Hidden landslide semantic segmentation method and system enhanced by quasi-pixel contrast learning

Abstract

The present invention relates to a method and system for latent landslide semantic segmentation enhanced by quasi-pixel contrast learning. In the method, a target feature is selected from the first dimensionality reduction feature output by a first projection head using an anchor selection strategy, and samples are selected from a category queue using a key value selection strategy. Positive and negative sample pairs are constructed based on the selected target feature and the selected samples, and a supervised contrast loss is calculated for the constructed positive and negative sample pairs. Backpropagation is used to update the query encoder and decoder, momentum is used to update the key encoder and second projection head, and the second dimensionality reduction feature output by the key encoder is added to the category queue until the mathematical model is trained to converge. The query encoder, key encoder, first projection head, and second projection head are removed from the converged mathematical model to obtain a landslide recognition model for landslide identification. This method can improve the ability to extract significant features for landslide identification and reduce the number of landslide samples required.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide identification, and in particular to a hidden landslide semantic segmentation method and system enhanced by quasi-pixel contrast learning. Background Art

[0002] Landslides are a frequent and devastating natural disaster, representing a serious form of natural disaster. Mountainous regions, particularly those with complex geological conditions, are prone to frequent geological disasters, often resulting in traffic disruptions, river and canal blockages, farmland damage, factory and mine destruction, village burials, and the crushing of people and livestock, resulting in significant human and economic losses. Therefore, studying the geological characteristics of landslides and leveraging information technology for wide-area landslide detection and early warning are crucial for disaster reduction and prevention.

[0003] The types of landslides that have occurred can be divided into two categories: ancient landslides and new landslides. Among them, new landslides refer to those that are in the process of repeated landslide activity or have just stopped landslide activity and are still at risk of activity; ancient landslides refer to landslides that occurred hundreds of years ago and remained inactive for a long time, but may revive again. New landslides are easily identified due to their significant differences from the environment due to their color and contour characteristics. They are called visually significant landslides. Ancient landslides occurred a long time ago, and the landslide surface has changed due to vegetation or natural erosion, human activities (agriculture, construction, etc.). It is difficult to distinguish them from the environment from the perspective of optical images. They are called hidden landslides. Reliable identification of ancient / old landslides with visual ambiguity is technically very challenging. This is specifically reflected in the small sample problem, specifically:

[0004] Hidden landslides are minimally differentiated from the background. Therefore, reliable landslide identification models must learn highly abstract semantic features, typically requiring a sufficient number of high-quality samples with a sufficient variety of types to adequately train the model. However, generating and labeling data samples for hidden landslides is particularly challenging. High-quality sample labeling relies on the time-consuming and labor-intensive interpretation of HRSI by experienced geological remote sensing experts, resulting in a limited number of high-quality hidden landslide samples. Therefore, it is necessary to carefully design a hidden landslide identification model to extract salient features for landslide identification and reduce the number of landslide samples required. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the deficiencies of the existing technology and provide a hidden landslide semantic segmentation method and system enhanced by quasi-pixel contrast learning.

[0006] The technical solution of the hidden landslide semantic segmentation method enhanced by quasi-pixel contrast learning of the present invention is as follows:

[0007] Establishing a mathematical model including a query encoder, a key encoder, a first projection head, a second projection head and a decoder;

[0008] The first projection head is used to: reduce the dimensions of features of different scales obtained by the query encoder based on the high-resolution remote sensing image data and the digital elevation model data to obtain a first reduced-dimensionality feature;

[0009] The second projection head is used to: perform dimensionality reduction on features of different scales obtained by the Key encoder based on high-resolution remote sensing image data and digital elevation model data to obtain second dimensionality reduction features;

[0010] The decoder is used to: splice the features of different scales obtained by the query encoder with the shallow features obtained by the query encoder based on high-resolution remote sensing image data and digital elevation model data to obtain a landslide identification result;

[0011] A target feature is selected from the first dimensionality reduction feature output by the first projection head using an anchor selection strategy, a sample is selected from the category queue using a key value selection strategy, a positive and negative sample pair is constructed based on the selected target feature and the selected sample, a supervised contrast loss of the constructed positive and negative sample pairs is calculated, the query encoder and the decoder are updated by backpropagation, the key encoder and the second projection head are updated by momentum, and the second dimensionality reduction feature output by the key encoder is added to the category queue until the mathematical model is trained to convergence;

[0012] The query encoder, the key encoder, the first projection head and the second projection head are removed from the converged mathematical model to obtain a landslide identification model for landslide identification.

