Geological disaster monitoring method based on deep learning and fused with multi-source remote sensing data

Abstract

The invention discloses a geological disaster monitoring method based on deep learning and fused with multi-source remote sensing data, and relates to the technical field of geological disaster monitoring, the geological disaster monitoring method based on deep learning and fused with multi-source remote sensing data mainly comprises the following steps: preprocessing multi-source remote sensing data to obtain preprocessed remote sensing data; fusing the preprocessed remote sensing data to obtain geological disaster remote sensing monitoring data; according to geological disaster remote sensing monitoring data, feature extraction is carried out by using a convolutional neural network introduced with adaptive multi-scale cavity convolution to obtain a multi-scale feature map; and performing target detection on various geological disasters by using an improved Faster R-CNN model according to the feature maps of the multiple scales to generate a geological disaster distribution map of the target area. By implementing the geological disaster monitoring method based on deep learning and fusing multi-source remote sensing data provided by the invention, the comprehensiveness and accuracy of automatic identification of geological disasters can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological disaster monitoring, more particularly to a geological disaster monitoring method based on deep learning and fusing multi-source remote sensing data. BACKGROUND

[0002] As one of the typical natural disasters, geological disasters cause casualties and property losses worldwide every year. Traditional geological disaster monitoring mainly relies on field monitoring stations and on-site surveys, which are costly and limited by complex geological environments and climate factors, and cannot achieve large-scale continuous monitoring. In recent years, with the characteristics of not being limited by surface environmental factors and being able to achieve large-scale continuous observation, earth observation technologies represented by optical remote sensing technology and synthetic aperture radar interferometry (InSAR) technology have begun to be more widely used in geological disaster monitoring.

[0003] Traditional remote sensing-based geological disaster monitoring technology still needs to manually determine geological disaster identification features and establish geological disaster identification rules. Due to the complex and diverse geological environment, the ground where geological disasters occur often has complex semantic features. Manually specified features are subjective and not comprehensive, especially in complex environments with multiple geological disasters, traditional technology cannot accurately locate disaster points and distinguish geological disaster categories. Deep learning methods rely on deep neural networks to automatically learn and extract target features without the need for human intervention to specify features based on prior knowledge, and have been widely used in automatic monitoring of landslides and other geological disasters.

[0004] However, existing deep learning-based geological disaster monitoring technology still has limitations. First, a single optical remote sensing technology or InSAR technology cannot achieve comprehensive monitoring of all types of geological disasters. Optical remote sensing can extract the optical features of the ground, so it can identify old landslides and new landslides caused by earthquakes, heavy rainfall, etc. that have obvious optical features. However, active landslides that are still developing or moving, as well as subsidence and ground subsidence caused by mining, which cannot be directly observed from the ground, cannot be accurately identified by optical remote sensing. InSAR can extract the deformation features of the ground and can monitor active landslides, subsidence caused by mining, and ground subsidence. However, for old landslides and ancient landslides that have basically stabilized, InSAR cannot monitor them. Current technology mostly focuses on only one aspect of optical remote sensing or InSAR, and few studies have fused the two for geological disaster monitoring.

[0005] Secondly, the causes of geological disasters are complex and diverse, and the existing geological disaster monitoring mostly only focuses on one of them, such as landslide or ground subsidence, and cannot comprehensively monitor and evaluate various types of geological disasters in a region. Moreover, the existing geological disaster monitoring technology still needs to manually determine the geological disaster identification features and establish the geological disaster identification rules. Due to the complex and diverse geological environment, the surface where the geological disaster occurs often has complex semantic features. The manually specified features are subjective and not comprehensive, especially in a complex environment with multiple geological disasters, the existing technology cannot accurately locate the disaster point and distinguish the geological disaster category.

