An optical remote sensing image landslide disaster automatic identification method and system

Abstract

The application provides a kind of optical remote sensing image landslide disaster automatic identification method and system, belong to image processing technical field.The method includes: multi-view feature extraction and fusion are carried out to input image, and robust feature representation is obtained;Based on this, initial landslide detection frame and semantic segmentation map are synchronously generated by integrated model;Further, an iterative optimization cycle containing segmentation quality evaluation, detection frame correction and segmentation map optimization guidance is executed, so that detection and segmentation tasks dynamically interact and enhance in inference process until the convergence condition is met;Finally, high-precision landslide positioning frame and pixel-level segmentation mask are output.The system is used to execute the method.The application adopts the above-mentioned optical remote sensing image landslide disaster automatic identification method and system, effectively solves the core contradiction that landslide positioning is not accurate and boundary recognition is rough and difficult to consider in the prior art, improves the recognition accuracy, robustness and regional adaptive ability of the model in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an automatic identification method and system for landslide disasters from optical remote sensing images. Background Technology

[0002] Landslides, as a common geological hazard, seriously threaten people's lives and property and the safety of major engineering projects. Automatic identification of large-scale, high-frequency landslide hazards using optical remote sensing imagery has become a key technological means for disaster prevention and mitigation. In recent years, with breakthroughs in deep learning technology, automatic landslide identification methods based on convolutional neural networks (CNNs) have made significant progress, gradually replacing earlier machine learning methods that relied on manual feature design. However, existing technologies still have the following major shortcomings, which limit their accuracy and practicality in actual operations.

[0003] Traditional classification methods based on pixels or objects. These methods (such as Support Vector Machines (SVM), Random Forests, etc.) heavily rely on hand-designed spectral, texture, and shape features. Their main drawbacks are: 1) Limited feature representation: Hand-designed features are difficult to capture the diverse visual patterns of landslides in complex terrain, resulting in weak model generalization ability; 2) Fragmented recognition results: Classification based on pixels or over-segmented objects often results in "salt and pepper noise" that cannot form complete and continuous landslide entities, and boundary extraction is extremely inaccurate.

[0004] Landslide recognition methods based on single-task deep learning are currently the mainstream research direction, mainly divided into two categories. The first category is methods based on object detection frameworks (such as Faster R-CNN and the YOLO series), which treat landslides as a whole object for localization. Its drawback is that it only outputs rectangular bounding boxes, failing to capture the fine, irregular shape contours of landslides, while accurate boundaries are crucial for disaster area estimation and risk assessment. The second category is methods based on semantic segmentation frameworks (such as U-Net and the DeepLab series), which can achieve pixel-level classification and output landslide contours. However, its drawbacks include: 1) high false negative rate for small-scale, insignificant landslides: the model tends to focus on large areas with obvious features; 2) high false positive rate against similar texture backgrounds such as bare soil and construction sites; 3) disconnect between detection and segmentation performance: pure detection models do not care about boundaries, and pure segmentation models do not optimize localization, failing to complement each other's advantages.

[0005] Dual-task models based on simple feature fusion. Some studies attempt to combine detection and segmentation, for example, by parallelizing two task heads on a shared backbone network or by performing shallow feature concatenation. The fundamental flaw of these methods lies in their static, unidirectional, and shallow collaboration. Detection and segmentation are simply combined at the network end or feature level, lacking a deep interactive feedback mechanism. This leads to potential interference between the two tasks during optimization, failing to resolve the core contradiction of "inaccurate detection box localization leading to misleading segmentation regions" and "segmentation noise affecting detection confidence assessment." Improvements in model performance face bottlenecks.

[0006] Data-driven models are highly dependent on labeled data. All the aforementioned deep learning methods require a large number of high-quality, pixel-level labeled landslide samples for training. Their drawbacks are: 1) extremely high labeling costs, as landslide labeling requires geological expertise; 2) poor model regional transferability, where a model trained in one region often experiences a sharp performance drop when directly applied to new regions with different topography and lighting conditions. Existing technologies lack efficient and stable model adaptation mechanisms.

