Landslide region detection method and system based on multi-level multi-scale feature enhancement

Abstract

The present application relates to the field of landslide detection, and more particularly to a landslide area detection method and system based on multi-level multi-scale feature enhancement, the method comprising: using a multi-level feature extraction model to extract features of a landslide area in a remote sensing image to obtain multiple levels of feature maps; unifying the channel numbers of the multiple levels of feature maps; inputting the feature maps with unified channel numbers into pyramid efficient multi-scale attention modules of different scales respectively to obtain multiple scales of attention feature maps with feature enhancement; and performing progressive upsampling processing on the multiple scales of attention feature maps to obtain a final remote sensing image landslide area detection result; the present application uses images containing spatial structures, texture features and abstract features and enhances them, which can better identify and distinguish landslide areas and avoid missing small-scale landslide areas; the progressive upsampling fuses high-level features with low-level features, achieves the effect of sufficient fusion of feature maps, and reduces the false alarm rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of landslide detection, in particular to a landslide area detection method and system based on multi-level multi-scale feature enhancement. BACKGROUND

[0002] Landslides are major natural disasters that pose a serious challenge to human survival in various regions around the world. Global climate change, population growth, and intensification of urbanization have further increased the frequency and severity of landslides. In the field of emergency management, the importance of landslide disaster site conditions is self-evident, as they can help emergency rescue personnel accurately locate areas with severe disasters and promptly carry out rescue missions. In addition, landslide area detection can also help experts assess disaster conditions and develop subsequent reconstruction plans. Therefore, extracting landslide disaster information from remote sensing images is an important direction in the field of remote sensing applications.

[0003] A landslide refers to an area that has shown signs of deformation and has a probability of disaster in the near future. There are two common methods for landslide identification. The first method is field investigation, which has high accuracy and strong controllability, but is low in efficiency and high in cost. The second method is remote sensing image processing and identification, including statistical methods, shallow machine learning methods, and deep learning methods. Commonly used data include interferometric synthetic aperture radar (InSAR), laser radar (LiDAR), and optical remote sensing images. InSAR has the advantages of high resolution and all-weather data acquisition, but is easily affected by factors such as observation angle, vegetation coverage, water vapor, and data processing technology. Existing technologies can successfully extract 17 unstable slopes within a watershed using the coherent scatterer InSAR (CSI) method. Stacking InSAR and deep learning technology have also been used to achieve intelligent identification of geological disaster hazards at a wide area scale. LiDAR can provide high-precision three-dimensional information, high-density high-precision sampling, and classification, but is relatively expensive and difficult to implement on a large scale. For example, using an object-oriented method to identify landslides covered by vegetation from LiDAR data successfully identified more than 70% of landslide bodies. Optical remote sensing images are also a common method for landslide identification. They have wide coverage and high resolution, and can effectively identify areas with obvious deformation signs over a large area. However, they are easily affected by cloud cover, vegetation coverage, and time sequence. Shallow machine learning methods for optical remote sensing imaging landslide identification require pre-extraction of features and interpretable factors present in the image, followed by the use of a classifier to perform classification calculations. This increases the complexity of algorithm design, limits the scalability of the algorithm, and the algorithm does not have the performance required for real-time applications.

[0004] Deep learning methods are also widely used in optical remote sensing target detection, semantic segmentation and image classification, and the convolutional neural network (CNN) is the main method of deep learning. This method has the ability of multi-layer nonlinear mapping, and can effectively extract the texture, pop, edge, color and other information of remote sensing images, and can better fit or learn the landslide features. In recent years, fully convolutional network (FCN) has been used in many remote sensing image recognition tasks, which can input images of any size and produce corresponding size output through effective inference and learning. In the architecture of FCN, the fully connected layer is replaced by a 1x1 convolutional layer, and the deconvolutional layer is used to restore the size of the feature map, so FCN becomes the standard for high-precision pixel-level labeling. This method provides a more efficient and low-cost landslide extraction strategy in the field of remote sensing, but its noise is large, so the details of the network output still need to be improved. In further development, U-net is proposed and becomes one of the most popular semantic segmentation methods. U-net is an improved FCN model that accepts the entire image as input and outputs pixel-level positioning classes. The model consists of two paths: the contraction path is used to capture contextual information, and the symmetric expansion path is used to obtain more accurate pixel-level positioning. In the prior art, DeepLabv3+ changes the resolution of the feature map output by the encoder by controlling the rate of the dilated convolution, increases the receptive field of the convolutional layer, and expands DeepLabv3 by adding a simple and effective decoder module to restore object boundaries and enhance the semantic segmentation capability of the model. Attention mechanism, dilated convolution and other methods help the semantic segmentation network better capture contextual details. How to effectively extract and fuse multi-scale features is the key to improving the detection capability of the model. A multi-scale feature fusion mechanism based on an encoding-decoding network (MFFENet) is used to extract and fuse multi-scale features of objects in remote sensing images. A novel and practical adaptive triangular fork (ATF) module is designed to effectively integrate useful features of different scales. The fusion of multi-scale features effectively improves the recognition ability of landslide areas. However, landslides in remote sensing images often have no fixed shape, and the size of the disaster is also different. The area of the landslide region in the remote sensing image can be large or small, and the general change detection method cannot well detect landslides of various sizes, which has caused the problem of missing small-scale landslides and incomplete detection of large-scale regions. SUMMARY

