Methods, systems and electronic devices for concealed landslide target identification and semantic segmentation

Abstract

This invention relates to the field of landslide recognition technology, and more particularly to a method, system, and electronic device for hidden landslide target recognition and semantic segmentation. It proposes a model collaboration method that iteratively optimizes the alternating training of a target-level contrastive learning landslide target classification network and a quasi-target-level contrastive learning landslide semantic segmentation network. The highly accurate target-level contrastive learning landslide target classification network is used to supervise the feature extraction of the highly complex quasi-target-level contrastive learning landslide semantic segmentation network, thereby improving the reliability of feature extraction. Experimental results show that this invention can simultaneously improve the accuracy of target-level landslide recognition and the accuracy of hidden landslide contour recognition.

Description

Technical Field

[0001] This invention relates to the field of landslide identification technology, and in particular to a method, system and electronic device for identifying and semantically segmenting concealed landslide targets. Background Technology

[0002] Landslides are a frequent natural disaster that causes severe personal and property damage as well as environmental destruction. Accurate prediction of potential landslides relies on analyzing a large number of existing landslides to fully understand their geological and geomorphological characteristics and triggering factors. Traditional landslide identification relies on manual field observation and mapping, which is time-consuming and labor-intensive, and cannot meet the requirements of high efficiency and automation for large-scale surveys. With the rapid development of remote sensing technology, due to its high precision, diverse data sources, and wide coverage, it has been widely used for landslide detection. Utilizing high-resolution remote sensing data for intelligent landslide identification has become the mainstream technology.

[0003] Landslides can be classified into three types based on the time of their occurrence: new landslides, ancient / old landslides, and landslide hazard. New landslides refer to those that occurred recently and may continue to move; their surface features differ significantly from the surrounding area in color and shape. Ancient / old landslides refer to those that occurred a long time ago but may reactivate under certain triggering factors. Their landslide characteristics are difficult to distinguish from the environment due to long-term effects of natural or human factors, making them the most difficult type of landslide to identify from remote sensing images; hence, they are also called hidden landslides. Landslide hazard refers to those that have begun surface subsidence but have not yet resulted in a landslide. Under the influence of external triggering factors such as rainfall and earthquakes, they may rapidly develop into landslide disasters, making them the most dangerous type of landslide.

[0004] In the intelligent identification of landslide targets, new landslides have color and morphological features that are significantly different from the environment, and landslide hazards have surface subsidence movement. Both are relatively easy to identify. However, ancient / old landslides, due to their long history and subsequent human modifications (reconstruction, reclamation) and natural erosion, have surface features that are extremely similar to the environment. At the same time, remote sensing image data lacks depth / height information, which reduces the accuracy of landslide identification. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a method, system, storage medium and electronic device for hidden landslide target identification and semantic segmentation landslide identification.

[0006] The technical solution of the concealed landslide target recognition and semantic segmentation landslide recognition method of the present invention is as follows:

[0007] S1. Establish a hidden landslide intelligent identification model that includes a landslide target classification network with target-level contrastive learning and a landslide semantic segmentation network with quasi-target-level contrastive learning. The landslide target classification network with target-level contrastive learning includes a classifier, and a first encoder and a second encoder with two identical twin network structures that share network parameters. The landslide semantic segmentation network with quasi-target-level contrastive learning includes a third encoder and a decoder. The structure of the third encoder is the same as that of the first encoder.

[0008] The target-level contrastive learning landslide target classification network is used to: extract semantic features of a first input image through the first encoder, extract semantic features of a second input image through the second encoder, perform channel concatenation on the semantic features of the first input image and the semantic features of the second input image to obtain a high-dimensional feature map representing the first input image and the second input image, and the classifier obtains the class confidence level that is commonly associated with the first input image and the second input image based on the high-dimensional feature map.

[0009] The quasi-target-level contrastive learning landslide semantic segmentation network is used to: extract semantic features of the third input image through the third encoder, calculate the class confidence of each pixel in the third input image based on the semantic features of the third input image using the decoder, and obtain pixel-level landslide recognition results based on the class confidence of each pixel.

[0010] S2. Based on the dataset, train the landslide target classification network of the target-level contrastive learning until it converges;

[0011] S3. Extract the network parameters of the first encoder or the second encoder in the landslide target classification network after convergence of the target-level contrastive learning, load them into the third encoder and freeze them, and then use the quasi-target-level contrastive learning strategy to train the decoder until convergence.

[0012] S4. Unlock the network parameters of the third encoder, and use the quasi-target-level contrastive learning strategy to train the quasi-target-level contrastive learning slippage semantic segmentation network until convergence.