[0006] Finally, the existing deep learning model has obvious deficiencies in monitoring multiple types of geological disasters. On the one hand, the existing deep learning model mostly uses fixed-size convolution blocks and fixed steps when extracting features, while geological disasters have significant scale differences. Using a single large-scale convolution will cause small landslides to be missed, while using a single small-scale convolution will cause large ground subsidence to be segmented and unable to extract complete and comprehensive features. On the other hand, the quality and quantity of the target samples used for training are important factors that determine the performance of the deep learning model. Ideal samples should be easy to learn and balanced among categories. However, due to the complexity of geological disasters, some samples do not have typical and easy-to-learn features, and such difficult samples usually only account for a small part of the samples and will not have a great impact on model convergence. This will cause these difficult samples to be difficult to train sufficiently after the same training process. Moreover, the distribution of geological disasters in the same region is generally not balanced, such as the coal reserves in Shanxi Province, China, which are extremely large, so the frequency of mining subsidence in this area is significantly higher than other geological disasters. This imbalance in sample categories makes it difficult for the existing deep learning model to detect multiple types of geological disasters with equal performance.

[0007] Therefore, there is an urgent need for a new geological disaster monitoring method that integrates multi-source remote sensing data and has the ability to automatically extract features and identify multiple types of geological disasters.

[0008] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0009] The purpose of the present application is to provide a deep learning-based geological disaster monitoring method that integrates multi-source remote sensing data, which can improve the comprehensiveness and accuracy of automatic identification of geological disasters.

[0010] The present application provides a deep learning-based geological disaster monitoring method that integrates multi-source remote sensing data, comprising the following steps: S1: Preprocessing multi-source remote sensing data to obtain preprocessed remote sensing data; S2: fusing the preprocessed remote sensing data to obtain geological disaster remote sensing monitoring data; S3: performing feature extraction on the geological disaster remote sensing monitoring data by using a convolutional neural network to obtain feature maps of multiple scales; S4: performing target detection on multiple geological disasters by using an improved Faster R-CNN model according to the feature maps of multiple scales to generate a geological disaster distribution map of a target region.

[0011] Further, the multi-source remote sensing data includes high-resolution optical remote sensing images, time-series InSAR data, and digital elevation model data.

[0012] Further, step S1 specifically includes: S11: obtaining slope data according to the digital elevation model data; S12: obtaining surface deformation information by using a phase gradient stacking method according to the time-series InSAR data; S13: obtaining preprocessed remote sensing data according to the slope data and the surface deformation information, wherein the preprocessed remote sensing data includes high-resolution optical remote sensing images, time-series InSAR data, digital elevation model data, slope data, and surface deformation information.

[0013] Further, step S12 specifically includes: S121: constructing a differential interferogram pair by using short baseline combination to obtain a differential interferogram; S122: obtaining directional phase gradients according to the differential interferogram; S123: stacking the directional phase gradients to obtain a stacking result; S124: filtering and denoising the stacking result to obtain a filtered result; S125: merging the filtered result to obtain merged directional phase gradients; S126: performing normalization processing on the merged directional phase gradients to obtain normalized directional phase gradients; S127: performing signal enhancement and weak signal rejection on the normalized directional phase gradients to obtain surface deformation information.

[0014] Further, step S122 specifically includes: obtaining directional phase gradients according to the differential interferogram, as shown in the formula: , wherein, is a phase gradient in the i-th differential interferogram, represents the horizontal and vertical coordinates, respectively, ​ and phase element, This indicates a phase wrapping operation. For the first Phase in a differential interferogram It is the step size for calculating the phase difference.

[0015] Further, step S123 specifically includes: stacking the phase gradients in each direction to obtain the stacking result, as shown in the formula: , in, Represents a pixel The stacking results at the location, The number of interferograms.

[0016] Further, step S124 specifically includes: filtering and denoising the stacking result to obtain the filtered result, as shown in the formula: , in, This represents the result after filtering. Indicates the stacking result. This indicates a median filtering operation.

