Site and relic detection and identification method and system

Abstract

The application belongs to the technical field of remote sensing detection and identification, and discloses a site and relic detection and identification method and system. The application constructs and trains a detection and identification network model. The detection and identification network model comprises a preprocessing unit, a basic feature extraction unit, a feature enhancement unit and a feature fusion unit connected in sequence. The feature enhancement unit is used for enhancing the basic features in the horizontal and vertical directions respectively. The collected image data and laser point cloud data are input into the trained detection and identification network model, and the detection and identification network model outputs site and relic detection and identification result information. The application can improve the completeness and automation of the site and relic detection and identification.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing detection and identification technology, and more specifically, relates to a method and system for detecting and identifying archaeological sites and relics. Background Technology

[0002] With the continuous development of remote sensing and spatial data processing technologies, the use of remote sensing images acquired from aerial, space, and ground-based platforms to detect and identify surface cultural heritage and archaeological sites has become an important means of archaeological research and cultural relic protection. However, in practical applications, many sites and surface remains have long been under vegetation cover, farmland reclamation, or natural erosion environments, and their surface features are often obscured by dense vegetation or soil deposits. This results in blurred boundaries and indistinct spectral features in traditional optical remote sensing images, making it difficult to effectively distinguish them using a single band or data source. First, dense vegetation reflects and scatters electromagnetic energy in the visible and near-infrared bands, obscuring the true reflective properties of the surface, making it difficult to directly observe the spectral anomalies of potential remains. Second, the temporal variations and differences introduced by different vegetation types and densities cause instability in surface spectral information, affecting the accuracy of subsequent feature extraction and identification of sites and remains.

[0003] With the continuous advancement of sensor technology, acquiring remote sensing data from various sensor sources, such as visible light, laser point clouds, and thermal infrared, has become increasingly easier. Existing technologies also attempt to use multi-source remote sensing data for the detection and identification of archaeological sites and relics. However, at the methodological model level, existing models mostly focus on global accuracy optimization, failing to adequately characterize the fine-grained weak signals of sites and relics and the differences in cross-regional occurrence environments. Under complex surface cover and environmental heterogeneity interference, semantic confusion and loss of detail easily occur. In terms of multimodal fusion, existing methods often only perform simple overlay or feature stitching at the data level, lacking an effective multimodal data fusion mechanism and failing to fully utilize the complementary characteristics of different sensors. Therefore, these methods have poor robustness to changes in illumination, noise interference, and topographic relief. Especially in areas with dense vegetation or severe surface disturbance, the texture, shape, and spectral features of relics in remote sensing images are often weakened, resulting in unclear boundaries and indistinct targets. The identification accuracy (i.e., the completeness of identification) and automation level are difficult to meet the needs of rapid surveys in large areas. Summary of the Invention

[0004] This invention provides a method and system for detecting and identifying archaeological sites and relics, which addresses the problem that the completeness and automation of archaeological site and relic detection and identification in the prior art need to be improved.

[0005] This invention provides a method for detecting and identifying archaeological sites and relics, comprising the following steps: Construct and train the detection and recognition network model; The detection and recognition network model includes a preprocessing unit, a basic feature extraction unit, a feature enhancement unit, and a feature fusion unit connected in sequence; the feature enhancement unit is used to enhance the basic features in both planar and vertical aspects. The acquired image data and laser point cloud data are input into the trained detection and recognition network model, which outputs the detection and recognition results of the site and relics.

[0006] Preferably, the preprocessing unit is used to preprocess the image data and the laser point cloud data; The preprocessing of laser point cloud data includes: classifying vegetation, separating and removing vegetation canopy points, and preserving the structure of the site and the micro-topography of the surface; constructing a vegetation mask based on the vegetation classification results; and using the vegetation mask to guide the elevation interpolation algorithm of the point cloud to generate a digital elevation model, thereby achieving a detailed description of the structure of the site and the micro-topography of the surface under vegetation shading conditions.

[0007] Preferably, the basic feature extraction unit is used to extract basic features; Specifically, for image data, a convolutional neural network is used as the backbone to extract multi-scale basic two-dimensional features from multiple different stages; for laser point cloud data, a three-dimensional point cloud deep learning network is used as the backbone network to extract multi-scale basic three-dimensional features from multiple different stages.

[0008] Preferably, the feature enhancement unit includes a planar feature enhancement module and a vertical feature enhancement module; The planar feature enhancement module uses strip convolutions of different widths to obtain different receptive field features and performs multi-scale feature fusion to obtain enhanced image planar features. The vertical feature enhancement module calculates the elevation difference of the point cloud, generates elevation difference attention weights, and embeds the elevation difference attention weights into the self-attention mechanism and the convolutional neural network to obtain the enhanced vertical features of the point cloud.

[0009] Preferably, the combined loss function of the detection and recognition network model includes a primary loss, an image-assisted loss, and a point cloud-assisted loss; the primary loss consists of Dice loss and cross-entropy loss.

