A dam identification method and device based on remote sensing pictures

Abstract

The application provides a dam identification method and device based on remote sensing pictures. The method comprises the following steps: obtaining a remote sensing picture to be identified; pre-processing the remote sensing picture to obtain a pre-processed remote sensing picture; inputting the pre-processed remote sensing picture into a preset dam identification model to obtain a candidate target region; extracting the candidate target region to obtain a flow cut-off node on a long flow basin and geographical environment information; and determining a dam position according to the flow cut-off node on the long flow basin and the geographical environment information. The application can accurately identify the position of the dam by pre-processing the remote sensing picture to be identified, then using the preset dam identification model to obtain the candidate target region, and then extracting the flow cut-off node on the long flow basin and the geographical environment information, thereby providing data support for dam management, monitoring and planning fields, and having the advantages of accurate identification and high effectiveness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dam identification, and further relates to a dam identification method and device based on remote sensing pictures. BACKGROUND

[0002] A dam is a kind of artificial structure, usually built on rivers or other water bodies. As a key water infrastructure, the dam not only plays a key role in water resource allocation, power generation, etc., but also plays an irreplaceable role in flood control and disaster prevention. Therefore, accurately obtaining the number, location and spatial distribution information of the dam is of great importance to the rational management of water resources, the reduction of flood risk, the promotion of renewable energy development, the protection of the ecological environment and the promotion of sustainable development. Although the importance of the dam has been recognized, due to the differences in local policies and financial resources, it is not easy to accurately obtain the spatial distribution of the dam. The existing distribution of the dam comes from field survey and construction data, which requires a large amount of manpower, material resources and financial resources, and has a high cost.

[0003] With the development of remote sensing technology, it is possible to extract ground objects from remote sensing images. However, most of the dams in the remote sensing images are located in complex and changeable geographical environments, and are surrounded by buildings or other ground objects. The low resolution may not be enough to clearly identify small or hidden dams. In addition, remote sensing data will inevitably be disturbed by factors such as clouds, atmosphere and reflectivity, thereby reducing the quality of the data and increasing the difficulty of data preprocessing and feature extraction, which directly affects the effect of dam identification and monitoring. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a dam identification method and device based on remote sensing pictures, so as to solve the problems of high cost and poor identification result of the existing dam identification method.

[0005] To solve the above technical problems, the technical scheme of the present application is as follows:

[0006] A dam identification method based on remote sensing pictures, comprising:

[0007] obtaining a remote sensing picture to be identified;

[0008] preprocessing the remote sensing picture to obtain a preprocessed remote sensing picture;

[0009] inputting the preprocessed remote sensing picture into a preset dam identification model to obtain a candidate target region; the preset dam identification model is obtained by training a preset network model according to a dam body picture containing a dam;

[0010] extracting the candidate target region to obtain a flow cut-off node on a long watershed and geographical environment information;

[0011] According to the cut-off node and geographical environment information on the long river basin, the dam position is determined.

[0012] Optionally, the remote sensing picture is preprocessed to obtain a preprocessed remote sensing picture, including:

[0013] The remote sensing picture is preprocessed by one or more of atmospheric correction, orthorectification and mosaic clipping to obtain a preprocessed remote sensing picture.

[0014] Optionally, the preprocessed remote sensing picture is input into a preset dam identification model to obtain a candidate target region, including:

[0015] The preprocessed remote sensing picture is subjected to convolution processing to obtain a shared feature map;

[0016] The shared feature map is input into a region generation network for region selection to obtain a candidate region;

[0017] The candidate region is subjected to object classification and bounding box regression processing to obtain a candidate target region.

[0018] Optionally, the training process of the preset dam identification model includes:

[0019] A plurality of dam body pictures containing a dam are obtained according to a preset condition;

[0020] The dam body pictures containing a dam are subjected to enhancement processing to obtain enhanced dam body pictures;

[0021] The enhanced dam body pictures are screened and labeled to obtain an effective sample set;

[0022] The effective sample set is divided according to a preset ratio to obtain training samples and verification samples;

[0023] The training samples and the verification samples are used to train and verify a preset network model, respectively, to obtain a preset dam identification model.

[0024] Optionally, the training samples and the verification samples are used to train and verify a preset network model, respectively, to obtain a preset dam identification model, including:

[0025] The training samples are used to train a plurality of preset network models to obtain a plurality of trained dam identification models;

[0026] The verification samples are used to verify a plurality of the trained dam identification models to obtain a plurality of verification results;

[0027] According to the verification results, a preset dam identification model is determined.

[0028] Optionally, the candidate target region is extracted to obtain a flow cut-off node on a long flow basin, comprising:

[0029] The candidate target region is extracted to obtain a reflectance of a green band, a reflectance of a red band and elevation data;

[0030] According to the reflectance of the green band, the reflectance of the red band and a formula A normalized difference water index is obtained; wherein NDWI is a normalized difference water index, ρ (Green) is a reflectance of a green band, ρ (NIR) is a reflectance of a red band;

[0031] According to the normalized difference water index and the elevation data, an aspect ratio of a circumscribed rectangle is obtained;

[0032] According to the aspect ratio of the circumscribed rectangle, a flow cut-off node on a long flow basin is obtained.

[0033] Optionally, the candidate target region is extracted to obtain geographical environment information, comprising:

[0034] The candidate target region is extracted to obtain elevation data;

[0035] The elevation data is used to obtain a ground roughness and a ground undulation degree.

[0036] Optionally, the method further comprises:

[0037] The elevation data is used to obtain a dam proportion detection value;

[0038] According to the flow cut-off node on the long flow basin, the geographical environment information and the dam proportion detection value, a dam position is determined.

[0039] Optionally, according to the flow cut-off node on the long flow basin, the geographical environment information and the dam proportion detection value, a dam position is determined, comprising:

[0040] According to a preset weight, the water body information, geographical environment information and the dam proportion detection value, a dam position is determined.