[0017] Further, step S125 specifically includes: merging the filtered results to obtain the merged phase gradients in each direction, as shown in the formula: , in, This represents the phase gradient in each direction after merging. , , , These represent the stacked phase gradients after filtering in the four directions: north, northeast, east, and southeast, respectively.

[0018] Further, step S126 specifically includes: normalizing the merged phase gradients in each direction to obtain normalized phase gradients in each direction, as shown in the formula:

[0019] in, For the normalized phase gradient in each direction, It is the phase gradient value of the current pixel. and These represent the maximum and minimum values ​​of the phase gradient, respectively. This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned deep learning-based geological disaster monitoring method that fuses multi-source remote sensing data.

[0020] The deep learning-based multi-source remote sensing data fusion geological disaster monitoring method provided by the present application has the following beneficial effects: The present application overcomes the problems of limited recognition ability of single monitoring means and inability to automatically identify multiple types of geological disasters in the prior art. It fuses high-resolution optical images, InSAR deformation information, digital elevation model and other multi-source remote sensing data, extracts surface deformation information using the phase gradient stacking method, and constructs geological disaster remote sensing monitoring data combined with slope data. By introducing adaptive multi-scale hollow convolution and feature pyramid, multi-scale semantic features are extracted to obtain feature maps of multiple scales to detect geological disasters of different scales. The improved Faster R-CNN model is used to realize automatic identification of multiple types of geological disasters such as landslides, ground subsidence and mining subsidence, and a geological disaster distribution map of the target area is generated. This method does not require human setting of feature rules, realizes automatic feature extraction and disaster identification, improves the comprehensiveness and accuracy of automatic geological disaster identification, and is suitable for large-scale geological disaster monitoring in complex geological environments, providing data support for disaster emergency response and early warning systems. BRIEF DESCRIPTION OF DRAWINGS

[0021] The present application will be further described below with reference to the accompanying drawings and examples. In the drawings: Figure 1 is a flowchart of the deep learning-based multi-source remote sensing data fusion geological disaster monitoring method provided by the present application; Figure 2 is a structure diagram of the Faster R-CNN model provided by the present application. DETAILED DESCRIPTION

[0022] In order to have a clearer understanding of the technical features, objectives and effects of the present application, the specific embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0023] Figure 1 shows a schematic diagram of the deep learning-based multi-source remote sensing data fusion geological disaster monitoring method of the present embodiment. In the present embodiment, the deep learning-based multi-source remote sensing data fusion geological disaster monitoring method includes the following steps: S1: pre-processing the multi-source remote sensing data to obtain pre-processed remote sensing data; In an exemplary embodiment, the multi-source remote sensing data includes high-resolution optical remote sensing images, time-series InSAR data and digital elevation model data. In an exemplary embodiment, step S1 specifically includes: S11: obtaining slope data according to the digital elevation model data; S12: Based on the time-series InSAR data, the surface deformation information is obtained using the phase gradient stacking method; In one exemplary embodiment, step S12 specifically includes: S121: Construct differential interferometric image pairs using short baseline combinations to obtain differential interferograms; S122: Based on the differential interferogram, obtain the phase gradient in each direction; In one exemplary embodiment, step S122 specifically includes: obtaining the phase gradient in each direction based on the differential interferogram, as shown in the formula: , in, For the first Phase gradient in a differential interferogram The x and y coordinates are respectively and phase element, This indicates a phase wrapping operation. For the first Phase in a differential interferogram It is the step size for calculating the phase difference; S123: Stack the phase gradients in each direction to obtain the stacked result; In one exemplary embodiment, step S123 specifically includes: stacking the phase gradients in each direction to obtain a stacking result, as shown in the formula: , in, Represents a pixel The stacking results at the location, The number of interferograms; S124: Filter and denoise the stacking result to obtain the filtered result; In one exemplary embodiment, step S124 specifically includes: filtering and denoising the stacking result to obtain the filtered result, as shown in the formula: , in, This represents the result after filtering. Indicates the stacking result. This indicates a median filtering operation; S125: Combine the filtered results to obtain the combined phase gradients in each direction; In one exemplary embodiment, step S125 specifically includes: merging the filtered results to obtain the merged phase gradients in each direction, as shown in the formula: , in, denote the merged phase gradient of each direction, 、 、 、 denote the stacked phase gradient filtered in the north, northeast, east, and southeast directions, respectively; S126: normalizing the merged phase gradient of each direction to obtain a normalized phase gradient of each direction; In an exemplary embodiment, step S126 specifically includes normalizing the merged phase gradient of each direction to obtain a normalized phase gradient of each direction, as shown in the following formula:

[0024] wherein, is the normalized phase gradient of each direction, is the phase gradient value of the current pixel, and represent the maximum and minimum values of the phase gradient, respectively; S127: performing signal enhancement and weak signal elimination on the normalized phase gradient of each direction to obtain surface deformation information; S13: obtaining preprocessed remote sensing data from the slope data and the surface deformation information, wherein the preprocessed remote sensing data includes high-resolution optical remote sensing images, time-series InSAR data, digital elevation model data, slope data, and surface deformation information; S2: fusing the preprocessed remote sensing data to obtain geological disaster remote sensing monitoring data; As an exemplary embodiment, in step S2, the optical images, InSAR deformation data, elevation data, and slope data obtained in step 1 are fused to obtain geological disaster remote sensing monitoring data containing surface optical features, deformation features, and terrain features; Geological disasters include landslides, ground subsidence, mining subsidence, etc. S3: performing feature extraction using a convolutional neural network according to the geological disaster remote sensing monitoring data to obtain feature maps of multiple scales; As an exemplary embodiment, the convolutional neural network includes an improved ResNet50 network and a feature pyramid network; As an exemplary embodiment, the improved ResNet50 network is obtained by replacing the 3x3 standard convolutional layer of each residual module in Conv3 to Conv5 of ResNet50 with an adaptive multi-scale dilated convolutional module; For an exemplary embodiment, in step S3, the remote sensing monitoring data of geological disasters is subjected to feature extraction based on a convolutional neural network, complex and rich semantic features are extracted, and a feature map is generated; the convolutional neural network is an improved ResNet50 network, and an adaptive multi-scale hollow convolution and a feature pyramid network are introduced to extract feature maps of different scales; S4: according to the feature maps of the multiple scales, a target detection of multiple geological disasters is performed by using an improved FasterR-CNN model, and a geological disaster distribution map of a target region is generated; In an exemplary embodiment, the improved FasterR-CNN model is obtained by using an improved loss function on the FasterR-CNN model; the improved loss function is as follows:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] wherein, is an improved loss function, is a classification loss, is a regression loss, are weights of the classification loss and the regression loss, respectively, is a smooth L1 weighted loss, is a class prediction probability, is an index for adjusting a focal point; is an intersection over union of a predicted box and a real box, is an Euclidean distance between centers of two boxes, is a diagonal length of a minimum circumscribed rectangle, is a balance factor, is a length-width ratio difference of a box, is a width of a real box, is a height of a real box, is a width of a predicted box, is a height of a predicted box, is a sample confidence, is a smooth L1 loss, is a residual error between the predicted value and the true value, is a classification confidence of the current sample, is an average value of the classification loss, is an adjustment coefficient; As an exemplary embodiment, in step S4, the feature map is subjected to target detection based on the improved Faster R-CNN model, various geological disasters are identified, and a geological disaster distribution map is generated, so as to realize automatic identification of geological disasters; the Faster R-CNN model generates a plurality of candidate frames through a region generation network, and optimizes the positions of the candidate frames through a boundary frame regression, and finally outputs target categories and boundary frame coordinates through a full connection layer.

