Intelligent ancient dam identification method and system based on historical image and deep learning

Abstract

The invention relates to the technical field of ancient dam recognition, in particular to an intelligent recognition method and system for an ancient dam based on a historical guard film and deep learning, and provides the following scheme: collecting the historical guard film, generating landform partitions, and generating a feature template in combination with time sequence analysis and a clustering algorithm; through a YOLOv5 model, by using a CSP architecture, a coordinate attention mechanism and a deformable convolution module, the ability of the model to extract ancient dam features is improved, the bounding box prediction precision is optimized in combination with a generalized intersection-to-union ratio, and efficient and accurate ancient dam identification and model dynamic optimization are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of ancient dam identification, and in particular to an intelligent identification method and system for ancient dams based on historical satellite images and deep learning. Background Art

[0002] Traditional archaeological methods and visual interpretation of remote sensing images of ancient dams have significant deficiencies in recognition efficiency and accuracy. Traditional technologies cannot effectively deal with feature extraction of irregular targets and complex backgrounds, and are prone to false detection and missed detection during the labeling and classification process. Therefore, combining deep learning technology, historical satellite image data and automated feature extraction methods has become a key direction for improving the intelligent recognition capabilities of ancient dams.

[0003] For example, the patent application with publication number CN110211173A discloses a method for locating and identifying paleontological fossils based on deep learning. By establishing a standard library for fossil image classification and using the standard library images to train a deep learning target detection model, automatic positioning and recognition of fossil images can be achieved, which can effectively assist human experts in batch complex fossil identification work and improve work efficiency. The method includes five steps: Step 1, establish a standard library for paleontological fossil images, and annotate the location category of each image in the library. Step 2, build a target detection network and load the pre-trained weight file. Step 3, use the constructed target detection network to train and annotate fossil images. Step 4, perform model testing and evaluation on the trained network model. Step 5, use the final training model to locate and identify actual paleontological fossil images. Automatic frame selection, positioning and classification recognition of multiple fossil individuals in an image or video are achieved with high recognition accuracy.

[0004] For example, a patent application with publication number CN118570656A discloses a method for finding ancient dams using historical remote sensing images and terrain data. The method includes: processing the historical remote sensing images before interpretation to obtain processed images that are easy to interpret. Distinguishing between landforms such as mountains, water bodies, and roads, and identifying key investigation areas by analyzing the spatial distribution patterns and identification features of dam sites in remote sensing images. According to the characteristics of the seven major interpretation elements, interpret the key investigation areas that have been identified in the image to identify the characteristic locations where dam sites may exist. Analyze the height of the study area in combination with elevation data to further determine the location of the dam. This technical solution can efficiently obtain information such as the location, boundaries, structure, shape, and relationship with the environment of the dam, and formulate more targeted investigation and excavation plans based on this, thereby improving the accuracy and efficiency of dam site identification.

[0005] The above technical solution has the problems raised by this background technology: the recognition efficiency is low under complex terrain background and vegetation coverage, and false detection and missed detection are prone to occur. To solve the above problems, this application discloses an intelligent recognition method and system for ancient dams based on historical satellite images and deep learning. Summary of the invention

[0006] The technical problem to be solved by the present invention is to address the deficiencies of the prior art, and to provide an intelligent identification method and system for ancient dams based on historical satellite images and deep learning, collect historical satellite images and generate geomorphic partitions, and generate feature templates in combination with time series analysis and clustering algorithms. Through the YOLOv5 model, using the CSP architecture, coordinate attention mechanism and deformable convolution module, the model's ability to extract ancient dam features is improved, and the bounding box prediction accuracy is optimized in combination with the generalized intersection over union (GIoU), achieving efficient and accurate ancient dam identification and model dynamic optimization.

[0007] To achieve the above object, the present invention provides the following technical solutions: An intelligent identification method for ancient dams based on historical satellite images and deep learning, the intelligent identification method includes: Collect the image data to be identified; Inputting the image data into a preset ancient dam recognition model, the ancient dam recognition model outputs the ancient dam pattern annotation in the image data, wherein the ancient dam recognition model is formed by training with multiple historical satellite images; The ancient dam pattern annotation is compared with a preset feature template to calculate the optimization parameters of the ancient dam recognition model.

[0008] The training steps of the ancient dam identification model include: Collecting a plurality of historical satellite images, and dividing the plurality of historical satellite images into a training set, a validation set, and a test set; Constructing an initial deep learning model, and training the initial deep learning model through a training set and a validation set; The test set is input into the trained initial deep learning model, the model parameters are updated according to the test results, and the final initial deep learning model is used as the ancient dam identification model.