[0013] S5. Extract the network parameters of the third encoder in the converged quasi-target-level contrastive learning landslide semantic segmentation network, load them into the current first encoder and second encoder, return to execute S2, until a quasi-target-level contrastive learning landslide semantic segmentation network that meets the preset conditions is obtained, and a trained hidden landslide intelligent identification model is obtained.

[0014] S6. Use the trained intelligent identification model for hidden landslides to identify landslides.

[0015] The technical solution of the concealed landslide target recognition and semantic segmentation landslide recognition system of the present invention is as follows:

[0016] It includes a model building module, a first training submodule, a second training submodule, a third training submodule, a fourth training submodule, and a recognition module;

[0017] The model building module is used to: build a hidden landslide intelligent identification model including a landslide target classification network with target-level contrastive learning and a landslide semantic segmentation network with quasi-target-level contrastive learning. The landslide target classification network with target-level contrastive learning includes a classifier, and a first encoder and a second encoder with two identical twin network structures that share network parameters. The landslide semantic segmentation network with quasi-target-level contrastive learning includes a third encoder and a decoder. The structure of the third encoder is the same as that of the first encoder.

[0018] The target-level contrastive learning landslide target classification network is used to: extract semantic features of a first input image through the first encoder, extract semantic features of a second input image through the second encoder, perform channel concatenation on the semantic features of the first input image and the semantic features of the second input image to obtain a high-dimensional feature map representing the first input image and the second input image, and the classifier obtains the class confidence level that is commonly associated with the first input image and the second input image based on the high-dimensional feature map.

[0019] The quasi-target-level contrastive learning landslide semantic segmentation network is used to: extract semantic features of the third input image through the third encoder, calculate the class confidence of each pixel in the third input image based on the semantic features of the third input image using the decoder, and obtain pixel-level landslide recognition results based on the class confidence of each pixel.

[0020] The first training submodule is used to: train the landslide target classification network of the target-level contrastive learning to converge based on the dataset;

[0021] The second training submodule is used to: extract the network parameters of the first encoder or the second encoder in the converged target-level contrastive learning landslide target classification network, load them into the third encoder and freeze them, and then use the quasi-target-level contrastive learning strategy to train the decoder until convergence;

[0022] The third training submodule is used to: unlock the network parameters of the third encoder, and train the quasi-target-level contrastive learning slippery semantic segmentation network until convergence using the quasi-target-level contrastive learning strategy.

[0023] The fourth training submodule is used to: extract the network parameters of the third encoder in the converged quasi-target-level contrastive learning landslide semantic segmentation network, load them into the current first encoder and second encoder, and re-call the first training submodule, the second training submodule and the third training module until a quasi-target-level contrastive learning landslide semantic segmentation network that meets the preset conditions is obtained, and a trained hidden landslide intelligent identification model is obtained.

[0024] The identification module is used to identify landslides using the trained intelligent identification model for concealed landslides.

[0025] The present invention provides a storage medium storing instructions, which, when read by a computer, cause the computer to execute the concealed landslide target identification and semantic segmentation landslide identification method described in any of the preceding claims.

[0026] An electronic device according to the present invention includes a processor and the above-described storage medium, wherein the processor executes instructions in the storage medium.

[0027] The technical effects of this invention are as follows:

[0028] A collaborative model method is proposed, which involves alternating training and iterative optimization of a landslide target classification network based on target-level contrastive learning and a landslide semantic segmentation network based on quasi-target-level contrastive learning. The landslide target classification network based on target-level contrastive learning, which has high accuracy, is used to supervise the feature extraction of the landslide semantic segmentation network based on quasi-target-level contrastive learning, thereby improving the reliability of feature extraction. Experimental results show that this invention can simultaneously improve the accuracy of target-level landslide recognition and the accuracy of hidden landslide contour recognition. Attached Figure Description

[0029] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0030] Figure 1 This is a flowchart illustrating a method for identifying concealed landslide targets and semantically segmenting landslides according to an embodiment of the present invention.

[0031] Figure 2 A schematic diagram of the structure of a hidden landslide intelligent identification model, which includes a landslide target classification network based on target-level contrastive learning and a landslide semantic segmentation network based on quasi-target-level contrastive learning.

[0032] Figure 3 This is a schematic diagram of the structure of a concealed landslide target recognition and semantic segmentation landslide recognition system according to an embodiment of the present invention. Detailed Implementation

[0033] like Figure 1As shown in the figure, a method for identifying concealed landslide targets and semantically segmenting landslides according to an embodiment of the present invention includes the following steps:

[0034] S1. Establish a hidden landslide intelligent identification model that includes a landslide target classification network based on target-level contrastive learning and a landslide semantic segmentation network based on quasi-target-level contrastive learning, such as... Figure 2 As shown, the target-level contrastive learning landslide target classification network includes a classifier, and a first encoder and a second encoder with two identical Siamese network structures that share network parameters. The quasi-target-level contrastive learning landslide semantic segmentation network includes a third encoder and a decoder. The structure of the third encoder is the same as that of the first encoder.