[0033] The embodiment provides a geological disaster automatic identification system based on deep learning, comprising: a data acquisition module configured to acquire high-resolution optical remote sensing images, time-series InSAR data, and digital elevation model (DEM) data; a data fusion module configured to fuse the remote sensing data to generate geological disaster remote sensing monitoring data; a feature extraction module configured to extract features of the geological disaster remote sensing monitoring data based on a convolutional neural network to generate a feature map; a target detection module configured to detect targets of the feature map based on an improved Faster R-CNN model to realize automatic identification of geological disasters and generate a geological disaster distribution map.

[0034] In some embodiments, the geological disaster monitoring method based on deep learning and fusion of multi-source remote sensing data can also be implemented in the following manner.

[0035] In the embodiment, the geological disaster monitoring method based on deep learning and fusion of multi-source remote sensing data comprises the following steps: I. Data acquisition and preprocessing 1. Acquire high-resolution optical remote sensing images, time-series InSAR data, and digital elevation model (DEM) data.

[0036] 2. Calculate slope data based on DEM data.

[0037] 3. Obtain surface deformation information based on InSAR data using a phase gradient stacking method, which mainly comprises the following steps: (1) Generate a differential interferogram: to ensure the coherence of the differential interferogram and reduce the interference of the terrain phase, a differential interferogram image pair is constructed based on short baseline combination to form a differential interferogram.

[0038] (2) Calculate the phase gradient in each direction: Calculate the phase gradient in the north, northeast, east, and southeast directions for all interferograms in steps of 1. The formula for calculating the phase gradient is as follows:

[0039] where, is the phase in the i-th difference interferogram, is the step size for calculating the phase difference, which is usually taken as 1 in actual calculations. is the phase unwrapping operation, which is mainly to ensure that all phase values are within the range of [0, 2π], providing a controllable data range for subsequent gradient stacking, normalization, etc.

[0040] (3) Phase gradient stacking: Calculate the phase gradient of the amplitude interferogram at each pixel and take the average, i.e., perform phase gradient stacking, to reduce noise and enhance the deformation signal. The specific calculation formula is as follows:

[0041] (4) Median filtering: Perform median filtering on the stacking results in each direction to remove noise.

[0042]

[0043] where, denotes the median filtering operation.

[0044] (5) Merge the phase gradients in each direction: Since the motion of the landslide has directionality, a single direction phase gradient may lead to missed detection. Therefore, the phase gradients in each direction are merged, and the specific calculation formula is as follows:

[0045] where, , , , represent the filtered and stacked phase gradients in the north, northeast, east, and southeast directions, respectively.

[0046] Normalize the merged results to [0, 1] to facilitate subsequent processing, and the specific calculation formula is as follows:

[0047] where, is the phase gradient value of the current pixel, and represent the maximum and minimum values of the phase gradient, respectively. ​​​​​

[0048] (6) Signal enhancement and weak signal elimination: set the minimum threshold to eliminate weak deformation signals less than the threshold; set the maximum threshold to enhance the pixels greater than the threshold to 1. The specific threshold size should be determined by the quality of the interferogram.

[0049] In an exemplary embodiment, the minimum threshold is 0.07 and the maximum threshold is 0.80. 4. Fusing the optical image, InSAR deformation data, elevation data and slope data obtained in the above steps to obtain geological disaster remote sensing monitoring data with rich surface information, i.e., the image obtained after fusion.

[0050] II. Feature extraction The pre-trained convolutional neural network (such as ResNet50) is used for feature extraction of the image obtained after fusion, which mainly includes the following steps: 1. Input the fused image obtained in the above steps.

[0051] 2. Extract features through the convolutional neural network: the improved ResNet50 overall structure includes five stages (Conv1-5), each stage is composed of multiple standard residual modules (Bottleneck Block) or adaptive multi-scale dilated convolution modules.