[0010] Preferably, the feature fusion unit includes a feature alignment module and a feature fusion module; The feature alignment module is used to align the enhanced planar features of the image with the enhanced vertical features of the point cloud in terms of spatial location, scale, and channel dimension; wherein, key feature channels are selected through channel attention, and significant areas of the site and remains are aligned through spatial attention. The feature fusion module is used to perform a global-level cross-fusion operation of aligned image planar features and point cloud vertical features to obtain fused features.

[0011] Preferably, the image data includes one or more of the following: visible light and near-infrared optical image data, synthetic aperture radar data, and multispectral / hyperspectral image data.

[0012] Preferably, after obtaining the detection and identification result information, the method further includes: verifying and evaluating the detection and identification result information.

[0013] On the other hand, the present invention provides a site and relic detection and identification system, comprising: The model building unit is used to build a detection and identification network model and train the detection and identification network model using sample data labeled with site and relic tags; The data acquisition unit is used to acquire image data and laser point cloud data; The detection and identification unit includes a trained detection and identification network model. The detection and identification unit is used to detect and identify the site and relics based on the input image data and laser point cloud data using the trained detection and identification network model, and output the detection and identification result information. The site and relic detection and identification system is used to perform the steps in the site and relic detection and identification method described above.

[0014] Preferably, the site and relic detection and identification system further includes: a verification and evaluation unit; the verification and evaluation unit is used to verify and evaluate the detection and identification results information.

[0015] One or more technical solutions provided in this invention have at least the following technical effects or advantages: (1) In this invention, a detection and recognition network model is constructed and trained. The acquired image data and laser point cloud data are input into the trained detection and recognition network model, and the detection and recognition network model outputs the detection and recognition results of the archaeological site. The detection and recognition network model in this invention includes a preprocessing unit, a basic feature extraction unit, a feature enhancement unit, and a feature fusion unit connected in sequence. The feature enhancement unit is used to enhance the basic features in both planar and vertical aspects. In view of the problems of weak features of archaeological sites, scattered and discontinuous spatial distribution, and mixture with terrain features and dense vegetation cover, this invention takes into account the characteristics of archaeological sites in both planar and vertical aspects. Through the complementary fusion of multimodal features, it improves the integrity and automation of the detection and recognition of archaeological sites.

[0016] (2) The planar feature enhancement module in this invention uses strip convolution with different widths to obtain features of different receptive fields and performs multi-scale feature fusion to obtain enhanced planar features of the image. That is, this invention uses strip shape convolution operations with different receptive field sizes to constrain the process of extracting planar features of sites and relics. Through multi-stage, multi-level local, global and planar strip feature adaptive extraction, it can enhance the representation ability of weak signals of relics in two-dimensional planes, thereby improving the integrity of site and relic detection and identification.

[0017] (3) The vertical feature enhancement module in this invention calculates the elevation difference of the point cloud, generates elevation difference attention weights, and embeds these elevation difference attention weights into a self-attention mechanism and a convolutional neural network to obtain the enhanced vertical features of the point cloud. In other words, based on the advantages of three-dimensional laser point clouds in representing archaeological remains, this invention proposes a novel elevation difference attention mechanism to achieve adaptive representation of weak signals in the vertical structure of remains, thereby improving the completeness of site and remains detection and identification.

[0018] (4) When preprocessing laser point cloud data, this invention first classifies vegetation, separates and removes vegetation canopy points, and retains the structure of archaeological sites and the micro-topography of the surface. Then, a vegetation mask is constructed based on the vegetation classification results. Finally, the vegetation mask is used to guide the elevation interpolation algorithm of the point cloud to generate a digital elevation model, thereby achieving a detailed description of the structure of archaeological sites and the micro-topography of the surface under vegetation shading conditions. Compared with the filtering algorithm used in the prior art, the above-mentioned scheme of this invention can ensure the retention rate of micro-topography point clouds such as archaeological sites and effectively improve the detailed description of the micro-topography of archaeological sites under vegetation shading conditions, thereby improving the detection capability of archaeological sites under vegetation shading.

[0019] (5) The feature fusion unit in this invention includes a feature alignment module and a feature fusion module. The feature alignment module is used to align the enhanced planar features of the image with the enhanced vertical features of the point cloud in terms of spatial location, scale, and channel dimension. Among them, key feature channels are selected through channel attention, and significant areas of the site are aligned through spatial attention. The feature fusion module is used to perform a global-level cross-fusion operation on the aligned planar features of the image and the vertical features of the point cloud to obtain fused features. This invention fully utilizes the complementary advantages of high-resolution texture of the image and stereoscopic elevation of the point cloud to accurately capture the planar distribution and vertical structure of archaeological sites, thereby improving the accuracy of site detection under complex vegetation occlusion. That is, the feature alignment module of this invention, based on channel and spatial attention mechanisms, realizes the selection of multimodal features in the local area of ​​archaeological sites, and the feature fusion module, based on cross-attention mechanisms, performs global-level cross-fusion and complementary perception of features from different sources of the image and point cloud, solving the problem of modal heterogeneity between the image and the point cloud, thereby achieving the ability to densely and accurately detect and identify the boundary of the site under vegetation occlusion. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for detecting and identifying archaeological sites and relics provided in Embodiment 1 of the present invention; Figure 2 A feature extraction and enhancement process for embedding multi-scale feature constraints of planar strips of archaeological remains; Figure 3 The process for extracting and enhancing vertical features of relics to take into account elevation differences and attention mechanisms; Figure 4 A network framework for archaeological site detection and recognition based on the cross-fusion of planar and vertical multimodal features; Figure 5 This is a flowchart of a method for detecting and identifying archaeological sites and relics provided in Embodiment 2 of the present invention. Detailed Implementation