[0005] In order to solve the problem of missing detection caused by the large size difference and the indefinite shape of the landslide in the prior art, the first aspect of the present application proposes a landslide area detection method based on multi-level multi-scale feature enhancement, comprising:

[0006] A multi-level feature extraction model was used to extract features from landslide areas in remote sensing images to obtain feature maps at multiple levels.

[0007] Unify the number of channels in feature maps across multiple levels;

[0008] The feature maps with unified channel counts are input into pyramid-based efficient multi-scale attention modules of different scales to obtain multi-scale attention feature maps with enhanced features.

[0009] The final remote sensing image landslide area detection results are obtained by progressively upsampling attention feature maps at multiple scales.

[0010] Preferably, the step of using a multi-level feature extraction model to extract features from the landslide area in the remote sensing image to obtain feature maps at multiple levels specifically involves:

[0011] The number of output layers and the number of output layers of the multi-level feature extraction model are determined based on the detection accuracy.

[0012] The landslide area in the remote sensing image is input into the multi-level feature extraction model, and feature maps of the corresponding levels are obtained in each output layer.

[0013] In this case, the size of the image of each output layer is one-quarter of that of the previous output layer.

[0014] Preferably, the multi-level feature extraction model is the MobileNetV2 model, which includes multiple bottleneck layers connected in sequence.

[0015] Preferably, the number of output layers is 4, and the output layers are the third, sixth, thirteenth and seventeenth layers of the bottleneck layer.

[0016] Preferably, the step of unifying the number of channels in the feature maps across multiple levels specifically involves:

[0017] The same convolutional layer is used to unify the number of channels in feature maps across multiple layers.

[0018] Preferably, the step of inputting the feature map with a unified number of channels into a pyramid-based efficient multi-scale attention module of different scales to obtain multi-scale attention feature maps with enhanced features is specifically as follows:

[0019] The feature maps of multiple levels after channel unification are input into the pyramid-efficiency multi-scale attention module of different scales;

[0020] Each pyramid's efficient multi-scale attention module sequentially performs multi-scale segmentation, efficient multi-scale attention extraction, concatenation, and convolution on the input feature map to obtain the output feature map. The input feature map and the output feature map are then superimposed to obtain the corresponding attention feature map.

[0021] Preferably, the step of performing progressive upsampling processing on attention feature maps at multiple scales to obtain the final remote sensing image landslide area detection result includes:

[0022] The upsampling module is used to perform stepwise upsampling of attention feature maps at multiple scales from high scale to obtain the final remote sensing image landslide area detection results;

[0023] The upsampling module includes multiple feature fusion units and multiple convolutional layers.

[0024] Preferably, the feature fusion unit includes: an upsampling layer, Convolutional layers, batch normalized layers, and the ReLU activation function.

[0025] A second aspect of the present invention provides a landslide area detection system based on multi-level, multi-scale feature enhancement, comprising:

[0026] A multi-level feature extraction model is used to extract features from landslide areas in remote sensing images to obtain feature maps at multiple levels.

[0027] The convolution module is used to unify the number of channels in feature maps across multiple levels.

[0028] The pyramid-shaped high-efficiency multi-scale attention module is used to obtain attention feature maps at multiple scales that capture changes in feature map information.

[0029] The upsampling module is used to perform step-by-step upsampling processing on attention feature maps at multiple scales to obtain the final landslide area detection results from remote sensing images.

[0030] Optionally, the loss function of the landslide area detection system based on multi-level, multi-scale feature enhancement is:

[0031]

[0032] in, The loss function is for a landslide area detection system based on multi-level, multi-scale feature enhancement. The binary cross-entropy loss function is... · represents the dot product operation. c The final results of landslide area detection from remote sensing imagery. g For real labels, The Dice loss function, , This is the first normal form.

[0033] Optionally, the multi-level feature extraction model extracts features from the landslide area in the remote sensing image to obtain feature maps at multiple levels, specifically:

[0034] The number of output layers and the number of output layers of the multi-level feature extraction model are determined based on the detection accuracy.