[0052] The standard residual module contains three standard convolution layers, and the specific structure is as follows: 1×1 convolution (dimension reduction) → 3×3 convolution (feature extraction) → 1×1 convolution (dimension increase) The adaptive multi-scale dilated convolution module is used to replace the 3×3 standard convolution layer of each residual module in Conv3-Conv5 of ResNet50. The adaptive multi-scale dilated convolution module extracts multi-scale features through parallel dilated convolution operations, and introduces an adaptive channel attention mechanism to weight and fuse each scale feature, thereby effectively improving the perception ability and feature expression ability of the model to different scale targets. The structure of the adaptive multi-scale dilated convolution module is as follows: (1) Parallel dilated convolution branch: three kinds of dilated rates are used to perform dilated convolution operations on the input feature map, and the dilated rates are r = 1, 3, and 5. The calculation formula of the dilated convolution is as follows:

[0053] Where ω represents the convolution kernel, y represents the output feature map, x represents the input feature map, r represents the dilated rate, and K represents the half size of the kernel. For a 3×3 convolution kernel, K = 1.

[0054] (2) Feature splicing and adaptive fusion The results obtained by the different scale hole convolutions are spliced in the channel dimension. In order to better fuse the multi-scale features, a channel attention mechanism is introduced to dynamically adjust the weight of each hole convolution branch according to its importance. First, a global average pooling (GAP) operation is performed on each convolution branch to obtain the global feature representation of each branch :

[0055] Then, each is input to a fully connected layer to generate the corresponding weight The weight is normalized by a softmax function:

[0056] wherein, represents the output of the fully connected layer corresponding to the i-th hole convolution branch.

[0057] Then, according to the adaptive weight of each branch, the features of different scales are weighted and summed to obtain the output feature map :

[0058] wherein, represents the output result of the i-th hole convolution branch after convolution.

[0059] Finally, the channel number of is adjusted by 1x1 convolution to obtain the final output feature map.

[0060] The main structure of ResNet50 and the input and output of each stage are as follows: (1) Conv 1: the input image size is 512x512x4, which is subjected to 7x7 standard convolution and 3x3 max pooling, and the output image size is 128x128x64.

[0061] (2) Conv 2: the input image size is 128x128x64, which is subjected to 3 Bottleneck modules, and the output image size is 64x64x256.

[0062] (3) Conv 3: the input image size is 64x64x256, which is subjected to 4 adaptive multi-scale hole convolution modules, and the output image size is 32x32x512.

[0063] (4) Conv 4: the input image size is 32x32x512, which is subjected to 6 adaptive multi-scale hole convolution modules, and the output image size is 16x16x1024.​​

[0064] (5) Conv 5: the input image size is 16x16x1024, and after passing through 3 adaptive multi-scale dilated convolution modules, the output image size is 8x8x2048.

[0065] Finally, global average pooling is performed, and the final output image size is 1x1x2048.

[0066] 3. Feature fusion through feature pyramid: the convolutional neural network extracts features from low-level spatial features to high-level semantic features through layer-by-layer convolution and pooling, and the features become more and more abstract, and the pixel size also becomes larger and larger. The traditional ResNet only analyzes the feature map of the deepest layer, and due to the large difference in scale between different types of geological disasters, it is difficult to detect landslides and mining collapses with smaller scales, so a feature pyramid network is introduced.

[0067] That is, the feature map extracted by ResNet is output layer by layer, and then up-sampling is performed layer by layer. The feature map obtained by up-sampling is spliced and convolved with the ResNet feature map of the same size, so that multiple scale feature maps can be obtained to detect geological disasters of different scales.

[0068] III. Automatic identification of geological disasters; Based on the improved Faster R-CNN model, the target is detected for multiple geological disasters to obtain the spatial distribution and category information of the target area geological disaster, and automatic identification of geological disasters is realized. Compared with the traditional Faster R-CNN model, the present application introduces a new loss function DTW-CAF Loss (Dynamic Task-aware Weighting Confidence-aware Adaptive Focal-IoU Loss), which effectively solves the problems of class imbalance, difficult sample learning and low-quality label of geological disasters, thereby improving the accuracy and robustness of target detection.