[0013] The technical solution of the hidden landslide semantic segmentation system enhanced by quasi-pixel contrast learning of the present invention is as follows:

[0014] Including establishment module, training module and removal module;

[0015] The establishment module is used to establish a mathematical model including a Query encoder, a Key encoder, a first projection head, a second projection head and a decoder; the first projection head is used to reduce the dimensions of features of different scales obtained by the Query encoder based on high-resolution remote sensing image data and digital elevation model data to obtain first reduced-dimensionality features; the second projection head is used to reduce the dimensions of features of different scales obtained by the Key encoder based on high-resolution remote sensing image data and digital elevation model data to obtain second reduced-dimensionality features; the decoder is used to splice the features of different scales obtained by the Query encoder with the shallow features obtained by the Query encoder based on high-resolution remote sensing image data and digital elevation model data to obtain a landslide identification result;

[0016] The training module is used to: select a target feature from the first dimensionality reduction feature output by the first projection head using an anchor selection strategy, select samples from the category queue using a key value selection strategy, construct positive and negative sample pairs based on the selected target feature and the selected samples, calculate the supervised contrast loss of the constructed positive and negative sample pairs, update the query encoder and the decoder by backpropagation, update the key encoder and the second projection head by momentum, and add the second dimensionality reduction feature output by the key encoder to the category queue until the mathematical model is trained to convergence;

[0017] The removal module is used to remove the query encoder, the key encoder, the first projection head and the second projection head from the converged mathematical model to obtain a landslide identification model for landslide identification.

[0018] A storage medium of the present invention stores instructions, and when a computer reads the instructions, the computer is caused to execute any one of the above-mentioned methods for latent landslide semantic segmentation enhanced by quasi-pixel contrast learning.

[0019] An electronic device of the present invention includes a processor and the above-mentioned storage medium, wherein the processor executes instructions in the storage medium.

[0020] The beneficial effects of the present invention are as follows:

[0021] The query encoder, key encoder, first projection head, and second projection head improve the ability to extract salient features for landslide identification, reducing the number of landslide samples required. To alleviate the small sample size issue, a category queue is used to construct positive and negative sample pairs across the entire dataset, enabling cross-image comparison. Furthermore, anchor and key selection strategies are used to further focus the trained landslide recognition model on salient features. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0023] Figure 1 Schematic diagram of a process for latent landslide semantic segmentation enhanced by quasi-pixel contrast learning according to an embodiment of the present invention;

[0024] Figure 2 It is the structural diagram of the mathematical model;

[0025] Figure 3 This is the structural diagram of the lightweight feature extraction and fusion unit;

[0026] Figure 4This is a structural diagram of the combination of an abstract feature extraction unit and a multi-scale channel attention unit using SENet;

[0027] Figure 5 A schematic diagram of maintaining queue length by dequeuing and enqueuing;

[0028] Figure 6 This is the structural diagram of the Query encoder and Key encoder;

[0029] Figure 7 The architecture of the learnable decoder;

[0030] Figure 8 Schematic diagram of the structure of a hidden landslide semantic segmentation system enhanced by quasi-pixel contrast learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] like Figures 1 and 2 As shown, a latent landslide semantic segmentation method enhanced by quasi-pixel contrast learning according to an embodiment of the present invention includes the following steps:

[0032] S1. Establish a mathematical model including a query encoder, a key encoder, a first projection head, a second projection head, and a decoder;

[0033] The first projection head is used to reduce the dimensions of features of different scales obtained by the query encoder based on the high-resolution remote sensing image data and the digital elevation model data to obtain a first reduced-dimensionality feature;

[0034] The second projection head is used to reduce the dimensions of features of different scales obtained by the Key encoder based on high-resolution remote sensing image data and digital elevation model data to obtain second reduced-dimensional features;

[0035] The decoder is used to combine the features of different scales obtained by the query encoder with the shallow features obtained by the query encoder based on high-resolution remote sensing image data and digital elevation model data to obtain landslide identification results.