[0007] In summary, the core shortcomings of existing technologies can be attributed to the irreconcilable contradiction between the accuracy of target-level localization and the precision of pixel-level boundaries in current landslide identification models. Furthermore, the practical robustness of these models is limited by high annotation costs and weak adaptability to new regions. Therefore, there is an urgent need for a novel identification method that can deeply integrate the advantages of detection and segmentation tasks, achieve dynamic collaborative optimization, and continuously evolve using limited data. Summary of the Invention

[0008] The purpose of this invention is to provide an automatic identification method and system for landslide disasters from optical remote sensing images, in order to solve the technical problems in the prior art where it is difficult to balance the accuracy of landslide identification and positioning with boundary details, as well as the model's reliance on a large amount of labeled data and its weak regional migration capability.

[0009] To achieve the above objectives, the present invention provides an automatic identification method for landslide hazards from optical remote sensing images, comprising the following steps: Step S1: Perform geometric transformation on the input optical remote sensing image to generate a multi-view image, and fuse its depth features to obtain a robust feature representation. Step S2: Based on robust feature representation, the target detection and semantic segmentation tasks are executed synchronously through an integrated model to generate an initial set of landslide detection boxes and an initial semantic segmentation map, which serve as the initial input for iterative optimization. Step S3: Starting from the initial input, execute the following iterative optimization loop: Step S31: Based on the semantic segmentation map of the current iteration, evaluate the quality of the landslide region in the corresponding detection box; Step S32: Based on the quality assessment results, correct the confidence and geometric position of the detection box in the current iteration; Step S33: Use the corrected detection box to guide optimization and generate the semantic segmentation map for the next iteration. Step S4: When the iterative optimization loop meets the predetermined termination condition, output the set of landslide detection boxes generated in the current iteration and its corresponding pixel-level segmentation mask.

[0010] Preferably, step S1 specifically includes the following steps: Step S11: Process the input image application Different affine transformations Generate multi-view image groups ; Step S12: Through a feature encoding network with shared weights Extract feature maps from images at each viewpoint. ; Step S13: Concatenate the multi-view feature maps along the channel dimension and pass them through a convolutional layer. By performing fusion and dimensionality reduction, robust feature representations are obtained. The calculation formula is as follows: .

[0011] Preferably, the integrated model used in step S2 is used to handle the characteristics of the landslide target, specifically including: Step S21: Represent robustness features Input Region Proposal Network Generate a set of candidate landslide area proposals ; Step S22, Detection and Segmentation Co-processing: a) Detection branch: For each proposal Through region of interest pooling layer from Extract fixed-size features and input them into the detection head. Predict the initial detection box and its confidence level ; b) Branch segmentation: Representing robustness features Input a semantic segmentation network with skip connections Jump connections are used to fuse shallow landslide boundary texture features with deep landslide semantic features; at the same time, candidate landslide region proposals are utilized. Generate spatial attention map This allows computational resources to be focused on candidate regions in the initial stage to generate an initial semantic segmentation map. .

[0012] Preferably, in step S31, for the first The detection box in the next iteration and its corresponding segmentation region mask Its overall quality score The calculations are used to distinguish between landslide and non-landslide areas: Step S311: Calculate the first morphological index : ; in, The function represents the intersection-union ratio calculation function. This represents a predefined morphological compaction operation function; Step S312: Calculate the second statistical indicator : ; in, This represents the function for calculating standard deviation. Indicates the first In the nth iteration, the 1st The area mask corresponding to each detection box The set of predicted probability values ​​for landslide categories for all pixels within the range; Step S313, Indicator Adjustment: Using the second statistical indicator For the first form indicator Make corrections to obtain the corrected morphological index. : ; in, This represents the preset morphological reliability adjustment factor, which is a constant greater than zero; Step S314: Calculate the overall quality score : .

[0013] Preferably, the correction of the detection box in step S32 is related to the overall quality score. Adaptive association: Step S321: Based on the overall quality score For the The original confidence of each detection box Make corrections to obtain the corrected confidence level. The calculation formula is as follows: ; Among them, the regulating factor ; Step S322: Based on the overall quality score For the The original geometric position of each detection box Make corrections to obtain the corrected geometric position. The calculation formula is as follows: ; in, To calculate the minimum bounding rectangle of the segmented region mask, adjust the coefficients. , This is the preset base adjustment rate.

