Remote sensing extraction method and system for damaged house in extreme rainstorm and flood disasters

Abstract

The invention provides a remote sensing extraction method and system for a damaged house under extreme rainstorm and flood disasters, and relates to the field of remote sensing image intelligent processing, and the method comprises the steps: obtaining remote sensing image data with consistent time-space reference before and after a key target house is affected by a disaster through multi-source high-resolution remote sensing data; constructing a house key target multi-modal remote sensing feature set fusing spectrum and texture features, and realizing dimensionality reduction and fusion of multiple features by adopting a principal component method and a dynamic weight distribution method; for different house types, a high-resolution remote sensing house sample marking method considering a geometric structure is provided, a Gaussian mixture model is adopted to realize house sample probability measurement, and a house element sample library is formed; training the features by adopting an improved deep learning model, and constructing a house remote sensing intelligent extraction model; and in combination with the training model and multi-modal remote sensing data, automatically extracting key targets of the house before and after the disaster, and determining a damaged house target by adopting a change analysis detection method.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image intelligent processing, and in particular relates to a remote sensing extraction method and system for houses damaged by extreme rainstorm and flood disasters. Background Art

[0002] In recent years, extreme rainstorms and floods have occurred frequently, causing tremendous loss of life and property. These disasters are characterized by suddenness and widespread impact, causing severe damage to homes and posing a serious threat to residents' property safety and post-disaster production and life.

[0003] Traditional monitoring of flood-damaged homes relies primarily on manual on-the-ground inspections. However, these surveys are limited in scope and fail to fully and accurately reflect the actual extent of damaged homes across the entire affected area. Furthermore, traditional monitoring methods lag in the timeliness of data acquisition, making them inadequate for timely on-site emergency response to extreme rainstorms and floods.

[0004] Satellite remote sensing technology provides a new approach for monitoring damaged homes during extreme rainstorms and flooding. It can periodically capture large-scale, pre- and post-disaster surface imagery, visually demonstrating changes in homes within affected areas. However, traditional methods for extracting information about water bodies and damaged homes suffer from low accuracy due to the existence of different spectra for the same object, the same spectra for different objects, and interference from complex scenes in high-resolution imagery. This makes it difficult to meet the precision requirements for on-site emergency response to extreme rainstorms and flooding.

[0005] In recent years, deep learning, with its powerful feature learning capabilities, has been able to deeply mine hidden features of damaged buildings in remote sensing imagery, effectively overcoming the subjectivity and limitations of manual feature extraction in traditional methods. Deep learning has been widely used in the field of remote sensing image recognition and classification, and it has unique advantages in processing complex, high-resolution remote sensing image data.

[0006] In order to solve the problem of rapid extraction of damaged houses in extreme rainstorms and floods, deep learning is applied to remote sensing extraction of damaged houses in extreme rainstorms and floods. A multimodal remote sensing dataset is constructed, remote sensing features of houses are extracted, and a typical sample library of house elements is built. Deep learning methods are combined to realize house extraction before and after disasters and change detection of damaged houses. This method can quickly and accurately identify the location of damaged houses from massive remote sensing images, providing key data support for disaster emergency response, loss assessment, and post-disaster reconstruction, greatly improving the scientificity and effectiveness of flood disaster management, and has broad application prospects. Summary of the Invention

[0007] In response to the above-mentioned deficiencies in the prior art, the present invention provides a remote sensing extraction method and system for houses damaged by extreme rainstorms and floods, which solves the problem of difficulty in quickly extracting houses damaged by extreme rainstorms and floods.

[0008] To achieve the above objectives, the present invention adopts a technical solution: a remote sensing extraction method for houses damaged by extreme rainstorm and flood disasters, comprising the following steps: Utilize multi-source high-resolution remote sensing data to obtain remote sensing image data of key target houses with consistent temporal and spatial benchmarks before and after the disaster; Based on the acquired remote sensing image data, a multimodal remote sensing feature set of key house targets is constructed by integrating spectral and texture features; For different house types, high-resolution remote sensing house samples are labeled based on geometric structures. High-resolution remote sensing house sample labeling is performed to calculate the similarity probability of multiple house types. Using labeled high-resolution remote sensing house samples and combining them with a multimodal remote sensing feature set of key house targets, we train and build a deep learning-based intelligent house extraction model. The trained intelligent house extraction model is used to determine the condition of damaged houses after extreme rainstorms and floods, and complete the remote sensing extraction of damaged houses.

[0009] Furthermore, the multimodal remote sensing feature set of key housing targets that integrates spectral and texture features is constructed as follows: Based on the acquired remote sensing image data, determine the pixel point of each remote sensing image The spectrum vector ; Calculate the normalized vegetation separately and normalized architecture spectral characteristics; The spectrum vector and normalized vegetation and normalized architecture The spectral feature combination is the house area spectral feature vector , complete the extraction of remote sensing spectral characteristics of the housing area; Extract image texture features based on acquired remote sensing image data , complete the extraction of texture features of the housing area; Based on the extracted spectral characteristics of the house area and the texture characteristics of the house area, for each candidate house area , extract the spectral feature vectors of the house area respectively and texture feature vector ,construct regional level features through feature splicing; Based on the regional-level features constructed by feature splicing, feature priority is performed to obtain the prioritized regional-level features; Prioritized regional-level features Forming a structured feature set ; Based on structured feature sets , calculate the information entropy of each feature ; Based on adaptive adjustment factor and feature information entropy , calculate the weight of each feature ; According to the weight of each feature Sorting, taking the feature with the largest weight as the house feature, completing the construction of the multimodal remote sensing feature set of the key targets of the house.