[0035] The landslide target classification network based on target-level contrastive learning is used to: extract semantic features of the first input image through the first encoder, extract semantic features of the second input image through the second encoder, concatenate the semantic features of the first input image and the second input image through channel concatenation to obtain a high-dimensional feature map that represents the first input image and the second input image, and the classifier obtains the class confidence that the first input image and the second input image are jointly corresponding to based on the high-dimensional feature map.

[0036] The quasi-target-level contrastive learning landslide semantic segmentation network is used to: extract semantic features of the third input image through the third encoder, use the decoder to calculate the class confidence of each pixel in the third input image based on the semantic features of the third input image, and obtain the pixel-level landslide recognition result based on the class confidence of each pixel.

[0037] Wherein, the first encoder employs a deep convolutional neural network, and the process of extracting semantic features from the first input image through the first encoder includes:

[0038] S10. Extract semantic features of the first input image through multi-layer stacked convolution operations in a deep convolutional neural network, combined with an attention mechanism.

[0039] Since the first encoder has the same structure as the second and third encoders, the semantic features of the second input image and the semantic features of the third input image can be obtained similarly.

[0040] The decoder includes an upsampling block, which comprises a convolutional layer and a transposed convolutional layer arranged sequentially. The decoder calculates the class confidence score of each pixel in the third input image based on the semantic features of the third input image, including:

[0041] S11. The semantic features of the third input image are restored to the size of the third input image through upsampling blocks, and the class confidence of each pixel in the third input image is calculated. This invention proposes a model collaboration method that alternately trains and iteratively optimizes a landslide target classification network based on target-level contrastive learning and a landslide semantic segmentation network based on quasi-target-level contrastive learning. The highly accurate target-level contrastive learning landslide target classification network is used to supervise the feature extraction of the highly complex quasi-target-level contrastive learning landslide semantic segmentation network, thereby improving the reliability of feature extraction. Experimental results show that this invention can simultaneously improve the accuracy of target-level landslide recognition and the accuracy of hidden landslide contour recognition.

[0042] The intelligent identification model for hidden landslides is explained as follows:

[0043] 1) The upper branch, namely the landslide target classification network of target-level contrastive learning, includes a classifier, and a first encoder and a second encoder with two identical twin network structures that share network parameters. Both the first encoder and the second encoder use a deep convolutional neural network, such as the ResNet 101 network, which has excellent semantic feature extraction performance, as the backbone network. The semantic features in the input image are extracted by combining the multi-layer stacked convolution operation in the deep convolutional neural network with the attention mechanism.

[0044] Simultaneously, the first encoder and the second encoder employ a Siamese network structure with shared network parameters to process a set of input images. Each set of input images includes two input images, namely the first input image and the second input image. The first encoder extracts the semantic features of the first input image, and the second encoder extracts the semantic features of the second input image. The semantic features of the first input image and the second input image are concatenated by channels to obtain a high-dimensional feature map representing the first input image and the second input image. The classifier maps the high-dimensional feature map into a feature vector and uses an activation function to obtain the class confidence that is common to both the first input image and the second input image.

[0045] The upper branch inputs two randomly selected first and second input images. If the input image includes a landslide area, it is a positive sample; if the input image does not include a landslide area, it is a negative sample. The first and second input images can form four classification results: positive sample-positive sample, positive sample-negative sample, negative sample-positive sample, and negative sample-negative sample.

[0046] The loss function of the landslide target classification network in the target-level contrastive learning can use the traditional cross-entropy loss function. The first encoder and the second encoder independently extract target-level semantic features from the first input image and the second input image, respectively. By enhancing the features of similar samples and separating the features of dissimilar samples, the extraction of global salient features of hidden landslides is achieved more reliably and efficiently. Similar samples refer to either both the first input image and the second input image being positive samples or both being negative samples. Dissimilar samples refer to either one of the first input image and the second input image being a positive sample and the other being a negative sample.

[0047] 2) The lower branch, namely the quasi-target-level contrastive learning landslide semantic segmentation network, includes a third encoder and a decoder. The structure of the third encoder is the same as that of the first encoder. During the training phase, both cross-entropy loss and supervised contrastive loss are introduced to update the network. The structure of the third encoder is the same as that of the first and second encoders in the upper branch, realizing the extraction of high-dimensional semantic features from the input image, i.e., the third input image. The decoder consists of upsampling blocks composed of convolutional layers and transposed convolutional layers. It is responsible for restoring the high-dimensional semantic features of the third input image to the size of the third input image, calculating the class confidence of each pixel in the third input image, and finally outputting the pixel-level landslide recognition result required by the task.