[0014] Preferably, the predetermined termination condition in step S4 is first satisfied by one of the following two conditions: a) Number of iterations Reaching the maximum value preset for the landslide identification task; b) The average change in the overall quality score of all retained landslide candidate boxes between two adjacent iterations. Less than the threshold ,in: ; in, For the first The number of detection boxes in the next iteration.

[0015] Preferably, the ensemble model is trained using a loss function specifically designed for landslide identification. : Get the Step detection loss and segmentation loss ; Calculate the consistency loss of predictions between adjacent steps. And assign adaptive weights to them. ,in For the first The average confidence level of the step segmentation prediction The preset maximum weight constant; Calculate the total loss: ; in, This represents the total number of iterations simulated during training.

[0016] Preferably, during the model training phase, a class balancing strategy is adopted, and focus loss weighting is applied to difficult pixel samples in the segmentation loss that are similar to the spectral or texture features of the landslide.

[0017] The present invention also provides an automatic identification system for landslide hazards from optical remote sensing images, used to perform the automatic identification method for landslide hazards from optical remote sensing images as described above, including: The multi-view feature extraction module is used to perform geometric transformations on the input optical remote sensing images to generate multi-view images and fuse their depth features to obtain robust feature representations. The collaborative perception and initial solution generation module is used to generate an initial set of landslide detection boxes and an initial semantic segmentation map by simultaneously performing target detection and semantic segmentation tasks through an integrated model based on robust feature representation, which serve as the initial input for iterative optimization. The solution space iterative optimization module is used to execute an iterative optimization loop starting from the initial input; The convergence result output module is used to output the set of landslide detection boxes and their corresponding pixel-level segmentation masks generated in the current iteration when the iterative optimization loop meets the predetermined termination condition.

[0018] Preferably, the solution space iterative optimization module includes: The segmentation quality assessment unit is used to assess the quality of the landslide region in the corresponding detection box based on the semantic segmentation map of the current iteration round. The detection box correction unit is used to correct the confidence and geometric position of the detection box in the current iteration based on the quality assessment results. The segmentation graph optimization guidance unit is used to guide optimization using the corrected detection boxes to generate the semantic segmentation graph for use in the next iteration round.

[0019] Therefore, the present invention employs the above-mentioned method and system for automatic identification of landslide hazards using optical remote sensing images, and the beneficial technical effects are as follows: (1) By introducing an iterative optimization loop consisting of segmentation quality assessment, detection box confidence and geometric position correction, and segmentation map optimization guidance, the detection task and the segmentation task can dynamically verify and enhance each other during the inference process. Specifically, the quality assessment based on the internal consistency and morphological regularity of the segmentation map can reliably screen and correct the detection boxes; and the more accurate detection boxes after correction can guide the next round of segmentation optimization through the spatial attention mechanism, so that the localization boundary of the landslide target continuously converges to the true contour in the iteration, thereby simultaneously improving the recall rate of detection and the boundary accuracy of segmentation.

[0020] (2) On the one hand, through multi-view feature extraction and fusion, a robust feature representation that is insensitive to changes in illumination and viewpoint is constructed, which enhances the basic generalization ability of the model. On the other hand, in the comprehensive quality scoring mechanism, the statistical index reflecting internal consistency is used to dynamically correct the index reflecting morphological regularity. This design enables the model to more accurately evaluate real landslides that are irregular in shape but homogeneous in internal structure, while reducing false alarms for noise in regular shapes. Thus, higher recognition accuracy and lower false alarm rate are achieved in complex mountainous backgrounds.