[0035] The landslide area in the remote sensing image is input into the multi-level feature extraction model, and feature maps of the corresponding levels are obtained in each output layer.

[0036] In this case, the size of the image of each output layer is one-quarter of that of the previous output layer.

[0037] Optionally, the multi-level feature extraction model is a MobileNetV2 model, which includes multiple bottleneck layers connected in sequence.

[0038] Optionally, the multi-level feature extraction model has four output layers, namely the third, sixth, thirteenth, and seventeenth layers of the bottleneck layer.

[0039] Optionally, the convolution module unifies the number of channels in the feature maps across multiple levels, specifically as follows:

[0040] The same convolutional layer is used to unify the number of channels in feature maps across multiple layers.

[0041] Optionally, the pyramid-based efficient multi-scale attention module obtains attention feature maps at multiple scales for image enhancement, specifically:

[0042] The feature maps of multiple levels after channel unification are input into the pyramid-efficiency multi-scale attention module of different scales;

[0043] Each pyramid's efficient multi-scale attention module sequentially performs multi-scale segmentation, efficient multi-scale attention extraction, concatenation, and convolution on the input feature map to obtain the output feature map. The input feature map and the output feature map are then superimposed to obtain the corresponding attention feature map.

[0044] Optionally, the upsampling module performs step-by-step upsampling processing on attention feature maps at multiple scales to obtain the final remote sensing image landslide area detection result, including:

[0045] The upsampling module is used to perform stepwise upsampling of attention feature maps at multiple scales from high scale to obtain the final remote sensing image landslide area detection results;

[0046] The upsampling module includes multiple feature fusion units and multiple convolutional layers.

[0047] Optionally, the feature fusion unit in the upsampling module includes: an upsampling layer, Convolutional layers, batch normalized layers, and the ReLU activation function.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] This invention provides a landslide area detection method and system based on multi-level, multi-scale feature enhancement. The method includes: extracting features from landslide areas in remote sensing images using a multi-level feature extraction model to obtain feature maps at multiple levels; unifying the number of channels in the feature maps at multiple levels; inputting the feature maps with unified channel numbers into pyramid-style high-efficiency multi-scale attention modules at different scales to obtain multi-scale attention feature maps with enhanced features; and performing stepwise upsampling on the attention feature maps at multiple scales to obtain the final landslide area detection result in the remote sensing image. The multi-level feature maps contain spatial structure and texture features as well as abstract features. The use of multi-scale attention further enhances the expressive power of the features, enabling better identification and differentiation of landslide areas and avoiding the missed detection of small-scale landslide areas. Stepwise upsampling effectively fuses high-level features with low-level features, achieving a full fusion of feature maps at various levels and reducing the false alarm rate. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the structure of the landslide area detection method based on multi-level and multi-scale feature enhancement proposed in this application;

[0051] Figure 2 This is a schematic diagram illustrating the steps of the landslide area detection method based on multi-level and multi-scale feature enhancement proposed in this invention.

[0052] Figure 3 This is a schematic diagram of the bottleneck layer structure proposed in this invention;

[0053] Figure 4 The present invention proposes Figure 2 A detailed step diagram of step S3;

[0054] Figure 5 This is a schematic diagram of the pyramid-shaped high-efficiency multi-scale attention module corresponding to the lowest layer feature map of the backbone network proposed in this invention.

[0055] Figure 6 This is a schematic diagram of the efficient multi-attention structure proposed in this invention;

[0056] Figure 7 This invention presents a pyramid-shaped, multi-scale attention module of different scales.

[0057] Figure 8 This is a schematic diagram of the upsampling module proposed in this invention;

[0058] Figure 9 This is a schematic diagram comparing the detection accuracy of the landslide area detection method based on multi-level and multi-scale feature enhancement proposed in this invention.

[0059] Figure 10 This is a schematic diagram of the landslide area detection system based on multi-level and multi-scale feature enhancement proposed in this invention. Detailed Implementation

