An extreme rainstorm flood disaster damaged house remote sensing extraction method and system

Abstract

The application provides a kind of extreme rainstorm flood disaster damaged house remote sensing extraction method and system, it is related to the field of intelligent processing of remote sensing image, the method includes using multi-source high-resolution remote sensing data, obtains the remote sensing image data of the consistent space-time reference before and after the key target house disaster;Build the house key target multi-modal remote sensing feature set of fusion spectrum and texture characteristics, realize the dimension reduction and fusion of multiple characteristics using principal component method and dynamic weight distribution method;For different house types, a high-resolution remote sensing house sample marking method is proposed considering geometric structure, a Gaussian mixture model is used to realize the probability measure of house samples, and a house element sample library is formed;Using an improved deep learning model, the characteristics are trained, and a house remote sensing intelligent extraction model is constructed;Combined with the training model and multi-modal remote sensing data, the pre-disaster and post-disaster house key targets are automatically extracted, and the change analysis detection method is used to determine the damaged house target.

Description

Technical Field

[0001] This invention belongs to the field of intelligent remote sensing image processing technology, and in particular relates to a method and system for remote sensing extraction of houses damaged by extreme rainstorms and floods. Background Technology

[0002] Traditional monitoring of flood-damaged houses relies primarily on manual on-site verification. However, these surveys have limited scope and cannot comprehensively and accurately reflect the actual situation of damaged houses throughout the entire disaster area. Furthermore, traditional monitoring methods suffer from timeliness in data acquisition, making it difficult to meet the timeliness requirements of on-site emergency response to extreme rainstorms and floods.

[0003] Satellite remote sensing technology provides a new approach for monitoring houses damaged by extreme rainstorms and floods. It can acquire large-area, periodic surface images before and after disasters, visually displaying changes in houses in the affected area. However, when extracting information on water bodies and damaged houses, the existence of phenomena such as different spectra for the same object and the same spectra for different objects, as well as interference from complex scenes in high-resolution images, results in low accuracy in extracting damaged houses using traditional methods, making it difficult to meet the accuracy requirements for emergency response to extreme rainstorms and floods.

[0004] 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 images, effectively overcoming the subjectivity and limitations of manual feature extraction in traditional methods. Deep learning has been widely applied in the field of remote sensing image recognition and classification, demonstrating unique advantages in processing complex high-resolution remote sensing image data.

[0005] To address the challenge of rapidly extracting damaged houses from extreme rainstorms and floods, deep learning is applied to remote sensing extraction of such damage. A multimodal remote sensing dataset is constructed, remote sensing features of houses are extracted, and a typical sample library of house elements is built. Combined with deep learning methods, this approach enables the extraction of houses before and after the disaster, and the detection of changes in damaged houses. It can quickly and accurately identify the location of damaged houses from massive amounts of remote sensing imagery, providing crucial data support for disaster emergency response, loss assessment, and post-disaster reconstruction. This significantly improves the scientific rigor and effectiveness of flood disaster management and has broad application prospects. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a remote sensing extraction method and system for houses damaged by extreme rainstorms and floods, solving the problem of the difficulty in quickly extracting houses damaged by extreme rainstorms and floods.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for remote sensing extraction of houses damaged by extreme rainstorms and floods, comprising the following steps:

[0008] By utilizing multi-source high-resolution remote sensing data, we can obtain remote sensing image data of key target buildings with consistent spatiotemporal references before and after the disaster.

[0009] Based on the acquired remote sensing image data, a multimodal remote sensing feature set of key building targets was constructed by fusing spectral and texture features;

[0010] For different housing types, high-resolution remote sensing housing samples that take into account geometric structure are labeled. In this process, the high-resolution remote sensing housing samples are labeled to calculate the similarity probability of multiple housing types.

[0011] Using labeled high-resolution remote sensing samples of houses, and combining them with multimodal remote sensing feature sets of key house targets, a deep learning-based intelligent house extraction model was trained and constructed.

[0012] Using a trained intelligent housing extraction model, the extent of damage to houses after extreme rainstorms and floods is determined, and remote sensing data of the damaged houses is extracted.

[0013] Furthermore, the construction of the multimodal remote sensing feature set for key building targets, which integrates spectral and texture features, specifically involves:

[0014] Based on the acquired remote sensing image data, determine each remote sensing image pixel. spectral vector ;

[0015] Calculate normalized vegetation separately and standardized architecture Spectral characteristics;

[0016] spectral vector and normalized vegetation and standardized architecture The spectral features are combined to form the spectral feature vector of the building area. This completes the remote sensing extraction of the spectral features of the building area;

[0017] Based on the acquired remote sensing image data, image texture features are extracted. This completes the extraction of texture features from the building area;

[0018] Based on the extracted spectral and texture features of the housing areas from remote sensing, for each candidate housing area... Extract the spectral feature vectors of the building areas respectively. and texture feature vector Regional features are constructed by feature concatenation;

[0019] Based on the regional features constructed by feature concatenation, feature priority is applied to obtain the preferred regional features;

[0020] Based on priority-based regional features Forming a structured feature set ;

[0021] Based on structured feature sets Calculate the entropy of each feature. ;

[0022] Based on adaptive adjustment factor and feature information entropy Calculate the weights of each feature. ;

[0023] According to the weight of each feature The features with the highest weights are sorted and used as the building features to complete the construction of a multimodal remote sensing feature set for key building targets.

[0024] Furthermore, the calculation of the similarity probability of multiple types of houses specifically involves:

[0025] Construct a hierarchical geometric feature space, where geometric features include primary features representing the area of ​​a house, secondary features representing the aspect ratio of a house, and tertiary features representing the shape index of a house.

