Submerged simulation-oriented DSM-DEM collaborative optimization method

Abstract

A DSM-DEM collaborative optimization method for submerging simulation comprises the following steps: based on a satellite image and a historical optical image, key ground feature elements influencing a submerging simulation effect are extracted by using a deep learning technology, and the key ground feature elements comprise roads, dams, trees and residential areas; after surface feature elements influencing flood simulation are extracted, FABDEM is optimized based on the surface feature elements, residual abnormal values including unremoved buildings, tree coverage, viaducts, dams, roads and the like are identified and processed, a DEM abnormal value problem is converted into a slope abnormal value problem, an abnormal slope is identified by using a slope filtering technology, and a flood simulation result is obtained. Carrying out processing in combination with the extracted ground feature elements and an open source tree coverage data set; while removing the abnormal elevation data, emphatically identifying and retaining dam data, and correcting the problem of overhigh dam elevation caused by tree coverage; and for a data vacancy area, filling is carried out by applying multiple interpolation methods so as to realize DEM optimization facing flood submerging simulation.

Description

Technical Field

[0001] The invention relates to the technical field of disaster information simulation, and in particular to a DSM-DEM collaborative optimization method for flooding simulation. Background Art

[0002] DEM is the key data for flood simulation, and the quality of data determines the quality of simulation. The current open source DEM data is not optimized for flood simulation, such as unclear river channels and missing dam details. Although some DEM data preserve river channel and dam details, they regard roads such as viaducts as "long embankments". In addition, due to resolution issues, there are no gaps between buildings, and they exist in a continuous piece. These shortcomings hinder the further improvement of flood simulation results.

[0003] The existing technology for flood and inundation simulation usually does not distinguish between the three types of data, DEM, DSM, and DTM, and collectively refers to them as DEM. Here, the concepts of the three types of data are distinguished to facilitate a clearer understanding of the characteristics and application scenarios of each type of data. This application also uses DEM to refer to DEM, DSM, and DTM. DEM represents the exposed surface of the earth, removing all natural and architectural features, and providing a relatively smooth terrain surface. This model is mainly used to analyze terrain changes and geomorphic features, and is suitable for basic representation of terrain in flood simulation. DTM contains vector features of natural terrain, such as rivers and ridges, which can be regarded as an enhancement of DEM, and DEM can be generated by interpolation. DTM provides more detailed and accurate terrain information and is suitable for applications that require high-precision terrain analysis. DSM contains natural and artificial features, such as trees, buildings, etc. This model is suitable for applications that need to consider all surface features, such as urban planning and environmental monitoring. In flood simulation, DSM can provide more detailed terrain information, but it will be affected by surface features. Next, the commonly used open source DEM data for flood and inundation simulation will be introduced in detail.

[0004] In order to solve the current shortcomings of FABDEM, this application will optimize it through two steps: feature extraction and DEM outlier identification and removal. In the feature extraction part, based on Landsat-8 satellite images and ESRI historical optical images, deep learning technology can be used to extract key features that affect the flooding simulation effect, such as roads (embankments), trees, and residential areas, to provide an identification basis for the next step of DEM outlier identification and optimization. Summary of the invention

[0005] In view of the above problems in the prior art, this application proposes a DSM-DEM collaborative optimization method for flooding simulation, comprising the following steps:

[0006] S1. Based on satellite images and historical optical images, deep learning technology is used to extract key features that affect flooding simulation effects. Key features include roads, dams, trees, and residential areas.

[0007] S2. After extracting the features that affect flood simulation, FABDEM (Forest And Buildings removed Copernicus DEM) is optimized based on these features. The remaining outliers, including buildings, tree cover, viaducts, dams and roads, are identified and processed. The DEM outlier problem is converted into a slope outlier problem. The slope filtering technology is used to identify abnormal slopes, and the extracted features and open source tree cover datasets are combined for processing. While removing abnormal elevation data, the dam data is identified and retained, and the problem of excessive dam elevation caused by tree cover is corrected. For areas with missing data, a variety of interpolation methods are used to fill them in to achieve DEM optimization for flood inundation simulation.

[0008] In one embodiment, a method for extracting residential areas includes: selecting images of cities surrounding a target area as a fine-tuning data set, segmenting images with equal length and width as a fine-tuning data set, and splitting them into a training set, a validation set, and a test set; then manually annotating the residential areas in the images to generate formatted data required for model fine-tuning; segmenting the target area images and inputting them into the fine-tuned model to obtain extraction results.