[0009] The main structure of the initial deep learning model is a YOLOv5 model, and the YOLOv5 model includes a feature extraction network, a feature fusion network and a prediction network; The feature extraction network includes a CSP architecture and a coordinate attention mechanism, and outputs a feature map; The loss function of the feature fusion network is GIoR; A deformable convolution module is embedded in front of each convolution residual block in the YOLOv5 model, and the deformable convolution module is used to adjust the sampling position of the convolution kernel.

[0010] The adjusting the sampling position of the convolution kernel includes: Predicting the offset of each sampling point according to the local area in the feature map, wherein the output of each sampling point includes 2N deviation values, where N represents the number of sampling points of the convolution kernel; The sampling position of the convolution kernel is adjusted according to the offset.

[0011] The dividing the plurality of historical satellite images into a training set, a validation set and a test set comprises: Registering, correcting and splicing the multiple historical satellite images to generate a historical image data set; Performing spatial distribution and time series analysis on the historical image data set to generate positive sample feature templates containing ancient dam marked locations and negative sample feature templates containing non-ancient dam features; Randomly extract data from positive samples and negative samples to construct training sets, validation sets, and test sets.

[0012] The spatial distribution and time series analysis of the historical image dataset includes: Generate geomorphic partitions according to the geographic coordinate information in the historical image data set; Calculate the changing characteristics of each geomorphic division at different times; Matching the change characteristics with the geomorphic change characteristics in the manually annotated ancient dam template to obtain potential areas that meet the characteristics of the ancient dam; According to the landform distribution characteristics of the potential area and the non-potential area, a feature template is generated by a clustering algorithm.

[0013] The generating of the feature template by the clustering algorithm comprises: Extract the texture features, gray value distribution and boundary shape information of potential areas and non-potential areas, and calculate the landform distribution characteristics of each area; According to DBSCAN and K-means clustering algorithms, the geomorphic distribution characteristics of potential areas and non-potential areas are processed respectively to generate positive sample feature templates and negative sample feature templates.

[0014] The optimization parameters of the calculation of the ancient dam identification model include: Calculating an error based on a bounding box annotated with the ancient dam pattern and a bounding box in a preset feature template; The loss function of the ancient dam identification model is optimized according to the error.

[0015] An intelligent recognition system for ancient dams based on historical satellite images and deep learning, the system comprising a data acquisition module, a model training module and an ancient dam recognition module; The data acquisition module is used to collect the current image data to be identified and historical satellite images; The model training module is used to construct an initial deep learning model, train the initial deep learning model through a training set and a validation set, input the test set into the trained initial deep learning model, update the model parameters according to the test results, and use the final initial deep learning model as the ancient dam identification model; The ancient dam identification module is used to input the image data to be identified into the trained ancient dam identification model to complete the ancient dam pattern labeling.

[0016] The model training module includes: The data processing unit is used to register, correct and splice the collected historical satellite images to generate a unified historical image data set, perform spatial distribution and time series analysis, extract positive and negative sample feature templates, and thus construct training sets, validation sets and test sets; The model building unit is used to build the initial deep learning model, including the YOLOv5 network framework, embedded CSP architecture, coordinate attention mechanism and deformable convolution module; The training unit trains the model based on the training set and the validation set, and uses GIoU as the loss function to optimize the bounding box positioning accuracy; The optimization unit is used to input the test set into the model, evaluate the error feedback of the model in the ancient dam identification task, iteratively update the model parameters according to the error feedback, and optimize the positive and negative sample classification and bounding box prediction performance.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention enhances the model's adaptability and boundary positioning capabilities for complex backgrounds and irregular targets through the coordinate attention mechanism, deformable convolution module, and generalized intersection-over-union loss function; 2. The present invention utilizes time series analysis and clustering algorithms to generate feature templates, optimizes the model's training process and dynamic adjustment capabilities, realizes efficient and accurate intelligent identification of dams, and provides technical support for large-scale archaeological surveys. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings: Figure 1 This is a flow chart of an intelligent identification method for ancient dams based on historical satellite images and deep learning according to Embodiment 1 of the present invention; Figure 2 This is a structural diagram of an ancient dam identification model according to Embodiment 1 of the present invention; Figure 3 This is a flow chart of an ancient dam identification model training method according to Embodiment 2 of the present invention; Figure 4This is a flow chart of the clustering algorithm of Embodiment 2 of the present invention; Figure 5 This is a module diagram of an intelligent identification system for ancient dams based on historical satellite images and deep learning according to Example 3 of the present invention. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0020] Example 1 See also Figure 1 The present invention provides an embodiment: an intelligent identification method for ancient dams based on historical satellite images and deep learning, the specific steps of the method are as follows: S1: Collect the image data to be recognized; S2: inputting the image data into a preset ancient dam recognition model, and the ancient dam recognition model outputs the ancient dam pattern annotation in the image data; In this embodiment, the ancient dam recognition model implemented by the improved YOLOv5 model is used to extract features, fuse and predict the input image data. The CSP architecture and coordinate attention mechanism are introduced in the feature extraction stage to enhance the perception of the key feature areas of the ancient dam; at the same time, the deformable convolution module is embedded to dynamically adjust the convolution kernel sampling position to adapt to complex boundaries. In the prediction stage, the generalized intersection-over-union (GIoU) ​​is used as the loss function to optimize the bounding box and improve the recognition accuracy of irregularly shaped ancient dams. This improvement significantly reduces the probability of missed detection and false detection.