[0026] Based on the constructed hierarchical geometric feature space, a house type discrimination rule is constructed;

[0027] Based on the constructed house type discrimination rules, a Gaussian mixture model is built for each house type, and the Gaussian mixture model is trained to model the geometric prototypes of multiple house types.

[0028] Based on the modeling results of geometric prototypes of various building types, the calculation is performed on each candidate region in the new remote sensing image. Each candidate region is obtained. eigenvectors The posterior probability of belonging to each class;

[0029] Based on the maximum a posteriori probability principle, house type labels are assigned to obtain the predicted house type labels. ;

[0030] Get real house type tags and label the actual house type With predicted house type labels By comparing the results, the Gaussian mixture model is corrected, and the similarity probability of multiple types of houses is calculated.

[0031] Furthermore, the expression for the housing type discrimination rule is as follows:

[0032]

[0033]

[0034]

[0035]

[0036] in, This indicates the rules for determining housing type. Indicating the shape index of the house. Indicates the length-to-width ratio of the house. This indicates the area of ​​the house, with values ​​ranging from 50 square meters to 5000 square meters. Indicates the length of the house. Indicates the width of the house. Indicates the perimeter of the house;

[0037] The predicted house type label The expression is as follows:

[0038]

[0039]

[0040] in, Represents a set The specific types of houses in the text, C Let represent the set of house types, containing all house categories to be classified, and argmax represent the option that maximizes the posterior probability. value, This represents the posterior probability, i.e., the probability of finding a known candidate region's feature vector. Under these conditions, this area belongs to the housing type. The probability, This represents the prior probability, i.e., the type of housing. The prior distribution probability in the entire dataset. This represents the likelihood probability, i.e., when the housing type is... When the eigenvector is observed The probability, Represents a set C Except for the current category c Any type of house outside, Indicates when the housing type is When the eigenvector is observed The probability, Indicates housing type c The prior probability of ′.

[0041] Furthermore, the training and construction of the deep learning-based intelligent house extraction model specifically involves:

[0042] The labeled high-resolution remote sensing house samples were divided into training and validation sets.

[0043] By combining multimodal remote sensing feature sets of key building targets, a deep learning-based intelligent building extraction model is constructed using the training set. This model is trained by learning the mapping relationship between image features and labels, and then based on the predicted building type labels... Probabilistic house extraction;

[0044] The parameters of the trained intelligent house extraction model are fine-tuned using the validation set, thus completing the construction of the intelligent house extraction model.

[0045] Furthermore, the expression for the loss function of the intelligent house extraction model is as follows:

[0046]

[0047]

[0048]

[0049]

[0050] in, This represents the loss function of the smart house extraction model. Indicates the weighting coefficient. This represents the category-sensitive weighted cross loss function. Represents the boundary loss function. N Indicates the total number of pixels. i Indicates pixel index, Indicates category weight, Indicates the actual housing type label, Indicates the predicted house type label, Indicates housing type The sample proportion, This represents a local minimum introduced to avoid the denominator being zero. Represents the boundary sensitivity coefficient. p This represents the probability that the predicted pixel belongs to the house boundary. Y Represents the actual house boundary mask. Represents partial derivatives, Represents the predicted pixels p The probability of belonging to a certain category of housing. Represents pixels p The true label.

[0051] Furthermore, the automatic extraction of the extent of housing damage before and after extreme rainstorms and floods, and the determination of changes in damaged housing, specifically involves:

[0052] High-resolution remote sensing image data of the same region before and after extreme rainstorm and flood disasters were acquired;

[0053] High-resolution remote sensing imagery data from before extreme rainstorms and floods are input into a trained intelligent house extraction model to obtain the bounding box coordinates of each house. Information on housing types and the number of houses in the area before the disaster. ;

[0054] High-resolution remote sensing imagery data following extreme rainstorms and floods is input into a trained intelligent house extraction model to obtain the bounding box coordinates of each house. Information on housing types and the number of houses in the area affected by the disaster. ;

[0055] For each house extracted before the extreme rainstorm and flood disaster, determine its positional relationship with the houses extracted after the extreme rainstorm and flood disaster, and measure the overlap of the two bounding boxes; 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.

[0056] Based on the number of houses before the disaster and the number of houses after the disaster Calculate the rate of change in the number of houses;

[0057] 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;

[0058] For each house, calculate the house area before and after the extreme rainstorm and flood disaster;

[0059] Calculate the rate of change in house area based on the house area before and after extreme rainstorm and flood disasters;

[0060] Based on the rate of change of building area, the changes in building area are determined, and the remote sensing data of damaged buildings is extracted.

[0061] This invention also provides a remote sensing extraction system for houses damaged by extreme rainstorms and floods, comprising:

[0062] The first processing module is used to acquire remote sensing image data of key target houses with consistent spatiotemporal reference before and after the disaster using multi-source high-resolution remote sensing data.

[0063] The second processing module is used to construct a multimodal remote sensing feature set of key building targets by fusing spectral and texture features based on the acquired remote sensing image data;

[0064] The third processing module is used to label high-resolution remote sensing house samples based on geometric structure for different house types. The labeling of high-resolution remote sensing house samples is used to calculate the similarity probability of multiple house types.

[0065] The fourth processing module is used to train and build a deep learning-based intelligent house extraction model by utilizing labeled high-resolution remote sensing house samples and combining them with multimodal remote sensing feature sets of key house targets.

[0066] The fifth processing module is used to determine the condition of damaged houses after extreme rainstorms and floods by using the trained intelligent house extraction model, and to complete the remote sensing extraction of damaged houses.