[0036] Among them, both the Query encoder and the Key encoder include a lightweight feature extraction and fusion unit, an abstract feature extraction unit, and a multi-scale channel attention unit;

[0037] At present, the automatic identification technology of hidden landslides of ancient and old landslide types based on high-resolution remote sensing image data mainly faces the following problems in data and methods:

[0038] Low-level visual features and high-level semantic features of landslides can be extracted from high-resolution remote sensing image data, namely HRSI. The low-level features are mainly the texture features formed by the back wall and side wall of the landslide on the two-dimensional optical image, as well as the color differences of the landslide boundary. High-level semantic features refer to abstract features such as the morphology of the landslide, including the tongue-shaped or horseshoe-shaped morphology of the landslide deposit. Digital elevation model data, namely DEM, provides terrain information such as slope, aspect and height changes. High-resolution remote sensing image data and digital elevation model data contain completely different semantic features. In the landslide identification task, it is necessary to design a suitable feature fusion method to achieve the organic fusion of heterogeneous information and provide more comprehensive and reliable semantic features for landslide identification. Therefore, the lightweight feature extraction and fusion unit proposed in the present invention is used to extract and fuse features from high-resolution remote sensing image data and digital elevation model data.

[0039] The lightweight feature extraction and fusion unit is used to obtain a feature map including low-level spatial position features and high-level abstract semantic features based on high-resolution remote sensing image data and digital elevation model data. Specifically:

[0040] A deep convolutional network is used to obtain high-level feature maps of abstract semantic features of high-resolution remote sensing image data and digital elevation model data, and channel stitching is performed to obtain a stitched feature map.

[0041] The coordinate attention mechanism is used to learn the global channel and spatial information of the spliced ​​feature map, and the splicing operation and convolution operation are performed with the low-level feature map of the deep convolutional network to obtain a feature map including low-level spatial position features and high-level abstract semantic features. Figure 3 As shown, the lightweight feature extraction and fusion unit includes: an optical feature extractor, a terrain feature extractor, a coordinate attention module, and a fusion feature extractor, wherein the optical feature extractor and the terrain feature extractor have the same deep convolutional neural network architecture.

[0042] The high-resolution remote sensing image data is input into the optical feature extractor, and a high-level feature map with abstract semantic features such as landslide texture, morphology, and color is output, and an intermediate-layer feature map is output. From the intermediate-layer feature map, low-level spatial position features including spatial position are obtained. The digital elevation model data is input into the terrain feature extractor, and a high-level feature map with abstract semantic features such as landslide height, slope, and slope direction is output, and an intermediate-layer feature map is output. From the intermediate-layer feature map, low-level spatial position features including spatial position are obtained.

[0043] The high-level feature map of the abstract semantic features of the high-resolution remote sensing image data and the high-level feature map of the abstract semantic features of the digital elevation model data are channel-spliced ​​to obtain a spliced ​​feature map.

[0044] The abstract feature extraction unit is used to: extract semantic features from the feature map;

[0045] The abstract feature extraction unit includes: Residual Network ResNet101. ResNet101 has excellent semantic feature extraction capabilities, and its structure is as follows Figure 4 As shown on the left, there are scale 1 hole residual blocks, scale 2 hole residual blocks, scale 3 hole residual blocks, and scale 4 hole residual blocks. Each hole residual block has a different feature map resolution.

[0046] The feature map output by the lightweight feature extraction and fusion unit is input into the abstract feature extraction unit. In order to retain more detailed features, the output feature map of the scale-2 dilated residual block is concatenated with the output feature map of the upsampled scale-4 dilated residual block in the channel dimension and then input into the multi-scale attention unit to obtain adaptively weighted multi-scale semantic features.

[0047] The multi-scale channel attention unit is used to: utilize the channel attention mechanism to extract semantic features from the feature map according to the abstract feature extraction unit to obtain adaptive weighted multi-scale features, and extract features of different scales in the adaptive weighted multi-scale features through the atrous spatial pooling pyramid (ASPP).