[0041] Another aspect of the application provides a dam identification device based on a remote sensing picture, comprising:

[0042] An acquisition module is configured to acquire a remote sensing picture to be identified;

[0043] A preprocessing module is configured to preprocess the remote sensing picture to obtain a preprocessed remote sensing picture;

[0044] An input module is configured to input the preprocessed remote sensing picture into a preset dam identification model to obtain a candidate target region; the preset dam identification model is obtained by training a preset network model according to a dam body picture containing a dam;

[0045] An extraction module is configured to extract the candidate target region to obtain a flow-cutting node on a long flow basin and geographical environment information;

[0046] A determination module is configured to determine a dam position according to the flow-cutting node on the long flow basin and the geographical environment information.

[0047] The above scheme of the present application at least has the following beneficial effects:

[0048] The above scheme of the present application can accurately identify the position of the dam in the remote sensing picture to be identified by preprocessing the remote sensing picture to be identified, then obtaining the candidate target region by using the preset dam identification model, and then extracting the flow-cutting node on the long flow basin and the geographical environment information, thereby providing data support for dam management, monitoring and planning fields, and having the advantages of accurate identification and high effectiveness. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is a flow diagram of the dam identification method based on a remote sensing picture in the embodiment of the present application;

[0050] Figure 2 is a specific flow diagram of the dam identification method based on a remote sensing picture in the embodiment of the present application;

[0051] Figure 3 is a structural diagram of the dam identification device based on a remote sensing picture in the embodiment of the present application. DETAILED DESCRIPTION

[0052] Exemplary embodiments of the present application will be described herein below with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided so that the present application can be more thoroughly understood and the scope of the present application can be accurately conveyed to those skilled in the art.

[0053] As shown in Figure 1 , an embodiment of the present application proposes a dam identification method based on a remote sensing picture, comprising:

[0054] Step 11, obtaining a remote sensing picture to be identified;

[0055] Step 12, preprocessing the remote sensing picture to obtain a preprocessed remote sensing picture;

[0056] Step 13, input the pre-processed remote sensing picture into a preset dam identification model to obtain a candidate target region; the preset dam identification model is obtained by training a preset network model according to dam body pictures containing a dam;

[0057] Step 14, extracting the candidate target region to obtain a flow cut-off node and geographical environment information on a long flow basin;

[0058] Step 15, determining a dam position according to the flow cut-off node and geographical environment information on the long flow basin.

[0059] The dam identification method based on a remote sensing picture provided by the embodiment of the application can accurately identify whether a remote sensing picture to be identified contains a dam and the position of the dam by pre-processing the remote sensing picture to be identified, using a preset dam identification model to obtain a candidate target region, and extracting the candidate target region to obtain a flow cut-off node and geographical environment information on a long flow basin, thereby providing data support for dam management, monitoring and planning, and having the advantages of high identification accuracy and high effectiveness.

[0060] In an optional embodiment of the application, step 12 comprises:

[0061] The remote sensing picture is pre-processed by one or more of atmospheric correction, orthorectification and mosaic clipping to obtain a pre-processed remote sensing picture.

[0062] Specifically, most of the dams in remote sensing images are located in complex and changeable geographical environments and are surrounded by buildings or other ground obstructions, and a low resolution may not be sufficient to clearly identify small or hidden dams, and remote sensing data is inevitably disturbed by factors such as clouds, atmosphere and reflectivity, thereby reducing the data quality. Therefore, before identifying the remote sensing picture to be identified, the remote sensing picture to be identified is pre-processed, and the pre-processing mode can be one or more of atmospheric correction, orthorectification, mosaic clipping and other operations. The image specifications of different sensors are different, and in specific implementation, the type of the sensor used to obtain the remote sensing picture needs to be considered. After pre-processing the remote sensing picture to be identified, an image without offset and with high resolution can be obtained, which is conducive to improving the identification accuracy and work efficiency of the method for the dam.

[0063] In an optional embodiment of the application, the training process of the preset dam identification model in step 13 comprises:

[0064] Step 131, obtaining a plurality of dam body pictures containing a dam according to a preset condition;

[0065] Specifically, the preset condition can be that a dam vector buffer zone is generated with dam vector point data as the center, and a plurality of dam body pictures containing the dam are obtained from Google Earth Pro (a geographic information system software) as the condition;

[0066] Step 132, performing enhancement processing on the dam body pictures containing the dam to obtain enhanced dam body pictures;

[0067] A large number of training samples are required for training the preset network model, therefore, the sample data set is enhanced by using methods such as mirror reflection, scale transformation, color transformation, noise disturbance, brightness transformation and position transformation, so as to enhance the robustness and generalization ability of the model and avoid model overfitting;

[0068] Step 133, screening and labeling the enhanced dam body pictures to obtain an effective sample set;

[0069] The dam in the enhanced dam body pictures is labeled, and unqualified samples such as repeated, deformed, incomplete and unrecognizable samples are removed, and finally the effective sample set is obtained;

[0070] Step 134, dividing the effective sample set according to a preset proportion to obtain training samples and verification samples;

[0071] Specifically, 70% of the effective sample set is used as the training samples, and 30% is used as the verification samples;

[0072] Step 135, training and verifying the preset network model by using the training samples and the verification samples respectively to obtain a preset dam identification model.

[0073] The preset network model is trained by using the training samples, and the trained preset network model is verified by using the verification samples, and in specific implementation, a loss function can be set to adjust the parameters and weights of the preset network model, so as to increase the accuracy of the model in dam identification.

[0074] In an optional embodiment of the present application, step 135 comprises:

[0075] Step 1351, training a plurality of preset network models by using the training samples to obtain a plurality of trained dam identification models;

[0076] Step 1352, verifying a plurality of the trained dam identification models by using the verification samples to obtain a plurality of verification results;

[0077] Step 1353, determining the preset dam identification model according to the verification results.

[0078] In the implementation, the plurality of preset network models can be trained by using the training samples, the plurality of trained dam identification models can be verified by using the verification samples, and the preset network model with the best identification result can be selected as the preset dam identification model, thereby improving the identification efficiency and accuracy of the dam in the remote sensing picture to be identified.

[0079] In an optional embodiment of the present application, step 13 comprises:

[0080] Step 136, performing convolution processing on the preprocessed remote sensing picture to obtain a shared feature map;

[0081] Step 136, performing convolution processing on the preprocessed remote sensing picture to obtain a shared feature map;

[0082] Step 136, performing convolution processing on the preprocessed remote sensing picture to obtain a shared feature map;

[0083] The shared feature map is input into a region generation network to select a region, and a candidate region is obtained.