[0067] The beneficial effects of this invention are:

[0068] (1) A remote sensing extraction method for houses damaged by rainstorms and floods was constructed, breaking through the technical bottleneck that it is difficult to quickly extract houses damaged by extreme rainstorms and floods; (2) A multimodal remote sensing feature set of key house targets with spectral texture features was constructed to realize the 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 developed to realize the calculation of similarity probability of multiple types of houses; (4) A house remote sensing intelligent extraction model based on deep learning was constructed to realize the automatic extraction of the house range before and after the disaster and the determination of the changes in damaged houses. Attached Figure Description

[0069] Figure 1 This is a flowchart of the method of the present invention.

[0070] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0071] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0072] Example 1

[0073] like Figure 1 As shown, this invention provides a method for remote sensing extraction of houses damaged by extreme rainstorms and floods, the implementation method of which is as follows:

[0074] S1. Utilize multi-source high-resolution remote sensing data to obtain remote sensing image data of key target buildings with consistent spatiotemporal references before and after the disaster.

[0075] In this embodiment, based on the emergency application requirements of extreme rainstorms and floods, high-resolution satellite remote sensing data is acquired. This data includes optical remote sensing data with a spatial resolution better than 2 meters, including data from Gaofen-1, Gaofen-2, and Gaofen-6. High-quality cloud-free and snow-free remote sensing data is selected, and geometric correction, atmospheric correction, and cloud masking are performed to obtain spatiotemporally consistent remote sensing imagery of key target buildings before and after the disaster.

[0076] S2. Based on the acquired remote sensing image data, construct a multimodal remote sensing feature set for key building targets that integrates spectral and texture features, specifically as follows:

[0077] Based on the acquired remote sensing image data, determine each remote sensing image pixel. spectral vector ; Calculate normalized vegetation separately and standardized architecture Spectral characteristics; spectral vectors and normalized vegetation and standardized architecture The spectral features are combined to form the spectral feature vector of the building area. The system extracts the spectral features of the building area using remote sensing; based on the acquired remote sensing image data, it extracts image texture features. This process extracts the texture features of the housing areas. The extracted spectral and texture features include those of the housing areas themselves, as well as other land cover types such as farmland. Based on these extracted spectral and texture features, for each candidate housing area... Extract the spectral feature vectors of the building areas respectively. and texture feature vector The spectral feature vector extracted here and texture feature vector Includes only residential areas, excluding other land cover types mentioned above. Regional features are constructed through feature stitching. Based on these regional features, feature priority is applied to obtain prioritized regional features. Based on these prioritized regional features... Forming a structured feature set Based on structured feature sets Calculate the entropy of each feature. Based on adaptive adjustment factor and feature information entropy Calculate the weights of each feature. According to the weight of each feature The features with the highest weights are sorted and used as the building features to complete the construction of a multimodal remote sensing feature set for key building targets.

[0078] In this embodiment, the remote sensing extraction of the building's spectral features is as follows:

[0079] Each pixel in the remote sensing image The spectral vector is represented as:

[0080]

[0081] in, Represents the spectral vector. Indicates the first c Each band at the pixel point ( x , y Spectral intensity value at ) This represents a set of bands.

[0082] To enhance discriminative power, normalized vegetation cover was calculated for each house. and standardized architecture Spectral characteristics:

[0083]

[0084]

[0085] in, Represents pixels The near-infrared spectral intensity value at that location, Represents pixels The red light band spectral intensity value at that location, Represents pixels The shortwave infrared spectral intensity value at that location.

[0086] Furthermore, the spectral vector, normalized vegetation, and normalized building features are combined to form the building spectral feature vector. ,in, k This indicates the number of newly added spectral features.

[0087] In this embodiment, the extraction of house texture features uses a gray-level co-occurrence matrix and a Gabor filter to extract image texture features: first, the image is converted into a grayscale image. Calculate the co-occurrence matrix for different directions and distances. P Then count the energy separately. Contrast and correlation feature:

[0088]

[0089]

[0090]

[0091] in, This represents the gray values ​​in the Gray-Level Co-occurrence Matrix (GLCM). and The normalized probability of occurrence of pixel pairs. Represents the grayscale values ​​of an image. In the normalized image, Indicates pure black. It represents pure white. The mean values ​​in the row direction and column direction of the gray-level co-occurrence matrix. This represents the standard deviation in the row direction and the standard deviation in the column direction of the gray-level co-occurrence matrix.

[0092] comprehensive n Statistical vectors of texture features in each direction are used to construct texture feature vectors. .

[0093] Gabor filters are used to provide texture features for remote sensing images. The Gabor filter function is:

[0094]

[0095]

[0096]

[0097] in, Represents the filter function. Represents the spatial coordinates of the image. This indicates that the original coordinates are adjusted according to the direction angle. The new coordinates after rotation The standard deviation of the Gaussian envelope is used to control the bandwidth of the Gaussian window. A larger value means a wider filter coverage area, but a lower resolution. The wavelength representing the sinusoidal component is used to control the period of the sinusoidal component in the filter and determines the scale of the texture. This indicates a phase shift, used to adjust the initial phase of the sinusoidal component, affecting the symmetry of the texture. This represents the direction angle of the filter, used to define the detection direction of the filter.

[0098] In multiscale and parameter angle Take one set of images from each of the above, and extract the mean and variance as texture feature statistical vectors. Final texture features for:

[0099]

[0100] in, This represents the spatial wavelength of the sinusoidal component in the filter. Indicates the first p One spatial wavelength, n This indicates the feature dimension extracted by the Gabor filter. m This indicates the dimension of the texture features extracted by GLCM. This represents the texture feature vector extracted from the Gray-Level Co-occurrence Matrix (GLCM). This represents the multi-scale, multi-directional texture feature vector extracted by the Gabor filter. The total dimension of the final texture feature vector is . .

[0101] In this embodiment, a dynamic weight allocation method is used to achieve multi-feature optimization:

[0102] For each candidate housing area Extract the spectral feature vectors of the building areas respectively. and texture feature vector Regional features are constructed by feature concatenation. :

[0103]

[0104] in, This indicates a feature concatenation operation. Indicates housing area medium pixel Spectral feature set, Housing area medium pixel The texture feature set.