[0069] As shown in Figure 2 Faster R-CNN is a two-stage target detection algorithm based on convolutional neural network, mainly including the following steps: 1. Input the feature maps of different scales obtained in the above steps.

[0070] 2. Generate candidate boxes based on the region generation network (RPN): RPN first generates multiple anchor boxes of different sizes at each pixel on the feature map, then judges whether each anchor box contains a target through convolution, and finally adjusts the position of the anchor box through boundary box regression to make the anchor box position closer to the true boundary. The output results of RPN include the boundary box coordinates of each position and the probability of containing a target.

[0071] 3. ROI Pooling: The ROI Pooling layer maps the generated candidate boxes of different sizes to the feature map of the original image, and unifies their corresponding feature maps to a fixed size for subsequent fully connected layer processing.

[0072] 4. Fully connected layer: The features output by the ROI Pooling layer are integrated, converted, and a final decision is made. The output of the fully connected layer includes the coordinates of the candidate box and the probability distribution of the target class.

[0073] The traditional Faster R-CNN uses standard cross-entropy loss and regression loss, but these loss functions fail to effectively address the problems of class imbalance, low-quality labels, and insufficient training of difficult samples in geological disaster detection. To solve these problems, the present invention proposes a new DTW-CAF Loss loss function and introduces it into Faster R-CNN. DTW-CAF Loss optimizes in the following three aspects: The specific form of the DTW-CAF Loss function is:

[0074] The calculation formulas and purposes of each loss function are as follows: Focal Loss, which is a classification loss, aims to reduce the impact of easy-to-classify samples on training and focus on difficult-to-classify samples. Its calculation formula is as follows:

[0075] where: is the class prediction probability. γ is the index that adjusts the focus, and the commonly used value is 2.

[0076] CIoU Loss, which is a regression loss, aims to optimize the positional relationship between the predicted box and the real box. Its calculation formula is as follows:

[0077] where: IoU is the intersection over union of the predicted box and the real box.

[0078] is the Euclidean distance between the centers of the two boxes.

[0079] c is the diagonal length of the minimum bounding rectangle, i.e., the diagonal of the smallest rectangle containing the two boxes.

[0080] ν is the aspect ratio difference of the box, and the calculation formula is:

[0081] is a balance factor that controls the influence of the aspect ratio term on the loss function.

[0082] is a loss function that weights Smooth L1 Loss according to the prediction confidence of each sample, aiming to further enhance the contribution of difficult samples in the model training process.

[0083]

[0084] wherein: is the sample confidence.

[0085] Smooth L1 Loss is a commonly used loss function to calculate the difference between the predicted value and the true value, and its calculation formula is as follows:

[0086] wherein x is the residual between the predicted value and the true value.

[0087] And the weight of each loss function in DTW-CAF Loss is dynamically adjusted according to the task difficulty of each training sample. The calculation of the weight needs to rely on the classification confidence and regression accuracy of each sample. The calculation formula is as follows:

[0088]

[0089] wherein: p is the classification confidence of the current sample.

[0090] is the average value of the classification loss, is the adjustment coefficient.

[0091] IoU is the intersection over union between the predicted box and the true box.

[0092] The embodiment provides a computer program product, comprising a computer program which, when executed by a processor, implements the steps of the deep learning-based geological disaster monitoring method for fusing multi-source remote sensing data.

[0093] The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are merely illustrative, but not restrictive, and a person of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, and these all belong to the protection of the present application.

Claims

1. A deep learning based method for monitoring geological disasters by fusing multi-source remote sensing data, characterized in that, The method comprises the following steps: S1: preprocessing multi-source remote sensing data to obtain preprocessed remote sensing data; S2: fusing the preprocessed remote sensing data to obtain geological disaster remote sensing monitoring data; S3: performing feature extraction on the geological disaster remote sensing monitoring data by using a convolutional neural network to obtain feature maps of multiple scales; S4: performing target detection on multiple geological disasters by using an improved Faster R-CNN model according to the feature maps of multiple scales to generate a geological disaster distribution map of a target region.