[0010] Furthermore, the similarity probability of multiple types of houses is calculated as follows: Constructing a hierarchical geometric feature space, where the geometric features include a first-level feature representing the house area, a second-level feature representing the house's aspect ratio, and a third-level feature representing the house's shape index; Based on the constructed hierarchical geometric feature space, the house type discrimination rules are constructed; Based on the constructed house type discrimination rules, a Gaussian mixture model is constructed for each house type, and the Gaussian mixture model is trained to model the geometric prototypes of multiple house types; Based on the modeling results of multi-type house geometric prototypes, each candidate area in the new remote sensing image is calculated , and get each candidate region The eigenvector of the posterior probability of belonging to each class; According to the principle of maximum a posteriori probability, the house type label is assigned to obtain the predicted house type label ; Get the real house type label , and label the actual house type and the predicted house type labels Make a comparison, use the comparison results to modify the Gaussian mixture model, and complete the calculation of the similarity probability of multiple types of houses.

[0011] Furthermore, the expression of the house type discrimination rule is as follows:

[0012]

[0013]

[0014]

[0015] in, Indicates the house type identification rule, represents the shape index of the house, The ratio of the length to the width of the house. Indicates the area of ​​the house. The value range of the area is between 50 square meters and 5000 square meters. Indicates the length of the house. Indicates the width of the house. Indicates the perimeter of the house; The predicted house type label The expression is as follows:

[0016]

[0017] in, Representing a collection The specific housing type in C Represents a set of house types, including all house categories to be classified, and argmax represents the maximum posterior probability. value, Represents the posterior probability, that is, the feature vector of the known candidate region Under the condition of The probability of Represents the prior probability, that is, the house type The prior distribution probability in the entire data set, Represents the likelihood probability, that is, when the house type is When the eigenvector The probability of Representing a collection C Except the current category c Any housing type except Indicates that when the house type is When the eigenvector The probability of Indicates the type of house c The prior probability of ′.

[0018] Furthermore, the training and construction of the deep learning-based intelligent house extraction model is specifically as follows: Divide the labeled high-scoring remote sensing house samples into training and validation sets; Combined with the multimodal remote sensing feature set of key housing targets, the training set is used to build a housing intelligent extraction model based on deep learning, and the mapping relationship between image features and labels is learned for training. Probabilistic extraction of housing; The validation set is used to tune the parameters of the trained house intelligent extraction model to complete the construction of the house intelligent extraction model.

[0019] Furthermore, the loss function of the intelligent house extraction model is expressed as follows:

[0020]

[0021]

[0022]

[0023] in, represents the loss function of the house intelligent extraction model, represents the weight coefficient, represents the category-sensitive weighted cross loss function, represents the boundary loss function, N Indicates the total number of pixels, i represents the pixel index, represents the category weight, represents the actual house type label, represents the predicted house type label, Indicates the type of house The proportion of samples Indicates the minimum value introduced to avoid the denominator being zero, represents the boundary sensitivity coefficient, p represents the probability that the predicted pixel belongs to the house boundary, Y represents the true house boundary mask, represents the partial derivative, Represents the predicted pixel p The probability of belonging to a certain category of housing, Represents pixels p The true label.

[0024] Furthermore, the automatic extraction of the housing range before and after the extreme rainstorm and flood disaster and the determination of the changes in damaged houses are specifically as follows: Obtain high-resolution remote sensing image data before and after extreme rainstorm and flood disasters occur in the same area; Input high-resolution remote sensing image data before the extreme rainstorm and flood disaster into the trained house intelligent extraction model to obtain the bounding box coordinates of each house and housing category information, as well as statistics on the number of houses in the area before the disaster ; Input high-resolution remote sensing image data after extreme rainstorm and flood disasters into the trained house intelligent extraction model to obtain the bounding box coordinates of each house and housing category information, as well as statistics on the number of houses in the area after the disaster ; For each house extracted before the extreme rainstorm and flood disaster, its positional relationship with the house extracted after the extreme rainstorm and flood disaster is determined, and the overlap of the two bounding boxes is measured. If the overlap is greater than a preset threshold, the house has not changed; if the overlap is less than the preset threshold, the house has been displaced and / or damaged. According to the number of houses before the disaster and the number of houses after the disaster , calculate the rate of change of the number of houses; Based on the rate of change in the number of houses, the extent of damage to houses caused by extreme rainstorms and floods can be determined by judging the increase or decrease in house data; For each house, the house area before and after the extreme rainstorm flooding disaster was calculated; Calculate the rate of change of housing area based on the housing area before and after the extreme rainstorm flood disaster; The change of house area is determined according to the rate of change of house area, and the remote sensing extraction of damaged houses is completed.

[0025] The present invention also provides a remote sensing extraction system for houses damaged by extreme rainstorm and flood disasters, comprising: The first processing module is used to use multi-source high-resolution remote sensing data to obtain remote sensing image data of key target houses with consistent temporal and spatial references before and after the disaster; The second processing module is used to construct a multimodal remote sensing feature set of key house targets that integrates spectral and texture features based on the acquired remote sensing image data; The third processing module is used to mark high-resolution remote sensing house samples based on geometric structures according to different house types, wherein the high-resolution remote sensing house sample marking is performed to calculate the similarity probability of multiple types of houses; The fourth processing module is used to train and build a deep learning-based intelligent house extraction model by using labeled high-resolution remote sensing house samples and combining them with a multimodal remote sensing feature set of key house targets; The fifth processing module is used to use the trained house intelligent extraction model to determine the condition of damaged houses after extreme rainstorms and floods, and complete the remote sensing extraction of damaged houses.

