-channel information reconstruction on the gated fusion feature map by improving the fast dual-branch feature extraction module to obtain a fast reconstructed feature map includes: Spatial information is extracted from the first channel of the gated fusion feature map by using partial convolutional layers to obtain a partial spatial feature map; The second channel of the gated fusion feature map is used as a bypass feature map, wherein the first channel and the second channel are all channels of the gated fusion feature map; By performing cross-channel information fusion and dimensional transformation on the partial spatial feature map through a first bottleneck convolutional layer and a second bottleneck convolutional layer in series, a bottleneck fusion feature map is obtained. The partial spatial feature map and the bottleneck fusion feature map are added together to obtain a fast reconstructed feature map.

7. The method as described in claim 1, characterized in that, The step of performing landslide target detection and confidence determination on the enhanced fusion feature map using a detection head to obtain landslide screening results includes: The enhanced fusion feature map is subjected to landslide category probability prediction through the category prediction branch to obtain the category prediction result; The enhanced fused feature map is regressed by the location regression branch to obtain the target bounding box coordinate regression and the location prediction result. Based on the category prediction results and the location prediction results, a confidence level is determined, and prediction results with a confidence level greater than a preset confidence threshold are identified as valid landslide detection results. A landslide screening result is generated based on the effective landslide detection results, wherein the landslide screening result includes slice identification information containing landslide targets.

8. A lightweight landslide image screening device integrating gating, characterized in that, The device includes: The acquisition module is used to acquire optical satellite image slices to be screened; The data preprocessing module is used to preprocess the optical satellite image slices to obtain standardized satellite images; The result module is used to input the standardized satellite image into a preset gated lightweight model for processing to obtain landslide screening results. The preset gated lightweight model includes a backbone network, a neck network, and a detection head. The backbone network includes convolutional downsampling layers, continuous deep feature extraction layers, and enhanced spatial pyramid pooling layers. The neck network includes a high- and low-order feature gating fusion module and an improved fast dual-branch feature extraction module. The detection head includes a category prediction branch and a location regression branch. The module is also used to perform multi-scale feature extraction on the standardized satellite image through the backbone network to obtain multi-level feature maps; to perform high- and low-order feature fusion and fast feature recombination on the multi-level feature maps through the neck network to obtain enhanced fused feature maps; and to perform landslide target detection and confidence determination on the enhanced fused feature maps through the detection head to obtain landslide screening results. The filtering module is used to output a set of satellite image slices containing landslide targets based on the landslide filtering results.

9. A lightweight landslide image screening device integrating gating, characterized in that, The device includes: a memory, a processor, and a fusion-gated lightweight landslide image screening program stored on the memory and running on the processor, the fusion-gated lightweight landslide image screening program being configured to implement the steps of the fusion-gated lightweight landslide image screening method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a fusion-gated lightweight landslide image screening program, which, when executed by a processor, implements the steps of the fusion-gated lightweight landslide image screening method as described in any one of claims 1-7.

Images