[0105] Based on the regional features constructed by feature concatenation, a feature-prioritized approach is used to obtain the preferred regional features. For example, prioritizing based on weights:

[0106]

[0107] in, This indicates the total number of candidate housing areas. Indicates the first k Each region Represents the horizontal and vertical coordinates of a pixel.

[0108] Forming a structured feature set :

[0109]

[0110] in, This represents the region-level feature vector of the Mth region.

[0111] Further calculate the entropy of each feature. , The calculation formula is as follows:

[0112]

[0113] in, Indicates the number of features. Indicates the feature number, No. The probability of a feature taking a value under a certain probability distribution.

[0114] The adaptive adjustment factor can dynamically adjust the weight calculation based on the degree of dispersion of the features; the greater the dispersion (the greater the variance), the better. The larger the value, the more it will affect the weight calculation. Adaptive adjustment factor. The calculation formula is as follows:

[0115]

[0116]

[0117] in, Indicates the number of features. Indicates the first The variance values ​​of each feature, Representation of features The number of samples, Indicates the first The first feature One value, Indicates the first The average value of each feature.

[0118] Calculate the weights of each feature. :

[0119] in, Indicates the first j Information entropy of each feature j Indicates all The first of the features j One characteristic.

[0120] The features are ranked according to their weights, with the features having the highest priority being designated as the house features. , This represents the Nth house feature.

[0121] S3. Label high-resolution remote sensing housing samples that take into account geometric structure for different housing types. Specifically, labeling high-resolution remote sensing housing samples is used to calculate the similarity probability of multiple housing types.

[0122] A hierarchical geometric feature space is constructed, where geometric features include primary features representing building area, secondary features representing building aspect ratio, and tertiary features representing building shape index. Based on the constructed hierarchical geometric feature space, building type discrimination rules are developed. Based on the constructed building type discrimination rules, a Gaussian mixture model is constructed for each building type, and the Gaussian mixture model is trained to model the geometric prototypes of multiple building types. Based on the modeling results of the geometric prototypes of multiple building types, each candidate region in the new remote sensing image is calculated. Each candidate region is obtained. eigenvectors The posterior probability of belonging to each class; based on the maximum posterior probability principle, assign house type labels to obtain the predicted house type labels. Get real house type labels and label the actual house type With predicted house type labels By comparing the results, the Gaussian mixture model is corrected, and the similarity probability of multiple types of houses is calculated.

[0123] In this embodiment, the geometric prototype modeling of multiple types of houses is as follows:

[0124] Establishing a hierarchical geometric feature space includes:

[0125] Primary features ,in, This indicates the area of ​​the house, with values ​​ranging from 50 square meters to 5000 square meters.

[0126] Secondary features: ,in, Indicates the length-to-width ratio of the house. Indicates the length of the house. This indicates the width of the house. The aspect ratio ranges from 0 to 2.5. This feature reflects whether the house's shape is more square or rectangular, and can help determine the layout and space utilization of the house.

[0127] Third-level features: ,in, Indicating the shape index of the house. Indicates the perimeter of the house. The shape index represents the area of ​​a house. It further describes the shape characteristics of a house by analyzing the relationship between perimeter and area, providing a more detailed description of the house's geometric features.

[0128] Construction type identification rules:

[0129]

[0130] Let the set of house types be For each class, a Gaussian Mixture Model (GMM) is established:

[0131]

[0132] in, Indicates when the housing type is When the eigenvector is observed The probability, Indicates the first j Gaussian components Indicates the number of Gaussian components. Indicates the first The mixed weights of the components, Represents the Gaussian distribution function. Represents the eigenvector. and Let represent the mean and covariance matrices, respectively.

[0133] For each Gaussian mixture model constructed, the geometric feature data of training samples belonging to that house type are used for training, and the EM algorithm (Expectation-Maximization algorithm) is employed. The EM algorithm is an iterative algorithm used to estimate parameters in models containing latent variables. In Gaussian mixture models, the EM algorithm is used to estimate the model parameters. , , Various types of houses were obtained through training. Gaussian mixture model This allows us to determine the distribution boundaries of various housing characteristics, thereby enabling more accurate classification and identification of housing types.

[0134] In this embodiment, the region type determination and label generation are as follows:

[0135] Each candidate region in the new image is calculated using Bayes' theorem. Calculate its eigenvectors Posterior probability of belonging to each class:

[0136]

[0137] House type labels are assigned based on the maximum posterior probability principle:

[0138]

[0139] in, Represents a set The specific types of houses in the text, C Let represent the set of house types, containing all house categories to be classified, and argmax represent the option that maximizes the posterior probability. value, This represents the posterior probability, i.e., the probability of finding a known candidate region's feature vector. Under these conditions, this area belongs to the housing type. The probability, This represents the prior probability, i.e., the type of housing. The prior distribution probability in the entire dataset. This represents the likelihood probability, i.e., when the housing type is... When the eigenvector is observed The probability, Represents a set C Except for the current category c Any other type of house besides Indicates when the housing type is When the eigenvector is observed The probability, Indicates housing type c The prior probability of ′.

[0140] If the maximum posterior probability is lower than the threshold If the result is uncertain, it will be marked as "uncertain" and submitted for manual review or automatically entered into the iterative learning pool.

[0141] In this embodiment, automatically labeled samples are fed into a lightweight validation model (1D-CNN) and reviewed by human experts. 1D-CNN is suitable for processing data with sequential features, automatically extracting feature patterns from the data through convolutional operations, enabling rapid sample validation; human expert review, on the other hand, leverages human expertise and experience to make detailed judgments on the samples. This two-pronged approach obtains genuine feedback. With system prediction By comparison, the Gaussian Mixture Model (GMM) was corrected.