[0048] The structure of the multi-scale channel attention unit is as follows Figure 4 As shown on the right, it includes the channel attention mechanism (SENet) and the dilated spatial pooling pyramid ASPP. The multi-scale channel attention unit uses the dilated spatial pooling pyramid ASPP to extract features of different scales from the adaptively weighted multi-scale features, including:

[0049] The global context information in the adaptive weighted multi-scale features is integrated by average pooling of image-level features to obtain features of different scales in the adaptive weighted multi-scale features.

[0050] The feature map output by the abstract feature extraction unit is input into the multi-scale channel attention unit. The channel features of the feature map are first weighted by SENet to obtain adaptive weighted multi-scale features. Then, the hollow spatial pooling pyramid ASPP is used to extract features of different scales from the adaptive weighted multi-scale features by using hollow convolution kernels with different expansion rates and global average pooling with image-level features. The global context information is integrated to obtain features of different scales in the adaptive weighted multi-scale features. Among them, the first projection head, the second projection head and the decoder are explained as follows:

[0051] 1) The first projection head is a single-layer perceptron that reduces the dimensions of the features of different scales obtained by the query encoder based on high-resolution remote sensing image data and digital elevation model data in the channel dimension to obtain the first reduced-dimensional features.

[0052] 2) The second projection head is also a single-layer perceptron that reduces the dimensions of the features of different scales obtained by the Key encoder based on the high-resolution remote sensing image data and the digital elevation model data in the channel dimension to obtain a second reduced-dimensionality feature;

[0053] 3) If Figure 7 As shown, the decoder includes two identically structured learnable upsampling modules, each comprising a convolutional layer, a rectified linear unit (ReLU) activation function, dropout, a learnable transposed convolutional layer, and batch normalization. The decoder is a learnable decoder that concatenates the low-level spatial features extracted by the scale-1 atrous residual block of ResNet101 in the query encoder's abstract feature extraction unit with the features of different scales in the adaptively weighted multi-scale features output by the query encoder's multi-scale channel attention unit. This combines low-level spatial details with high-level semantic abstract features, and uses transposed convolutional layers for upsampling to reconstruct the resolution of high-resolution remote sensing image data.

[0054] Moreover, the decoder uses the ReLU activation function after the convolutional layer to improve the nonlinear fitting ability of the network, introduces Dropout to avoid information redundancy and prevent overfitting, and introduces Batch Norm to constrain data distribution to avoid gradient explosion and gradient disappearance and help accelerate network convergence.

[0055] S2. Use the anchor selection strategy to select the target feature from the first dimensionality reduction feature output by the first projection head, use the key selection strategy to select samples from the category queue, construct positive and negative sample pairs based on the selected target feature and the selected samples, calculate the supervised contrast loss of the constructed positive and negative sample pairs, backpropagate to update the query encoder and decoder, use momentum to update the key encoder and the second projection head, and add the second dimensionality reduction feature output by the key encoder to the category queue until the mathematical model is trained to convergence;

[0056] The explanations of the anchor selection strategy and key value selection strategy are as follows:

[0057] 1) Anchor selection strategy is the anchor selection strategy:

[0058] Each batch selects target features, or anchors, from the first dimensionality reduction feature map. The DEM data and labels are downsampled, and points within the landslide whose height is greater than the median height and on the landslide edge are selected as anchors for the landslide category. This comparison optimizes the query encoder's feature space, ensuring that similar features are close together and heterogeneous features are separated. Homogeneous features are landslides and slopes, while heterogeneous features are landslides and slopes. During training, all filtered landslide anchors are selected for each batch, and an equal number of slope anchors are randomly selected.

[0059] In order to reduce computational complexity and improve learning efficiency, the present invention innovatively designs an external storage data structure of a supervised category queue to store the keys of landslides and slopes, i.e., the second dimension reduction features in each batch of queues. The structure is as follows: Figure 5 The advantage of using an external queue data structure is that it decouples the queue length from the batch size. Keys of features in the second dimensionality reduction feature map that share the same position as anchors are selected and queued. This process updates the category queue through enqueueing and dequeuing, saving significant computational resources. The category queue contains enqueued keys selected from the entire dataset, allowing for a global comparison of the landslide and slope anchors in the current batch of images with the keys of the entire dataset.