[0009] In one embodiment, the road extraction method includes encoding, training, reasoning and decoding stages; in the encoding stage, the model receives high-resolution satellite images containing road networks as true value data, and converts these images into high-dimensional tensor representations through a graph tensor encoding algorithm; this process can retain the spatial distribution information of the roads, as well as the encoded topological structure and connection relationship; in the training stage, the encoded graph tensor is used to pair with the original satellite image for learning, so that the neural network can extract key features from the original image and map it to the corresponding graph tensor representation; through repeated iterations, the model gradually optimizes its parameters; in the reasoning stage, when the model faces new, unlabeled satellite images, the trained neural network can automatically generate the corresponding graph tensor encoding; in the decoding stage, the model converts the inferred graph tensor encoding back to a visualized road network structure.

[0010] In one embodiment, a method for extracting tree cover includes: randomly selecting multiple areas in a tree crown height map dataset, calculating the intersection-union ratio of the tree cover range and the manually drawn tree cover patches, judging their quality, and quantifying the degree of overlap between them by calculating the ratio of the intersection and union of two areas.

[0011] In one embodiment, by overlaying residential area and tree cover area data with outlier data, overlapping outliers are identified and removed to ensure that buildings and tree cover do not interfere with flood simulation; by creating river buffers and calculating the average relative height of roads based on COPDEM, river embankments with impoundment functions are identified and retained; and by performing spatial intersection analysis between roads and outliers, roads are identified and removed.

[0012] In one embodiment, the residential area data is used as a mask file to overlay the slope grid, and the grid values ​​that overlap with the residential area and have a slope value greater than a predetermined threshold are deleted from the slope grid, and the new mask file is recorded as the first slope mask; the first slope mask is then overlaid with FABDEM to delete abnormal elevation data, thereby generating a first DEM file after optimizing the residential area.

[0013] In one embodiment, outliers are removed by superimposing tree cover areas based on the first slope mask and the first DEM to generate a second slope mask and a second DEM with abnormal slopes and abnormal elevations of buildings and trees removed, and natural neighbor interpolation is used to fill irregular gaps in the second DEM.

[0014] In one embodiment, the extracted road data is subjected to spatial intersection analysis with the second slope mask, and the outliers intersecting with the roads in the second DEM are removed to obtain a third slope mask that removes outliers of ordinary roads and elevated roads; the third slope mask is superimposed with FABDEM, and the abnormal elevation data is removed to obtain an optimized third DEM.

[0015] In one embodiment, first, a buffer zone is created using river surface data provided by the national basic geographic information database as a preliminary range for the distribution of potential dams; then, the buffer zone outside the river is used as a mask to extract a slope raster, and abnormal slope values ​​are extracted; the extracted road data is loaded, and if there are multiple roads in the buffer zone outside the river, the road with the highest average relative height is selected as the dam road; the dam road and the outliers are spatially analyzed, and the outliers that intersect with the dam road are the real dams, and other outliers are deleted to obtain the dam information of the river.

[0016] In one implementation, the DEM and forest height data are first read and aligned; then, a block strategy is used to process the data. For each data block, the forest height is first subtracted and then a median filter is applied; the median filter process includes three steps: filling the edge of the data, applying a 5x5 filter window, and outputting the results; finally, an optimized DEM is generated by iteratively processing all data blocks.

[0017] The above technical features can be combined in various suitable ways or replaced by equivalent technical features as long as the purpose of the present invention can be achieved.

[0018] The DSM-DEM collaborative optimization method for flooding simulation provided by the present invention has at least the following beneficial effects compared with the prior art:

[0019] The shortcomings of current open source DEM data in flooding simulation are solved. FABDEM is selected as the optimization object. By introducing Landsat-8 satellite images and ESRI historical optical images, deep learning technology is used to extract key land features. Based on the overall slope characteristic of 0.0002 in the study area, the DEM outlier problem is innovatively transformed into a slope outlier problem. The slope filtering technology is used to identify abnormal slopes, and the extracted land features and open source tree cover datasets are combined for processing. While removing abnormal elevation data, the embankment data is focused on identifying and retaining, and the problem of excessive embankment elevation caused by tree cover is corrected. For data vacant areas, a variety of interpolation methods are used to fill them. Through this collaborative optimization method, the study effectively improves the applicability of DEM in flooding simulation, greatly improves the quality of DEM for flooding simulation, and provides important support for improving the accuracy and reliability of flooding simulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be described in more detail below based on embodiments and with reference to the accompanying drawings, wherein:

[0021] Figure 1 The U-net architecture diagram of the present invention is shown;

[0022] Figure 2 The result map of residential area extraction based on Landsat-8 image is shown;

[0023] Figure 3 The Sat2Graph model workflow diagram is shown;

[0024] Figure 4 A sample graph of road extraction based on Sat2Graph is shown;

[0025] Figure 5 A sample map of tree-covered areas is shown;

[0026] Figure 6 The slope outlier detection result graph based on FABDEM is shown;

[0027] Figure 7 A flow chart of levee extraction is shown;

[0028] Figure 8 The DEM and slope map with all outliers removed are shown;

[0029] Fig. 9 The COPDEM dam false height phenomenon map is shown;

[0030] Fig.10 A flow chart for dam height optimization is shown;

[0031] Fig.11 The corrected levee map is shown. DETAILED DESCRIPTION

[0032] The present invention will be further described below in conjunction with the accompanying drawings.