[0084] Step 138, performing object classification and bounding box regression processing on the candidate region to obtain a candidate target region.

[0085] The feature representation of each candidate region is extracted, and then each candidate region is mapped to a fixed-size feature map by using a pooling operation, and object classification and bounding box regression processing are performed by using a full connection layer to obtain a candidate target region, in which the position of the dam exists.

[0086] In the embodiment, a cascade structure is used, and each level has an independent target detector to perform screening at different stages; the target detector at each level first performs target detection to assign a confidence score to the candidate target region, and then different loss functions are used for training, and the preset dam identification model allows end-to-end training, so that the detectors at different levels can share the feature extractor, accelerate the training process, and improve the detection accuracy in stages.

[0087] In an optional embodiment of the present application, step 14 comprises:

[0088] Step 141, extracting the candidate target region to obtain the reflectance of the green band, the reflectance of the red band, and the height data.

[0089] Step 142, obtaining a normalized difference water index according to the reflectance of the green band, the reflectance of the red band, and the formula wherein, NDWI is the normalized difference water index, ρ (Green) is the reflectance of the green band, ρ (NIR)reflectance of red band;

[0090] Step 143, obtaining a length-width ratio of the circumscribed rectangle according to the normalized difference water index and the elevation data;

[0091] Step 144, obtaining a cut-off node on a long watershed according to the length-width ratio of the circumscribed rectangle.

[0092] Specifically, the remote sensing picture contains elevation data (DEM data), the normalized difference water index is used to extract water body information from the candidate target region of the remote sensing picture, and then the elevation data is used to identify and extract the river network in the candidate target region of the picture. According to the relative size, geographical position and flow, etc. Characteristics are classified and graded. Finally, the length-width ratio of the circumscribed rectangle is calculated to determine the cut-off node on the long watershed. The calculation formula of the length-width ratio of the circumscribed rectangle is: AR=L / W, wherein AR is the length-width ratio of the circumscribed rectangle, L represents the length of the circumscribed rectangle, and W represents the width of the circumscribed rectangle. The length and width of the circumscribed rectangle can be obtained from the elevation data.

[0093] In an optional embodiment of the present application, step 14 comprises:

[0094] Step 145, extracting the candidate target region to obtain elevation data;

[0095] Step 146, obtaining ground roughness and ground relief degree by using the elevation data.

[0096] When determining the position of the dam, the terrain feature can provide the information of the geographical environment, which helps to identify the correct dam position and effectively eliminate the influence of background information. The ground roughness is the ratio of the earth surface area in a specific area to its projection area, which is also a macro index reflecting the landform. The elevation data is usually presented in the form of a grid, and each grid cell contains a height value. According to the definition of ground roughness (M), the ratio of the surface area of each grid cell to its projection area is: M=(AC×AB) / (AC×AC)=1 / cosα. ABC is the longitudinal profile of a grid cell, and α is the slope of the sub-grid cell. Then the area of AB (ABAC) is the surface area of the grid, and the area of AC (ACAC) is the projection area of the grid, wherein cosα=AC / AB. The ground relief degree is the maximum relative height difference per unit area, which can reflect the relative height difference of the ground and is a quantitative index for describing the landform. The formula is: R=Hmax-Hmin, wherein R represents the ground relief degree, Hmax represents the maximum height value per unit area, and Hmin represents the minimum height value per unit area. The maximum height value per unit area and the minimum height value per unit area can be obtained from the elevation data.

[0097] In an optional embodiment of the present application, step 14 further comprises: obtaining a dam proportion detection value by using the elevation data.

[0098] Step 15 further comprises: determining the dam location according to the flow-cutting node on the long watershed, the geographic environment information and the dam proportion detection value.

[0099] A proper buffer zone is established around the center point of the dam candidate area, and whether the candidate area is a dam is determined by the ratio of the water area on both sides of the dam candidate area. Thus, important information about water resources and terrain features is provided. The calculation formula of the dam proportion detection value is: Sp=U / D, wherein Sp represents the dam proportion detection value, U represents the larger water area in the buffer zone, and D represents the smaller water area in the buffer zone. The larger water area in the buffer zone and the smaller water area in the buffer zone can be obtained from the elevation data.

[0100] In an optional embodiment of the present application, step 15 comprises:

[0101] The dam location is determined according to the preset weight, the water body information, the geographic environment information and the dam proportion detection value.

[0102] The dam location is determined according to the preset weight, the water body information, the geographic environment information and the dam proportion detection value.

[0103] As shown in FIG. 1, one specific embodiment of the dam identification method based on remote sensing pictures provided by the present application is as follows: Figure 2

[0104] ​Step 21, making a sample set; the currently disclosed dam samples are all ordinary digital images taken at close range, there is no public remote sensing image data set, and the design standards of dams in different countries are also different. Therefore, a sample data set for training a preset dam recognition model needs to be made. A circular buffer area is generated around a dam point vector in a global data set containing more than 38000 geographically referenced dams, and dam body pictures are obtained from Google Earth Pro (a geographic information system software) or a global dam database to obtain a preliminary data training set of the dam. The training set includes dam pictures of different types, sizes, and geographical locations, which helps to capture the diversity of dams. Considering that deep learning requires a large number of training samples, the sample data set is enhanced using methods such as mirror flipping, scale transformation, color transformation, noise disturbance, brightness transformation, and position transformation to enhance the robustness and generalization ability of the model, avoid model overfitting, and finally store and manage using YOLO data set, VOC data set, and COCO data set formats. The dam in the sample set is labeled after the sample enhancement process, and the quality of the sample is checked during the labeling process, and the unqualified samples such as duplicates, deformations, incompleteness, and unidentifiable samples are filtered and labeled, and finally an effective sample set is obtained. The labeled sample set is randomly divided into a training sample set and a test sample set, of which the training sample accounts for 70% and the validation sample accounts for 30%.