[0060] 2) Key value selection strategy, that is, Key selection strategy:

[0061] The length of the landslide queue and the slope queue in the category queue are both L. For each selected anchor, for any anchor, K keys of the same type and K keys of different types are selected from the category queue. Similar types refer to landslides and slopes, and different types refer to landslides and slopes.

[0062] Among them, positive and negative sample pairs are constructed based on the selected target features, namely Anchors, and the selected samples, namely keys. Specifically:

[0063] First, the similarity between the anchor and all keys in the two queues, namely the landslide queue and the slope queue, is calculated. The 10%K similar keys with the lowest similarity are taken as hard positive samples and the 10% different keys with the highest similarity are taken as hard negative samples. The remaining 90%K positive and negative samples are randomly selected from the landslide queue and the slope queue to construct positive and negative sample pairs. The positive and negative sample pairs include: one landslide anchor corresponds to K landslide keys, forming K positive sample pairs, and one landslide anchor corresponds to K slope keys, forming K negative sample pairs.

[0064] Calculate the supervised contrast loss of the constructed positive and negative sample pairs. The specific implementation process is as follows: Use the formula: Calculate the supervised contrast loss of each anchor of the positive and negative sample pairs, where i represents: the i-th anchor selected from the current batch, P i Indicates: the same queue as the i-th anchor selected from the current batch, i + Indicates: From P i The positive samples selected from i Represents: a queue that is heterogeneous with the i-th anchor selected from the current batch, i - Represents from N i Negative samples selected from .

[0065] The query encoder and decoder are updated by back propagation, specifically: the query encoder, the first projection head and the decoder parameters are updated by back propagation.

[0066] The process of momentum updating the Key encoder and the second projection head is as follows:

[0067] The overall structure of the Key encoder and the second projection head is defined as η. In order to avoid the asynchronous update problem caused by the keys entering the category queue at different times being encoded by η with large parameter differences at different times, a method for momentum updating the parameters of the Key encoder and the second projection head is designed, so that all keys in the category queue are encoded by similar η, avoiding the encoding consistency problem caused by asynchronous update.

[0068] The query encoder and key encoder use the same structure, and the first and second projection heads use the same structure and are initialized with the same parameters. The key encoder and the first and second projection heads only participate in training and are discarded when making predictions.

[0069] like Figure 6 As shown, the parameters of the Key encoder and the second projection head are updated using momentum update. The momentum update formula is as follows:

[0070] θ k ←mθ k +(1-m)θ q

[0071] P k ←mP k +(1-m)P q

[0072] Among them, θ k is the parameter of the Key encoder, θ qis the parameter of Query encoder, P k is the parameter of the first projection head, P q is the parameter of the second projection head. m is close to 1, for example, m = 0.999, and the updated Key encoder parameter becomes: 0.999 times the current batch Key encoder parameter + 0.001 times the updated Query encoder parameter, so that the Key encoder parameter is updated slowly.

[0073] S3. Remove the query encoder, key encoder, first projection head, and second projection head from the converged mathematical model to obtain a landslide recognition model for landslide recognition, and use the landslide recognition model to perform landslide recognition on the image to be recognized.

[0074] This method was validated on a dataset of ancient landslides from the Loess Plateau, with a HRSI resolution of 2 meters per pixel and a DEM resolution of 30 meters per pixel. The dataset contains 141 samples, divided into training, validation, and test sets at a ratio of 6:2:2. Data augmentation was performed on both the training and test sets, using a 5-fold cross-validation approach.

[0075] The trained semantic segmentation model is evaluated using five indicators: mean intersection over union (mIoU), slide-IoU, recall, precision, and F1 score. The evaluation indicator formula is as follows:

[0076]

[0077]

[0078]

[0079]

[0080]

[0081] Among them, TP, TN, FP and FN are the proportion of accurately predicted landslide pixels, the proportion of accurately predicted non-landslide pixels, the proportion of landslide pixels mistakenly identified as non-landslide pixels and the proportion of non-landslide pixels mistakenly identified as landslide pixels, respectively. See Table 1 for details.