[0033] The present invention provides a DSM-DEM collaborative optimization method for flooding simulation, extracts residential area data from Landsat-8 images based on the U-net model, and is implemented by the deep learning environment provided by ArcGIS Pro. Environment configuration: operating system is Windows 11, CPU is AMD R7 5800H, and graphics card is RTX 30606G.

[0034] 1. Model architecture and principle. The U-Net model consists of an encoder (downsampling path, implemented by maximum pooling) and a decoder (upsampling path). It has a symmetrical structure and connects the corresponding layers of the encoder and decoder through skip connections to retain high-resolution feature information, such as Figure 1 shown.

[0035] The encoder part consists of a series of convolutional layers and maximum pooling layers. Each convolutional layer uses a 3x3 convolution kernel to extract feature information from the input image. Subsequently, the maximum pooling layer downsamples the feature map through a 2x2 pooling kernel, reducing the spatial resolution while still retaining important image features. The Dropout layer is used to prevent overfitting and is added after each convolution block. The encoder part outputs a low-resolution but high-dimensional feature map that contains the feature information of the image.

[0036] The skip connection directly transfers the feature map of each layer in the encoder to the corresponding layer in the decoder, and improves the accuracy of image segmentation by splicing the high-resolution features with the low-resolution feature maps in the decoder. The skip connection method can not only retain high-resolution feature information, but also solve the gradient disappearance problem, making the model training more stable.

[0037] The decoder can restore the feature map to the size of the original image, and gradually restore the spatial resolution of the feature map through a series of upsampling layers and convolution layers and perform pixel-level classification. All upsampling layers use 2x2 upsampling kernels to restore low-resolution feature maps to high resolution. In addition, each layer of the decoder also receives the jump connection feature map from the corresponding layer of the encoder, and merges the high-resolution feature map with the feature map of the current layer through a splicing operation. Finally, the multi-channel feature map is mapped to the required number of categories through a 1x1 convolution kernel to achieve pixel-level classification.

[0038] 2. Model fine-tuning. The basic U-Net model provided by ArcGIS Pro did not include Chinese data sets during training. In order to improve the performance of the model in this application area, local data must be used for fine-tuning. This application selects Landsat-8 images of cities around City S as fine-tuning datasets, and cuts out 500 images with a resolution of 30 meters and a length and width of 7.68 km as fine-tuning datasets, and splits them into training sets, validation sets, and test sets in a ratio of 7:2:1. Then manually annotate the residential areas in the images to generate the formatted data required for model fine-tuning.

[0039] After the data was labeled, the U-Net model was fine-tuned using the "Train Deep Learning Model" tool in ArcGIS Pro. The maximum number of iterations for fine-tuning was set to 100, the batch size was set to 16, and the early stopping mechanism was set. The model finally converged and terminated early after about 60 rounds, with an overall accuracy of 92%. The fine-tuned model not only retains the basic feature extraction capabilities of the original model, but also further improves the versatility of the model in new geographic areas.

[0040] 3. Data input and processing. The input of the fine-tuned model is the Landsat8 image with a resolution of 30 meters, which is obtained from the GEE platform. Considering that the production date of COPDEM is 2011-2015, in order to keep the time consistent, the Landsat-8 image of August 2015 was selected. The GEE platform has performed radiation correction, geometric correction, cloud removal and terrain correction on the Landsat-8 data, and can be used directly.

[0041] 4. Extract results. The Landsat-8 image of City S is segmented and input into the U-net model. The segmentation results are as follows: Figure 2 shown. Figure 2 (a) in the figure is the Landsat-8 image of S City. Figure 2 (b) is the binary extraction result of residential areas based on Landsat-8 images, and green represents the extracted residential areas. By comparing the images, it can be found that the model can accurately identify the residential areas in the application area, with clear boundaries, especially in densely populated urban areas, where the model shows high extraction accuracy.

[0042] After extracting the residential area, this application uses the Sat2Graph model to extract roads and dams from the ESRI 2015 map image. Roads and dams appear the same in satellite images, so all roads can be extracted first. In the DEM optimization part, the spatial topological relationship between roads and rivers and the elevation difference between them and the ground will be used to determine whether a certain section of road is a dam.