[0105] Step 22, determining a dam recognition model; a plurality of preset network models are trained using the training samples, and the trained preset network models are verified using the validation samples, and the best model is selected as the preset dam recognition model, as shown in Figure 2 The accuracy, recall rate, and inference speed can be used to evaluate the recognition effect of each preset network model. The accuracy reflects the ratio of the number of correctly detected dam bounding boxes to the total number of detected bounding boxes, while the recall rate indicates the ratio of the number of correctly detected dams by the preset network model in the validation data set to the actual number of existing dams. The calculation formulas of the accuracy, recall rate, and inference speed are as follows:

[0106]

[0107]

[0108]

[0109]

[0110] Wherein, IOU is the intersection over union, used to judge the model, the greater the value of IOU, the more accurate the model prediction, area(detection∩groundttruth) is the intersection area of two bounding boxes, area(detection∪groundtruth) is the union area of two bounding boxes; P represents the accuracy, R represents the recall, TP represents the prediction is 1, the prediction is correct, that is, the actual is 1; FP represents the prediction is 1, the prediction is wrong, that is, the actual is 0; FN represents the prediction is 0, the prediction is wrong, that is, the actual is 1; AP is the inference speed, the higher the AP value, the more accurate detection results the model can provide under various IOU thresholds. In the embodiment, the selected preset network model adopts a cascade structure, and each level has an independent target detector for screening at different stages. The target detector at each level first performs target detection, configures a confidence score for the candidate target region, and then uses different loss functions for training; the model allows end-to-end training, so that the detectors at different levels can share the feature extractor, speed up the training process, and improve the accuracy of detection in stages. The P-R curve shows the accuracy and recall changes under different intersection over union (IOU) thresholds. The area (AP) enclosed by the P-R curve is a comprehensive index considering the precision and recall of the model under different IOU threshold conditions. A high AP value indicates that the model can provide high-precision detection results under various IOU thresholds, and the selected preset network model in the embodiment has relatively small fluctuations and the largest enclosed area under two IOU thresholds, indicating that the model can reliably find the target under various conditions.

[0111] Alternatively, the extraction rate, the missing extraction rate and the false extraction rate can be used to comprehensively evaluate the extraction effect of various models. The extraction rate represents the ratio of the number of correctly extracted dams to the actual number of dams, and the formula is: E=R / T, wherein E is the extraction rate, R is the number of correctly extracted dams, and T is the actual number of dams; the missing extraction rate represents the ratio of the number of unextracted dams to the actual number of dams, and the formula is: M=(T-R) / T, wherein M is the missing extraction rate, R is the number of correctly extracted dams, and T is the actual number of dams; the false extraction rate represents the ratio of the number of incorrectly extracted dams to the actual number of dams, and the formula is: F=K / T, wherein F is the false extraction rate, K is the number of incorrectly extracted dams, and T is the actual number of dams. The higher the extraction rate and the lower the missing extraction rate and the false extraction rate, the better the effect of the model, and the best preset network model is selected as the preset dam identification model.

[0112] Step 23, determining a candidate target region; after obtaining the remote sensing picture to be recognized, basic pretreatment is performed, such as atmospheric correction, orthorectification, tiling and cutting, to obtain an image without deviation and with high resolution. The image specifications of different sensors are different, and the pretreatment method needs to be specific to the selected image type; the remote sensing picture after pretreatment is input into the determined preset dam recognition model to obtain a remote sensing picture containing a candidate target region. The candidate target region is a plurality of point positions for bounding the possible dam in the remote sensing picture.

[0113] Step 24, spatial constraint condition; the dam is an important water conservancy element, and hydrological analysis plays a crucial role in the dam location selection and design process. To determine the dam location, first, the normalized difference water index (NDWI) is used to extract water body information from the candidate target region. The calculation formula of the normalized difference water index is: wherein, NDWI is the normalized difference water index, p (Green) is the reflectance of the green band, p (NIR) is the reflectance of the red band, second, the river network vector data in the region is identified and extracted using the elevation data, the river network grading diagram and the circumscribed matrix length-width ratio are used to determine the long watershed with large flow, and the long watershed (i.e. the main watershed) is determined according to the relative size, geographical location and flow characteristics, etc. Finally, the circumscribed rectangle length-width ratio is calculated to determine the flow cut-off node on the long watershed. The calculation formula of the circumscribed rectangle length-width ratio (AR) is: AR = L / W, wherein L represents the length of the circumscribed rectangle, and W represents the width of the circumscribed rectangle. The length of the circumscribed rectangle and the width of the circumscribed rectangle can be obtained from the elevation data. Using 200 square meters as the threshold value can effectively remove most of the ponds, reservoirs, hot springs and other water areas where dams are unlikely to exist;

[0114] When determining the location of the dam, the terrain feature can provide information about the geographical environment, help identify the correct dam location and effectively eliminate the influence of background information. Topographic index is the most basic natural geographical element, and the natural factor that has the greatest impact on human production and life. The extraction of topographic index plays an important role in the research of soil erosion, land use, land resource evaluation, urban planning, etc. The extraction of topographic index based on ArcGIS is mostly completed by using elevation data. According to the different scales of the research area, there are many factors of topographic index, mainly including slope, aspect, terrain relief degree, and ground roughness. Ground roughness is the ratio of the earth's surface area to its projected area in a specific region, and it is also a macro index reflecting the topography of the earth's surface; according to the common sense of dam, the ground roughness of the dam area is higher, while the ground roughness of the urban area is lower, so the ground roughness can be used to remove the interference points in the candidate target area; the terrain of the dam area changes sharply, and the corresponding layer difference will be large; elevation data is usually presented in the form of a grid, and each grid cell contains a height value. According to the definition of ground roughness, the ratio of the surface area of each grid cell to its projected area is H=(AC×AB) / (AC×AC)=1 / cosα. ABC is the longitudinal profile of a grid cell, and α is the slope of the sub-grid cell, then the area of AB face (ABAC) is the surface area of this grid, and the area of AC face (ACAC) is the projected area of this grid, where cosα=AC / AB. The ground relief degree is the maximum relative height difference in unit area, which can reflect the relative height difference of the ground and is a quantitative index to describe the geomorphic form, and the formula is G=Hmax-Hmin, where G represents the ground relief degree, Hmax represents the maximum height value in unit area, and Hmin represents the minimum height value in unit area. Among them, the maximum height value in unit area and the minimum height value in unit area can be obtained from the elevation data.