[0142] Construct the confidence function:

[0143]

[0144] Update model weights :

[0145]

[0146] in, Indicates the first i Confidence value for each sample c Represents a set of housing types Specific categories, Indicates sample Category The posterior probability, Indicates the first i The known true labels of each sample. Representing a given feature Under the condition of, predict as label The probability of.

[0147] By repeatedly performing the above verification, confidence calculation, and model weight update processes, the Gaussian Mixture Model (GMM) is iteratively optimized. As the number of iterations increases, the GMM can continuously learn new knowledge, gradually correct erroneous predictions, and thus improve the accuracy of sample labeling. This makes the GMM more accurate and reliable in region type discrimination and label generation tasks.

[0148] S4. Using labeled high-resolution remote sensing samples of houses, and combining them with multimodal remote sensing feature sets of key house targets, a deep learning-based intelligent house extraction model is trained and constructed, specifically as follows:

[0149] High-resolution remote sensing samples of labeled houses were divided into training and validation sets. Combining multimodal remote sensing feature sets of key house targets, a deep learning-based intelligent house extraction model was constructed using the training set. This model was trained by learning the mapping relationship between image features and labels, and then applied to the predicted house type labels. Houses are extracted probabilistically. The training set is input into an improved U-net model, and the weights of the intelligent house extraction model are optimized through backpropagation. The mapping relationship between remote sensing image features and house type labels and bounding boxes is learned. The parameters of the trained intelligent house extraction model are fine-tuned using a validation set to complete the construction of the intelligent house extraction model.

[0150] In this embodiment, the labeled multi-type house samples are divided into training and validation sets. Combining house feature extraction and optimization methods, features of each method are extracted to 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 category probabilities. The model parameters are then fine-tuned using a validation set, and accurate house extraction results are obtained through post-processing. The following sections describe the construction method of the intelligent house remote sensing extraction model from three aspects: the U-net deep learning model algorithm, the loss function, and model training.

[0151] (1) Model Algorithm

[0152] The model consists of an encoder and a decoder. The intermediate layer fuses the house feature maps from the encoder and the decoder through a skip connection to retain more spatial detail information.

[0153] 1) Encoder

[0154] The encoder consists of multi-level downsampling modules. Each module performs double convolution and max pooling to progressively compress the spatial size and increase the number of channels, extracting multi-level features. Each layer contains two 3×3 convolution operations, followed by a ReLU activation function, as shown in the following formula:

[0155]

[0156]

[0157] in, Indicates the first The input house feature map for layer 1 is the training sample set, and the input house feature map for layer 2 is the training sample set. Input house feature map of the layer It is the first The output of the layer, , and , These represent the kernel weights and biases, respectively. This represents the convolution operation. This represents the output house feature map after the first-level convolutional module. After every two convolutional layers, downsampling is performed through a 2×2 max-pooling layer, as shown in the following formula:

[0158]

[0159] in, This indicates that the output should be a feature map of the house. This represents the output house feature map after passing through the second-level convolutional module.

[0160] After max pooling, the size of the output feature map is halved, while the number of channels in the house feature map is increased. Global semantic information of the image is extracted step by step, and the output house feature map is used as the input for the next intermediate layer.

[0161] 2) Intermediate layer

[0162] The intermediate layer receives the final stage output of the encoder, which shows the house feature map. Deeper semantic features are further extracted through two convolutions:

[0163] )+ ).

[0164] 3) Decoder

[0165] The decoder receives the house feature map output from the intermediate layer. Resolution is gradually restored through deconvolution and skip connections. The decoder takes a house feature map as input at each stage. Upsampled from the previous stage output or an intermediate layer via a 2×2 transposed convolution:

[0166]

[0167] Upsampled house feature map House feature map corresponding to the encoder layer The formula for stitching along the channel dimension is as follows:

[0168]

[0169] in, This represents the composite feature map after splicing. This represents the building feature map of the corresponding layer of the encoder. The stitched feature map is then subjected to two 3×3 convolution operations to gradually restore the image's detailed information.

[0170] To improve the building remote sensing intelligent extraction model, the convolution kernels are optimized to target the geometric features of buildings. In the shallow network (levels 1-2), a heterogeneous convolution kernel combination method is used for optimization, employing a 3×3 standard convolution kernel in parallel with a 1×3 or 3×1 rectangular kernel to extract features, enhancing sensitivity to building edges and extracting detailed textures. In the deep network (levels 3-4), a large receptive field convolution method is used for optimization, employing 5×5 and 7×7 convolution kernels to expand the receptive field and capture the global distribution patterns of buildings, thereby improving the accuracy of building extraction.

[0171] In this embodiment, the training objective of the intelligent house extraction model in this invention is to minimize the difference between the predicted house segmentation result and the actual house label. Therefore, cross-entropy loss function and Dice coefficient are typically used as loss functions. Traditional cross-entropy loss function tends to bias the model towards the majority class in cases of class imbalance, thus affecting the house extraction accuracy. Dice loss function measures the overlap between the predicted segmentation result and the actual label, and is particularly suitable for segmentation tasks with class imbalance, but it has low sensitivity to boundary regions. Therefore, the intelligent house extraction model proposes a classification-boundary co-optimization loss function, specifically designed as follows:

[0172] Category-Sensitive Weighted Cross-Entropy (CS-WCE):

[0173]

[0174] in, , =10 −5 By using inverse frequency weighting, the attention given to a minority category (damaged houses) can be increased.

[0175] Boundary-enhanced Dice loss (BE-Dice):

[0176]

[0177] in, =0.3 is the boundary sensitivity coefficient. By introducing a boundary distance penalty term, the model's fitting accuracy to the house outline is improved.