[0043] The extraction of road networks belongs to computer vision tasks, and the extraction methods can be roughly divided into pixel segmentation methods and graph-based methods. Pixel segmentation methods generate road networks by predicting whether each pixel belongs to a road, while graph-based methods construct roads by iteratively adding vertices and edges. Pixel segmentation methods have a large receptive field, but rely on non-graph intermediate representations, which often leads to large noise and low accuracy in road networks. Graph-based methods directly learn graph structures and can generate more connected road networks, but they usually rely on local information and lack a global perspective. Sat2Graph is an image parsing method based on deep learning and graph neural networks. It combines the advantages of pixel segmentation methods and graph-based methods, and encodes the road network graph into a tensor representation through graph tensor encoding. It can accurately extract linear objects from high-resolution remote sensing images and achieve efficient extraction of complex road structures.

[0044] 1. Model architecture principle. Sat2Graph converts remote sensing images into graph structures and combines node features and edge features for feature extraction and classification, thereby improving the accuracy of linear feature extraction, such as Figure 3 As shown. The core of the model is GTE. The method first constructs a graph structure by extracting road pixels in the image. Each node in the graph represents a road pixel, and the edge represents the connection relationship between road pixels. Subsequently, the method creates a multidimensional tensor to capture the complex structure and characteristics of the graph. The tensor construction process involves multiple dimensions: the first dimension corresponds to node features, such as position coordinates and color information; the second dimension represents edge features, including attributes such as edge length and direction. The third dimension represents the use of a vector to encode the graph information at each spatial position. This vector contains several elements. The first element represents the probability of whether there is a vertex at that position, and the subsequent elements encode the information of the edges starting from that position, including the probability of the existence of the edge and the relative position.

[0045] In the tensor filling stage, the features of the nodes and edges in the graph are mapped to the corresponding tensor elements. Each tensor element corresponds to a node or edge in the graph, and its value reflects the features of the node or edge. For a road network graph of a specified area, the GTE mechanism uses a three-dimensional tensor containing the grid of the area to store the encoded information of the graph. The first two dimensions of the tensor correspond to the spatial coordinates on the two-dimensional plane, and the third dimension contains a vector representing the graph information. For example, a simplified road network consisting of three intersections and four roads, the three-dimensional tensor of the network is represented as follows: the first dimension corresponds to three nodes, the second dimension represents four edges, and the third dimension describes the relationship between node pairs, such as connectivity. Therefore, GTE can not only represent the location and shape of the road, but also encode information such as the connection relationship between roads and the topological structure of the road, so that the model can learn more comprehensive road features.

[0046] 2. Model workflow. As shown in the figure above, the model workflow can be divided into encoding, training, inference, and decoding stages. In the encoding stage, the model receives high-resolution satellite images containing road networks as true data, and converts these images into high-dimensional tensor representations through a graph tensor encoding algorithm. This process can preserve the spatial distribution information of roads, as well as the encoded topological structure and connection relationships. In the training stage, the encoded graph tensor is paired with the original satellite image for learning, so that the neural network can extract key features from the original image and map them to the corresponding graph tensor representation. Through repeated iterations, the model gradually optimizes its parameters and improves its ability to recognize and encode road features.

[0047] The inference phase refers to the time when the model is faced with new, unlabeled satellite images, and the trained neural network can automatically generate the corresponding graph tensor encoding. In the decoding phase, the model converts the inferred graph tensor encoding back into a visualized road network structure. It can be found that from inputting the original satellite image, to generating the graph tensor encoding through the neural network, and then finally decoding it into a road network graph, the Sat2Graph model implements an end-to-end processing flow, which improves the extraction efficiency while ensuring the consistency of the results.

[0048] 3. Road extraction. The model input data requires 8-bit, 3-band, 1-meter resolution satellite images. This application selects ESRI satellite images of S City in 2015 and slices them as input data. The images are obtained from the Esri Wayback LivingAtlas platform. The Sat2Graph model has strong generalization ability, and the overall accuracy is 76% after testing. It can be used in different regions, so there is no need to retrain.

[0049] 4. Extract the results. Figure 4 As shown in the figure, the upper left is COPDEM, the upper right is the 2015 Wayback image, and the yellow box indicates the sample area. Area 1 is a dense urban road network, and the river embankment on the far right is the river embankment. The model can accurately identify and extract small streets and intersections, and also extract the roads corresponding to the embankment. Area 2 is a complex road network with viaducts. The model successfully extracts the road structure, including roundabouts and multiple intersections, proving its ability to extract viaducts. Area 3 is a rural road, and there are also viaducts above it. The area is mainly straight roads. The model can accurately extract the main skeleton of the rural road and the viaduct above it. Area 4 is a complex road network with viaducts and embankments. It can be found that the model can effectively extract urban roads, viaducts and embankments, showing the robustness of the model.

[0050] In general, the Sat2Graph model has good robustness in various complex road environments. The model can accurately extract the road network structure in different environments such as cities, rural areas and farmlands. It can not only identify and extract subtle streets and intersections, but also handle complex road networks including viaducts and embankments. After the road extraction work is completed, it is still necessary to further divide the long embankments and viaducts, which will be carried out in the DEM optimization part.