[0115] With the center point of the dam candidate area as the origin, a proper buffer zone is established around it, and whether the candidate target area in the remote sensing picture is a dam is judged by the ratio of the water area on both sides of the dam candidate area, thereby providing important information about water resources and terrain features. The formula for calculating the dam proportion detection value is: Sp=U / D, where Sp represents the dam proportion detection value within a certain range of the buffer zone, U represents the larger of the water area within the buffer zone, and D represents the smaller of the water area within the buffer zone. Both the larger of the water area within the buffer zone and the smaller of the water area within the buffer zone can be obtained from the elevation data. The dam proportion detection value is calculated, and if the dam proportion detection value is greater than a certain threshold value, it is considered that the target is a dam. In this embodiment, the threshold value is set to 1 (if the original image of the dam is relatively complete, it can be appropriately relaxed), and the dam identification needs to calculate the area ratio of the upstream and downstream of the dam river, especially in the case of a buffer zone 300 meters away from the dam. If the upstream area ratio is similar to the downstream area ratio, it may indicate that the area is not a dam; but if the dam proportion detection value ratio is greater than 1, it may be a sign of a dam. As shown in Table 1, according to the dam proportion detection value of each point, the point with a ratio of 1 can be excluded.

[0116] Table 1: Area ratio of water area on both sides of different points

[0117]

[0118] Step 25, Analytic Hierarchy Process; After excluding the point with a ratio of 1 according to the dam proportion detection value of each point, the break flow node on the long river basin, the ground roughness, the ground relief degree, and the dam proportion detection value need to be assigned different attributes through the Analytic Hierarchy Process (AHP), and compared with each other to determine the judgment matrix, as shown in Table 2. Finally, the unimportant point in the candidate target area is deleted, which helps to improve efficiency and accuracy; in the terrain analysis, the terrain undulation of some areas is not easy to intuitively distinguish, therefore, the weight of the ground roughness (A2) and the ground relief degree (A3) is mainly used to reduce the interference point in the candidate target area.

[0119] Firstly, the judgment matrix is constructed according to the break flow node on the long river basin, the ground roughness, the ground relief degree, and the dam proportion detection value, as shown in Table 2; the maximum eigenvalue Consistency index Consistency ratio CR=CI / RI judges whether the judgment matrix meets the consistency requirement, when CR<0.1, it means that the matrix is reasonable, otherwise, the matrix needs to be adjusted; through the judgment matrix of Table 2, CR=0.02318<0.1, which means that the judgment matrix is reasonable. Wherein, λmax is the maximum eigenvalue, N is the dimension of the matrix, A is the matrix, ω is the eigenvector of the matrix, CI is the consistency index, CR is the consistency ratio, RI is the consistency check value, and RI can be obtained by table, as shown in Table 3.

[0120] Table 2 judgment matrix

[0121]

[0122] Table 3 consistency check RI value

[0123]

[0124] Step 26, determine the dam position; as shown in Table 4, the importance ratio (correlation in the table) of multiple point positions, A2 is the ground roughness, A3 is the ground undulation degree, the weight in the table can be set according to the actual situation; compare the importance ratio with the preset value, select the point position, delete the point position with the importance ratio less than the preset value; as shown in Table 4, 15 point positions are sorted according to the correlation from large to small, then the longitude and latitude of the point position are input from the Maponline plug-in, and the point position is matched on the map. And find two point positions (the adjacent two points are dam point and non-dam point respectively), then the correlation of the non-dam point position is the threshold value of the excluded interference point position. According to the value of the correlation in Table 4, it can be inferred that the irrelevant point positions are 1, 3, 6, 8 and the like, so these interference points can be excluded. The point position after excluding the interference point is the position where the dam may exist, and the point position is determined as the dam position.

[0125] Table 4 importance ratio result of different point positions

[0126] Z Weight Point 1 Point 2 Point 3 Point 4 Point 5 Point 6 …… Point 14 Point 15 A2 0.10732 0.7543 0.5981 0.7124 0.6789 0.5213 0.1890 …… 0.3289 0.4390 A3 0.06125 0.3678 0.4529 0.1437 0.2876 0.4967 0.6542 …… 0.7654 0.6012 Correlation 0.4738 0.6127 0.2958 0.7461 0.5183 0.3872 …… 0.6519 0.8291

[0127] The dam identification method based on remote sensing pictures provided by the embodiment of the application first acquires dam pictures from a GOODD (global dam geographic reference database) to construct a dam data set; secondly, the data set is used to train a plurality of preset network models respectively, and high-security dam position information is extracted based on a parameter threshold. In addition, three spatial constraint strategies and an analytic hierarchy process (AHP) are used to further reduce the influence of other factors in the background area. In order to verify the applicability and efficiency of the method, the Q province is taken as an experimental area for testing, and dam images in a Google Earth Pro (geographic information system software) database are taken as detection standards. The experimental results show that the recognition accuracy of the dam identification method based on remote sensing pictures provided by the embodiment of the application reaches 94.73%. At the same time, 6 unrecorded dams in the Q province are detected by the experiment, and 4 false dams are detected. The reservoirs and dams are distinguished, and the global dam geological reference database is further improved and supplemented. On the basis of open geographic data products, the overall framework based on the organic fusion of deep learning target detection technology and spatial constraint strategy can realize more efficient and accurate dam remote sensing intelligent identification.

[0128] The accuracy of dam extraction not only depends on the accuracy of the dam identification model, but also is closely related to factors such as the quality of the remote sensing picture, the water body extraction effect and the post-processing accuracy. The data training set made by the embodiment of the present application includes single-type high-quality dam pictures of different types, sizes and geographical locations. The accuracy of image quality and category can not only help improve the performance of the dam identification model, but also more clearly reflect the texture features and morphological features of the dam. The dam scale detection value proposed by the present application reduces the range of dam discrimination, effectively removes the interference points in the study area, and makes a great contribution to the extraction of the dam candidate region. The present application has reliability in dam target detection, can effectively identify and locate the target dam, not only overcomes the limitations and challenges in the traditional method, but also can discover unknown dams, thereby providing new insights and possibilities for dam management, monitoring and planning fields.