[0048] S2. Based on the dataset, train the landslide target classification network of target-level contrastive learning until convergence;

[0049] S3. Extract the network parameters of the first or second encoder in the landslide target classification network after convergence of target-level contrastive learning, load them into the third encoder and freeze them, and then use the quasi-target-level contrastive learning strategy to train the decoder until convergence.

[0050] S4. Unlock the network parameters of the third encoder, and use a quasi-target-level contrastive learning strategy to train the quasi-target-level contrastive learning slippery semantic segmentation network until convergence.

[0051] S5. Extract the network parameters of the third encoder in the converged quasi-target-level contrastive learning landslide semantic segmentation network, load them into the current first encoder and second encoder, return to execute S2, until a quasi-target-level contrastive learning landslide semantic segmentation network that meets the preset conditions is obtained, and a trained hidden landslide intelligent identification model is obtained.

[0052] The preset condition is that the validation loss of the slippery semantic segmentation network in the quasi-target level contrastive learning will no longer decrease. Other preset conditions can also be set according to the actual situation.

[0053] S6. Use the trained intelligent identification model for hidden landslides to identify landslides. Specifically, use the trained intelligent identification model for hidden landslides to identify high-resolution remote sensing images and determine the landslide identification results.

[0054] The classification task achieved by the upper branch has high reliability in identifying hidden landslide targets; the lower branch is a semantic segmentation network for hidden landslides, which shares the feature extraction network with the upper branch and introduces a quasi-target contrast learning module to further optimize the feature extraction module; the upper and lower branches are trained and iteratively optimized to enable the upper branch to guide the semantic segmentation network of the lower branch, thereby reducing the missed and false judgments in the identification of hidden landslide targets.

[0055] Currently, the automatic identification technology for hidden landslides of ancient and old landslide types based on high-resolution remote sensing image data mainly faces two problems: data and methodology.

[0056] 1) Missing information:

[0057] Landslide identification by experts mainly relies on a combination of digital elevation information and surface texture features to identify the concave-convex morphological features of the landslide body and the sudden drop in height at the edge of the landslide body. These features all depend on the height change information of the landslide body. However, high-resolution remote sensing image data lacks height / depth information. The texture features generated by projecting height information onto 2D images are very similar to the natural shadows of the mountain, making it extremely difficult to distinguish between landslides and non-landslides from HRSI data.

[0058] 2) Small sample size problem:

[0059] While high-resolution remote sensing data is abundant, accurate landslide labeling remains a challenge. On one hand, accurate landslide identification relies heavily on expert experience; on the other hand, delineating landslides for semantic segmentation in academic papers is time-consuming and labor-intensive, resulting in a lack of large-scale public datasets. The small sample size problem severely restricts the performance of landslide identification methods.

[0060] The quasi-target-level contrastive learning strategy proposed in this invention can solve the above two problems, and can further improve the accuracy of target-level landslide identification and the accuracy of hidden landslide contour identification. The explanation of the quasi-target-level contrastive learning strategy is as follows:

[0061] In S3, a quasi-target-level contrastive learning strategy is employed to train the decoder until convergence, including:

[0062] S30. Construct a quasi-target-level positive and negative sample set, and make the quasi-target-level positive and negative sample set correspond to the semantic feature embedding of the semantic features currently extracted by the third encoder. Use the supervised contrastive loss function to assist the cross-entropy loss function to update the third encoder and train the decoder to convergence.

[0063] Optionally, in the above technical solution, in S4, the quasi-target-level contrastive learning strategy trains the quasi-target-level contrastive learning slippery semantic segmentation network to convergence, including:

[0064] S4. Construct a quasi-target-level positive and negative sample set, and make the quasi-target-level positive and negative sample set correspond to the semantic feature embedding of the semantic features currently extracted by the third encoder. Use the supervised contrastive loss function to assist the cross-entropy loss function to update the third encoder, and train the quasi-target-level contrastive learning slippery semantic segmentation network until convergence.

[0065] The quasi-target-level contrastive learning strategy includes the following:

[0066] 1) Construct a quasi-target level set of positive and negative samples:

[0067] Using fixed-size rectangles, the binary classification labels of the training images in the dataset are bounded, resulting in multiple label blocks. Each label block represents a local label at a corresponding location in the training image. All label blocks are then filtered, specifically:

[0068] Label blocks containing pixels related to landslides within a preset range are identified as Class 1 samples of the landslide boundary class; the preset range is 10% to 90%, and can also be set according to actual conditions; label blocks containing pixels related to landslides below the lower limit of the preset range are identified as Class 0 samples of the landslide boundary class, for example, the preset range is 10% to 90%, and the lower limit of the preset range is 10%.