[0021] S3: comparing the ancient dam pattern annotation with a preset feature template, and calculating the optimization parameters of the ancient dam recognition model; In this embodiment, the feature templates of positive and negative samples are first generated based on texture features, grayscale distribution and boundary shape through clustering algorithms (such as DBSCAN and K-means). Then the identified ancient dam pattern annotations are compared with the templates, and the model performance is evaluated by calculating the error. For areas with large deviations, the adaptability and robustness of the model are improved by optimizing the loss function, dynamically adjusting the weights of the coordinate attention mechanism and updating the sampling offset of the deformable convolution.

[0022] Specifically, ancient dams are often irregular in shape and may be affected by geological erosion or vegetation cover, making their boundaries blurred and discontinuous in the image. At the same time, ancient dams are similar to surrounding features (such as ridges and houses) in texture and grayscale distribution, which can easily lead to model misjudgment. Historical satellite images have low resolution, small targets and limited information, which further increases the difficulty of identification.

[0023] In this embodiment, an improved YOLOv5 model is used, in which a deformable convolution module is embedded, and the adaptability of the model to complex targets is enhanced by dynamically calculating the sampling point offset. Deformable convolution can adjust the convolution kernel sampling position in real time according to the local features of the input image to capture the key areas and subtle features of the ancient dam. This dynamic sampling mechanism improves the model's ability to capture irregular boundaries. For example, when the model detects that the boundary of the ancient dam is curved or broken, the deformable convolution adjusts the sampling point position so that it is distributed along the boundary direction, thereby extracting complete shape information and avoiding the omission of important features.

[0024] Furthermore, the generalized intersection-over-union (GIoU) ​​is used as a loss function to ensure that even if the predicted box does not overlap with the true box, an effective gradient optimization signal can be provided. For example, when the model's predicted bounding box of an ancient dam is partially misplaced due to shape differences, GIoU guides the model to gradually adjust the position and shape of the predicted box by considering the minimum enclosing area of ​​the bounding box. Compared with traditional IoU, the GIoU optimization process is more stable and helps to converge quickly, especially in low-resolution images.

[0025] For example, taking the identification of an ancient dam ruins around the ancient city as an example, the dam in the target area has curved and broken boundaries due to long-term erosion and vegetation coverage, which is similar to the texture of the surrounding ridges, resulting in blurred boundaries and difficulty in distinguishing. After the image is input into the improved YOLOv5 model, the feature extraction network enhances the model's attention to the dam boundary area through the coordinate attention mechanism, while suppressing interference features such as ridges. The deformable convolution module dynamically calculates the offset to redistribute the convolution kernel sampling points along the changing trend of the dam boundary, accurately capturing the curved contour of the dam. GIoU guides the model to gradually adjust the shape and position of the bounding box. The annotated boundary generated by the model is highly consistent with the actual dam outline, avoiding the problem of false detection caused by blurred boundaries.

[0026] See also Figure 2 ,Structural diagram of the ancient dam identification model in an embodiment of the present invention,In this embodiment, the ancient dam identification model is based on the YOLOv5 framework and is divided into three main parts: Backbone (feature extraction network), Neck (feature fusion network) and Head (prediction network).

[0027] Specifically, the feature extraction network is used to extract multi-level features of the input image, using the CSP architecture and coordinate attention mechanism, and embedding a deformable convolution module; Furthermore, the feature extraction network extracts the basic features of the input parameters through a preliminary convolution operation. The convolution kernel size is 3×3 and the step size is 1. The BN (Batch Normalization) and SiLU activation functions are used to enhance the feature expression ability, thereby generating an initial feature map.