[0129] As shown in Figure 3 The embodiment of the present application provides a dam identification device 100 based on a remote sensing picture, which comprises:

[0130] The acquisition module 101 is used for acquiring a remote sensing picture to be identified.

[0131] The preprocessing module 102 is used for preprocessing the remote sensing picture to obtain a preprocessed remote sensing picture.

[0132] The input module 103 is used for inputting the preprocessed remote sensing picture into a preset dam identification model to obtain a candidate target region. The preset dam identification model is obtained by training a preset network model according to a dam body picture containing a dam.

[0133] The extraction module 104 is used for extracting the candidate target region to obtain a flow-cutting node on a long watershed and geographical environment information.

[0134] The determination module 105 is used for determining the dam position according to the flow-cutting node on the long watershed and the geographical environment information.

[0135] The dam identification device based on a remote sensing picture proposed by the embodiment of the present application can accurately identify whether the remote sensing picture to be identified contains a dam and the position of the dam by preprocessing the remote sensing picture to be identified, then using the preset dam identification model to obtain the candidate target region, and then extracting the flow-cutting node on the long watershed and the geographical environment information. The present application provides data support for dam management, monitoring and planning fields, and has the advantages of accurate identification and high effectiveness.

[0136] Optionally, the preprocessing of the remote sensing picture to obtain the preprocessed remote sensing picture comprises:

[0137] The remote sensing picture is preprocessed by one or more of atmospheric correction, orthorectification and mosaic cropping to obtain a preprocessed remote sensing picture.

[0138] Optionally, the preprocessed remote sensing picture is input into a preset dam identification model to obtain a candidate target region, including:

[0139] The preprocessed remote sensing picture is subjected to convolution processing to obtain a shared feature map.

[0140] The shared feature map is input into a region generation network for region selection to obtain a candidate region.

[0141] The candidate region is subjected to object classification and bounding box regression processing to obtain a candidate target region.

[0142] Optionally, the training process of the preset dam identification model includes:

[0143] A plurality of dam body pictures containing a dam are obtained according to a preset condition.

[0144] The dam body pictures containing a dam are subjected to enhancement processing to obtain enhanced dam body pictures.

[0145] The enhanced dam body pictures are screened and labeled to obtain an effective sample set.

[0146] The effective sample set is divided according to a preset ratio to obtain training samples and verification samples.

[0147] The preset network model is trained and verified using the training samples and the verification samples, respectively, to obtain a preset dam identification model.

[0148] Optionally, the preset network model is trained and verified using the training samples and the verification samples, respectively, to obtain a preset dam identification model, including:

[0149] A plurality of preset network models are trained using the training samples to obtain a plurality of trained dam identification models.

[0150] The plurality of trained dam identification models are verified using the verification samples to obtain a plurality of verification results.

[0151] The preset dam identification model is determined according to the verification results.

[0152] Optionally, the candidate target region is extracted to obtain a flow interruption node on a long watershed, including:

[0153] The candidate target region is extracted to obtain reflectance of a green band, reflectance of a red band and elevation data.

[0154] According to the reflectance of the green wave band, the reflectance of the red wave band and the formula to obtain a normalized difference water index; wherein, NDWI is the normalized difference water index, ρ (Green) is the reflectance of the green wave band, ρ (NIR) is the reflectance of the red wave band;

[0155] According to the normalized difference water index and the elevation data, obtain an aspect ratio of a circumscribed rectangle;

[0156] According to the aspect ratio of the circumscribed rectangle, obtain a cut-off node on a long watershed.

[0157] Optionally, the candidate target region is extracted to obtain geographical environment information, including:

[0158] The candidate target region is extracted to obtain elevation data;

[0159] The elevation data is used to obtain ground roughness and ground relief degree.

[0160] Optionally, the extraction module is further configured to use the elevation data to obtain a dam proportion detection value;

[0161] The determination module is further configured to determine a dam location according to the cut-off node on the long watershed, the geographical environment information and the dam proportion detection value.

[0162] Optionally, determining a dam location according to the cut-off node on the long watershed, the geographical environment information and the dam proportion detection value, includes:

[0163] According to a preset weight, the water body information, the geographical environment information and the dam proportion detection value, determine a dam location.

[0164] It should be noted that the device corresponds to the above method, and all implementation manners in the above method embodiments are applicable to the embodiments of the device, and the same technical effects can be achieved. In this embodiment, no longer be described.

[0165] The above is the preferred embodiment of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should be considered within the scope of the present application.