[0021] (3) The multi-task loss function that integrates adaptive consistency constraints and the optimization strategy for difficult samples adopted in the training phase of this invention improves the stability and efficiency of model learning. More importantly, the entire iterative optimization framework and the adaptive correction mechanism based on quality scoring essentially enable the model to have a self-calibration capability during the inference phase. This capability allows the model to self-optimize and adjust its output through internal iteration when applied to new areas with different terrain features, thereby reducing the stringent requirements on the completeness of the training data distribution to a certain extent and improving the transfer applicability and robustness of the technology in actual business. Attached Figure Description

[0022] Figure 1 This is a flowchart of an automatic identification method for landslide hazards from optical remote sensing images according to the present invention; Figure 2 To iteratively optimize the internal workflow diagram of the loop; Figure 3 Comparison of the convergence curves of the loss function during the training process. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0025] Example 1 This embodiment provides a specific implementation of a method for automatic identification of landslide hazards using optical remote sensing images. (Refer to...) Figure 1 The flowchart shown illustrates this method, implemented on a GPU-equipped server using the PyTorch deep learning framework. The specific steps include: Data preparation and preprocessing.

[0026] Obtain publicly available optical remote sensing image datasets of landslides (such as LLSD). The images contain RGB three bands and have a spatial resolution of approximately 0.5 m to 2 m. Perform the following processing on each training image and its corresponding pixel-level landslide annotation mask: Cropping the image into fixed-size slices (e.g., 512×512 pixels).

[0027] The image is standardized by normalizing the pixel values ​​to the range of [0, 1].

[0028] Online data augmentation strategies were employed, including random horizontal / vertical flipping, small rotations (±15°), and color jitter, to increase the diversity of training samples.

[0029] Step S1: Multi-view feature extraction.

[0030] Geometric transformations are performed on the input optical remote sensing images to generate multi-view images, and their depth features are fused to obtain robust feature representations.

[0031] Step S11: Process the input image application Different affine transformations Generate multi-view image groups In this embodiment, three different affine transformations are set, corresponding to: 1) identity transformation, 2) clockwise rotation of 30 degrees, and 3) scaling to 0.9 times the original image, to generate a multi-view image set. .

[0032] Step S12: Use a ResNet-50 network pre-trained on ImageNet as the feature encoding network with shared weights. Feature maps of images from each viewpoint are extracted respectively. In this embodiment, its fully connected layer is removed, and the output of its fourth stage (stage4) is used as the feature map. The following steps are then performed: enter Three corresponding feature maps were obtained. , , The size of each feature map is ,in , .

[0033] Step S13: Concatenate the three feature maps along the channel dimension to obtain a concatenation array of size S13. The feature map. Then, through a... Convolutional layer By performing fusion and dimensionality reduction, robust feature representations are obtained. The calculation formula is as follows: .

[0034] Step S2: Collaborative sensing and initial solution generation.

[0035] Based on robust feature representation, the object detection and semantic segmentation tasks are executed synchronously by an integrated model to generate an initial set of landslide detection boxes and an initial semantic segmentation map, which serve as the initial input for iterative optimization.

[0036] The ensemble model used in step S2 is used to handle the characteristics of the landslide target, specifically including: Step S21, Region Proposal: Represent robustness features Input Region Proposal Network The RPN pre-sets a set of anchors whose scale and aspect ratio are set according to the statistical distribution of landslide targets in the training set (e.g., areas ranging from 32...). 2 up to 512 2 (Pixels, aspect ratio 0.5:1 to 2:1). RPN outputs a set of candidate landslide area proposals. Each proposal includes its location and a preliminary score as a prospect (slope).

[0037] Step S22, Detection and Segmentation Co-processing: a) Detection branch: For each proposal Through region of interest pooling layer from Extract a fixed size ( The characteristics of ) are input into the detection head. Predict the initial detection box and its confidence level .

[0038] b) Branch segmentation: Representing robustness features Input a semantic segmentation network with skip connections Jump connections are used to fuse shallow landslide boundary texture features with deep landslide semantic features; at the same time, candidate landslide region proposals are utilized. Generate spatial attention map (The value in this image is 1 within the proposal region and 0 outside the region) to focus computational resources on candidate regions in the initial stage to generate an initial semantic segmentation map. .

[0039] In this embodiment, The U-Net architecture is adopted, with its encoder part and feature encoding network... With shared weights (i.e., ResNet-50), the decoder part fuses shallow high-resolution features and deep semantic features from the encoder through skip connections to simultaneously recover details and context.

[0040] Step S3: Solution space iterative optimization.