[0060] To address the issue of missed detection caused by the significant differences in scale and shape of landslides, this invention proposes a landslide area detection method and system based on multi-level, multi-scale feature enhancement. The aim is to utilize post-disaster UAV impact data to achieve rapid extraction of landslide areas. This method can be termed a multi-level, multi-scale attention neural network, whose core idea is to improve the accuracy of landslide area identification through multi-level and multi-scale feature extraction, enhancement, and fusion. Figure 1 As shown, the multi-level, multi-scale attention neural network uses a multi-level feature extraction model (such as MobileNetV2) as its backbone network. After feature unification, the input image yields multi-level feature maps d2, d3, d4, and d5. These are then input into a multi-scale pyramid-efficiency multi-scale attention module to obtain feature-enhanced attention feature maps Z2, Z3, Z4, and Z5. Z2, Z3, Z4, and Z5 are then progressively upsampled to obtain the final detection result. In particular, MobileNetV2 is a lightweight convolutional neural network with fewer parameters and lower computational complexity, making it well-suited for tasks requiring application in resource-constrained environments. Using MobileNetV2 for feature extraction from UAV remote sensing images can efficiently capture key features in the image, especially in high-resolution UAV remote sensing images, effectively extracting texture and boundary features of different ground objects. After obtaining the multi-level features, the pyramid-efficiency multi-scale attention module is used to process these features, enhancing spatial and channel features. Specifically, a pyramid-based efficient multi-scale attention module is applied to sub-regions of different scales on the same feature map, enhancing the expressive power of the features. This allows for better identification and differentiation of different spatial structures and texture features within landslide areas, preventing small-scale landslides from being missed, and resisting interference from irrelevant factors such as lighting. Finally, an upsampling module is used to decode the extracted and enhanced features. By progressively fusing features at different levels, the upsampling module maps the features back to the original image dimensions, ultimately generating the predicted landslide area. The design of the upsampling module fully considers the multi-level characteristics of features, effectively fusing features at different levels to achieve accurate localization of landslide areas.

[0061] Example 1:

[0062] A landslide area detection method based on multi-level, multi-scale feature enhancement, such as... Figure 2 As shown, the method includes steps S1 to S4.

[0063] S1: A multi-level feature extraction model is used to extract features from the landslide area in the remote sensing image to obtain feature maps at multiple levels.

[0064] In a further preferred embodiment, the step of using a multi-level feature extraction model to extract features from the landslide area in the remote sensing image to obtain feature maps at multiple levels specifically involves:

[0065] The number of output layers and the number of output layers of the multi-level feature extraction model are determined based on the detection accuracy.

[0066] The landslide area in the remote sensing image is input into the multi-level feature extraction model, and feature maps of the corresponding levels are obtained in each output layer.

[0067] In this case, the size of the image of each output layer is one-quarter of that of the previous output layer.

[0068] In a further preferred embodiment, the multi-level feature extraction model is the MobileNetV2 model, such as... Figure 1 As shown, the MobileNetV2 model includes multiple bottleneck layers connected in sequence.

[0069] The MobileNetV2 model is a lightweight neural network designed for mobile and embedded devices. It delivers high accuracy while maintaining a low number of parameters and low computational cost. The main structure of the MobileNetV2 model consists of a bottleneck layer (such as...). Figure 3 As shown), it is an inverse residual structure with two ends being The convolutional layer has a ReLU6 activation function at the input (limiting the maximum output to 6), but no activation function at the output; the layer in between has a ReLU6 activation function. Depth-separable convolutions.

[0070] The MobileNetV2 model can be a pre-trained MobileNetV2 model downloaded from the PyTorch website. In actual use, only the first 17 bottleneck layers were selected as the backbone network of the model. In this embodiment, the outputs of the third, sixth, thirteenth, and seventeenth bottleneck layers are used as feature maps at different levels. Features after the first bottleneck layer were not used because the lowest level feature map contains too much noise and redundant information, and the feature map at this level is large in scale, resulting in high computational cost. The feature map of the third bottleneck layer is one-quarter the length and width of the original image, i.e., one-sixteenth the size of the original image, with 24 channels. The feature map of the sixth bottleneck layer is one-quarter the size of the feature map of the third bottleneck layer, with 32 channels. The feature map of the thirteenth bottleneck layer is one-quarter the size of the feature map of the sixth bottleneck layer, with 96 channels. The feature map of the seventeenth bottleneck layer is one-quarter the size of the feature map of the thirteenth bottleneck layer, with 320 channels.

[0071] S2: Unify the number of channels in feature maps across multiple levels.

[0072] The feature maps from multiple levels output by different bottleneck layers are input into the convolutional layer to unify the number of channels, for example, all of them go through a single... After convolution, the number of channels in each convolutional layer is adjusted to 64. The feature maps of each layer are aligned in the channel dimension to facilitate subsequent operations.

[0073] S3: Input the feature maps with unified channel counts into the pyramid-based efficient multi-scale attention modules of different scales to obtain multi-scale attention feature maps with enhanced features.

[0074] In further optimized solutions, such as Figure 4 As shown, step S3 includes S31 and S32:

[0075] S31: Input the feature maps of multiple levels after channel unification into the pyramid-style high-efficiency multi-scale attention module of different scales respectively;

[0076] S32: Each pyramid's efficient multi-scale attention module sequentially performs multi-scale segmentation, efficient multi-scale attention extraction, concatenation, and convolution on the input feature map to obtain the output feature map. The input feature map and the output feature map are then superimposed to obtain the corresponding attention feature map.