[0026] Beneficial effects of the present invention: (1) A remote sensing extraction method for houses damaged by rainstorms and floods was constructed, breaking through the technical bottleneck of rapid extraction of houses damaged by extreme rainstorms and floods; (2) A multimodal remote sensing feature set of key house targets that integrates spectral texture features was constructed to achieve quantitative extraction of house features; (3) A high-resolution remote sensing house sample labeling method that takes into account house type and geometric structure was used to calculate the similarity probability of multiple types of houses; (3) A deep learning-based house remote sensing intelligent extraction model was constructed to achieve automatic extraction of house ranges before and after disasters and determination of changes in damaged houses. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Flow chart of the method of the present invention.

[0028] Figure 2 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0029] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0030] Example 1 like Figure 1 As shown, the present invention provides a remote sensing extraction method for houses damaged by extreme rainstorm and flood disasters, and its implementation method is as follows: S1. Utilize multi-source high-resolution remote sensing data to obtain remote sensing image data of key target houses with consistent temporal and spatial references before and after the disaster; In this embodiment, high-resolution satellite remote sensing data is acquired to meet the needs of emergency response to extreme rainstorms and floods. This remote sensing data includes optical remote sensing data with a spatial resolution better than 2 meters, including data from the Gaofen-1, Gaofen-2, and Gaofen-6 satellites. High-quality cloud- and snow-free data is selected for this remote sensing data. This data is then processed through geometric correction, atmospheric correction, and cloud mask processing to obtain remote sensing imagery of key target houses with consistent temporal and spatial references before and after the disaster.

[0031] S2. Based on the acquired remote sensing image data, a multimodal remote sensing feature set of key housing targets is constructed by integrating spectral and texture features. Specifically, the set is as follows: Based on the acquired remote sensing image data, determine the pixel point of each remote sensing image The spectrum vector ; Calculate the normalized vegetation and normalized architecture Spectral characteristics; spectral vector and normalized vegetation and normalized architecture The spectral feature combination is the house area spectral feature vector , complete the extraction of spectral characteristics of the housing area by remote sensing; extract image texture features based on the acquired remote sensing image data , complete the extraction of the housing area texture features, where the extracted housing area spectral features and housing area texture features include housing areas and other types of land features, such as farmland. Based on the extracted housing area spectral features and housing area texture features, for each candidate housing area , extract the spectral feature vectors of the house area respectively and texture feature vector , where the spectral feature vector extracted here is and texture feature vector Only the housing area is included, excluding the other landform types mentioned above. Regional features are constructed by feature splicing. Based on the regional features constructed by feature splicing, feature priority is performed to obtain the prioritized regional features. Based on the prioritized regional features Forming a structured feature set , based on structured feature sets , calculate the information entropy of each feature ; Based on adaptive adjustment factor and feature information entropy , calculate the weight of each feature ; According to the weight of each feature Sorting, taking the feature with the largest weight as the house feature, completing the construction of the multimodal remote sensing feature set of the key targets of the house.

[0032] In this embodiment, the spectral characteristics of the house are extracted by remote sensing as follows: Each remote sensing image pixel The spectral vector of is expressed as:

[0033] in, represents the spectral vector, Indicates the c Bands at pixel points ( x , y ) at the spectral intensity value, Represents a collection of bands.

[0034] To enhance the discrimination, the normalized vegetation of the houses is calculated separately. and normalized architecture Spectral characteristics:

[0035]

[0036] in, Represents pixel points The near-infrared band spectral intensity value at Represents pixel points The red light band spectral intensity value at Represents pixel points The spectral intensity value of the short-wave infrared band at .

[0037] The spectral vector, normalized vegetation and normalized building features are further combined into the house spectral feature vector ,in, k Indicates the number of newly added spectral features.

[0038] In this embodiment, the house texture feature extraction uses the gray level co-occurrence matrix and Gabor filter to extract the image texture features: First, the image is converted into a grayscale image. , calculate the co-occurrence matrix of different directions and distances P , and then calculate the energy separately , contrast and correlation feature:

[0039]

[0040]

[0041] in, Indicates that the gray value in the gray level co-occurrence matrix (GLCM) is and The normalized probability of the pixel pair appearing is Represents the grayscale value of the image. In the normalized image, Indicates pure black, Means pure white, The mean of the row direction and the mean of the column direction in the gray level co-occurrence matrix, Represents the standard deviation in the row direction and column direction in the gray-level co-occurrence matrix.

[0042] comprehensive n Texture feature statistics vector in each direction, construct texture feature vector .

[0043] The Gabor filter is used to provide the texture features of remote sensing images. The Gabor filter function is:

[0044]

[0045]

[0046] in, represents the filter function, represents the image space coordinates, Indicates that the original coordinates are converted into direction angles The new coordinates after rotation, Indicates the standard deviation of the Gaussian envelope, which is used to control the bandwidth of the Gaussian window. The larger the value, the wider the filter coverage area, but the lower the resolution. Represents the wavelength of the sine component, which is used to control the sine component period of the filter and determine the scale of the texture. Represents the phase offset, which is used to adjust the initial phase of the sine component and affect the symmetry of the texture. Indicates the direction angle of the filter, which is used to define the detection direction of the filter.

[0047] At multiple scales and parameter angle Take a group of images from each of the above, extract the mean and variance as the texture feature statistical vector Final texture features for:

[0048] in, represents the spatial wavelength of the sinusoidal component in the filter, Indicates the p spatial wavelengths, n represents the feature dimension extracted by Gabor filter, m represents the texture feature dimension extracted by GLCM, represents the texture feature vector extracted by the gray-level co-occurrence matrix (GLCM), represents the multi-scale, multi-directional texture feature vector extracted by the Gabor filter, The total dimension of the final texture feature vector is .