[0041] Starting from the initial input, execute the following iterative optimization loop ( Figure 2 ): Step S31, Segmentation quality assessment: Based on the semantic segmentation map of the current iteration, assess the quality of the landslide region in the corresponding detection box.

[0042] In step S31, for the first The detection box in the next iteration and its corresponding segmentation region mask Its overall quality score The calculations are used to distinguish between landslide and non-landslide areas: Step S311: Calculate the first morphological index : ; in, The function represents the intersection-union ratio calculation function. This represents a preset morphological compaction operation function used to fill mask holes and smooth boundaries.

[0043] This indicator is used to assess the morphological deviation of a landslide area from its compact shape, in order to distinguish the irregular boundaries of a landslide from regular artificial features.

[0044] Step S312: Calculate the second statistical indicator : ; in, This represents the function for calculating standard deviation. Indicates the first In the nth iteration, the 1st The area mask corresponding to each detection box The set of predicted probability values ​​for landslide categories for all pixels within the range; This index is used to assess the spectral and textural consistency within a landslide area to distinguish between a homogeneous landslide body and a heterogeneous clutter background.

[0045] Step S313: Utilize the second statistical indicator (Internal consistency) for the first-form indicator (Form compactness) is corrected to obtain the corrected form index. : ; in, This represents the preset morphological reliability adjustment factor, which is a constant greater than zero.

[0046] Step S314: Calculate the overall quality score : ; This score comprehensively reflects the credibility of the candidate area as a landslide.

[0047] Step S32: Based on the quality assessment results, correct the confidence and geometric position of the current iteration detection box.

[0048] Step S321: Based on the overall quality score For the The original confidence of each detection box Make corrections to obtain the corrected confidence level. The calculation formula is as follows: ; Among them, the regulating factor In this embodiment, it is set to 2.0 to apply a more significant attenuation to the confidence level of low-scoring regions.

[0049] Step S322: Based on the overall quality score For the The original geometric position of each detection box Make corrections to obtain the corrected geometric position. The calculation formula is as follows: ; in, To calculate the minimum bounding rectangle of the segmented region mask, adjust the coefficients. , The preset base adjustment rate is a constant between 0 and 1, and is set to 0.7 in this embodiment.

[0050] Step S33, Segmentation Map Optimization Guidance: Using the Corrected Detection Box Set and its corresponding corrected confidence level Generate spatial attention map Unlike the initial binary attention, this time each box contributes to the attention map in a way that the intensity is proportional to the box center. A proportional two-dimensional Gaussian distribution. This soft attention map is more refined. With features Combined, and then input into the segmentation network again. At this point, some parameters can be fine-tuned (or different weights can be used than the initial segmentation) to generate an optimized segmentation map that focuses more on high-confidence target regions for the next iteration.

[0051] Step S4: Output the convergence result.

[0052] When the iterative optimization loop meets the predetermined termination condition, the set of landslide detection boxes generated in the current iteration and its corresponding pixel-level segmentation mask are output.

[0053] The predetermined termination condition in step S4 is that one of the following two conditions must be met first: a) Number of iterations The maximum value preset for the landslide identification task has been reached.

[0054] b) The average change in the overall quality score of all retained landslide candidate boxes between two adjacent iterations. Less than the threshold ,in: ; in, For the first The number of detection boxes in the next iteration.

[0055] The ensemble model is trained using a loss function specifically designed for landslide identification. : Obtain the base loss: Obtain the first Step detection loss and segmentation loss .

[0056] Constructing an adaptive consistency loss: Calculating the consistency loss of predictions between adjacent steps. And assign adaptive weights to them. ,in For the first The average confidence level of the step segmentation prediction This is the preset maximum weight constant.

[0057] Calculate the total loss: ; in, This represents the total number of iterations simulated during training.

[0058] During the model training phase, a class balancing strategy is adopted, and focus loss is applied to difficult pixel samples in the segmentation loss that are similar to the spectral or texture features of landslides.

[0059] The invention will be further illustrated below with specific examples.

[0060] 1. Data sources and preprocessing.