[0077] Multi-level multi-scale attention neural networks employ aggregated multi-scale contextual attention to enhance the ability to capture information about changes in detail. Therefore, based on the introduction of efficient multi-scale attention as an attention mechanism, multi-level multi-scale attention neural networks form the Pyramid Efficient Multi-Scale Attention (PEMA) module, which can aggregate spatiotemporal attention contexts of different scales to obtain multi-scale attention features.

[0078] like Figure 5 The diagram shows the structure of the pyramid-shaped efficient multi-scale attention module corresponding to the lowest layer feature map of the backbone network. It has four different branches, each of which divides the feature map into four equal parts. Sub-regions , representing the four scales of the pyramid. For the input feature map Where C is the number of channels in the feature map, H is the height of the feature map, and W is the width of the feature map; in this branch with scale N, the feature... Figure X Each subregion can be represented as , , For each sub-region of these four branches, an Efficient Multi-Scale Attention (EMA) method is used to obtain the feature maps after processing the four branches. , Then, the feature maps of the four branches are concatenated along the channel dimension, and a convolutional layer is used to reduce the number of channels to the level of the input features. Figure X The same size yields the feature map. Finally, the elements in X and Y are superimposed to obtain the attention feature map Z at the corresponding scale. The superposition formula is shown below:

[0079]

[0080] Among them, Efficient Multi-Scale Attention (EMA) is a novel and efficient attention mechanism designed to effectively capture multi-scale spatial information and improve the performance of convolutional neural networks in computer vision tasks. This method prevents the channel dimension from being obscured by grouping channel-dimensional information into multiple sub-features. Furthermore, as... Figure 6As shown, efficient multi-scale attention employs a parallel processing strategy, utilizing three parallel paths to extract attention weight descriptors for grouped feature maps. The use of three parallel paths enables EMA to not only encode cross-channel information to adjust the importance of different channels, but also to retain precise spatial structure information in the channels. Finally, EMA uses a cross-spatial information aggregation method to model long-range dependencies and embeds precise location information into EMA, fusing contextual information at different scales so that the neural network can generate better pixel-level attention for feature maps.

[0081] EMA establishes long-range dependencies by aggregating cross-spatial information and embedding precise location information within them. EMA fuses contextual information at different scales, enabling convolutional neural networks to generate better pixel-level attention for high-level feature maps. By parallelizing convolutional kernels, EMA can handle both short-range and long-range dependencies using cross-spatial information aggregation, compared to asymptotic behavior that forms a limited receptive field. and Convolution can better utilize contextual information between intermediate feature maps.

[0082] like Figure 7 As shown, the number of branches designed in PEMA differs for feature maps extracted from different levels of the backbone network. The example above uses low-level feature maps, which are located in the lower layers of the backbone network and contain more detailed information, so they are divided into 4 branches. Features at a level higher than the lowest level are divided into 3 branches. Because the feature map at this level has a smaller resolution and contains less detailed information, it is not necessary to divide it into smaller-scale regions; the feature map at the next higher level only has two branches, and each sub-region is divided into a size equal to or one-quarter of the size of the original feature map, i.e. Similarly, there are up to two branches, and each sub-region is divided into portions equal to or one-quarter the size of the original feature map. Similarly, the highest-level feature map contains more abstract features and has the smallest height and width, so it is divided into only one branch. .

[0083] By applying pyramid-based efficient multi-scale attention to feature maps at different levels, the spatial information between pixels at different levels can be fully explored. This allows for the capture of changing locations within the feature maps at different levels, leading to better pixel localization during subsequent upsampling. Therefore, in this embodiment, the pyramid-based efficient multi-scale attention module adds four branches. These four branches process feature maps at different levels extracted by the backbone network, locating landslide pixels within each level of the feature map, thus achieving more accurate landslide detection.

[0084] S4: The attention feature maps at multiple scales are upsampled stepwise to obtain the final remote sensing image landslide area detection results.

[0085] Step S4 specifically involves using an upsampling module to perform stepwise upsampling processing on attention feature maps at multiple scales from the high scale to obtain the final remote sensing image landslide area detection result.

[0086] Among them, such as Figure 8 As shown, the upsampling module includes multiple feature fusion units and multiple convolutional layers.

[0087] The feature fusion unit includes: an upsampling layer, Convolutional layers, batch normalized layers, and the ReLU activation function.