[0049] In this embodiment, a dynamic weight allocation method is used to achieve multi-feature optimization: For each candidate housing area , extract the spectral feature vectors of the house area respectively and texture feature vector , construct regional level features by feature splicing :

[0050] in, represents the feature concatenation operation, Indicates housing area Medium pixel The spectral feature set of Housing area Medium pixel Texture feature set.

[0051] Based on the regional features constructed by feature splicing, feature priority is performed to obtain the prioritized regional features. , such as prioritizing based on weight:

[0052] in, represents the total number of candidate housing areas, Indicates the k regions, Indicates the horizontal and vertical coordinates of pixels.

[0053] Forming a structured feature set :

[0054] in, Represents the region-level feature vector of the Mth region.

[0055] Further calculate the information entropy of each feature , The calculation formula is as follows:

[0056] in, represents the number of features, Indicates the feature number, No. The probability of a feature taking a value under a certain probability distribution.

[0057] The adaptive adjustment factor can dynamically adjust the weight calculation according to the discrete degree of the feature. The greater the discrete degree (the greater the variance), The larger the value of may be, the different effects on the weight calculation will be produced. The calculation formula is as follows:

[0058]

[0059] in, represents the number of features, Indicates the The variance of the features, Representation characteristics The number of samples, Indicates the The first feature values, Indicates the The average value of the features.

[0060] Calculate the weight of each feature :

[0061] in, Indicates the j The information entropy of the features, j Indicates all The first of the features j Features.

[0062] Sort by the weight of each feature, and the features with the highest priority will be used as the house features. , Represents the Nth house feature.

[0063] S3. Labeling high-resolution remote sensing house samples based on geometric structure for different house types. The high-resolution remote sensing house sample labeling is performed to calculate the similarity probability of multiple house types, which is specifically as follows: A hierarchical geometric feature space is constructed, wherein the geometric features include a primary feature representing the house area, a secondary feature representing the length-to-width ratio of the house, and a tertiary feature representing the house shape index; based on the constructed hierarchical geometric feature space, a house type discrimination rule is constructed; based on the constructed house type discrimination rule, a Gaussian mixture model is constructed for each house type, and the Gaussian mixture model is trained to model the geometric prototypes of multiple house types; based on the modeling results of the geometric prototypes of multiple house types, each candidate area in the new remote sensing image is calculated , and get each candidate region The eigenvector of The posterior probability of belonging to each category; according to the maximum posterior probability principle, the house type label is assigned to obtain the predicted house type label ; Get the real house type label , and label the actual house type and the predicted house type labels Make a comparison, use the comparison results to modify the Gaussian mixture model, and complete the calculation of the similarity probability of multiple types of houses.

[0064] In this embodiment, the geometric prototypes of multiple types of houses are modeled as follows: Establishing a hierarchical geometric feature space includes: First-level features ,in, Indicates the area of ​​the house, and the value range is between 50 square meters and 5000 square meters; Secondary features: ,in, The ratio of the length to the width of the house. Indicates the length of the house. Indicates the width of the house. The value range for the aspect ratio is between 0 and 2.5. This feature can reflect whether the house shape is more square or rectangular, which can be used to determine the layout and space utilization of the house.

[0065] Level 3 features: ,in, represents the shape index of the house, The perimeter of the house. Indicates the area of ​​a house. The shape index further characterizes the shape characteristics of a house through the relationship between perimeter and area, and can describe the geometric characteristics of a house in more detail.

[0066] Build type discrimination rules:

[0067] Let the set of house types be , build a Gaussian mixture model (GMM) for each class:

[0068] in, Indicates that when the house type is When the eigenvector The probability of Indicates the j Gaussian components, represents the number of Gaussian components, Indicates the The mixing weights of the components, represents the Gaussian distribution function, represents the feature vector, and denote the mean and covariance matrix respectively.

[0069] For each type of Gaussian mixture model constructed, the geometric feature data of the training samples belonging to that type of house are used to train it through the EM algorithm (Expectation Maximization algorithm). The EM algorithm is an iterative algorithm used to estimate parameters in models with latent variables. In the Gaussian mixture model, the EM algorithm is used to estimate the model parameters. 、 、 . Through training, various types of houses are obtained Gaussian mixture model , the distribution boundaries of various types of house characteristics can be determined, thereby more accurately classifying and identifying house types.

[0070] In this embodiment, region type identification and label generation are as follows: According to the Bayesian formula, each candidate region in the new image is calculated , calculate its eigenvector The posterior probability of belonging to each class:

[0071] Assign house type labels based on the maximum a posteriori probability principle:

[0072] in, Representing a collection The specific housing type in C Represents a set of house types, including all house categories to be classified, and argmax represents the maximum posterior probability. value, Represents the posterior probability, that is, the feature vector of the known candidate region Under the condition of The probability of Represents the prior probability, that is, the house type The prior distribution probability in the entire data set, Represents the likelihood probability, that is, when the house type is When the eigenvector The probability of Representing a collection C Except the current category c Any other housing type except Indicates that when the house type is When the eigenvector The probability of Indicates the type of house c The prior probability of ′.

[0073] If the maximum posterior probability is lower than the threshold , it is marked as "uncertain" and sent for manual review or automatically enters the iterative learning pool.

[0074] In this example, the automatically labeled samples are fed into a lightweight verification model (1D-CNN) and combined with human experts for review. 1D-CNN is suitable for processing data with sequence characteristics. It automatically extracts characteristic patterns in the data through convolution operations and can quickly verify the samples. Human expert review can make detailed judgments on the samples based on human expertise and experience. Real feedback is obtained through two methods. and system prediction For comparison, the modified Gaussian mixture model GMM.