[0069] Label blocks in which the proportion of pixels containing landslides is higher than the upper limit of a preset range are identified as samples inside the landslide and defined as irrelevant samples. For example, the preset range is 10% to 90%, and the upper limit of the preset range is 90%.

[0070] Samples of class 1 and class 0 are selected to participate in the calculation of the supervised contrastive loss function, while irrelevant samples are discarded and are not included in the calculation of the supervised contrastive loss function.

[0071] Randomly select a preset number of samples from the selected Class 1 and Class 0 samples, or select them according to actual needs, to form a quasi-target level positive and negative sample set. The preset number can be 10, 100, etc., and can be set according to the actual situation.

[0072] 2) Semantic feature embedding and supervised contrastive loss function:

[0073] The semantic features of the third input image output by the third encoder are normalized, and the normalized semantic features are then split into multiple quasi-target-level feature blocks in the spatial dimension. Specifically:

[0074] For the semantic features of the third input image, spatial dimensions are split at each feature point in both the horizontal and vertical directions, resulting in quasi-target-level feature blocks equal to the horizontal size multiplied by the vertical size. Each quasi-target-level feature block can be approximated as the semantic features within the corresponding region of the third input image and can correspond to a set of quasi-target-level positive and negative samples constructed for that region. The quasi-target-level feature blocks in a batch are flattened, and a fixed number of equal numbers of feature blocks corresponding to class 1 samples and class 0 samples are randomly selected from all quasi-target-level feature blocks. The loss of the lower branch is calculated using the supervised contrastive learning loss function, which is:

[0075]

[0076] Among them, P i Let i represent the total number of samples in the quasi-target level sample set, and let i represent the features of the current sample. + The characteristics of Class 1 samples representing landslides, i - Features representing class 0 samples, Represents the supervised contrastive learning loss function, where τ represents the temperature coefficient.

[0077] In addition to the supervised contrastive learning loss function mentioned above, the quasi-target-level contrastive learning strategy also includes another loss function, namely the cross-entropy loss function. The supervised contrastive learning loss function and the cross-entropy loss function are related by a function... To conduct collaborative training, including training on the next branch, where L... SEG Let i represent the total loss function of the quasi-target-level contrastive learning strategy, and let i represent the features of the current sample. The cross-entropy loss function represents the features of the current sample, where λ represents the scaling factor. The supervised contrastive learning loss function represents the features of the current sample.

[0078] 2) A model collaboration method is proposed, which involves alternating training and iterative optimization of a landslide target classification network based on target-level contrastive learning and a landslide semantic segmentation network based on quasi-target-level contrastive learning. The highly reliable landslide target classification network based on target-level contrastive learning guides the highly complex semantic feature extraction of the landslide semantic segmentation network based on quasi-target-level contrastive learning. In the landslide target classification network based on target-level contrastive learning, a first encoder and a second encoder with identical structures and shared network parameters are designed to improve the efficiency and accuracy of semantic feature extraction.

[0079] The upper branch classification network employs two identical Siamese network structures—a first encoder and a second encoder—that share network parameters, focusing on target-level landslide features. The lower branch uses a quasi-target-level contrastive learning strategy, focusing on the salient semantic features of the landslide boundary. The first, second, and third encoders have identical structures and are trained iteratively and alternately, enabling the lower branch to effectively reduce the target-level misclassification rate while ensuring the accuracy of the predicted contour.

[0080] The technical effects of the present invention are illustrated by the following embodiments:

[0081] The dataset used in the performance testing experiment is an ancient / old landslide dataset from the Loess Plateau region of Northwest my country. These landslides occurred a long time ago, and many have been transformed into farmland or residential areas. Their shapes and textures are highly similar to the surrounding environment, making them difficult to identify from high-resolution remote sensing images. The original image resolution was 2m / pixel. These images were cropped to obtain 304 landslide samples of 512×512 pixels. The samples were then divided into training, testing, and validation sets in a 6:2:2 ratio. Data augmentation was applied to the training and validation sets by adding one-third of the landslide samples as class 0 samples, resulting in the final dataset, as shown in Table 1.

[0082] Table 1:

[0083] training set Validation set test set 1460 486 80

[0084] The classic DeepLabV3+ model was used as the benchmark model, and the performance of the method proposed in this patent for landslide identification was compared. The experimental results are shown in the table below. As can be seen from the experimental results, the method designed in this patent has the best performance in all performance metrics. Semantic segmentation performance comparison is shown in Table 2, landslide target classification performance comparison is shown in Table 3, and the confusion matrix of target classification performance after the first iteration is shown in Table 4. At this point, the average accuracy is 0.8375, precision is 0.9607, recall is 0.8167, and F1 score is 0.8846. The confusion matrix of the target classification performance of the proposed method after the second iteration is shown in Table 5. At this point, the accuracy is 0.8875, precision is 0.9473, recall is 0.9, and F1 score is 0.9230.