[0082] Table 1:

[0083]

[0084] In the above embodiments, although the steps are numbered S1, S2, etc., these are only specific embodiments given in this application. Those skilled in the art can adjust the execution order of S1, S2, etc. according to actual conditions, which is also within the scope of protection of the present invention. It can be understood that in some embodiments, some or all of the above embodiments may be included.

[0085] like Figure 8 As shown, a hidden landslide semantic segmentation system 200 enhanced by quasi-pixel contrast learning according to an embodiment of the present invention includes a building module 210, a training module 220 and a removal module 230;

[0086] The establishment module 210 is used to establish a mathematical model including a query encoder, a key encoder, a first projection head, a second projection head, and a decoder; the first projection head is used to reduce the dimensions of the features of different scales obtained by the query encoder based on the high-resolution remote sensing image data and the digital elevation model data to obtain a first reduced-dimensionality feature; the second projection head is used to reduce the dimensions of the features of different scales obtained by the key encoder based on the high-resolution remote sensing image data and the digital elevation model data to obtain a second reduced-dimensionality feature; the decoder is used to combine the features of different scales obtained by the query encoder with the shallow features obtained by the query encoder based on the high-resolution remote sensing image data and the digital elevation model data to obtain a landslide identification result;

[0087] The training module 220 is used to: select a target feature from the first dimensionality reduction feature output by the first projection head using the anchor selection strategy, select samples from the preset category queue using the key value selection strategy, construct positive and negative sample pairs based on the selected target feature and the selected samples, calculate the supervised contrast loss of the constructed positive and negative sample pairs, update the query encoder and decoder by backpropagation, update the key encoder and the second projection head by momentum, and add the second dimensionality reduction feature output by the key encoder to the preset category queue until the mathematical model is trained to convergence;

[0088] The removal module 230 is used to remove the query encoder, the key encoder, the first projection head, and the second projection head from the converged mathematical model to obtain a landslide identification model for landslide identification.

[0089] Optionally, in the above technical solution, the query encoder and the key encoder both include a lightweight feature extraction and fusion unit, an abstract feature extraction unit, and a multi-scale channel attention unit;

[0090] The lightweight feature extraction and fusion unit is used to obtain a feature map including low-level spatial position features and high-level abstract semantic features based on high-resolution remote sensing image data and digital elevation model data;

[0091] The abstract feature extraction unit is used to: extract semantic features from the feature map;

[0092] The multi-scale channel attention unit is used to: utilize the self-attention mechanism to extract semantic features from the feature map according to the abstract feature extraction unit to obtain adaptive weighted multi-scale features, and extract features of different scales in the adaptive weighted multi-scale features through the void space pooling pyramid.

[0093] Optionally, in the above technical solution, the lightweight feature extraction and fusion unit is specifically used to:

[0094] A deep convolutional network is used to obtain high-level feature maps of abstract semantic features of high-resolution remote sensing image data and digital elevation model data, and channel stitching is performed to obtain a stitched feature map.

[0095] The coordinate attention mechanism is used to learn the global channel and spatial information of the spliced ​​feature map, and splicing and convolution operations are performed with the low-level feature map of the deep convolutional network to obtain a feature map that includes low-level spatial position features and high-level abstract semantic features.

[0096] Optionally, in the above technical solution, the process of extracting features of different scales from the adaptively weighted multi-scale features by the multi-scale channel attention unit using the dilated spatial pooling pyramid includes:

[0097] The global context information in the adaptive weighted multi-scale features is integrated by average pooling of image-level features to obtain features of different scales in the adaptive weighted multi-scale features.

[0098] The above-mentioned parameters and steps for each unit module to implement corresponding functions in the latent landslide semantic segmentation system enhanced by quasi-pixel contrast learning of the present invention can be referred to the parameters and steps in the embodiment of the latent landslide semantic segmentation method enhanced by quasi-pixel contrast learning above, and will not be repeated here.

[0099] A storage medium according to an embodiment of the present invention stores instructions, and when a computer reads the instructions, the computer is caused to execute any one of the aforementioned methods for latent landslide semantic segmentation enhanced by quasi-pixel contrast learning.

[0100] An electronic device according to an embodiment of the present invention includes a processor and the aforementioned storage medium, wherein the processor executes instructions in the storage medium.