Claims

1. A dam identification method based on remote sensing images, characterized in that: include: Obtain remote sensing images to be identified; Preprocessing the remote sensing image to obtain a preprocessed remote sensing image; Inputting the pre-processed remote sensing image into a preset dam recognition model to obtain a candidate target area; The preset dam recognition model is obtained by training a preset network model based on a dam body picture containing the dam; Extracting the candidate target area to obtain the flow-breaking nodes and geographical environment information on the long river basin; Determine the location of the dam based on the flow-breaking nodes and geographical environment information on the long river basin; The pre-processed remote sensing image is input into a preset dam recognition model to obtain candidate target areas, including: Inputting the preprocessed remote sensing image into a convolutional network for convolution processing to obtain a shared feature map; The shared feature map is input into a region generation network for region selection to obtain a candidate region. Specifically, a sliding window mechanism and an anchor frame operation are used to define a bounding box and perform region selection on the dam position in the shared feature map to obtain a candidate region. The candidate regions are subjected to object classification and bounding box regression processing to obtain candidate target regions. Specifically, feature representations of each candidate region are extracted, and then each candidate region is mapped to a fixed-size feature map using a pooling operation. Object classification and bounding box regression processing are performed through a fully connected layer to obtain candidate target regions. The candidate target regions contain locations of dams. The candidate target regions include multiple points. The training process of the preset dam identification model includes: Acquire multiple images containing the dam according to pre-defined conditions. The pre-defined conditions are to generate a dam vector buffer centered around the dam vector point data. A circular buffer is generated centered around a dam point vector from a global dataset containing over 38,000 georeferenced dams. Dam images are obtained from GIS software or the Global Dam Database. Performing enhancement processing on the dam body image containing the dam to obtain an enhanced dam body image; wherein the enhancement processing on the dam body image containing the dam is performed using mirror flipping, scale transformation, color transformation, noise perturbation, brightness transformation and position transformation; Screening and labeling the enhanced dam body images to obtain a valid sample set; wherein, labeling the dams in the enhanced dam body images, removing duplicate, deformed, incomplete, and unrecognizable unqualified samples to obtain a valid sample set; The valid sample set is divided according to a preset ratio to obtain training samples and verification samples; wherein 70% of the valid sample set is used as training samples and 30% is used as verification samples; Using the training samples and the verification samples to train and verify the preset network model respectively, to obtain a preset dam recognition model; The preset network model is trained and verified using the training samples and the verification samples to obtain a preset dam recognition model, including: Using the training samples to train multiple preset network models respectively to obtain multiple trained dam recognition models; Using the verification samples to verify the plurality of trained dam identification models, to obtain a plurality of verification results; Determine a preset dam identification model based on the verification result; wherein the preset dam identification model is determined by any of the following methods: The recognition effects of the multiple trained dam recognition models are evaluated according to the intersection-over-union ratio, accuracy, recall rate and inference speed, and the preset dam recognition model is determined; the calculation formula of the intersection-over-union ratio is: , the accuracy calculation formula is , the calculation formula of recall rate is , the calculation formula for inference speed is , where IOU is the intersection-over-union ratio, is the intersection area of ​​the two bounding boxes, is the union area of ​​the two bounding boxes; P represents precision, R represents recall, TP represents prediction of 1, the prediction is correct, that is, the actual value is 1; FP represents prediction of 1, the prediction is wrong, that is, the actual value is 0; FN represents prediction of 0, the prediction is wrong, that is, the actual value is 1; AP is the inference speed; or The extraction effect of various models was comprehensively evaluated using three indicators: extraction rate, omission rate, and false extraction rate, and the preset dam identification model was determined. The extraction rate formula is E=R / T, where E is the extraction rate, R is the number of correctly extracted dams, and T is the actual number of dams. The omission rate formula is M=(TR) / T, where M is the omission rate, R is the number of correctly extracted dams, and T is the actual number of dams. The false extraction rate formula is F=K / T, where F is the false extraction rate, K is the number of incorrectly extracted dams, and T is the actual number of dams. The candidate target area is extracted to obtain the flow-breaking nodes and geographical environment information on the long river basin, including: Extracting the candidate target area to obtain the reflectance of the green band, the reflectance of the red band and elevation data; According to the reflectance of the green band, the reflectance of the red band and the formula , get the normalized difference water index; where NDWI is the normalized difference water index, is the reflectance of the green band, is the reflectance of the red band; using the normalized difference water index to extract water body information from the candidate target area; Obtaining a circumscribed rectangle aspect ratio according to the normalized difference water index and the elevation data; wherein the circumscribed rectangle aspect ratio is obtained according to AR=L / W, wherein AR is the circumscribed rectangle aspect ratio, L represents the length of the circumscribed rectangle, and W represents the width of the circumscribed rectangle, and both the length and width of the circumscribed rectangle are obtained from the elevation data; According to the aspect ratio of the circumscribed rectangle, a flow-breaking node on the long flow basin is obtained; Extracting the candidate target area to obtain elevation data; The elevation data is used to obtain ground roughness and ground relief. Ground roughness is the ratio of the earth's surface area to its projected area in a specific area. The elevation data is presented in a grid format, and each grid cell contains a height value. According to the definition of ground roughness, the ratio of the surface area of ​​each grid cell to its projected area is as follows: M = (AC × AB) / (AC × AC) = 1 / cosα, where ABC is the longitudinal section of a grid cell, M is the ground roughness, and α is the slope of the grid cell. The area of ​​the AB surface is the surface area of ​​the grid cell, and the area of ​​the AC surface is the projected area of ​​the grid cell, where cosα = AC / AB. Ground relief is the maximum relative elevation difference per unit area, and is as follows: R = Hmax - Hmin, where R represents ground relief, Hmax represents the maximum elevation value per unit area, and Hmin represents the minimum elevation value per unit area. Both the maximum elevation value per unit area and the minimum elevation value per unit area can be obtained from the elevation data. The method further comprises: Using the elevation data, a dam proportion detection value is obtained; the calculation formula for the dam proportion detection value is: Sp = U / D, where Sp represents the dam proportion detection value, U represents the larger area of ​​the water area within the buffer zone, and D represents the smaller area of ​​the water area within the buffer zone; wherein the larger area of ​​the water area within the buffer zone and the smaller area of ​​the water area within the buffer zone can both be obtained from the elevation data; Determining a dam location based on a flow-breaking node on the long river basin, the geographical environment information, and the dam ratio detection value; The method of determining the location of the dam according to the cut-off node on the long river basin, the geographical environment information, and the dam ratio detection value includes: Determining the dam location based on preset weights, the water body information, the geographical environment information, and the dam ratio detection value; The remote sensing image is preprocessed to obtain a preprocessed remote sensing image, including: The remote sensing image is preprocessed by performing one or more of atmospheric correction, orthorectification and mosaic cropping to obtain a preprocessed remote sensing image.