[0075] Construct a confidence function:

[0076] Update model weights :

[0077] in, Indicates the i The confidence value of the sample, c Represents a set of house types Specific categories in Representation sample Belong to category The posterior probability of Indicates the i The known true labels of samples, Represents a given feature Under the condition of probability.

[0078] By repeatedly repeating the above process of verification, confidence calculation, and model weight update, the Gaussian mixture model (GMM) is iteratively optimized. As the number of iterations increases, the GMM continuously learns new knowledge and gradually corrects incorrect predictions, thereby improving the accuracy of sample labeling and making the GMM more accurate and reliable in region type identification and label generation tasks.

[0079] S4. Using labeled high-resolution remote sensing house samples and combining them with a multimodal remote sensing feature set of key house targets, we train and build a deep learning-based intelligent house extraction model. Specifically: The labeled high-resolution remote sensing house samples are divided into a training set and a validation set. Combined with the multimodal remote sensing feature set of key house targets, the training set is used to build a house intelligent extraction model based on deep learning, and the model is trained by learning the mapping relationship between image features and labels. The predicted house type label is used as the basis for the extraction. Probabilistic house extraction; the training set is input into the improved U-net model, and the weights of the house intelligent extraction model are optimized through back propagation, and the mapping relationship between remote sensing image features and house type labels and bounding boxes is learned; the validation set is used to tune the parameters of the trained house intelligent extraction model to complete the construction of the house intelligent extraction model.

[0080] In this example, labeled multi-type house samples are divided into a training set and a validation set. A combination of house feature extraction and optimization methods is used to extract features from each method and construct a U-net deep learning model. The U-net deep learning model is trained by learning the mapping relationship between image features and labels, extracting houses based on predicted class probabilities. The validation set is then used to fine-tune the model parameters, and post-processing yields accurate house extraction results. The following describes the method for constructing a remote sensing intelligent house extraction model from three aspects: the U-net deep learning model algorithm, the loss function, and the model training.

[0081] (1) Model algorithm The model includes an encoder and a decoder. The intermediate layer fuses the encoder's house feature map with the decoder's house feature map through skip connections to retain more spatial detail information.

[0082] 1) Encoder The encoder consists of multiple downsampling modules. Each module performs double convolution and max pooling, gradually compressing the spatial size and increasing the number of channels to extract multi-level features. Each layer consists of two 3×3 convolution operations, each followed by a ReLU activation function, as shown in the following formula:

[0083]

[0084] in, Indicates the The input house feature map of the first layer is the training sample set, Input house feature map of the layer It is The output of the layer, 、 and 、 Represent the convolution kernel weight and bias respectively, ∗ represents the convolution operation, It represents the output house feature map after the first level convolution module. After each two convolution layers, it is downsampled through a 2×2 max pooling layer. The formula is as follows:

[0085] in, represents the output housing feature map, Represents the output house feature map after the second-level convolution module.

[0086] After the maximum pooling operation, the size of the output feature map is halved, and the number of channels of the house feature map is increased. The global semantic information of the image is gradually extracted, and the house feature map is output as the input of the next intermediate layer.

[0087] 2) Middle layer The middle layer receives the house feature map output by the encoder at the final stage , , further extract deep semantic features through two convolutions: )+ ).

[0088] 3) Decoder The decoder receives the house feature map output by the intermediate layer , gradually restore the resolution through deconvolution and jump connection. Each level of the decoder inputs the house feature map From the previous level output or intermediate layer, upsampled by 2×2 transposed convolution:

[0089] The upsampled house feature map House feature map corresponding to the encoder layer Splicing along the channel dimension, the formula is as follows:

[0090] in, represents the composite feature map after splicing, The concatenated feature maps are then subjected to two 3×3 convolution operations to gradually restore the image details.

[0091] The intelligent house remote sensing extraction model was improved by optimizing convolution kernels based on the geometric features of houses. In the shallow network (levels 1-2), a heterogeneous convolution kernel combination method was used for optimization. A 3×3 standard convolution kernel was used in parallel with 1×3 or 3×1 rectangular kernels to extract features, enhancing sensitivity to house edges and extracting detailed textures. In the deep network (levels 3-4), a large receptive field convolution method was used for optimization, using 5×5 and 7×7 convolution kernels to expand the receptive field, capture the global distribution pattern of houses, and improve the accuracy of house extraction.

[0092] In this embodiment, the training goal of the intelligent house extraction model of the present invention is to minimize the difference between the predicted house segmentation results and the actual house labels. Therefore, the cross entropy loss function and the Dice coefficient are usually used as loss functions. The traditional cross entropy loss function easily causes the model to be biased towards the majority class in the case of class imbalance, thereby affecting the extraction accuracy of the house; the Dice loss function is used to measure the overlap between the predicted segmentation results and the actual label. It is particularly suitable for segmentation tasks with class imbalance, but it has low sensitivity to boundary areas. To this end, the intelligent house extraction model proposes a classification-boundary collaborative optimization loss function, which is specifically designed as follows: Class-Sensitive Weighted Cross Entropy (CS-WCE):

[0093] in, , =10 −5 , the attention of the minority class (damaged houses) is increased through inverse frequency weighting.

[0094] Boundary Enhanced Dice Loss (BE-Dice):

[0095] in, =0.3 is the boundary sensitivity coefficient. By introducing the boundary distance penalty term, the model's fitting accuracy for the house outline is improved.