[0088] High-level features contain abstract semantic information, enabling better localization of changing regions; low-level features contain more details, better representing object boundaries and texture information. Effectively fusing high-level and low-level features can more accurately represent the specific details of changing regions. Therefore, this invention relates to an upsampling module for progressively fusing multi-level feature maps to achieve a thorough fusion of feature maps from different levels, thereby reducing the false alarm rate.

[0089] In this embodiment, the upsampling module includes three identical feature fusion units, whose function is to fuse feature maps at different levels. Each feature fusion unit includes an upsampling layer and a... The system consists of convolutional layers, a batch normalization layer, and a ReLU activation function. The high-level feature map is first upsampled to the size of the low-level feature map through double linear interpolation, then the two are concatenated along the channel dimension, and then... The convolutional layer reduces the dimensionality of the mechanical energy channel in the concatenated feature map to achieve feature fusion. After convolution, batch normalization layers and ReLU activation functions are added to accelerate network training and improve generalization ability. For example, the attention feature map Z5 is first subjected to linear interpolation, upsampled by two times to the size of feature map Z4, and then the two are concatenated. The convolutional layer achieves feature fusion between the two, followed by batch normalization and activation functions. In the actual model, the number of channels for Z5, Z4, Z3, and Z2 is all 64. The convolutional input has 128 channels, and the output has 64 channels. After fusing the feature maps at each stage, the final fused feature map undergoes linear interpolation, upsampling to the size of the input image, and then passes through a... The convolutional layer and a The convolutional layers then perform further feature fusion to obtain the final output feature map. Finally, the feature map is passed through a sigmoid function to obtain the predicted result map. Regions with pixel values ​​greater than 0.5 are considered landslides, while those less than 0.5 are considered background.

[0090] The upsampling module can avoid upsampling high-level features. Figure one The method of multiplying features reduces feature distortion and prevents the addition of false changes in pixels. By gradually fusing feature maps at different levels, the method can better achieve the fusion of features at different levels and reduce the false alarm rate.

[0091] Experiments verify the accuracy of the method of the present invention.

[0092] (1) Dataset

[0093] Training and testing utilized unmanned remote sensing data from the Chinese Academy of Sciences (CAS) Landslide Dataset. This large-scale multi-sensor dataset for landslide detection, developed by the Artificial Intelligence Group of the Institute of Mountain Hazards and Environment (CAS), CAS, aims to address the challenges encountered in landslide identification. With the increasing incidence of landslides due to climate change and earthquakes, there is a growing need for an accurate and comprehensive dataset to support rapid and effective landslide identification. Compared to existing datasets with limitations in dataset size, coverage, sensor type, and resolution, the CAS Landslide Dataset contains 20,865 images, integrating satellite and UAV data from nine regions. This embodiment only used UAV data from the CAS Landslide Dataset, totaling 13,443 images, and randomly divided each subset into training and validation sets in a 7:3 ratio.

[0094] (2) Evaluation indicators

[0095] To accurately and effectively evaluate the accuracy of the method, this invention uses overall precision, accuracy, recall, and F1 score as evaluation metrics. In overall precision, TP (True Positive) represents the number of positive class samples correctly predicted as positive class samples, FP (False Positive) represents the number of negative class samples incorrectly predicted as positive class samples, FN (False Negative) represents the number of positive class samples incorrectly predicted as negative class samples, and TN (True Negative) represents the number of negative class samples correctly predicted as negative class samples.

[0096] Precision refers to the proportion of samples predicted as positive by a detection method that are actually positive. It measures the accuracy of the detection method in predicting positive samples. Recall refers to the proportion of all positive samples that are successfully predicted as positive. It measures the coverage of positive samples, i.e., how many positive samples can be detected. High recall means that positive samples can be captured well. The F1 score is the harmonic mean of precision and recall. It comprehensively considers both accuracy and comprehensiveness and is a relatively comprehensive indicator. The F1 score ranges from 0 to 1, and a higher value indicates better model performance. In summary, the formulas for calculating overall precision, precision, recall, and F1 score are as follows:

[0097]

[0098]

[0099] in,

[0100] (3) Experimental equipment

[0101] Regarding parameter settings, the detection method of this application was trained on an Nvidia RTX4060 GPU, using PyTorch as the framework, version 2.1.1, CUDA version 12.6, the Adam optimizer, the learning rate of 0.005, the momentum of 0.9, the weight decay of 0.0005, the batch size of 8, and the epoch of 200.

[0102] (4) Ablation analysis

[0103] The results of the ablation experiment are shown in Table 1 below. It can be seen that the accuracy and F1 score of the detection method in this application have been improved, while the recall rate has decreased slightly, and the accuracy of detection in landslide areas has been improved.