[0101] Those skilled in the art will appreciate that the present invention may be implemented as a system, method or computer program product.

[0102] Therefore, the present disclosure may be embodied in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, the present disclosure may be embodied in the form of a computer program product embodied in one or more computer-readable media, wherein the computer-readable media contains computer-readable program code.

[0103] Any combination of one or more computer-readable media can 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, device, or component, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device, or device.

[0104] 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 latent landslide semantic segmentation method enhanced by quasi-pixel contrast learning, characterized by: include: Establishing a mathematical model including a query encoder, a key encoder, a first projection head, a second projection head and a decoder; The first projection head is used to: perform dimensionality reduction on features of different scales obtained by the query encoder based on the high-resolution remote sensing image data and the digital elevation model data to obtain first dimensionality reduction features; The second projection head is used to: perform dimensionality reduction on features of different scales obtained by the Key encoder based on high-resolution remote sensing image data and digital elevation model data to obtain second dimensionality reduction features; The decoder is used to: splice the features of different scales obtained by the query encoder with the shallow features obtained by the query encoder based on high-resolution remote sensing image data and digital elevation model data to obtain a landslide identification result; A target feature is selected from the first dimensionality reduction feature output by the first projection head using an anchor selection strategy, a sample is selected from the category queue using a key value selection strategy, a positive and negative sample pair is constructed based on the selected target feature and the selected sample, a supervised contrast loss of the constructed positive and negative sample pairs is calculated, the query encoder and the decoder are updated by backpropagation, the key encoder and the second projection head are updated by momentum, and the second dimensionality reduction feature output by the key encoder is added to the category queue until the mathematical model is trained to convergence; Removing the query encoder, the key encoder, the first projection head, and the second projection head from the converged mathematical model to obtain a landslide identification model for landslide identification; The anchor selection strategy is as follows: each batch selects the target feature from the first dimensionality reduction feature map as the anchor point; the digital elevation model data and labels are downsampled, and the points with a height value greater than the median height of the landslide and on the edge of the landslide are selected as the anchor points of the landslide category; Among them, the key value selection strategy is: the length of the landslide queue and the slope queue in the category queue are both L. For each selected anchor, for any anchor, K keys of the same type and K keys of different types will be selected from the category queue; the same type refers to: landslide and landslide, slope and slope, and different types refers to: landslide and slope.

2. The hidden landslide semantic segmentation method enhanced by quasi-pixel contrast learning according to claim 1 is characterized in that: The query encoder and the key encoder both include a lightweight feature extraction and fusion unit, an abstract feature extraction unit, and a multi-scale channel attention unit; The lightweight feature extraction and fusion unit is used to obtain a feature map including low-level spatial position features and high-level abstract semantic features based on high-resolution remote sensing image data and digital elevation model data; The abstract feature extraction unit is used to: extract semantic features from the feature map; The multi-scale channel attention unit is used to: utilize the self-attention mechanism to extract semantic features from the feature map according to the abstract feature extraction unit to obtain adaptive weighted multi-scale features, and extract features of different scales in the adaptive weighted multi-scale features through a void space pooling pyramid.

3. The hidden landslide semantic segmentation method enhanced by quasi-pixel contrast learning according to claim 2 is characterized in that: The lightweight feature extraction and fusion unit is specifically used for: A deep convolutional network is used to obtain high-level feature maps of abstract semantic features of high-resolution remote sensing image data and digital elevation model data, and channel stitching is performed to obtain a stitched feature map. The coordinate attention mechanism is used to learn the global channel and spatial information of the spliced ​​feature map, and splicing and convolution operations are performed with the low-level feature map of the deep convolutional network to obtain a feature map that includes low-level spatial position features and high-level abstract semantic features.

4. The hidden landslide semantic segmentation method enhanced by quasi-pixel contrast learning according to claim 3 is characterized in that: The process of extracting features of different scales in the adaptive weighted multi-scale features by the multi-scale channel attention unit using the dilated spatial pooling pyramid includes: Global context information in the adaptive weighted multi-scale features is integrated by average pooling of image-level features to obtain features of different scales in the adaptive weighted multi-scale features.