[0096] The total losses are as follows:

[0097] in, represents the loss function of the house intelligent extraction model, represents the category-sensitive weighted cross loss function, represents the boundary loss function, N Indicates the total number of pixels, i represents the pixel index, represents the category weight, represents the actual house type label, represents the predicted house type label, represents the category cThe sample ratio of represents the minimum value introduced to avoid the denominator being zero, such as, represents the boundary sensitivity coefficient, p represents the probability that the model predicts that the pixel belongs to the house boundary, Y represents the actual house boundary mask, represents the partial derivative, and represents the predicted pixel p The probability of belonging to a certain category of housing, representing the pixel p The true label, Represents the weight coefficient, which decays exponentially. , achieving the staged optimization goal of focusing on classification accuracy in the early stage of training and strengthening boundary matching in the later stage.

[0098] In this embodiment, the house intelligent extraction model is trained as follows: The constructed U-net model is trained using a high-resolution remote sensing house sample training set data of multiple types of houses. The Adam optimization algorithm is used to optimize the model parameters. First, the first-order moment estimate is calculated. and second-order moment estimates :

[0099]

[0100] in, and All represent hyperparameters, Set to 0.9 to balance the stability of gradient updates, Set to 0.999 to accelerate model convergence. Represents the gradient.

[0101] Then the first-order moment estimate and the second-order moment estimate are corrected:

[0102]

[0103] in, and denote the modified first-order moment estimate and second-order moment estimate, respectively. and They represent the first-order and second-order moment decay rates after training iteration t steps respectively.

[0104] Finally update the model parameters :

[0105] in, Represents the updated U-net model parameters, Represents the learning rate, usually in the range of 10 -5 to 10 -3 , is a minimum value (such as ), used to prevent the denominator from being 0.

[0106] During the training process, the model training was optimized and improved. First, the data enhancement strategy was optimized based on the characteristics of house extraction. Specifically, it included: random rotation, angle range of −10 ∘ to 10 ∘ to increase the model's robustness to the rotation invariance of the house; random flipping, including horizontal flipping and vertical flipping, to enhance the model's ability to recognize the symmetry of the house; random scaling, with a scaling ratio between 0.8 and 1.2, to increase the model's ability to extract houses of different scales; brightness and contrast adjustment, random brightness and contrast adjustments are performed on the input image to enhance the model's adaptability to lighting changes and improve the model's generalization ability.

[0107] Secondly, an early stopping mechanism is used. Training is stopped when the accuracy on the validation set stops improving for 10 consecutive epochs. Alternatively, when the loss function value on the validation set decreases by less than a certain threshold (such as 0.001), the model is considered to have converged and training is stopped to prevent overfitting. Furthermore, the model weight parameters are regularly saved so that the best-performing model can be selected for subsequent evaluation and application after training.

[0108] The trained intelligent house extraction model is applied, combined with high-resolution satellite remote sensing images before and after the rainstorm and flood disaster with consistent temporal and spatial benchmarks, to extract the house boundary range before and after the flood disaster.

[0109] S5. Use the trained intelligent house extraction model to determine the damaged houses after the extreme rainstorm and flood disaster, and complete the remote sensing extraction of the damaged houses. Specifically: Obtain high-resolution remote sensing image data before and after extreme rainstorm and flood disasters in the same area; input the high-resolution remote sensing image data before the extreme rainstorm and flood disaster into the trained house intelligent extraction model to obtain the bounding box coordinates of each house and housing category information, as well as statistics on the number of houses in the area before the disaster ; Input high-resolution remote sensing image data after extreme rainstorm and flood disasters into the trained house intelligent extraction model to obtain the bounding box coordinates of each house and housing category information, as well as statistics on the number of houses in the area after the disaster For each house extracted before the extreme rainstorm and flood disaster, determine its positional relationship with the house extracted after the extreme rainstorm and flood disaster, and measure the overlap of the two bounding boxes; if the overlap is greater than the preset threshold, the house has not changed; if the overlap is less than the preset threshold, the house has been displaced and / or damaged; according to the number of houses before the disaster and the number of houses after the disaster , calculate the rate of change of the number of houses; according to the rate of change of the number of houses, determine the degree of damage to the houses caused by extreme rainstorms and floods by judging the increase or decrease of house data; for each house, calculate the house area before and after the extreme rainstorms and floods; according to the house area before and after the extreme rainstorms and floods, calculate the rate of change of the house area; according to the rate of change of the house area, determine the change of the house area, and complete the remote sensing extraction of the damaged houses.

[0110] In this embodiment, high-resolution satellite remote sensing images are obtained before and after an extreme rainstorm and flood disaster occurs in the same area. After the images are operated as described above, the house extraction results obtained by inputting the images into the model before and after the flood disaster are detected for changes in the number, location, and area of ​​houses, thereby obtaining statistics on house damage caused by the rainstorm and flood disaster.

[0111] (1) Model training and house extraction (preliminary work) Use the image input model before the flood disaster to obtain the house extraction boundary results and record the bounding box coordinates of each house and its category information (all housing categories), the number of houses in the area before the disaster was .

[0112] Similarly, we use the images after the flood disaster to train the deep model and extract the houses to obtain the house bounding box coordinates. and its category information, the number of houses in the area before the disaster was .

[0113] (2) Detection of changes in the location of damaged houses For each house extracted before the flood, calculate its positional relationship with the house extracted after the flood. The Intersection over Union (IoU) metric is used to measure the degree of overlap between the two bounding boxes. The calculation formula is:

[0114] Set an IoU threshold (Take 0.5), if the IoU between a house before the flood and a house after the flood is greater than the threshold , it is considered that the location of the house has basically not changed; if IoU is less than , it indicates that the house may have been displaced or damaged.

[0115] (3) Detection of changes in the number of damaged houses Count the number of houses extracted before and after the flood disaster and Calculate the rate of change of the number of houses , the formula is:

[0116] According to the rate of change of the number of houses The value of is used to judge the increase or decrease in the number of houses. A positive value indicates an increase in the number of houses, and a negative value indicates a decrease in the number of houses. The absolute value reflects the degree of change in quantity and can help determine the extent of damage to houses caused by floods.