[0104] Table 1

[0105]

[0106] (5) Comparison with other methods

[0107] Table 2 compares the proposed method with other methods, demonstrating superior performance. The experimental results in Table 2 show that the proposed method outperforms the other methods in terms of Precision, Recall, F1 score, and overall accuracy (OA), indicating higher accuracy and reliability in landslide extraction tasks.

[0108] Table 2

[0109]

[0110] The visualization results of the method in this application are as follows: Figure 9 As shown, the first row is the UAV remote sensing image, the second row is the real label, and the third row is the prediction result. It can be seen that the method in this application can effectively separate the landslide area from other background areas while maintaining a low detection rate and a low false detection rate.

[0111] Example 2:

[0112] Based on the same inventive concept, this invention also provides a landslide area detection system based on multi-level, multi-scale feature enhancement, comprising:

[0113] A multi-level feature extraction model is used to extract features from landslide areas in remote sensing images to obtain feature maps at multiple levels.

[0114] The convolution module is used to unify the number of channels in feature maps across multiple levels.

[0115] The pyramid-shaped high-efficiency multi-scale attention module is used to obtain attention feature maps at multiple scales that capture changes in feature map information.

[0116] The upsampling module is used to perform step-by-step upsampling processing on attention feature maps at multiple scales to obtain the final landslide area detection results from remote sensing images.

[0117] In the landslide extraction segmentation task, the area where landslides occur is much smaller than the area where no landslides have occurred, leading to class imbalance. To alleviate this problem and guide the network to learn from complex scenes, this application employs a hybrid loss method, including binary cross-entropy (BCE) loss and Dice loss. Therefore, the loss function of the landslide area detection system based on multi-level, multi-scale feature enhancement is:

[0118]

[0119] in, The loss function is for a landslide area detection system based on multi-level, multi-scale feature enhancement. The binary cross-entropy loss function is... · represents the dot product operation. c The final results of landslide area detection from remote sensing imagery. g For real labels, The Dice loss function, , This is the first normal form.

[0120] Optionally, the multi-level feature extraction model extracts features from the landslide area in the remote sensing image to obtain feature maps at multiple levels, specifically:

[0121] The number of output layers and the number of output layers of the multi-level feature extraction model are determined based on the detection accuracy.

[0122] The landslide area in the remote sensing image is input into the multi-level feature extraction model, and feature maps of the corresponding levels are obtained in each output layer.

[0123] In this case, the size of the image of each output layer is one-quarter of that of the previous output layer.

[0124] Optionally, the multi-level feature extraction model is a MobileNetV2 model, which includes multiple bottleneck layers connected in sequence.

[0125] Optionally, the multi-level feature extraction model has four output layers, namely the third, sixth, thirteenth, and seventeenth layers of the bottleneck layer.

[0126] Optionally, the convolution module unifies the number of channels in the feature maps across multiple levels, specifically as follows:

[0127] The same convolutional layer is used to unify the number of channels in feature maps across multiple layers.

[0128] Optionally, the pyramid-based efficient multi-scale attention module obtains attention feature maps at multiple scales for image enhancement, specifically:

[0129] The feature maps of multiple levels after channel unification are input into the pyramid-efficiency multi-scale attention module of different scales;

[0130] Each pyramid's efficient multi-scale attention module sequentially performs multi-scale segmentation, efficient multi-scale attention extraction, concatenation, and convolution on the input feature map to obtain the output feature map. The input feature map and the output feature map are then superimposed to obtain the corresponding attention feature map.

[0131] Optionally, the upsampling module performs step-by-step upsampling processing on attention feature maps at multiple scales to obtain the final remote sensing image landslide area detection result, including:

[0132] The upsampling module is used to perform stepwise upsampling of attention feature maps at multiple scales from high scale to obtain the final remote sensing image landslide area detection results;

[0133] The upsampling module includes multiple feature fusion units and multiple convolutional layers.

[0134] Optionally, the feature fusion unit in the upsampling module includes: an upsampling layer, Convolutional layers, batch normalized layers, and the ReLU activation function.