[0085] Table 2:

[0086]

[0087] Table 3:

[0088] landslide accuracy Slope accuracy Average accuracy Iteration 1 0.817 0.9 0.859 Iteration 2 0.9 0.85 0.875

[0089] Table 4:

[0090]

[0091]

[0092] Table 5:

[0093] Actual / Forecast landslide slope landslide 54 6 slope 3 17

[0094] Furthermore, the target classification performance of landslides and slopes after semantic segmentation was further evaluated. For landslides, if more than 400 target class pixels overlapped with the label in the semantic segmentation result, the landslide target classification was considered correct; otherwise, it was considered a missed classification error. For slopes, if fewer than 100 target class pixels appeared in the semantic segmentation prediction result, the slope classification was considered correct; otherwise, it was considered a misclassification error. The experimental results are shown in Table 6. From the experimental results, it can be seen that the proposed solution improves the classification accuracy by iteratively optimizing the semantic feature extraction network through upper and lower branches.

[0095] Table 6:

[0096] In the above embodiments, although the steps are numbered S1, S2, etc., they are only specific embodiments given in this application. Those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation, which is also within the protection scope of this invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0097] like Figure 3 As shown, a concealed landslide target recognition and semantic segmentation landslide recognition system 200 according to an embodiment of the present invention includes a model building module 210, a first training submodule 220, a second training submodule 230, a third training submodule 240, a fourth training submodule 250 and a recognition module 260.

[0098] Model building module 210 is used to: build an intelligent identification model for hidden landslides, including a landslide target classification network based on target-level contrastive learning and a landslide semantic segmentation network based on quasi-target-level contrastive learning.

[0099] landslide accuracy Slope accuracy Average accuracy DeepLabV3+ 0.73 0.55 0.67 Propose a model 0.73 0.9 0.82

[0100] The contrastive learning landslide target classification network includes a classifier, and a first encoder and a second encoder with two identical twin network structures that share network parameters. The quasi-target-level contrastive learning landslide semantic segmentation network includes a third encoder and a decoder, with the structure of the third encoder being the same as that of the first encoder.

[0101] The landslide target classification network based on target-level contrastive learning is used to: extract semantic features of the first input image through the first encoder, extract semantic features of the second input image through the second encoder, concatenate the semantic features of the first input image and the second input image through channel concatenation to obtain a high-dimensional feature map that represents the first input image and the second input image, and the classifier obtains the class confidence that the first input image and the second input image are jointly corresponding to based on the high-dimensional feature map.

[0102] The quasi-target-level contrastive learning landslide semantic segmentation network is used to: extract semantic features of the third input image through the third encoder, use the decoder to calculate the class confidence of each pixel in the third input image based on the semantic features of the third input image, and obtain the pixel-level landslide recognition result based on the class confidence of each pixel.

[0103] The first training submodule 220 is used to: train the landslide target classification network of target-level contrastive learning to convergence based on the dataset;

[0104] The second training submodule 230 is used to: extract the network parameters of the first encoder or the second encoder in the converged target-level contrastive learning landslide target classification network, load them into the third encoder and freeze them, and then use a quasi-target-level contrastive learning strategy to train the decoder until convergence.

[0105] The third training submodule 240 is used to: unlock the network parameters of the third encoder, adopt a quasi-target-level contrastive learning strategy, and train the quasi-target-level contrastive learning slippery semantic segmentation network until convergence.

[0106] The fourth training submodule 250 is used to: extract the network parameters of the third encoder in the converged quasi-target-level contrastive learning landslide semantic segmentation network, load them into the current first encoder and second encoder, and re-call the first training submodule, the second training submodule and the third training submodule until a quasi-target-level contrastive learning landslide semantic segmentation network that meets the preset conditions is obtained, and a trained hidden landslide intelligent identification model is obtained.

[0107] The identification module 260 is used to identify landslides using a trained intelligent identification model for concealed landslides.

[0108] Optionally, in the above technical solution, the second training submodule 230 is specifically used for:

[0109] Construct a first quasi-target level positive and negative sample set, and the quasi-target level positive and negative sample set corresponds to the semantic feature embedding of the semantic features currently extracted by the third encoder. Use the supervised contrastive loss function to assist the cross-entropy loss function to update the third encoder, and train the decoder to convergence.