[0028] Furthermore, the CSP (Cross Stage Partial) architecture divides the initial feature map into two paths for parallel processing, performing feature extraction and feature aggregation tasks respectively. This design can reduce the amount of computation, optimize memory usage, and ensure the integrity of feature information. In this model, the CSP architecture is mainly used to reduce feature redundancy and improve the expressiveness of features, especially in low-resolution historical satellite images, which significantly improves the efficiency of feature extraction. Among them, each CSP architecture contains multiple convolutional residual blocks, thereby ensuring the extraction of deep features while preventing the problem of gradient disappearance or explosion.

[0029] Furthermore, the coordinate attention mechanism encodes the output of the CSP architecture in the horizontal and vertical directions respectively, generating a directional weight vector, thereby enhancing the model's ability to capture spatial features. It can strengthen the feature response of the ancient dam's boundary area while suppressing background interference such as surrounding fields and houses. In this model, the CA mechanism makes the output feature map of the feature extraction network more focused on the key area of ​​the ancient dam, effectively improving the accuracy of subsequent detection.

[0030] Specifically, unlike the traditional attention mechanism that treats spatial information as a unified whole, the coordinate attention mechanism divides it into two different dimensions, allowing the model to identify more detailed features in a more sophisticated way. This enhanced feature representation helps to locate ancient dams more accurately, while improving the model's ability to distinguish dams that are spatially close but have different semantic features. In addition, the coordinate attention mechanism can also enhance the model's ability to handle occluded and high-density scene images, effectively reducing false positives and improving overall detection accuracy. At the same time, the coordinate attention mechanism has a lightweight design feature, which makes it an ideal complement to YOLOv5. Given that YOLOv5 itself is widely recognized for its fast inference speed, the introduction of a very small number of additional parameters and calculations allows the coordinate attention mechanism to improve its accuracy while maintaining YOLOv5's high performance, making it an ideal choice for real-time ancient dam detection applications.

[0031] Furthermore, a deformable convolution module (DCN) is embedded in front of each convolution residual block in the CSP architecture. DCN adapts to the shape changes of the ancient dam boundary by dynamically adjusting the sampling position of the convolution kernel. Its offset prediction module generates an offset value for each sampling point (including 2N offset values, N is the number of sampling points of the convolution kernel) based on the local area in the feature map. For example, for a 3×3 convolution kernel, an offset of 9 sampling points is generated, and the sampling position is dynamically adjusted so that the sampling point distribution fits the curved features of the dam boundary.

[0032] In this embodiment, the sampling points of DCN are dynamically generated by the offset prediction module. Unlike the traditional convolution kernel sampling at a fixed position, the sampling position of DCN is adjusted in real time according to the information of the local area in the feature map.

[0033] Specifically, the offset prediction module receives the input feature map and aggregates the information of the local area using a 3×3 convolution operation. For example, for an area in a feature map, the convolution kernel collects the feature values ​​of the surrounding 3×3 areas based on the center point. The offset prediction module outputs a feature map, and each pixel of the feature map contains 2N values ​​(N is the number of sampling points of the convolution kernel). These 2N values ​​represent the offset of the sampling point in the x and y directions respectively.

[0034] Furthermore, the offset of each sampling point is determined by the local information of the feature map. The convolution kernel adjusts the sampling position according to the offset so that the sampling points are concentrated in the key information area. First, the feature changes of the local area about the boundary gradient are detected. For the boundary area with higher gradient, the offset adjusts the sampling point close to the boundary to capture more detailed information. This is because the boundary of the ancient dam usually appears as an area with high grayscale changes, and DCN will offset the sampling points to these change points. In areas with small texture or gradient changes (such as background), the offset will reduce the distribution of sampling points in these areas. This is because in texture interference areas such as ridges, DCN will reduce the offset of the sampling points to avoid the influence of interfering features.

[0035] Specifically, the task of the feature fusion network is to fuse features of different scales to improve the detection capability of objects of different sizes. This part uses FPN (Feature Pyramid Network) and PAN (Path Aggregation Network).

[0036] Furthermore, FPN is responsible for extracting high-resolution information from shallow features, while PAN further fuses high-resolution information with deep features to generate multi-scale feature representations. In historical satellite images, ancient dam targets may be small and have blurred boundaries, so multi-scale feature fusion is particularly important.

[0037] Furthermore, the feature fusion network uses generalized intersection over union (IoU) as the main loss function. Traditional IoU loss has limitations, especially when the dam bounding boxes do not overlap, it cannot provide meaningful gradients. GIoU makes up for this defect to a certain extent. Even if the predicted box and the true box do not intersect at all, it can ensure that the model continues to learn and optimize. This not only speeds up the training process of dam recognition, but also makes the model converge faster.