[0135] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A landslide area detection method based on multi-level multi-scale feature enhancement, characterized in that, The application comprises the following steps: a multi-level feature extraction model is used to extract features of a landslide area in a remote sensing image to obtain multiple levels of feature maps; specifically, the number of output layers of the multi-level feature extraction model is determined according to the detection accuracy; the landslide area in the remote sensing image is input into the multi-level feature extraction model, and a feature map of a corresponding level is obtained at each output layer; the size of the image of each output layer is one quarter of the size of the image of the previous output layer; the number of channels of the multiple levels of feature maps is unified; the feature maps with unified channel numbers are input into pyramid efficient multi-scale attention modules of different scales to obtain multiple scales of attention feature maps with enhanced features; for feature maps of different levels, the number of branches of the pyramid efficient multi-scale attention modules designed for the feature maps of different levels is different, and the number of branches of the pyramid efficient multi-scale attention modules gradually decreases from low-level feature maps to high-level feature maps; the multiple scales of attention feature maps are processed by progressively upsampling from high scales to obtain a final remote sensing image landslide area detection result; specifically, the multiple scales of attention feature maps are processed by progressively upsampling from high scales by using an upsampling module to obtain a final remote sensing image landslide area detection result; the upsampling module comprises multiple feature fusion units and multiple convolutional layers; the multiple feature fusion units progressively upsample high-level feature maps to low-level feature maps, then splice and convolve the low-level feature maps; after the feature maps at all levels are fused, the last fused feature map is subjected to a linear interpolation operation, upsampled to the size of the input image, and then subjected to further feature fusion by a 3x3 convolutional layer and a 1x1 convolutional layer to obtain a final output feature map; finally, the feature map is subjected to a Sigmoid function to obtain a predicted result map. 2.The landslide area detection method based on multi-level multi-scale feature enhancement according to claim 1, characterized in that, The multi-level feature extraction model is a MobileNetV2 model, and the MobileNetV2 model comprises multiple bottleneck layers connected in sequence. 3.The landslide area detection method based on multi-level multi-scale feature enhancement according to claim 2, characterized in that, The number of output layers is four, and the output layers are the third layer, the sixth layer, the thirteenth layer and the seventeenth layer of the bottleneck layers. 4.The landslide area detection method based on multi-level multi-scale feature enhancement according to claim 1, characterized in that, The number of channels of the multiple levels of feature maps is unified by using the same convolutional layer. The multiple levels of feature maps with unified channel numbers are input into pyramid efficient multi-scale attention modules of different scales to obtain multiple scales of attention feature maps with enhanced features. 5.The landslide area detection method based on multi-level multi-scale feature enhancement according to claim 1, characterized in that, The multiple levels of feature maps with unified channel numbers are input into pyramid efficient multi-scale attention modules of different scales. Each pyramid efficient multi-scale attention module sequentially performs multi-scale segmentation, efficient multi-scale attention extraction, splicing and convolution on the input feature map to obtain an output feature map, and the input feature map and the output feature map are superimposed to obtain a corresponding attention feature map. The feature fusion unit comprises an upsampling layer, a convolutional layer, a batch normalization layer and a ReLU activation function. 6.The landslide area detection method based on multi-level multi-scale feature enhancement according to claim 1, characterized in that, The application comprises the following steps: 7.A landslide area detection system based on multi-level multi-scale feature enhancement, characterized in that, ​ The multi-level feature extraction model is used for feature extraction of a landslide area in a remote sensing image to obtain a plurality of levels of feature maps; specifically: the number of output layers of the multi-level feature extraction model is determined according to the detection accuracy; the landslide area in the remote sensing image is input into the multi-level feature extraction model, and a corresponding level of feature map is obtained at each output layer; wherein the size of the image of each output layer is one quarter of the determined previous output layer; The convolution module is used for unifying the number of channels of the plurality of levels of feature maps; The pyramid efficient multi-scale attention module is used for obtaining a plurality of scales of attention feature maps of image enhancement; for different levels of feature maps, the number of branches designed by the pyramid efficient multi-scale attention module is different, and the number of branches of the pyramid efficient multi-scale attention module corresponding to the feature maps of the low level to the feature maps of the high level gradually decreases; The up-sampling module is used for performing step-by-step up-sampling processing on the plurality of scales of attention feature maps to obtain a final remote sensing image landslide area detection result, the up-sampling module includes a plurality of feature fusion units and a plurality of convolution layers, the plurality of feature fusion units gradually up-sample the feature maps of the high level to the feature maps of the low level, then splice and convolutionally fuse; after fusing the feature maps at each level, the last fused feature map is subjected to a linear interpolation operation, up-sampled to the size of the input image, and then subjected to a further feature fusion through a 3*3 convolution layer and a 1*1 convolution layer to obtain a final output feature map, finally, the feature map is subjected to a Sigmoid function to obtain a predicted result map.

8. The multi-level multi-scale feature enhancement based landslide area detection system of claim 7, wherein, The loss function of the landslide area detection system based on multi-level multi-scale feature enhancement is: wherein, is a loss function of the landslide area detection system based on multi-level multi-scale feature enhancement, is a binary cross-entropy loss function, is a dot product operation, c is the final remote sensing image landslide area detection result, g is a true label, is a Dice loss function, , is the first norm.

Images