5. A latent landslide semantic segmentation system enhanced by quasi-pixel contrast learning, characterized by: Including establishment module, training module and removal module; The establishment module is used to establish a mathematical model including a Query encoder, a Key encoder, a first projection head, a second projection head and a decoder; the first projection head is used to reduce the dimensions of features of different scales obtained by the Query encoder based on high-resolution remote sensing image data and digital elevation model data to obtain first reduced-dimensionality features; the second projection head is used to reduce the dimensions of features of different scales obtained by the Key encoder based on high-resolution remote sensing image data and digital elevation model data to obtain second reduced-dimensionality features; the decoder is used to splice the features of different scales obtained by the Query encoder with the shallow features obtained by the Query encoder based on high-resolution remote sensing image data and digital elevation model data to obtain a landslide identification result; The training module is used to: select a target feature from the first dimensionality reduction feature output by the first projection head using an anchor selection strategy, select samples from the category queue using a key value selection strategy, construct positive and negative sample pairs based on the selected target feature and the selected samples, calculate the supervised contrast loss of the constructed positive and negative sample pairs, update the query encoder and the decoder by backpropagation, update the key encoder and the second projection head by momentum, and add the second dimensionality reduction feature output by the key encoder to the category queue until the mathematical model is trained to convergence; The removal module is used to remove the query encoder, the key encoder, the first projection head, and the second projection head from the converged mathematical model to obtain a landslide identification model for landslide identification; The anchor selection strategy is as follows: each batch selects the target feature from the first dimensionality reduction feature map as the anchor point; downsamples the digital elevation model data and labels, and selects the point whose height value inside the landslide is greater than the median height of the landslide and is on the edge of the landslide as the anchor point of the landslide category. Among them, the key value selection strategy is: the length of the landslide queue and the slope queue in the category queue are both L. For each selected anchor, for any anchor, K keys of the same type and K keys of different types will be selected from the category queue; the same type refers to: landslide and landslide, slope and slope, and different types refers to: landslide and slope.

6. The latent landslide semantic segmentation system enhanced by quasi-pixel contrast learning according to claim 5 is characterized in that: The query encoder and the key encoder both include a lightweight feature extraction and fusion unit, an abstract feature extraction unit, and a multi-scale channel attention unit; The lightweight feature extraction and fusion unit is used to obtain a feature map including low-level spatial position features and high-level abstract semantic features based on high-resolution remote sensing image data and digital elevation model data; The abstract feature extraction unit is used to: extract semantic features from the feature map; The multi-scale channel attention unit is used to: utilize the self-attention mechanism to extract semantic features from the feature map according to the abstract feature extraction unit to obtain adaptive weighted multi-scale features, and extract features of different scales in the adaptive weighted multi-scale features through a void space pooling pyramid.

7. The latent landslide semantic segmentation system enhanced by quasi-pixel contrast learning according to claim 6 is characterized in that: The lightweight feature extraction and fusion unit is specifically used for: A deep convolutional network is used to obtain high-level feature maps of abstract semantic features of high-resolution remote sensing image data and digital elevation model data, and channel stitching is performed to obtain a stitched feature map. The coordinate attention mechanism is used to learn the global channel and spatial information of the spliced ​​feature map, and splicing and convolution operations are performed with the low-level feature map of the deep convolutional network to obtain a feature map that includes low-level spatial position features and high-level abstract semantic features.

8. The latent landslide semantic segmentation system enhanced by quasi-pixel contrast learning according to claim 7 is characterized in that: The process of extracting features of different scales in the adaptive weighted multi-scale features by the multi-scale channel attention unit using the dilated spatial pooling pyramid includes: Global context information in the adaptive weighted multi-scale features is integrated by average pooling of image-level features to obtain features of different scales in the adaptive weighted multi-scale features.

9. A storage medium, characterized in that: The storage medium stores instructions, and when a computer reads the instructions, the computer is caused to execute the latent landslide semantic segmentation method enhanced by quasi-pixel contrast learning according to any one of claims 1 to 4.

10. An electronic device, characterized in that: The device comprises a processor and the storage medium according to claim 9, wherein the processor executes instructions in the storage medium.

Images