[0038] Specifically, GIoU can provide a more stable and consistent loss gradient throughout the training process. In addition, incorporating GIoU into the loss function can take into account the bounding box area around the predicted box and the true box, which helps to improve the positioning of the bounding box and thus improve the accuracy of dam prediction. In the dam detection task, this enhanced positioning accuracy is particularly critical, because even a small difference in the bounding box coordinates can have a significant impact on the overall detection results.

[0039] Specifically, after performing feature extraction and feature fusion on the input image data, the prediction network finally generates the pattern annotation of the ancient dam through steps such as multi-scale detection, bounding box prediction, and target classification.

[0040] Furthermore, the prediction network receives feature maps of three resolutions from the feature fusion network (Neck), corresponding to small targets (such as dam remnants), medium-sized targets (partially preserved intact dams), and large targets (complete dam structures). Low-resolution feature maps provide semantically rich information and are suitable for detecting large targets; high-resolution feature maps provide more detailed information and are suitable for detecting small targets.

[0041] Furthermore, each pixel in the feature map corresponds to a candidate box, and the position and size of the candidate box are decoded through the regression algorithm.

[0042] Furthermore, the prediction network calculates the category probability of the target based on the candidate box using the channel information of the feature map. The Softmax activation function normalizes the category score to a probability distribution, and the category with the highest probability value is used as the final classification of the candidate box. Non-maximum suppression (NMS) is used to screen the candidate boxes to avoid multiple boxes repeatedly labeling the same dam.

[0043] Furthermore, by supporting multi-scale detection, the model can detect dams of different scales with the help of activation information at different depths in the network. This allows the detection of small objects (such as dam remains) that are more strongly activated in earlier layers, as well as large objects that are easier to identify in deeper layers. The multi-scale approach greatly improves the accuracy of the model, enabling it to distinguish and identify dams of different sizes with higher precision.

[0044] Furthermore, the incorporation of additional detection layers also enhances the robustness of YOLOv5, reduces the possibility of missed detection in cluttered or dense environments, and improves the overall performance of the model in complex scenarios. At the same time, the supplementary layer can also promote rapid convergence in the training phase, allowing the model to learn dam recognition in a more efficient way. It is worth noting that YOLOv5 with additional detection layers has stronger model output capabilities, and can optimize the estimation of the dam's position and size by predicting bounding boxes at multiple scales, thereby achieving more accurate positioning.

[0045] The specific steps of S3 are as follows: S3.1: Calculate the error based on the bounding box annotated by the ancient dam pattern and the bounding box in the preset feature template; Specifically, this operation compares the bounding box predicted by the model with the manually annotated template to analyze the deviations in position, size, and shape. For example, the model may have offset or inaccurate scaling when predicting the bounding box, resulting in an incomplete fit with the contour of the real ancient dam. By calculating the difference between the two, it is clear which aspects of the model's performance need to be improved.

[0046] In this embodiment, the generalized intersection-over-union (GIoU) ​​is used as the main error evaluation indicator. GIoU considers not only the intersection of the predicted box and the true box, but also introduces the information of the area enclosed by the two. By calculating the overlap between the predicted box and the true box and their distribution relative to the minimum enclosing box, GIoU can still provide an effective error signal when the two do not overlap. This design is particularly critical when dealing with ancient dam targets with blurred or broken boundaries in historical satellite images, and can effectively solve the problem that traditional IoU loss cannot provide gradient optimization in this scenario.

[0047] S3.2: Optimizing the loss function of the ancient dam identification model according to the error.

[0048] Specifically, the GIoU loss function feeds the above error back into the parameter update of the model, so that the model can gradually adjust the weight distribution of the convolution kernel weights, offset predictions, and attention mechanisms to reduce the bounding box bias. The optimized loss function guides the model to capture target features more accurately, especially for ancient dams with complex shapes and irregular boundaries. The model can dynamically adjust the sampling position and strengthen the focus on key areas. The design of this loss function greatly improves the bounding box prediction accuracy of the model, making it not only suitable for targets with regular shapes, but also well adapted to small dam remains in complex scenes.

[0049] Example 2 See also Figure 3The present invention provides an embodiment: a method for training an ancient dam identification model, the training steps of the ancient dam identification model are as follows: A1: Collect multiple historical satellite images and divide the multiple historical satellite images into a training set, a validation set and a test set; In this embodiment, a large amount of historical satellite image data from the 1940s to the 1970s is collected, registered, corrected, and spliced ​​to form a high-quality historical image data set. According to the spatial distribution characteristics and time series changes, the data set is divided into a training set, a validation set, and a test set. Potential areas and non-potential areas are extracted by classifying geographic coordinates, and DBSCAN and K-means clustering algorithms are used to generate positive sample feature templates and negative sample feature templates, respectively.