2. The deep learning-based geologic hazard monitoring method of fusing multi-source remote sensing data according to claim 1, characterized in that, The multi-source remote sensing data comprises high-resolution optical remote sensing images, time-series InSAR data and digital elevation model data.

3. The deep learning-based method for monitoring geological disasters by fusing multi-source remote sensing data according to claim 2, characterized in that, Step S1 specifically comprises: S11: obtaining slope data according to the digital elevation model data; S12: obtaining surface deformation information by using a phase gradient stacking method according to the time-series InSAR data; S13: obtaining preprocessed remote sensing data according to the slope data and the surface deformation information, wherein the preprocessed remote sensing data comprises high-resolution optical remote sensing images, time-series InSAR data, digital elevation model data, slope data and surface deformation information.

4. The deep learning-based method for monitoring geological disasters by fusing multi-source remote sensing data according to claim 3, characterized in that, Step S12 specifically comprises: S121: constructing a differential interferogram image pair by using short baseline combination to obtain a differential interferogram; S122: obtaining directional phase gradients according to the differential interferogram; S123: stacking the directional phase gradients to obtain a stacking result; S124: filtering and denoising the stacking result to obtain a filtered result; S125: merging the filtered result to obtain merged directional phase gradients; S126: normalizing the merged directional phase gradients to obtain normalized directional phase gradients; S127: performing signal enhancement and weak signal rejection on the normalized directional phase gradients to obtain surface deformation information.

5. The deep learning-based method for monitoring geological disasters by fusing multi-source remote sensing data according to claim 4, characterized in that, Step S122 specifically comprises: obtaining directional phase gradients according to the differential interferogram, as shown in the formula: ， wherein, is the phase gradient in the mth interferogram, denotes a phase element with horizontal and vertical coordinates and denotes a phase wrapping operation, is the phase in the mth interferogram, is the step size for calculating the phase difference.​ 6. The deep learning-based method for monitoring geological disasters by fusing multi-source remote sensing data according to claim 4, characterized in that, Step S123 specifically comprises: stacking the directional phase gradients to obtain a stacking result, as shown in the formula: ， wherein, represents the number of stacks, at the pixel, the number of interferograms.

7. The deep learning-based method for monitoring geological disasters by fusing multi-source remote sensing data according to claim 1, characterized in that, The convolutional neural network comprises an improved ResNet50 network and a feature pyramid network.

8. The deep learning-based method for monitoring geological disasters by fusing multi-source remote sensing data according to claim 7, characterized in that, The improved ResNet50 network is obtained by replacing the 3*3 standard convolution layer of each residual module in Conv3 to Conv5 of ResNet50 with an adaptive multi-scale hollow convolution module. 9.The deep learning based geologic hazard monitoring method of fusing multi-source remote sensing data according to claim 1, wherein, The improved Faster R-CNN model is obtained by using an improved loss function on the Faster R-CNN model; the improved loss function is as shown in the formula: wherein, is an improved loss function, is a classification loss, is a regression loss, are weights for the classification loss and the regression loss, respectively, is a smooth L1 weighted loss, is a category prediction probability, is an index for adjusting the focal point; is an intersection over union of a predicted box and a real box, is an Euclidean distance between two box center points, is a diagonal length of a minimum circumscribed rectangle, is a balance factor, is a length-width ratio difference of a box, is a width of a real box, is a height of a real box, is a width of a predicted box, is a height of a predicted box, is a sample confidence, is a smooth L1 loss, is a residual between a predicted value and a real value, is a classification confidence of a current sample, is an average of a classification loss, is an adjustment coefficient.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by a processor to implement the steps of the geological disaster monitoring method based on deep learning and fusing multi-source remote sensing data according to any one of claims 1-9.

Images