[0051] After attempts, the edge accuracy of extracting individual trees from sub-meter remote sensing images using deep learning models is usually low, and the quality of the extracted data is poor, which cannot meet the requirements of this application. The tree cover in remote sensing images has complex spectral characteristics. Due to the influence of factors such as different tree species, health conditions, seasonal changes and lighting conditions, the spectral response of the same tree species in different images is different. Secondly, the accuracy of tree extraction is affected by the spatial resolution and spectral resolution of the image. High-resolution images can achieve higher extraction accuracy, but data acquisition is more difficult. In addition, the training of deep learning models requires a large amount of labeled data, and the labeling process is time-consuming and labor-intensive. In practical applications, the generalization ability and robustness of the model are difficult to maintain stability in different geographical regions, climatic conditions and seasonal changes.

[0052] This application uses a global tree crown height map dataset developed by Meta and World Resources Institute with a spatial resolution of 1 meter, which covers satellite image data from 2009 to 2020, of which 80% of the data is from 2018 to 2020. This dataset is based on the DiNOv2 model and is generated by training 18 million natural color remote sensing images with a resolution of 0.5 meters. The self-supervised learning (SSL) method was used in the data production process to train the DiNOv2 model with unlabeled satellite images, enabling it to extract common image features. Subsequently, the model was combined with LIDAR ground truth data from the NEON dataset to train a tree crown height predictor. The predictor has a mean absolute error of 2.8 meters and can accurately predict the height of individual trees.

[0053] Since COPDEM was produced in 2015, but the dataset used by the global tree canopy height map dataset is concentrated in 2018-2020, direct use will cause tree coverage error problems. Considering that the common tree species in S City are mainly sycamore, Sophora japonica, Fraxinus chinensis, Koelreuteria paniculata and Zelkova, these trees grow slowly, so this application removes the coverage height data below 1 meter. 30 areas of 1.5 kilometers were randomly selected, and the intersection over Union (IoU) of the tree coverage above 1 meter and the manually drawn tree coverage patches in 2015 were calculated to judge their quality. IoU is a commonly used indicator to measure the similarity between two areas. The ratio of the intersection to the union of two areas is calculated to quantify the degree of overlap between them.

[0054] The calculation formula of intersection-over-union ratio is as follows.

[0055]

[0056] Among them, |A∩B| represents the intersection area of ​​the two comparison areas, and |A∪B| represents the union area of ​​the two areas. The IoU value is between 0 and 1. The larger the value, the greater the overlap of the two areas and the higher the similarity. When the IoU value is 1, it means that the two areas overlap completely; when the IoU value is 0, it means that the two areas do not overlap at all. Through calculation and testing of 30 images, the IoU of the drawn patches and the number coverage is 62.35%, which is of good quality. Data samples are as follows Figure 5 shown.

[0057] After extracting the features that affect flood simulation, such as residential areas, roads, and tree-covered areas, FABDEM needs to be optimized based on these features. Most of the residential areas, roads, and tree-covered areas have been removed from FABDEM, but in order to improve the quality of inundation simulation, it is still necessary to identify and process the remaining outliers, including buildings, tree cover, viaducts, dams, and ordinary roads that have not been removed, so as to achieve DEM optimization for flood inundation simulation.

[0058] The average slope of the application area is 0.0002, and the outliers are usually protruding or concave surfaces, so the outliers can be converted into abnormal slope identification. The application uses a slope filter to detect and identify abnormal slope values ​​from FABDEM. These slope values ​​include unremoved buildings, tree cover, viaducts, dams and ordinary roads. For buildings that have not been removed, the outliers overlapping with the residential area can be removed by superimposing the abnormal slope with the residential area. Similarly, the same method can be used to remove tree-covered outliers. In flood simulation, viaducts and ordinary roads are permeable to floods, and dams play a role in intercepting and storing water. It is necessary to distinguish and identify the extracted viaducts, dams and ordinary roads.

[0059] This application uses slope filtering for outlier detection. First, the slope value of FABDEM is calculated. The slope is a measure of the rate of change of pixel height, expressed as a percentage or degree. The specific calculation method is as follows.

[0060]

[0061] in, and are the elevation change rates of DEM in the x and y directions respectively. The calculation results are as follows Figure 6 shown.

[0062] The dark green area occupies most of the area in the figure, indicating that the quality of FABDEM is relatively good. There are obvious linear high slope features in the area, mostly yellow or orange, with slope values ​​between 1.6-7.7 degrees. These may represent rivers, trees, dams, ordinary roads, viaducts, and other man-made structures. In addition, there are scattered patches with slopes greater than 7.7 in the figure, which may be buildings or other man-made structures. When simulating floods, it is necessary to retain the details of rivers and dams, and remove the remaining parts of buildings, woods and roads. In summary, this application uses areas with slope values ​​of 1.6-19.5 as outliers of FABDEM.