[0050] A2: construct an initial deep learning model, and train the initial deep learning model using a training set and a validation set; In this embodiment, an initial deep learning model based on YOLOv5 is constructed, which includes: The CSP architecture and coordinate attention mechanism are used in the feature extraction network to enhance the model's ability to perceive key features; Embedding a deformable convolution module enables the convolution kernel to dynamically adjust the sampling position and better adapt to the irregularity of the shape of the ancient dam; The generalized intersection-over-union (GIoU) ​​is used as the loss function to optimize the prediction ability of the bounding box. The model is trained through the training set and the validation set, and the model parameters are updated repeatedly.

[0051] A3: Input the test set into the trained initial deep learning model, update the model parameters according to the test results, and use the final initial deep learning model as the ancient dam identification model.

[0052] In this embodiment, the test set is input into the trained initial deep learning model, and the model output is evaluated using the feature template. For the areas with incorrect detection, the model parameters are adjusted to optimize the coordinate attention weight, deformable convolution sampling offset, and GIoU gradient feedback. Finally, the model that has undergone multiple rounds of optimization is defined as the ancient dam recognition model.

[0053] The specific steps of A1 are as follows: A1.1: Register, correct and stitch the multiple historical satellite images to generate a historical image dataset; Specifically, in order to ensure that multiple historical satellite images have a unified coordinate reference system, the geographic coordinates of the satellite images are first aligned with the ground reference points using georeferencing technology. By introducing high-precision georeferenced layers (such as digital elevation models or geographic information system data), the distortion, scaling, and offset in the historical satellite images are corrected. Then, multiple time series images are merged into a complete geomorphic area image through image stitching technology to ensure seamless connection and pixel continuity of regional boundaries.

[0054] Furthermore, the above operations can eliminate the errors caused by the shooting angle and resolution differences of historical images, and improve the accuracy of subsequent analysis. The generated historical image dataset can truly reflect the geomorphic characteristics of the area where the ancient dam is located, ensuring the stability of subsequent feature extraction.

[0055] A1.2: Perform spatial distribution and time series analysis on the historical image dataset to generate positive sample feature templates containing the marked locations of ancient dams and negative sample feature templates containing non-ancient dam features; Specifically, based on the geographic coordinate information in the historical image data set, the study area is divided into multiple geomorphic zones, and geomorphic features are extracted for each zone. Using time series analysis technology, the texture features, grayscale value distribution and geomorphic morphological changes in images of different periods are compared to capture the unique characteristics of the ancient dam area in a specific time period, such as the rate of change of edge texture and the regularity of surface undulations. For known ancient dam areas, combined with manually annotated data, positive sample feature templates of marked locations are generated; for areas without ancient dams, negative sample feature templates of non-ancient dams are generated through the statistical distribution of their texture, shape and grayscale values. Strengthen the model's ability to distinguish ancient dam areas while reducing interference from complex backgrounds.

[0056] A1.3: Randomly extract data from positive and negative samples to construct training sets, validation sets, and test sets; Specifically, according to the ratio of the number of positive samples to negative samples, data are randomly selected to construct the training set, validation set, and test set, while ensuring that the sample distribution in each data set is representative in space and time. A combination of random sampling and stratified sampling is used to make the training data cover different types of landforms and ancient dam features, thereby improving the generalization ability of the model. The model performance is evaluated in real time through the validation set, and the final recognition effect is verified through the test set.

[0057] The specific steps of A1.2 are as follows: A1.2.1: Generate geomorphic partitions based on the geographic coordinate information in the historical image dataset; Specifically, the image area is classified according to the terrain undulation, river distribution and terrain units (such as hills, plains, and mountains), and the key partitions are marked in combination with the manually annotated ancient dam distribution data. This ensures that the characteristics of ancient dams under different geomorphic conditions are extracted independently, thereby improving the pertinence and accuracy of subsequent analysis. This achieves fine-grained management of features and avoids noise interference caused by complex terrain confusion.

[0058] A1.2.2: Calculate the changing characteristics of each geomorphic zone at different times; Specifically, within each geomorphic subregion, image change characteristics, including vegetation coverage, terrain elevation changes, and texture uniformity, were extracted through time series analysis.

[0059] Furthermore, the gray value change curve of the image is calculated to capture the fluctuation of vegetation cover, the surface elevation change is calculated through DEM (digital elevation model), and the texture continuity and direction change are extracted using the GLCM (gray level co-occurrence matrix) of the image. These change characteristics can reflect the existence and status of potential ancient dams in the area. For example, the vegetation cover near the dam may show regular growth and decline, and the local changes in the elevation model may also correspond to the historical sites of the dam.