[0178] The total losses are as follows:

[0179]

[0180] in, This represents the loss function of the smart house extraction model. This represents the category-sensitive weighted cross loss function. Represents the boundary loss function. N Indicates the total number of pixels. i Indicates pixel index, Indicates category weight, Indicates the actual housing type label, Indicates the predicted house type label, Indicate category c The sample proportion, This represents a local minimum introduced to avoid the denominator being zero. Represents the boundary sensitivity coefficient. p This represents the probability that the pixel predicted by the model belongs to the house boundary. Y Represents the actual house boundary mask. Represents partial derivatives, Represents the predicted pixels p The probability of belonging to a certain category of housing. Represents pixels p The true label, This represents the weighting coefficient, which decays exponentially. This achieves the phased optimization goal of focusing on classification accuracy in the early stages of training and strengthening boundary matching in the later stages.

[0181] In this embodiment, the intelligent house extraction model is trained as follows:

[0182] The constructed U-net model was trained using a training set of high-resolution remote sensing data of various types of buildings. The Adam optimization algorithm was then used to optimize and adjust the model parameters, starting with the calculation of first-order moment estimates. and second-order moment estimation :

[0183]

[0184]

[0185] in, and All of these represent hyperparameters. Set to 0.9 to balance the stability of gradient updates. Set it to 0.999 to accelerate model convergence. This represents the gradient.

[0186] Then, the first-order moment estimates and second-order moment estimates are corrected:

[0187]

[0188]

[0189] in, and These represent the corrected first-order moment estimates and second-order moment estimates, respectively. and These represent the decay rates of the first and second moments after t training iterations, respectively.

[0190] Finally, update the model parameters. :

[0191]

[0192] in, This represents the updated U-net model parameters. This represents the learning rate, which typically ranges from 10. -5 Up to 10 -3 , For a minimum value (e.g.) ), used to prevent the denominator from being 0.

[0193] During training, the model was optimized and improved. First, the data augmentation strategy was optimized to suit the characteristics of house extraction. Specifically, this included random rotation with an angle range of −10 degrees. ∘ Up to 10 ∘ To increase the model's robustness to house rotation invariance; random flipping, including horizontal and vertical flipping, to enhance the model's ability to recognize house symmetry; random scaling, with scaling ratios between 0.8 and 1.2, to increase the model's ability to extract houses at different scales; and brightness and contrast adjustment, which randomly adjusts the brightness and contrast of the input image to enhance the model's adaptability to changes in lighting and improve the model's generalization ability.

[0194] Secondly, an early stopping mechanism is employed: training is stopped when the accuracy on the validation set no longer improves for 10 consecutive epochs, or when the decrease in the loss function value on the validation set is less than a certain threshold (e.g., 0.001), the model is considered to have converged and training is stopped to prevent overfitting. Furthermore, the model's weight parameters are periodically saved so that the best-performing model can be selected for subsequent evaluation and application after training.

[0195] By applying the trained intelligent house extraction model and combining it with high-resolution satellite remote sensing images obtained with consistent spatiotemporal references before and after the rainstorm and flood disaster, the boundary range of houses before and after the flood disaster is extracted.

[0196] S5. Using the trained intelligent housing extraction model, determine the extent of damage to houses after extreme rainstorms and floods, and complete the remote sensing extraction of damaged houses. Specifically:

[0197] High-resolution remote sensing imagery of the same region before and after extreme rainstorms and floods were acquired. The high-resolution remote sensing imagery of the region before the extreme rainstorms and floods were then input into a trained intelligent house extraction model to obtain the bounding box coordinates of each house. Information on housing types and the number of houses in the area before the disaster. High-resolution remote sensing imagery data from extreme rainstorms and floods is input into a trained intelligent house extraction model to obtain the bounding box coordinates of each house. Information on housing types and the number of houses in the area affected by the disaster. For each house extracted before the extreme rainstorm and flood disaster, determine its positional relationship with houses extracted after the extreme rainstorm and flood disaster, and measure the overlap between the two bounding boxes; 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; based on the number of houses before the disaster... and the number of houses after the disaster The process involves calculating the rate of change in the number of houses; determining the extent of damage caused by extreme rainstorms and floods by judging the increase or decrease in house data based on the rate of change in the number of houses; calculating the house area before and after the extreme rainstorm and flood for each house; calculating the rate of change in house area based on the house area before and after the extreme rainstorm and flood; and determining the change in house area based on the rate of change in house area to complete the remote sensing extraction of damaged houses.

[0198] In this embodiment, high-resolution satellite remote sensing images of the same region before and after extreme rainstorm and flood disasters are acquired. After performing the operations described above on the images, the changes in the number, location, and area of ​​houses are detected by inputting the images into the model before and after the flood disaster, thereby obtaining statistics on the damage to houses caused by the rainstorm and flood disaster.

[0199] (1) Model training and house extraction (preliminary work)

[0200] Using pre-flood imagery as input, the bounding box coordinates of each house were recorded to obtain the house extraction bounding box boundaries. Information on the types and categories (all housing categories) was compiled, and the number of houses in the area before the disaster was counted. .

[0201] Similarly, deep models were trained using images from after floods and houses were extracted to obtain the coordinates of the house bounding boxes. Information on the types and categories of houses in the area, including the number of houses before the disaster. .

[0202] (2) Detection of changes in the location of damaged houses

[0203] For each house extracted before the flood disaster, its positional relationship with houses extracted after the flood disaster is calculated. The Intersection over Union (IoU) metric is used to measure the degree of overlap between the two bounding boxes, and the calculation formula is as follows:

[0204]

[0205] Set an IoU threshold (Take 0.5) If the IoU between a house before and after a flood is greater than the threshold. If the IoU is less than 1, then the location of the house is considered to have remained essentially unchanged; if the IoU is less than 1, then the location of the house is considered to have remained essentially unchanged. If so, it indicates that the house may have been displaced or damaged.