2. A dam identification device based on remote sensing images, characterized in that: include: An acquisition module is used to obtain remote sensing images to be identified; A preprocessing module, configured to preprocess the remote sensing image to obtain a preprocessed remote sensing image; An input module, configured to input the pre-processed remote sensing image into a preset dam recognition model to obtain a candidate target area; The preset dam recognition model is obtained by training a preset network model based on a dam body picture containing the dam; An extraction module is used to extract the candidate target area to obtain the flow-breaking nodes and geographical environment information on the long river basin; A determination module, configured to determine a dam location based on the flow-breaking nodes and geographical environment information on the long river basin; The pre-processed remote sensing image is input into a preset dam recognition model to obtain candidate target areas, including: Inputting the preprocessed remote sensing image into a convolutional network for convolution processing to obtain a shared feature map; The shared feature map is input into a region generation network for region selection to obtain a candidate region. Specifically, a sliding window mechanism and an anchor frame operation are used to define a bounding box and perform region selection on the dam position in the shared feature map to obtain a candidate region. The candidate regions are subjected to object classification and bounding box regression processing to obtain candidate target regions. Specifically, feature representations of each candidate region are extracted, and then each candidate region is mapped to a fixed-size feature map using a pooling operation. Object classification and bounding box regression processing are performed through a fully connected layer to obtain candidate target regions. The candidate target regions contain locations of dams. The candidate target regions include multiple points. The training process of the preset dam identification model includes: Acquire multiple images containing the dam according to pre-defined conditions. The pre-defined conditions are to generate a dam vector buffer centered around the dam vector point data. A circular buffer is generated centered around a dam point vector from a global dataset containing over 38,000 georeferenced dams. Dam images are obtained from GIS software or the Global Dam Database. Performing enhancement processing on the dam body image containing the dam to obtain an enhanced dam body image; wherein the enhancement processing on the dam body image containing the dam is performed using mirror flipping, scale transformation, color transformation, noise perturbation, brightness transformation and position transformation; Screening and labeling the enhanced dam body images to obtain a valid sample set; wherein, labeling the dams in the enhanced dam body images, removing duplicate, deformed, incomplete, and unrecognizable unqualified samples to obtain a valid sample set; The valid sample set is divided according to a preset ratio to obtain training samples and verification samples; wherein 70% of the valid sample set is used as training samples and 30% is used as verification samples; Using the training samples and the verification samples to train and verify the preset network model respectively, to obtain a preset dam recognition model; The preset network model is trained and verified using the training samples and the verification samples to obtain a preset dam recognition model, including: Using the training samples to train multiple preset network models respectively to obtain multiple trained dam recognition models; Using the verification samples to verify the plurality of trained dam identification models, to obtain a plurality of verification results; Determine a preset dam identification model based on the verification result; wherein the preset dam identification model is determined by any of the following methods: The recognition effects of the multiple trained dam recognition models are evaluated according to the intersection-over-union ratio, accuracy, recall rate and inference speed, and the preset dam recognition model is determined; the calculation formula of the intersection-over-union ratio is: , the accuracy calculation formula is , the calculation formula of recall rate is , the calculation formula for inference speed is , where IOU is the intersection-over-union ratio, is the intersection area of ​​the two bounding boxes, is the union area of ​​the two bounding boxes; P represents precision, R represents recall, TP represents prediction of 1, the prediction is correct, that is, the actual value is 1; FP represents prediction of 1, the prediction is wrong, that is, the actual value is 0; FN represents prediction of 0, the prediction is wrong, that is, the actual value is 1; AP is the inference speed; or The extraction effect of various models was comprehensively evaluated using three indicators: extraction rate, omission rate, and false extraction rate, and the preset dam identification model was determined. The extraction rate formula is E=R / T, where E is the extraction rate, R is the number of correctly extracted dams, and T is the actual number of dams. The omission rate formula is M=(TR) / T, where M is the omission rate, R is the number of correctly extracted dams, and T is the actual number of dams. The false extraction rate formula is F=K / T, where F is the false extraction rate, K is the number of incorrectly extracted dams, and T is the actual number of dams. The candidate target area is extracted to obtain the flow-breaking nodes and geographical environment information on the long river basin, including: Extracting the candidate target area to obtain the reflectance of the green band, the reflectance of the red band and elevation data; According to the reflectance of the green band, the reflectance of the red band and the formula , get the normalized difference water index; where NDWI is the normalized difference water index, is the reflectance of the green band, is the reflectance of the red band; using the normalized difference water index to extract water body information from the candidate target area; Obtaining a circumscribed rectangle aspect ratio according to the normalized difference water index and the elevation data; wherein the circumscribed rectangle aspect ratio is obtained according to AR=L / W, wherein AR is the circumscribed rectangle aspect ratio, L represents the length of the circumscribed rectangle, and W represents the width of the circumscribed rectangle, and both the length and width of the circumscribed rectangle are obtained from the elevation data; According to the aspect ratio of the circumscribed rectangle, a flow-breaking node on the long flow basin is obtained; Extracting the candidate target area to obtain elevation data; The elevation data is used to obtain ground roughness and ground relief. Ground roughness is the ratio of the earth's surface area to its projected area in a specific area. The elevation data is presented in a grid format, and each grid cell contains a height value. According to the definition of ground roughness, the ratio of the surface area of ​​each grid cell to its projected area is as follows: M = (AC × AB) / (AC × AC) = 1 / cosα, where ABC is the longitudinal section of a grid cell, M is the ground roughness, and α is the slope of the grid cell. The area of ​​the AB surface is the surface area of ​​the grid cell, and the area of ​​the AC surface is the projected area of ​​the grid cell, where cosα = AC / AB. Ground relief is the maximum relative elevation difference per unit area, and is as follows: R = Hmax - Hmin, where R represents ground relief, Hmax represents the maximum elevation value per unit area, and Hmin represents the minimum elevation value per unit area. Both the maximum elevation value per unit area and the minimum elevation value per unit area can be obtained from the elevation data. The extraction module is also used to: Using the elevation data, a dam proportion detection value is obtained; the calculation formula for the dam proportion detection value is: Sp = U / D, where Sp represents the dam proportion detection value, U represents the larger area of ​​the water area within the buffer zone, and D represents the smaller area of ​​the water area within the buffer zone; wherein the larger area of ​​the water area within the buffer zone and the smaller area of ​​the water area within the buffer zone can both be obtained from the elevation data; The Determine module is also used to: Determining a dam location based on a flow-breaking node on the long river basin, the geographical environment information, and the dam ratio detection value; The method of determining the location of the dam according to the cut-off node on the long river basin, the geographical environment information, and the dam ratio detection value includes: Determining the dam location based on preset weights, the water body information, the geographical environment information, and the dam ratio detection value; The remote sensing image is preprocessed to obtain a preprocessed remote sensing image, including: The remote sensing image is preprocessed by performing one or more of atmospheric correction, orthorectification and mosaic cropping to obtain a preprocessed remote sensing image.

Images