[0061] To verify the effectiveness of this invention, two publicly available optical remote sensing landslide identification datasets were selected for experiments: The LLSD dataset contains approximately 2,000 high-resolution (0.5m-2m) landslide images and corresponding pixel-level annotations, covering various landforms (mountains, hills), landslide types (new and old landslides), and lighting conditions. It is a widely used benchmark dataset in the field.

[0062] The GF2-LS dataset, sourced from my country's Gaofen-2 satellite, contains approximately 500 images with a spatial resolution of 1 meter. This dataset features smaller landslide targets against more complex backgrounds (including numerous farmlands, bare soil, roads, and other easily confused features), making it suitable for testing the model's generalization ability in more challenging scenarios.

[0063] Before the experiment, all images were uniformly cropped into 512×512 pixel slices and randomly divided into training, validation, and test sets in a 7:2:1 ratio. All slices were then standardized.

[0064] 2. Simulation test setup.

[0065] Hardware and software environment: The experiment was conducted on a server equipped with an NVIDIA Tesla V100 GPU (32GB of video memory), the operating system was Ubuntu 20.04, and the PyTorch 1.12.1 deep learning framework was used.

[0066] Parameter settings for the method of this invention: Multi-view enhancement: The transformations include: identity, rotation by 30°, and scaling by 0.9 times.

[0067] Feature encoding network: ResNet-50 pre-trained on ImageNet is used.

[0068] Iterative optimization: Maximum number of iterations Quality assessment parameters Confidence correction factor Position base adjustment rate Convergence threshold .

[0069] Training parameters: The Adam optimizer was used with an initial learning rate of 0.001, adjusted using cosine annealing. The batch size was 8, and the total number of training epochs was 100. In the loss function, the maximum weight of the adaptive consistency loss was 0.5.

[0070] Comparison Method Selection: To comprehensively evaluate performance, three representative existing technologies were selected as comparison baselines: Pure segmentation model: U-Net (based on ResNet-50 encoder), representing the current mainstream single-task method for landslide identification.

[0071] Pure detection model: Faster R-CNN (based on ResNet-50-FPN backbone network), representing a bounding box-based localization method.

[0072] Simple dual-task model: A shared ResNet-50 backbone network followed by parallel detection and segmentation heads (Dual-task Net), representing existing static feature fusion methods.

[0073] Evaluation indicators: For the segmentation task: the mean intersection-union ratio (mIoU) and F1 score are used, both of which take into account the accuracy of landslide category identification.

[0074] For the detection task: Average Precision (AP) is used, and the cross-union threshold is set to 0.5.

[0075] Comprehensive assessment: Recall is used to measure the model's ability to identify all landslides, and False Alarm Rate is used to measure the number of false alarms.

[0076] 3. Comparative experimental results and analysis.

[0077] As shown in Table 1, the method of this invention significantly outperforms the baseline method on all core metrics. Compared with the pure segmentation model (U-Net), mIoU is improved by 8.5 percentage points, demonstrating that iterative optimization greatly improves boundary accuracy. Compared with the pure detection model (Faster R-CNN), AP is improved by 7.3 percentage points, and the false alarm rate is significantly reduced, demonstrating that the quality assessment and correction mechanism effectively filters false alarm boxes. Compared with the simple dual-task model (Dual-task Net), this invention has significant advantages on all metrics, demonstrating the superiority of "dynamic iterative collaborative optimization" over the "static parallel" architecture.

[0078] Table 1 Overall performance comparison on the LLSD test set

[0079] As shown in Table 2, the method of this invention exhibits stronger robustness in identifying landslides of different scales, especially improving the mIoU of small-scale landslides by more than 10 percentage points. This is due to the localization guidance provided by the initial detection box and the enhanced attention paid to small targets with low confidence during the iteration process, effectively alleviating the problem of insufficient feature capture of small targets by pure segmentation models.

[0080] Table 2 Comparison of segmentation performance on landslide targets at different scales (mIoU, %)

[0081] As shown in Table 3, when the model is applied to new regions with different terrains and resolutions (GF2-LS), the performance of all baseline models shows a significant decline. In contrast, the method of this invention exhibits stronger cross-domain generalization ability and the smallest performance degradation.