[0206] (3) Monitoring of changes in the number of damaged houses

[0207] Statistics on the number of houses collected before and after the flood disaster and Calculate the rate of change in the number of houses. The formula is:

[0208]

[0209] Based on the rate of change in the number of houses The value indicates whether the number of houses has increased or decreased. 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.

[0210] (4) Detection of changes in damaged area of ​​a building based on its boundary

[0211] For each house, the area is calculated based on its bounding box coordinates. House area before flooding. The calculation formula is: House area after flood disaster The calculation formula is: .

[0212] Calculate the rate of change in building area :

[0213]

[0214] pass The value indicates the change in the house area. A positive value indicates an increase in area, which may be due to factors such as image errors or changes in the surrounding environment. A negative value indicates a decrease in area, which may be due to 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.

[0215] The above method for detecting changes in houses before and after floods based on deep learning models 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 floods, providing important evidence for flood disaster assessment and post-disaster reconstruction.

[0216] Example 2

[0217] like Figure 2 As shown, this invention provides a remote sensing extraction system for houses damaged by extreme rainstorms and floods, applied to any of the remote sensing extraction methods for houses damaged by extreme rainstorms and floods described in Example 1, comprising:

[0218] The first processing module is used to acquire remote sensing image data of key target houses with consistent spatiotemporal reference before and after the disaster using multi-source high-resolution remote sensing data.

[0219] The second processing module is used to construct a multimodal remote sensing feature set of key building targets by fusing spectral and texture features based on the acquired remote sensing image data;

[0220] The third processing module is used to label high-resolution remote sensing house samples based on geometric structure for different house types. The labeling of high-resolution remote sensing house samples is used to calculate the similarity probability of multiple house types.

[0221] The fourth processing module is used to train and build a deep learning-based intelligent house extraction model by utilizing labeled high-resolution remote sensing house samples and combining them with multimodal remote sensing feature sets of key house targets.

[0222] The fifth processing module is used to determine the condition of damaged houses after extreme rainstorms and floods by using the trained intelligent house extraction model, and to complete the remote sensing extraction of damaged houses.

[0223] like Figure 2 The remote sensing extraction system provided in the embodiment shown can execute the technical solution shown in the remote sensing extraction method of embodiment 1 above. Its implementation principle and beneficial effects are similar, and will not be repeated here.

[0224] In this embodiment, the functional units can be divided according to the remote sensing extraction method in Embodiment 1. For example, each function can be divided into its own functional units, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the division of units in this invention is illustrative and only represents a logical division; other division methods may be used in actual implementation.

[0225] In this embodiment, the remote sensing extraction system, in order to realize the principle and beneficial effects of the remote sensing extraction method, includes hardware structures and / or software modules corresponding to the execution of various functions. Those skilled in the art should readily recognize that, in conjunction with the illustrative units and algorithm steps described in the embodiments disclosed herein, the present invention can be implemented in hardware and / or a combination of hardware and computer software. Whether a function is executed by hardware or computer software depends on the specific application and design constraints of the technical solution. Different methods can be used to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

Claims

1. A method for remote sensing extraction of houses damaged by extreme rainstorms and floods, characterized in that, Includes the following steps: By utilizing multi-source high-resolution remote sensing data, we can obtain remote sensing image data of key target buildings with consistent spatiotemporal references before and after the disaster. Based on the acquired remote sensing image data, a multimodal remote sensing feature set of key building targets was constructed by fusing spectral and texture features; For different housing types, high-resolution remote sensing housing samples that take into account geometric structure are labeled. In this process, the high-resolution remote sensing housing samples are labeled to calculate the similarity probability of multiple housing types. The calculation of the similarity probability of multiple types of houses is specifically as follows: Construct a hierarchical geometric feature space, where geometric features include primary features representing the area of ​​a house, secondary features representing the aspect ratio of a house, and tertiary features representing the shape index of a house. Based on the constructed hierarchical geometric feature space, a house type discrimination rule is constructed; Based on the constructed house type discrimination rules, a Gaussian mixture model is built 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 geometric prototypes of various building types, the calculation is performed on each candidate region in the new remote sensing image. Each candidate region is obtained. eigenvectors The posterior probability of belonging to each class; Based on the maximum a posteriori probability principle, house type labels are assigned to obtain the predicted house type labels. ; Get real house type tags and label the actual house type With predicted house type labels By making comparisons and using the comparison results to correct the Gaussian mixture model, the similarity probability of multiple types of houses can be calculated. Using labeled high-resolution remote sensing samples of houses, and combining them with multimodal remote sensing feature sets of key house targets, a deep learning-based intelligent house extraction model was trained and constructed. Using a trained intelligent housing extraction model, the extent of damage to houses after extreme rainstorms and floods is determined, and remote sensing data of the damaged houses is extracted.

2. The method for remote sensing extraction of houses damaged by extreme rainstorms and floods according to claim 1, characterized in that, The construction of a multimodal remote sensing feature set for key building targets, which integrates spectral and texture features, specifically involves: Based on the acquired remote sensing image data, determine each remote sensing image pixel. spectral vector ; Calculate normalized vegetation separately and standardized architecture Spectral characteristics; spectral vector and normalized vegetation and standardized architecture The spectral features are combined to form the spectral feature vector of the building area. This completes the remote sensing extraction of the spectral features of the building area; Based on the acquired remote sensing image data, image texture features are extracted. This completes the extraction of texture features from the building area; Based on the extracted spectral and texture features of the housing areas from remote sensing, for each candidate housing area... Extract the spectral feature vectors of the building areas respectively. and texture feature vector Regional features are constructed by feature concatenation; Based on the regional features constructed by feature concatenation, feature priority is applied to obtain the preferred regional features; Based on priority-based regional features Forming a structured feature set ; Based on structured feature sets Calculate the entropy of each feature. ; Based on adaptive adjustment factor and feature information entropy Calculate the weights of each feature. ; According to the weight of each feature The features with the highest weights are sorted and used as the building features to complete the construction of a multimodal remote sensing feature set for key building targets.