[0060] A1.2.3: Match the above mentioned change characteristics with the geomorphic change characteristics in the manually annotated ancient dam template to obtain potential areas that meet the characteristics of ancient dams; Specifically, by constructing the feature templates of the ancient dams manually annotated in the historical data, the change patterns in these feature templates are extracted, such as the irregularity of the edges, the directional distribution of the texture, and the local uplift characteristics of the elevation. These features are then compared one by one with the change characteristics of the geomorphic partitions, and the matching degree between the features of each region and the template is calculated using a similarity measurement method (such as Euclidean distance or cosine similarity). The regions with a higher matching degree are judged as potential regions that meet the characteristics of the ancient dams.

[0061] A1.2.4: Generate a feature template by clustering algorithm based on the geomorphic distribution characteristics of the potential area and the non-potential area; In this embodiment, the geomorphic partitions that are not marked as potential areas are marked as non-potential areas; Specifically, in the potential area and non-potential area, texture features, gray value distribution, boundary shape and vegetation change rate are extracted to construct a multidimensional feature vector. The DBSCAN algorithm is used to perform density clustering on the feature vector to identify areas with high feature similarity; at the same time, the K-means clustering method is combined to optimize the distribution of the center points of the regional division, thereby generating positive sample feature templates and negative sample feature templates. The positive sample template is used to represent the area where the ancient dam may exist, and the negative sample template is used to exclude the area that does not match the characteristics of the ancient dam.

[0062] Furthermore, the texture features, gray value distribution and boundary shape information of the potential area and the non-potential area are extracted, and the landform distribution characteristics of each area are calculated; Furthermore, the geomorphic distribution characteristics of the potential area and the non-potential area are processed according to the DBSCAN and K-means clustering algorithms respectively to generate positive sample feature templates and negative sample feature templates.

[0063] See also Figure 4 , a flow chart of a clustering algorithm according to an embodiment of the present invention, firstly, the landform distribution features are preliminarily clustered in sequence by using the DBSCAN clustering algorithm, and core points are divided according to the preliminary clustering results; the number of cluster center clusters is determined; Further, the core point is used as the initial center cluster of the K-means clustering algorithm, the Euclidean distance from the landform distribution feature to each initial center cluster is calculated, clustering is performed according to the Euclidean distance, and the clustering average error of the initial center cluster is calculated; Update the central cluster according to the clustering average error, define the convergence condition, and judge whether the updated central cluster meets the convergence condition. If it does not meet the convergence condition, continue to update the central cluster. If it meets the convergence condition, output the clustering result.

[0064] Example 3 See also Figure 5 , the present invention provides an embodiment: an intelligent identification system for ancient dams based on historical satellite images and deep learning, the system comprising a data acquisition module, a model training module and an ancient dam identification module; The data acquisition module is used to acquire the current image data to be identified and historical satellite images; The model training module is used to construct an initial deep learning model, train the initial deep learning model through a training set and a validation set, input the test set into the trained initial deep learning model, update the model parameters according to the test results, and use the final initial deep learning model as the ancient dam identification model; The ancient dam identification module is used to input the image data to be identified into the trained ancient dam identification model to complete the ancient dam pattern labeling.

[0065] The model training module includes: The data processing unit is used to register, correct and splice the collected historical satellite images to generate a unified historical image data set, perform spatial distribution and time series analysis, extract positive and negative sample feature templates, and thus construct training sets, validation sets and test sets; The model building unit is used to build the initial deep learning model, including the YOLOv5 network framework, embedded CSP architecture, coordinate attention mechanism and deformable convolution module; The training unit is used to train the model based on the training set and the validation set, and optimize the bounding box positioning accuracy using GIoU as the loss function; The optimization unit is used to input the test set into the model, evaluate the error feedback of the model in the ancient dam identification task, iteratively update the model parameters according to the error feedback, and optimize the positive and negative sample classification and bounding box prediction performance.

[0066] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. An intelligent identification method for ancient dams based on historical satellite images and deep learning, characterized in that: The intelligent identification method comprises: Collect the image data to be identified; Inputting the image data into a preset ancient dam recognition model, the ancient dam recognition model outputs the ancient dam pattern annotation in the image data, wherein the ancient dam recognition model is formed by training with multiple historical satellite images; The ancient dam pattern annotation is compared with a preset feature template to calculate the optimization parameters of the ancient dam recognition model.