[0110] Optionally, in the above technical solution, the third training submodule 240 is used for:

[0111] A second quasi-target level positive and negative sample set is constructed, and the quasi-target level positive and negative sample set corresponds to the semantic feature embedding of the semantic features currently extracted by the third encoder. The third encoder is updated by using the supervised contrastive loss function to assist the cross-entropy loss function, and the quasi-target level contrastive learning slippery semantic segmentation network is trained to convergence.

[0112] Optionally, in the above technical solution, if the first encoder employs a deep convolutional neural network, then the process of extracting semantic features of the first input image through the first encoder includes:

[0113] Semantic features of the first input image are extracted through multi-layered stacked convolution operations in a deep convolutional neural network, combined with an attention mechanism.

[0114] Optionally, in the above technical solution, the decoder includes an upsampling block, which includes a convolutional layer and a transposed convolutional layer arranged sequentially. The decoder calculates the class confidence of each pixel in the third input image based on the semantic features of the third input image, including:

[0115] The semantic features of the third input image are restored to the size of the third input image by upsampling blocks, and the class confidence of each pixel in the third input image is calculated.

[0116] The parameters and steps of each unit module in the concealed landslide target identification and semantic segmentation landslide identification system 200 of the present invention described above can be referred to the parameters and steps in the embodiments of the concealed landslide target identification and semantic segmentation landslide identification method described above, and will not be repeated here.

[0117] An embodiment of the present invention provides a storage medium storing instructions, which, when read by a computer, cause the computer to execute any of the above-described methods for identifying concealed landslide targets and semantic segmentation landslides.

[0118] An electronic device according to an embodiment of the present invention includes a processor and the aforementioned storage medium. The processor executes instructions in the storage medium. The electronic device may be a computer, a mobile phone, or the like.

[0119] Those skilled in the art will know that this invention can be implemented as a system, method, or computer program product.

[0120] Therefore, this disclosure can be implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, the invention can also be implemented as a computer program product in one or more computer-readable media, the computer-readable medium containing computer-readable program code.

[0121] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0122] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for identifying and semantically segmenting concealed landslide targets, characterized in that, include: S1. Establish a hidden landslide intelligent identification model that includes a landslide target classification network with target-level contrastive learning and a landslide semantic segmentation network with quasi-target-level contrastive learning. The landslide target classification network with target-level contrastive learning includes a classifier, and a first encoder and a second encoder with two identical twin network structures that share network parameters. The landslide semantic segmentation network with quasi-target-level contrastive learning includes a third encoder and a decoder. The structure of the third encoder is the same as that of the first encoder. The target-level contrastive learning landslide target classification network is used to: extract semantic features of a first input image through the first encoder, extract semantic features of a second input image through the second encoder, perform channel concatenation on the semantic features of the first input image and the semantic features of the second input image to obtain a high-dimensional feature map representing the first input image and the second input image, and the classifier obtains the class confidence level that is commonly associated with the first input image and the second input image based on the high-dimensional feature map. The quasi-target-level contrastive learning landslide semantic segmentation network is used to: extract semantic features of the third input image through the third encoder, calculate the class confidence of each pixel in the third input image based on the semantic features of the third input image using the decoder, and obtain pixel-level landslide recognition results based on the class confidence of each pixel. S2. Based on the dataset, train the landslide target classification network of the target-level contrastive learning until it converges; S3. Extract the network parameters of the first encoder or the second encoder in the landslide target classification network after convergence of the target-level contrastive learning, load them into the third encoder and freeze them, and then use the quasi-target-level contrastive learning strategy to train the decoder until convergence. S4. Unlock the network parameters of the third encoder, and use the quasi-target-level contrastive learning strategy to train the quasi-target-level contrastive learning slippage semantic segmentation network until convergence. S5. Extract the network parameters of the third encoder in the converged quasi-target-level contrastive learning landslide semantic segmentation network, load them into the current first encoder and second encoder, return to execute S2, until a quasi-target-level contrastive learning landslide semantic segmentation network that meets the preset conditions is obtained, and a trained hidden landslide intelligent identification model is obtained. S6. Use the trained intelligent identification model for hidden landslides to identify landslides.

2. The method for identifying and semantically segmenting concealed landslide targets according to claim 1, characterized in that, The decoder is trained to convergence using the quasi-target-level contrastive learning strategy, including: A quasi-target-level positive and negative sample set is constructed, and the quasi-target-level positive and negative sample set corresponds to the semantic feature embedding of the semantic features currently extracted by the third encoder. The third encoder is updated using a supervised contrastive loss function assisted by a cross-entropy loss function, and the decoder is trained to convergence.