[0117] (4) Detection of changes in damaged housing area based on housing borders For each house, the area is calculated based on its bounding box coordinates. The calculation formula is: , housing area after flood disaster The calculation formula is: .

[0118] Calculate the rate of change of house area :

[0119] pass The value of is used to judge the change in the house area. A positive value indicates an increase in the area, which may be due to factors such as image error or changes in the surrounding environment; a negative value indicates a decrease in the area, which may be caused by damage to the house. The absolute value reflects the degree of change in the house area and helps to further analyze the damage to the house.

[0120] The above-mentioned house change detection method before and after flood disasters based on deep learning model extraction of house results can effectively utilize satellite remote sensing image data to quickly and accurately detect changes in the location, number and area of ​​houses before and after flood disasters, providing an important basis for flood disaster assessment and post-disaster reconstruction.

[0121] Example 2 like Figure 2As shown, the present invention provides a remote sensing extraction system for houses damaged by extreme rainstorm and flood disasters, which is applied to any of the remote sensing extraction methods for houses damaged by extreme rainstorm and flood disasters described in Example 1, comprising: The first processing module is used to use multi-source high-resolution remote sensing data to obtain remote sensing image data of key target houses with consistent temporal and spatial references before and after the disaster; The second processing module is used to construct a multimodal remote sensing feature set of key house targets that integrates spectral and texture features based on the acquired remote sensing image data; The third processing module is used to mark high-resolution remote sensing house samples based on geometric structures according to different house types, wherein the high-resolution remote sensing house sample marking is performed to calculate the similarity probability of multiple types of houses; The fourth processing module is used to train and build a deep learning-based intelligent house extraction model by using labeled high-resolution remote sensing house samples and combining them with a multimodal remote sensing feature set of key house targets; The fifth processing module is used to use the trained house intelligent extraction model to determine the condition of damaged houses after extreme rainstorms and floods, and complete the remote sensing extraction of damaged houses.

[0122] like Figure 2 The remote sensing extraction system provided in the illustrated embodiment can execute the technical solution shown in the remote sensing extraction method in the above-mentioned method embodiment 1. Its implementation principle and beneficial effects are similar and will not be repeated here.

[0123] In this embodiment, the present application can divide the functional units according to the remote sensing extraction method in Example 1. For example, each function can be divided into different functional units, or two or more functions can be integrated into a single processing unit. The above-mentioned integrated unit can be implemented in the form of hardware or software functional units. It should be noted that the division of units in the present invention is schematic and is only a logical division. In actual implementation, other division methods may be used.

[0124] In this embodiment, the remote sensing extraction system includes hardware structures and / or software modules for executing various functions in order to realize the principles and beneficial effects of the remote sensing extraction method. It should be readily apparent to those skilled in the art that, in combination with the various schematic units and algorithm steps described in the embodiments disclosed herein, the present invention can be implemented in the form of hardware and / or a combination of hardware and computer software. Whether a function is executed in a hardware- or computer software-driven manner depends on the specific application and design constraints of the technical solution. Different methods can be used for each specific application to implement the described function, but such implementation should not be considered to be beyond the scope of this application.

Claims

1. A remote sensing extraction method for houses damaged by extreme rainstorm and flood disasters, characterized in that: The following steps are involved: Utilize multi-source high-resolution remote sensing data to obtain remote sensing image data of key target houses with consistent temporal and spatial benchmarks before and after the disaster; Based on the acquired remote sensing image data, a multimodal remote sensing feature set of key house targets is constructed by integrating spectral and texture features; For different house types, high-resolution remote sensing house samples are labeled based on geometric structures. High-resolution remote sensing house sample labeling is performed to calculate the similarity probability of multiple house types. Using labeled high-resolution remote sensing house samples and combining them with a multimodal remote sensing feature set of key house targets, we train and build a deep learning-based intelligent house extraction model. The trained intelligent house extraction model is used to determine the condition of damaged houses after extreme rainstorms and floods, and complete the remote sensing extraction of damaged houses.

2. The remote sensing extraction method for houses damaged by extreme rainstorm and flood disasters according to claim 1 is characterized in that: The multimodal remote sensing feature set of key housing targets that fuses spectral and texture features is constructed as follows: Based on the acquired remote sensing image data, determine the pixel point of each remote sensing image The spectrum vector ; Calculate the normalized vegetation separately and normalized architecture spectral characteristics; The spectrum vector and normalized vegetation and normalized architecture The spectral feature combination is the house area spectral feature vector , complete the extraction of remote sensing spectral characteristics of the housing area; Extract image texture features based on acquired remote sensing image data , complete the extraction of texture features of the housing area; Based on the extracted spectral characteristics of the house area and the texture characteristics of the house area, for each candidate house area , extract the spectral feature vectors of the house area respectively and texture feature vector ,construct regional level features through feature splicing; Based on the regional-level features constructed by feature splicing, feature priority is performed to obtain the prioritized regional-level features; Prioritized regional-level features Forming a structured feature set ; Based on structured feature sets , calculate the information entropy of each feature ; Based on adaptive adjustment factor and feature information entropy , calculate the weight of each feature ; According to the weight of each feature Sorting, taking the feature with the largest weight as the house feature, completing the construction of the multimodal remote sensing feature set of the key targets of the house.