2. The intelligent identification method of ancient dams based on historical satellite images and deep learning according to claim 1 is characterized in that: The training steps of the ancient dam identification model include: Collecting a plurality of historical satellite images, and dividing the plurality of historical satellite images into a training set, a validation set, and a test set; Constructing an initial deep learning model, and training the initial deep learning model through a training set and a validation set; The test set is input into the trained initial deep learning model, the model parameters are updated according to the test results, and the final initial deep learning model is used as the ancient dam identification model.

3. The intelligent identification method of ancient dams based on historical satellite images and deep learning according to claim 2 is characterized in that: The main structure of the initial deep learning model is a YOLOv5 model, and the YOLOv5 model includes a feature extraction network, a feature fusion network and a prediction network; The feature extraction network includes a CSP architecture and a coordinate attention mechanism, and outputs a feature map; The loss function of the feature fusion network is GIoR; A deformable convolution module is embedded in front of each convolution residual block in the YOLOv5 model, and the deformable convolution module is used to adjust the sampling position of the convolution kernel.

4. The intelligent identification method of ancient dams based on historical satellite images and deep learning according to claim 3 is characterized in that: The adjusting the sampling position of the convolution kernel includes: Predicting the offset of each sampling point according to the local area in the feature map, wherein the output of each sampling point includes 2N deviation values, where N represents the number of sampling points of the convolution kernel; The sampling position of the convolution kernel is adjusted according to the offset.

5. The intelligent identification method of ancient dams based on historical satellite images and deep learning according to claim 2 is characterized in that: The dividing the plurality of historical satellite images into a training set, a validation set and a test set comprises: Registering, correcting and splicing the multiple historical satellite images to generate a historical image data set; Performing spatial distribution and time series analysis on the historical image data set to generate positive sample feature templates containing ancient dam marked locations and negative sample feature templates containing non-ancient dam features; Randomly extract data from positive samples and negative samples to construct training sets, validation sets, and test sets.

6. The intelligent identification method of ancient dams based on historical satellite images and deep learning according to claim 5 is characterized in that: The spatial distribution and time series analysis of the historical image dataset includes: Generate geomorphic partitions according to the geographic coordinate information in the historical image data set; Calculate the changing characteristics of each geomorphic division at different times; Matching the change characteristics with the geomorphic change characteristics in the manually annotated ancient dam template to obtain potential areas that meet the characteristics of the ancient dam; According to the landform distribution characteristics of the potential area and the non-potential area, a feature template is generated by a clustering algorithm.

7. The intelligent identification method of ancient dams based on historical satellite images and deep learning according to claim 6 is characterized in that: The generating of the feature template by the clustering algorithm comprises: Extract the texture features, gray value distribution and boundary shape information of potential areas and non-potential areas, and calculate the landform distribution characteristics of each area; According to DBSCAN and K-means clustering algorithms, the geomorphic distribution characteristics of potential areas and non-potential areas are processed respectively to generate positive sample feature templates and negative sample feature templates.

8. The intelligent identification method of ancient dams based on historical satellite images and deep learning according to claim 1 is characterized in that: The optimization parameters of the calculation of the ancient dam identification model include: Calculating an error based on a bounding box annotated with the ancient dam pattern and a bounding box in a preset feature template; The loss function of the ancient dam identification model is optimized according to the error.

9. An intelligent identification system for ancient dams based on historical satellite images and deep learning, used to implement the intelligent identification method for ancient dams based on historical satellite images and deep learning as described in any one of claims 1 to 8, characterized in that: The system includes a data acquisition module, a model training module and an ancient dam identification module; The data acquisition module is used to acquire the current image data to be identified and historical satellite images; The model training module is used to construct an initial deep learning model, train the initial deep learning model through a training set and a validation set, input the test set into the trained initial deep learning model, update the model parameters according to the test results, and use the final initial deep learning model as the ancient dam identification model; The ancient dam identification module is used to input the image data to be identified into the trained ancient dam identification model to complete the ancient dam pattern labeling.

10. The intelligent identification system for ancient dams based on historical satellite images and deep learning according to claim 9 is characterized in that: The model training module includes: The data processing unit is used to register, correct and splice the collected historical satellite images to generate a unified historical image data set, perform spatial distribution and time series analysis, extract positive and negative sample feature templates, and thus construct training sets, validation sets and test sets; The model building unit is used to build the initial deep learning model, including the YOLOv5 network framework, embedded CSP architecture, coordinate attention mechanism and deformable convolution module; The training unit trains the model based on the training set and the validation set, and uses GIoU as the loss function to optimize the bounding box positioning accuracy; The optimization unit is used to input the test set into the model, evaluate the error feedback of the model in the ancient dam identification task, iteratively update the model parameters according to the error feedback, and optimize the positive and negative sample classification and bounding box prediction performance.

Images