[0063] After completing the detection and preliminary analysis of slope outliers in the FABDEM data, it is necessary to identify these outliers and optimize the DEM accordingly. The optimization order is as follows: by superimposing the residential area and tree-covered area data with the outlier data, the overlapping outliers are identified and removed to ensure that buildings and tree coverage do not interfere with flood simulation; by creating river buffers and calculating the average relative height of roads based on COPDEM, river embankments with interception and storage functions are identified and retained; through the spatial intersection analysis of roads and outliers, roads are identified and removed. A large number of gaps will be generated in the process of outlier removal. This application uses a variety of interpolation methods to deal with DEM omissions after the outliers are removed. In addition, this application also restores the actual embankment coverage height based on the vegetation coverage height of the embankment to achieve further optimization of FABDEM.

[0064] 1. Remove building coverage outliers. Use the residential area data as a mask file to overlay the slope grid. Delete the grid values ​​that overlap with the residential area and have a slope value greater than 1.5 in the slope grid, and use it as a new mask file, recorded as slope mask 1. Then overlay FABDEM with mask 1 to delete abnormal elevation data. This step generates an optimized DEM file for the residential area, recorded as DEM 1.

[0065] Removing building outliers will produce rectangular or circular gaps. Inverse distance weighted interpolation is a method of interpolation based on inverse distance weights. The core idea is that the value of the interpolation point is the weighted average of the known points around it, where the weight is inversely proportional to the distance from the known point to the interpolation point. The inverse distance weighted interpolation method assumes that points that are close in space have higher similarity, and points that are closer have a greater impact on the interpolation result. Therefore, this application uses the inverse distance weighted interpolation method to fill the gap area after the building coverage outliers are removed.

[0066] 2. Tree cover outlier removal. Tree cover outlier removal is a serial task of building cover outlier removal. Based on slope mask 1 and DEM 1, outliers are removed by superimposing tree cover areas, generating slope mask 2 and DEM 2 with abnormal slopes and abnormal elevations of buildings and trees removed.

[0067] Irregular gaps will be generated after the tree cover outliers are removed. Natural neighbor interpolation is an interpolation method based on the Thiessen polygon principle. This method calculates the value of the interpolation point by determining the natural neighbors of the interpolation point (i.e., the neighboring points that share edges with the interpolation point in the Thiessen polygon) and using the known values ​​of the natural neighbors for weighted averaging. Natural neighbor interpolation can handle irregularly distributed data points, while taking into account the local characteristics of the original data, ensuring that the interpolation result is naturally smooth at the boundary without obvious boundary effects. Therefore, this application uses natural neighbor interpolation to fill the irregular gaps in DEM 2.

[0068] 3. Elevated and ordinary road removal. This task is also a serial task of tree cover outlier removal. The extracted road is essentially the centerline of the road, which does not contain width and elevation information. The extracted road data can be spatially intersected with the slope mask 2. The outliers that intersect with the road in DEM 2 will be removed to obtain the slope mask 3 that removes the outliers of ordinary roads and elevated roads. Slope mask 3 is superimposed with FABDEM, and the abnormal elevation data is removed to obtain the optimized DEM 3.

[0069] Linear gaps will be generated after the road outliers are removed. Spline interpolation uses piecewise polynomial functions to create smooth and continuous surfaces that can capture subtle changes in terrain. The cubic spline interpolation method is used in DEM processing. This interpolation method ensures the continuity of the first and second order derivatives at the data nodes by constructing a cubic polynomial with continuous first and second order derivatives between each data point. The core idea of ​​spline interpolation is to fit a low-order polynomial between data points while maintaining global smoothness. Therefore, this application uses the spline interpolation method to process the gaps in linear features such as roads in DEM 3.

[0070] 4. Levee extraction. Since FABDEM lacks levee details, this step is based on COPDEM. This task is parallel to the above three tasks, not serial. Levee extraction is performed based on river buffer and nearest road analysis. The process is as follows: Figure 7 As shown. First, a buffer zone with a width of 100 meters is created using the river surface data provided by the National Basic Geographic Information Database as the preliminary range of potential dam distribution. Then the buffer zone outside the river is used as a mask to extract the slope raster and the abnormal slope value. The extracted road data is loaded. If there are multiple roads in the buffer zone outside the river, the road with the highest average relative height is selected as the dam road. The dam road and the outliers are spatially analyzed. The outliers that intersect with the dam road are the real dams, and the other outliers are deleted. This step obtains the dam information within 100 meters of the river, which is recorded as DEM 4.