3. The method for identifying and semantically segmenting concealed landslide targets according to claim 1, characterized in that, The quasi-target-level contrastive learning strategy trains the quasi-target-level contrastive learning slippery semantic segmentation network to convergence, including: A quasi-target-level positive and negative sample set is constructed, and the quasi-target-level positive and negative sample set corresponds to the semantic feature embedding of the semantic features currently extracted by the third encoder. The third encoder is updated using a supervised contrastive loss function assisted by a cross-entropy loss function, and the quasi-target-level contrastive learning slippage semantic segmentation network is trained to convergence.

4. The method for identifying and semantically segmenting concealed landslide targets according to claim 1, characterized in that, If the first encoder employs a deep convolutional neural network, then the process of extracting semantic features from the first input image using the first encoder includes: The semantic features of the first input image are extracted through multi-layered stacked convolution operations in a deep convolutional neural network, combined with an attention mechanism.

5. The method for identifying and semantically segmenting concealed landslide targets according to claim 1, characterized in that, The decoder includes an upsampling block, which comprises a convolutional layer and a transposed convolutional layer arranged sequentially. The decoder calculates the class confidence of each pixel in the third input image based on the semantic features of the third input image, including: The semantic features of the third input image are restored to the size of the third input image through the upsampling block, and the class confidence of each pixel in the third input image is calculated.

6. A concealed landslide target identification and semantic segmentation system, characterized in that, It includes a model building module, a first training submodule, a second training submodule, a third training submodule, a fourth training submodule, and a recognition module; The model building module is used to: build a hidden landslide intelligent identification model including a landslide target classification network with target-level contrastive learning and a landslide semantic segmentation network with quasi-target-level contrastive learning. The landslide target classification network with target-level contrastive learning includes a classifier, and a first encoder and a second encoder with two identical twin network structures that share network parameters. The landslide semantic segmentation network with quasi-target-level contrastive learning includes a third encoder and a decoder. The structure of the third encoder is the same as that of the first encoder. The target-level contrastive learning landslide target classification network is used to: extract semantic features of a first input image through the first encoder, extract semantic features of a second input image through the second encoder, perform channel concatenation on the semantic features of the first input image and the semantic features of the second input image to obtain a high-dimensional feature map representing the first input image and the second input image, and the classifier obtains the class confidence level that is commonly associated with the first input image and the second input image based on the high-dimensional feature map. The quasi-target-level contrastive learning landslide semantic segmentation network is used to: extract semantic features of the third input image through the third encoder, calculate the class confidence of each pixel in the third input image based on the semantic features of the third input image using the decoder, and obtain pixel-level landslide recognition results based on the class confidence of each pixel. The first training submodule is used to: train the landslide target classification network of the target-level contrastive learning to converge based on the dataset; The second training submodule is used to: extract the network parameters of the first encoder or the second encoder in the converged target-level contrastive learning landslide target classification network, load them into the third encoder and freeze them, and then use the quasi-target-level contrastive learning strategy to train the decoder until convergence; The third training submodule is used to: unlock the network parameters of the third encoder, and train the quasi-target-level contrastive learning slippery semantic segmentation network until convergence using the quasi-target-level contrastive learning strategy. The fourth training submodule is used to: extract the network parameters of the third encoder in the converged quasi-target-level contrastive learning landslide semantic segmentation network, load them into the current first encoder and second encoder, and re-call the first training submodule, the second training submodule and the third training module until a quasi-target-level contrastive learning landslide semantic segmentation network that meets the preset conditions is obtained, and a trained hidden landslide intelligent identification model is obtained. The identification module is used to identify landslides using the trained intelligent identification model for concealed landslides.

7. The concealed landslide target identification and semantic segmentation system according to claim 6, characterized in that, The second training submodule is specifically used for: A first quasi-target level positive and negative sample set is constructed, and the quasi-target level positive and negative sample set corresponds to the semantic feature embedding of the semantic features currently extracted by the third encoder. The third encoder is updated using a supervised contrastive loss function assisted by a cross-entropy loss function, and the decoder is trained to convergence.

8. A concealed landslide target identification and semantic segmentation system according to claim 6, characterized in that, The third training submodule is used for: A second quasi-target level positive and negative sample set is constructed, and the quasi-target level positive and negative sample set corresponds to the semantic feature embedding of the semantic features currently extracted by the third encoder. The third encoder is updated using a supervised contrastive loss function assisted by a cross-entropy loss function, and the quasi-target level contrastive learning slippage semantic segmentation network is trained to convergence.

9. A storage medium, characterized in that, The storage medium stores instructions that, when read by a computer, cause the computer to execute a method for identifying and semantically segmenting concealed landslide targets as described in any one of claims 1 to 5.

10. An electronic device, characterized in that, It includes a processor and the storage medium of claim 9, wherein the processor executes instructions in the storage medium.

Images