3. The method for remote sensing extraction of houses damaged by extreme rainstorms and floods according to claim 1, characterized in that, The expression for the housing type discrimination rule is as follows: in, This indicates the rules for determining housing type. Indicating the shape index of the house. Indicates the length-to-width ratio of the house. This indicates the area of ​​the house, with values ​​ranging from 50 square meters to 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, Represents a set The specific types of houses in the text, C Let represent the set of house types, containing all house categories to be classified, and argmax represent the option that maximizes the posterior probability. value, This represents the posterior probability, i.e., the probability of finding a known candidate region's feature vector. Under these conditions, this area belongs to the housing type. The probability, This represents the prior probability, i.e., the type of housing. The prior distribution probability in the entire dataset. This represents the likelihood probability, i.e., when the housing type is... When the eigenvector is observed The probability, Represents a set C Except for the current category c Any type of house outside, Indicates when the housing type is When the eigenvector is observed The probability, Indicates housing type c The prior probability of ′.

4. The method for remote sensing extraction of houses damaged by extreme rainstorms and floods according to claim 1, characterized in that, The training and construction of the deep learning-based intelligent house extraction model specifically involves: The labeled high-resolution remote sensing house samples were divided into training and validation sets. By combining multimodal remote sensing feature sets of key building targets, a deep learning-based intelligent building extraction model is constructed using the training set. This model is trained by learning the mapping relationship between image features and labels, and then based on the predicted building type labels... Probabilistic house extraction; The parameters of the trained intelligent house extraction model are fine-tuned using the validation set, thus completing the construction of the intelligent house extraction model.

5. The method for remote sensing extraction of houses damaged by extreme rainstorms and floods according to claim 1, characterized in that, The expression for the loss function of the intelligent house extraction model is as follows: in, This represents the loss function of the smart house extraction model. Indicates the weighting coefficient. This represents the category-sensitive weighted cross loss function. Represents the boundary loss function. N Indicates the total number of pixels. i Indicates pixel index, Indicates category weight, Indicates the actual housing type label, Indicates the predicted house type label, Indicates housing type The proportion of samples, This represents a local minimum introduced to avoid the denominator being zero. Represents the boundary sensitivity coefficient. p This represents the probability that the predicted pixel belongs to the house boundary. Y Represents the actual house boundary mask. Represents partial derivatives, Represents the predicted pixels p The probability of belonging to a certain category of housing. Represents pixels p The true label.

6. The method for remote sensing extraction of houses damaged by extreme rainstorms and floods according to claim 1, characterized in that, The determination of the damage to houses after extreme rainstorms and floods specifically includes: High-resolution remote sensing image data of the same region before and after extreme rainstorm and flood disasters were acquired; High-resolution remote sensing imagery data from before extreme rainstorms and floods are input into a trained intelligent house extraction model to obtain the bounding box coordinates of each house. Information on housing types and the number of houses in the area before the disaster. ; High-resolution remote sensing imagery data following extreme rainstorms and floods is input into a trained intelligent house extraction model to obtain the bounding box coordinates of each house. Information on housing types and the number of houses in the area affected by the disaster. ; For each house extracted before the extreme rainstorm and flood disaster, determine its positional relationship with the houses extracted after the extreme rainstorm and flood disaster, and measure the overlap of the two bounding boxes; 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. Based on the number of houses before the disaster and the number of houses after the disaster Calculate the rate of change in 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, calculate the house area before and after the extreme rainstorm and flood disaster; Calculate the rate of change in house area based on the house area before and after extreme rainstorm and flood disasters; Based on the rate of change of building area, the changes in building area are determined, and the remote sensing data of damaged buildings is extracted.

7. A remote sensing extraction system for houses damaged by extreme rainstorms and floods, applied to the remote sensing extraction method for houses damaged by extreme rainstorms and floods as described in any one of claims 1-6, characterized in that, include: The first processing module is used to acquire remote sensing image data of key target houses with consistent spatiotemporal reference before and after the disaster using multi-source high-resolution remote sensing data. The second processing module is used to construct a multimodal remote sensing feature set of key building targets by fusing spectral and texture features based on the acquired remote sensing image data; The third processing module is used to label high-resolution remote sensing house samples based on geometric structure for different house types. Specifically, labeling the high-resolution remote sensing house samples is used to calculate the similarity probability of multiple house types. Construct a hierarchical geometric feature space, where geometric features include primary features representing the area of ​​a house, secondary features representing the aspect ratio of a house, and tertiary features representing the shape index of a house. Based on the constructed hierarchical geometric feature space, a house type discrimination rule is constructed; Based on the constructed house type discrimination rules, a Gaussian mixture model is built 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 geometric prototypes of various building types, the calculation is performed on each candidate region in the new remote sensing image. Each candidate region is obtained. eigenvectors The posterior probability of belonging to each class; Based on the maximum a posteriori probability principle, house type labels are assigned to obtain the predicted house type labels. ; Get real house type tags and label the actual house type With predicted house type labels By making comparisons and using the comparison results to correct the Gaussian mixture model, the similarity probability of multiple types of houses can be calculated. The fourth processing module is used to train and build a deep learning-based intelligent house extraction model by utilizing labeled high-resolution remote sensing house samples and combining them with multimodal remote sensing feature sets of key house targets. The fifth processing module is used to determine the condition of damaged houses after extreme rainstorms and floods by using the trained intelligent house extraction model, and to complete the remote sensing extraction of damaged houses.

Images