[0082] Table 3. Cross-domain generalization ability test on the GF2-LS dataset

[0083] All models were trained on the LLSD dataset and evaluated directly on the GF2-LS test set.

[0084] To further verify the optimization effect of the adaptive consistency loss function used in this invention on the model training process, the convergence curve of the validation set loss value during training was plotted, as shown below. Figure 3 As shown, the changes in loss values ​​between the baseline method (Dual-task Net) and the method of this invention are intuitively compared under the same number of training rounds.

[0085] Example 2 An automatic landslide hazard identification system based on optical remote sensing images includes: The multi-view feature extraction module is used to perform geometric transformations on the input optical remote sensing images to generate multi-view images and fuse their depth features to obtain robust feature representations. The collaborative perception and initial solution generation module is used to generate an initial set of landslide detection boxes and an initial semantic segmentation map by simultaneously performing target detection and semantic segmentation tasks through an integrated model based on robust feature representation, which serve as the initial input for iterative optimization. The solution space iterative optimization module is used to execute an iterative optimization loop starting from the initial input. The solution space iterative optimization module includes: The segmentation quality assessment unit is used to assess the quality of the landslide region in the corresponding detection box based on the semantic segmentation map of the current iteration round. The detection box correction unit is used to correct the confidence and geometric position of the detection box in the current iteration based on the quality assessment results. The segmentation graph optimization guidance unit is used to guide optimization using the corrected detection boxes to generate the semantic segmentation graph used in the next iteration round; The convergence result output module is used to output the set of landslide detection boxes and their corresponding pixel-level segmentation masks generated in the current iteration when the iterative optimization loop meets the predetermined termination condition.

[0086] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0087] Therefore, the present invention employs the above-mentioned method and system for automatic identification of landslide disasters using optical remote sensing images, which can effectively solve the problems of inaccurate target positioning, rough boundary identification, model dependence on a large amount of labeled data, and poor regional adaptability in the existing technology, providing an efficient and reliable technical means for rapid and accurate monitoring of geological disasters.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for automatic identification of landslide hazards from optical remote sensing images, characterized in that, Includes the following steps: Step S1: Perform geometric transformation on the input optical remote sensing image to generate a multi-view image, and fuse its depth features to obtain a robust feature representation. Step S2: Based on robust feature representation, the target detection and semantic segmentation tasks are executed synchronously through an integrated model to generate an initial set of landslide detection boxes and an initial semantic segmentation map, which serve as the initial input for iterative optimization. Step S3: Starting from the initial input, execute the following iterative optimization loop: Step S31: Based on the semantic segmentation map of the current iteration, evaluate the quality of the landslide region in the corresponding detection box; Step S32: Based on the quality assessment results, correct the confidence and geometric position of the detection box in the current iteration; Step S33: Use the corrected detection box to guide optimization and generate the semantic segmentation map for the next iteration. Step S4: When the iterative optimization loop meets the predetermined termination condition, output the set of landslide detection boxes generated in the current iteration and its corresponding pixel-level segmentation mask.

2. The method for automatic identification of landslide hazards from optical remote sensing images according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S11: Process the input image application Different affine transformations Generate multi-view image groups ; Step S12: Through a feature encoding network with shared weights Extract feature maps from images at each viewpoint. ; Step S13: Concatenate the multi-view feature maps along the channel dimension and pass them through a convolutional layer. By performing fusion and dimensionality reduction, robust feature representations are obtained. The calculation formula is as follows: 。 3. The method for automatic identification of landslide hazards from optical remote sensing images according to claim 2, characterized in that, The ensemble model used in step S2 is used to handle the characteristics of the landslide target, specifically including: Step S21: Represent robustness features Input Region Proposal Network Generate a set of candidate landslide area proposals ; Step S22, Detection and Segmentation Co-processing: a) Detection branch: For each proposal Through region of interest pooling layer from Extract fixed-size features and input them into the detection head. Predict the initial detection box and its confidence level ; b) Branch segmentation: Representing robustness features Input a semantic segmentation network with skip connections Jump connections are used to fuse shallow landslide boundary texture features with deep landslide semantic features; at the same time, candidate landslide region proposals are utilized. Generate spatial attention map This allows computational resources to be focused on candidate regions in the initial stage to generate an initial semantic segmentation map. .