[0071] After completing the above four processes, we can get DEM 3 without buildings, trees, ordinary roads and elevated outliers, and DEM 4 of the extracted embankment. DEM 4 is used as the updated data to mosaic DEM 3 to get the new DEM 5. DEM5 still has a few outliers because the tree coverage area, building coverage area and road intersection cannot completely cover all outliers. It is necessary to visually determine the outlier type in combination with the historical image map, and then manually delete or retain it, and fill it with the corresponding interpolation method. The final DEM is as follows: Figure 8 shown.

[0072] The dam details in FABDEM are lost, while the dam in COPDEM is relatively complete. However, the dam is artificially high because it is covered by trees, e.g. Fig. 9 shown.

[0073] The upper left image is the optical map image of S City in 2018, the upper right image is COPDEM, the middle image is the street view image of Baidu Map in 2018, and the bottom image is the tree coverage area. It can be found that there are no obviously elevated features around point 1 and point 2, but the elevation range within the box of point 1 and point 2 is 26-45 meters, which does not conform to the actual situation. Therefore, the actual elevation of the dam is significantly false due to the influence of trees. In flood simulation, the false height of the dam will affect the initial time of flood diffusion, so the dam elevation data needs to be corrected. Considering that the tree coverage data is from 2018 to 2020, but COPDEM was generated in 2015, the common tree species in S City include sycamore, Sophora japonica, Fraxinus chinensis, Koelreuteria paniculata, and Zelkova. The growth height of these trees in 3-5 years ranges from 0.2 to 1.5 meters, and the median value of 0.85 meters is taken as the growth height. The raster value of the tree coverage dataset is subtracted by 0.85 as the average tree coverage height in 2015.

[0074] In order to solve the problem of false height of dams, this application adopts a statistical filtering method to correct the dam elevation. This method combines forest height data with median filtering technology to eliminate the influence of surface objects on DEM accuracy. The core idea of ​​the method is based on two assumptions: first, subtracting the forest height from the DEM can obtain a preliminary estimate of the ground elevation; second, applying median filtering can further optimize this estimate, remove outliers and maintain terrain features. The basic idea of ​​median filtering is to replace the pixel value in the center of the window with the median of all pixel values ​​in the window. The advantage is that it can effectively remove noise while retaining edge features.

[0075] Processing Flow Fig.10 As shown in the figure. First, the DEM and forest height data are read and aligned. Subsequently, the data is processed using a block strategy to adapt to the needs of large-scale data sets. For each data block, the forest height is first subtracted, and then a median filter is applied. The median filter process includes three steps: filling the edge of the data, applying a 5x5 filter window, and outputting the results. Finally, the optimized DEM is generated by iteratively processing all data blocks.

[0076] The revised dam is as follows Fig.11 As shown: Fig.11 (a) shows the DEM data of the levee extent from COPDEM, which contains all the features of the surface, including topography and vegetation. Fig.11 (b) shows tree height data for the same area, with green areas indicating no tree cover and red areas representing taller trees. Fig.11 (c) shows the levee range of FABDEM, which shows overestimation of levees and lacks topographic details. Fig.11 (d) shows the revised levee range, which is obtained by subtracting the tree height from the original levee range and applying a median filter. The revised DEM significantly reduces the elevation anomaly caused by trees and more accurately shows the actual terrain details on the ground. This process effectively eliminates the phenomenon of false height of the levee and provides a more accurate levee model, which is of great significance for subsequent flooding simulation analysis.

[0077] This application proposes a DSM-DEM collaborative optimization method for inundation simulation to address the shortcomings of current open source DEM data in inundation simulation. This application selects FABDEM as the optimization object, introduces Landsat-8 satellite images and ESRI historical optical images, and uses deep learning technology to extract key land features. Based on the overall slope characteristic of 0.0002 in the application area, this application transforms the DEM outlier problem into a slope outlier problem, uses slope filtering technology to identify abnormal slopes, and combines the extracted land features and open source tree cover datasets for processing. While removing abnormal elevation data, focus on identifying and retaining embankment data, and correct the problem of excessive embankment elevation caused by tree cover. For data-free areas, a variety of interpolation methods are used to fill them. Through this collaborative optimization method, this application effectively improves the applicability of DEM in inundation simulation, greatly improves the quality of DEM for inundation simulation, and provides important support for improving the accuracy and reliability of inundation simulation.

[0078] Although the present invention is described in this application with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the features described in the various dependent claims and in this application may be combined in a manner different from that described in the original claims. It should also be understood that the features described in conjunction with a single embodiment may be used in other described embodiments.