3. The remote sensing extraction method for houses damaged by extreme rainstorm and flood disasters according to claim 2 is characterized in that: The calculation of similarity probabilities of multiple types of houses is specifically as follows: Constructing a hierarchical geometric feature space, where the geometric features include a first-level feature representing the house area, a second-level feature representing the house's aspect ratio, and a third-level feature representing the house's shape index; Based on the constructed hierarchical geometric feature space, the house type discrimination rules are constructed; Based on the constructed house type discrimination rules, a Gaussian mixture model is constructed for each house type, and the Gaussian mixture model is trained to model the geometric prototypes of multiple house types; Based on the modeling results of multi-type house geometric prototypes, each candidate area in the new remote sensing image is calculated , and get each candidate region The eigenvector of the posterior probability of belonging to each class; According to the maximum a posteriori probability principle, the house type label is assigned to obtain the predicted house type label ; Get the real house type label , and label the actual house type and the predicted house type labels Make a comparison, use the comparison results to modify the Gaussian mixture model, and complete the calculation of the similarity probability of multiple types of houses.

4. The remote sensing extraction method for houses damaged by extreme rainstorm and flood disasters according to claim 3 is characterized in that: The expression of the house type discrimination rule is as follows: in, Indicates the house type identification rule, represents the shape index of the house, The ratio of the length to the width of the house. Indicates the area of ​​the house. The value range of the area is between 50 square meters and 5000 square meters. Indicates the length of the house. Indicates the width of the house. Indicates the perimeter of the house; The predicted house type label The expression is as follows: in, Representing a collection The specific housing type in C Represents a set of house types, including all house categories to be classified, and argmax represents the maximum posterior probability. value, Represents the posterior probability, that is, the feature vector of the known candidate region Under the condition of The probability of Represents the prior probability, that is, the house type The prior distribution probability in the entire data set, Represents the likelihood probability, that is, when the house type is When the eigenvector The probability of Representing a collection C Except the current category c Any housing type except Indicates that when the house type is When the eigenvector The probability of Indicates the type of house c The prior probability of ′.

5. The remote sensing extraction method for houses damaged by extreme rainstorm and flood disasters according to claim 3 is characterized in that: The training and construction of the deep learning-based intelligent house extraction model is specifically as follows: Divide the labeled high-scoring remote sensing house samples into a training set and a validation set; Combined with the multimodal remote sensing feature set of key housing targets, the training set is used to build a housing intelligent extraction model based on deep learning, and the mapping relationship between image features and labels is learned for training. Probabilistic extraction of housing; The validation set is used to tune the parameters of the trained house intelligent extraction model to complete the construction of the house intelligent extraction model.

6. The remote sensing extraction method for houses damaged by extreme rainstorm and flood disasters according to claim 5 is characterized in that: The loss function of the intelligent house extraction model is expressed as follows: in, represents the loss function of the house intelligent extraction model, represents the weight coefficient, represents the category-sensitive weighted cross loss function, represents the boundary loss function, N Indicates the total number of pixels, i represents the pixel index, represents the category weight, represents the actual house type label, represents the predicted house type label, Indicates the type of house The sample proportion of Indicates the minimum value introduced to avoid the denominator being zero, represents the boundary sensitivity coefficient, p represents the probability that the predicted pixel belongs to the house boundary, Y represents the true house boundary mask, represents the partial derivative, Represents the predicted pixel p The probability of belonging to a certain category of housing, Represents pixels p The true label.

7. The remote sensing extraction method for houses damaged by extreme rainstorm and flood disasters according to claim 1 is characterized in that: The determination of damaged houses after extreme rainstorm and flood disasters is as follows: Obtain high-resolution remote sensing image data before and after extreme rainstorm and flood disasters occur in the same area; Input high-resolution remote sensing image data before the extreme rainstorm and flood disaster into the trained house intelligent extraction model to obtain the bounding box coordinates of each house and housing category information, as well as statistics on the number of houses in the area before the disaster ; Input high-resolution remote sensing image data after extreme rainstorm and flood disasters into the trained house intelligent extraction model to obtain the bounding box coordinates of each house and housing category information, as well as statistics on the number of houses in the area after the disaster ; For each house extracted before the extreme rainstorm and flood disaster, its positional relationship with the house extracted after the extreme rainstorm and flood disaster is determined, and the overlap of the two bounding boxes is measured. If the overlap is greater than a preset threshold, the house has not changed; if the overlap is less than the preset threshold, the house has been displaced and / or damaged. According to the number of houses before the disaster and the number of houses after the disaster , calculate the rate of change of the number of houses; Based on the rate of change in the number of houses, the extent of damage to houses caused by extreme rainstorms and floods can be determined by judging the increase or decrease in house data; For each house, the house area before and after the extreme rainstorm flooding disaster was calculated; Calculate the rate of change of housing area based on the housing area before and after the extreme rainstorm and flood disaster; The change of house area is determined according to the rate of change of house area, and the remote sensing extraction of damaged houses is completed.

8. A remote sensing extraction system for houses damaged by extreme rainstorm and flood disasters, applied to the remote sensing extraction method for houses damaged by extreme rainstorm and flood disasters according to any one of claims 1 to 7, characterized in that: include: The first processing module is used to use multi-source high-resolution remote sensing data to obtain remote sensing image data of key target houses with consistent temporal and spatial references before and after the disaster; The second processing module is used to construct a multimodal remote sensing feature set of key house targets that integrates spectral and texture features based on the acquired remote sensing image data; The third processing module is used to mark high-resolution remote sensing house samples based on geometric structures according to different house types, wherein the high-resolution remote sensing house sample marking is performed to calculate the similarity probability of multiple types of houses; The fourth processing module is used to train and build a deep learning-based intelligent house extraction model by using labeled high-resolution remote sensing house samples and combining them with a multimodal remote sensing feature set of key house targets; The fifth processing module is used to use the trained house intelligent extraction model to determine the condition of damaged houses after extreme rainstorms and floods, and complete the remote sensing extraction of damaged houses.

Images