4. The method for automatic identification of landslide hazards from optical remote sensing images according to claim 3, characterized in that, In step S31, for the first The detection box in the next iteration and its corresponding segmentation region mask Its overall quality score The calculations are used to distinguish between landslide and non-landslide areas: Step S311: Calculate the first morphological index : ； in, The function represents the intersection-union ratio calculation function. This represents a predefined morphological compaction operation function; Step S312: Calculate the second statistical indicator : ； in, This represents the function for calculating standard deviation. Indicates the first In the nth iteration, the 1st The area mask corresponding to each detection box The set of predicted probability values ​​for landslide categories for all pixels within the range; Step S313, Indicator Adjustment: Using the second statistical indicator For the first form indicator Make corrections to obtain the corrected morphological index. : ； in, This represents the preset morphological reliability adjustment factor, which is a constant greater than zero; Step S314: Calculate the overall quality score : 。 5. The method for automatic identification of landslide hazards from optical remote sensing images according to claim 4, characterized in that, The detection box correction in step S32 has a correction range that is related to the overall quality score. Adaptive association: Step S321: Based on the overall quality score For the The original confidence of each detection box Make corrections to obtain the corrected confidence level. The calculation formula is as follows: ； Among them, the regulating factor ; Step S322: Based on the overall quality score For the The original geometric position of each detection box Make corrections to obtain the corrected geometric position. The calculation formula is as follows: ； in, To calculate the minimum bounding rectangle of the segmented region mask, adjust the coefficients. , This is the preset base adjustment rate.

6. The method for automatic identification of landslide hazards from optical remote sensing images according to claim 5, characterized in that, The predetermined termination condition in step S4 is first satisfied if one of the following two conditions is met: a) Number of iterations Reaching the maximum value preset for the landslide identification task; b) The average change in the overall quality score of all retained landslide candidate boxes between two adjacent iterations. Less than the threshold ,in: ； in, For the first The number of detection boxes in the next iteration.

7. The method for automatic identification of landslide hazards from optical remote sensing images according to claim 6, characterized in that, The ensemble model is trained using a loss function specifically designed for landslide identification. : Get the Step detection loss and segmentation loss ; Calculate the consistency loss of predictions between adjacent steps. And assign adaptive weights to them. ,in For the first The average confidence level of the step segmentation prediction The preset maximum weight constant; Calculate the total loss: ； in, This represents the total number of iterations simulated during training.

8. The method for automatic identification of landslide hazards from optical remote sensing images according to claim 7, characterized in that, During the model training phase, a class balancing strategy is adopted, and focus loss weights are applied to difficult pixel samples in the segmentation loss that are similar to the spectral or texture features of landslides.

9. An automatic identification system for landslide hazards based on optical remote sensing images, characterized in that, The method for automatically identifying landslide hazards from optical remote sensing images as described in any one of claims 1-8 includes: The multi-view feature extraction module is used to perform geometric transformations on the input optical remote sensing images to generate multi-view images and fuse their depth features to obtain robust feature representations. The collaborative perception and initial solution generation module is used to generate an initial set of landslide detection boxes and an initial semantic segmentation map by simultaneously performing target detection and semantic segmentation tasks through an integrated model based on robust feature representation, which serve as the initial input for iterative optimization. The solution space iterative optimization module is used to execute an iterative optimization loop starting from the initial input; The convergence result output module is used to output the set of landslide detection boxes and their corresponding pixel-level segmentation masks generated in the current iteration when the iterative optimization loop meets the predetermined termination condition.

10. The automatic identification system for landslide hazards based on optical remote sensing images according to claim 9, characterized in that, include: The solution space iterative optimization module includes: The segmentation quality assessment unit is used to assess the quality of the landslide region in the corresponding detection box based on the semantic segmentation map of the current iteration round. The detection box correction unit is used to correct the confidence and geometric position of the detection box in the current iteration based on the quality assessment results. The segmentation graph optimization guidance unit is used to guide optimization using the corrected detection boxes to generate the semantic segmentation graph for use in the next iteration round.

Images