Claims

1. A DSM-DEM collaborative optimization method for flooding simulation, characterized in that: The following steps are involved: S1. Based on satellite images and historical optical images, deep learning technology is used to extract key features that affect flooding simulation effects. Key features include roads, dams, trees, and residential areas. S2. After extracting the features that affect flood simulation, FABDEM (Forest And Buildings removed Copernicus DEM) is optimized based on these features. The remaining outliers, including buildings, tree cover, viaducts, dams and roads, are identified and processed. The DEM outlier problem is converted into a slope outlier problem. The slope filtering technology is used to identify abnormal slopes, and the extracted features and open source tree cover datasets are combined for processing. While removing abnormal elevation data, the dam data is identified and retained, and the problem of excessive dam elevation caused by tree cover is corrected. For areas with missing data, a variety of interpolation methods are used to fill them in to achieve DEM optimization for flood inundation simulation.

2. The DSM-DEM collaborative optimization method for flooding simulation according to claim 1 is characterized in that: The method for extracting residential areas includes: selecting images of cities around the target area as fine-tuning data sets, segmenting images with equal length and width as fine-tuning data sets, and splitting them into training set, validation set and test set; then manually marking the residential areas in the images to generate formatted data required for model fine-tuning; segmenting the target area images and inputting them into the fine-tuned model to obtain the extraction results.

3. The DSM-DEM collaborative optimization method for flooding simulation according to claim 1 is characterized in that: The road extraction method includes encoding, training, inference and decoding stages. In the encoding stage, the model receives high-resolution satellite images containing road networks as true data, and converts these images into high-dimensional tensor representations through a graph tensor encoding algorithm. This process can retain the spatial distribution information of roads, as well as the encoded topological structure and connection relationship. In the training stage, the encoded graph tensor is paired with the original satellite image for learning, so that the neural network can extract key features from the original image and map it to the corresponding graph tensor representation. Through repeated iterations, the model gradually optimizes its parameters. In the inference stage, when the model faces new, unlabeled satellite images, the trained neural network can automatically generate the corresponding graph tensor encoding. In the decoding stage, the model converts the inferred graph tensor encoding back into a visualized road network structure.

4. The DSM-DEM collaborative optimization method for flooding simulation according to claim 1, characterized in that: The tree cover extraction method includes: randomly selecting multiple areas in the tree crown height map dataset, calculating the intersection and union ratio of the tree cover area and the manually drawn tree cover patches, judging their quality, and quantifying the degree of overlap between them by calculating the ratio of the intersection and union of two areas.

5. The DSM-DEM collaborative optimization method for flooding simulation according to claim 3, characterized in that: By overlaying residential and tree-covered area data with outlier data, the overlapping outliers are identified and removed to ensure that buildings and tree cover do not interfere with flood simulation; by creating river buffers and calculating the average relative height of roads based on COPDEM, river embankments with impoundment functions are identified and retained; and through spatial intersection analysis of roads and outliers, roads are identified and removed.

6. The DSM-DEM collaborative optimization method for flooding simulation according to claim 5, characterized in that: The residential area data is used as a mask file to overlay the slope raster. The raster values ​​that overlap with the residential area and have a slope value greater than a predetermined threshold are deleted from the slope raster, and the raster values ​​are used as a new mask file, recorded as the first slope mask. The first slope mask is then overlaid on FABDEM to delete abnormal elevation data and generate the first DEM file after residential area optimization.

7. The DSM-DEM collaborative optimization method for flooding simulation according to claim 6, characterized in that: Based on the first slope mask and the first DEM, outliers are removed by superimposing tree coverage areas to generate a second slope mask and a second DEM that remove abnormal slopes and abnormal elevations of buildings and trees. Natural neighbor interpolation is used to fill irregular gaps in the second DEM.

8. The DSM-DEM collaborative optimization method for flooding simulation according to claim 7, characterized in that: The extracted road data is spatially intersected with the second slope mask, and the outliers intersecting with the roads in the second DEM are removed to obtain a third slope mask that removes outliers of ordinary roads and elevated roads; the third slope mask is superimposed with FABDEM, and the abnormal elevation data is removed to obtain the optimized third DEM.

9. The DSM-DEM collaborative optimization method for flooding simulation according to claim 1, characterized in that: Firstly, a buffer zone was created using river surface data provided by the National Basic Geographic Information Database as the preliminary range of potential levee distribution. Then, the buffer zone outside the river was used as a mask to extract the slope raster and the abnormal slope values. The extracted road data was loaded. If there were multiple roads in the buffer zone outside the river, the road with the highest average relative height was selected as the levee road. The levee road and the outliers were spatially analyzed. The outliers that intersected with the levee road were the real levees. The other outliers were deleted to obtain the levee information of the river.

10. The DSM-DEM collaborative optimization method for flooding simulation according to claim 9, characterized in that: First, the DEM and forest height data are read and aligned. Then, the data is processed using a block strategy. For each data block, the forest height is first subtracted and then a median filter is applied. The median filter process includes three steps: filling the edge of the data, applying a 5x5 filter window, and outputting the results. Finally, all data blocks are iteratively processed to generate the optimized DEM.

Images