[0021] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0022] Example 1: Example 1 provides a method for detecting and identifying archaeological sites and remains. See [link to example]. Figure 1 This includes the following steps: Construct and train the detection and recognition network model; The detection and recognition network model includes a preprocessing unit, a basic feature extraction unit, a feature enhancement unit, and a feature fusion unit connected in sequence; the feature enhancement unit is used to enhance the basic features in both planar and vertical aspects. The acquired image data and laser point cloud data are input into the trained detection and recognition network model, which outputs the detection and recognition results of the site and relics.

[0023] Example 1 can first construct and train a detection and recognition network model, or first collect and acquire image data and laser point cloud data.

[0024] This invention can acquire multimodal remote sensing data of a target area through various remote sensing platforms and sensors. The image data includes visible and near-infrared optical image data, synthetic aperture radar (SAR) data, multispectral / hyperspectral image data, etc. The laser point cloud data is LiDAR (Light Laser Detection and Ranging) point cloud data.

[0025] The detection and recognition network model and its main units are described below.

[0026] (1) Preprocessing unit.

[0027] The preprocessing unit is used to preprocess image data and laser point cloud data.

[0028] The preprocessing of laser point cloud data includes: classifying vegetation, separating and removing vegetation canopy points, and preserving the site structure and surface micro-topography; constructing a vegetation mask based on the vegetation classification results; and using the vegetation mask to guide the elevation interpolation algorithm of the point cloud to generate a digital elevation model (DEM) to achieve a detailed description of the site structure and surface micro-topography under vegetation shading conditions.

[0029] Data preprocessing operations include geometric correction, vegetation classification, vegetation masking, elevation interpolation, and image cropping.

[0030] After preprocessing, the multimodal remote sensing data is standardized and samples are constructed to form a dataset for model training and site detection and identification.

[0031] (2) Basic feature extraction unit.

[0032] The basic feature extraction unit is used to extract basic features.

[0033] This invention employs different neural network backbone structures to extract basic features from remote sensing data based on the different modal types of multi-source remote sensing data.

[0034] Specifically, for image data, a convolutional neural network is used as the backbone to extract multi-scale basic two-dimensional features from multiple different stages. For laser point cloud data, a three-dimensional point cloud deep learning network is used as the backbone network to extract multi-scale basic three-dimensional features from multiple different stages.

[0035] (3) Feature enhancement unit.

[0036] The feature enhancement unit includes a planar feature enhancement module and a vertical feature enhancement module.

[0037] The planar feature enhancement module uses strip convolutions of different widths to obtain features with different receptive fields and performs multi-scale feature fusion to obtain enhanced planar features of the image. The vertical feature enhancement module calculates the elevation difference of the point cloud, generates elevation difference attention weights, and embeds the elevation difference attention weights into a self-attention mechanism and a convolutional neural network to obtain enhanced vertical features of the point cloud.

[0038] Specifically, this invention addresses the geometric contours, boundary connectivity, and spatial morphological features of archaeological sites and remains in the planar dimension. It designs planar feature and enhancement network modules to form a characteristic representation of the two-dimensional planar morphological distribution of archaeological sites and remains. Furthermore, it designs vertical dimension feature extraction and enhancement network modules to identify changes in terrain elevation and differences in underground structures, addressing the presence of surface undulations, stratigraphic accumulation, or burial anomalies in the vertical direction of archaeological sites and remains.

[0039] (4) Feature fusion unit.

[0040] The feature fusion unit includes a feature alignment module and a feature fusion module.

[0041] The feature alignment module is used to align the enhanced planar features of the image with the enhanced vertical features of the point cloud in terms of spatial location, scale, and channel dimension; wherein, key feature channels are selected through channel attention, and significant areas of the site and remains are aligned through spatial attention. The feature fusion module is used to perform a global-level cross-fusion operation of aligned image planar features and point cloud vertical features to obtain fused features.

[0042] (5) Combination loss function.

[0043] The combined loss function of the detection and recognition network model includes the main loss, image-assisted loss, and point cloud-assisted loss; the main loss consists of Dice loss and cross-entropy loss.

[0044] Overall, this invention employs an encoder-decoder structure and a multi-branch fusion structure to construct a unified network framework for detecting and identifying archaeological sites and relics, and designs a loss function. Utilizing complementary information about relics in both planar and vertical dimensions, a planar-vertical feature cross-fusion module (i.e., a feature fusion unit) is designed through feature attention weighting, mutual information constraints, or bidirectional gating mechanisms to achieve dynamic fusion of weak relic signals between the spatial and height domains. Using the weights of the network model trained on the training set, archaeological site relics are extracted from the test set. Furthermore, redundant and obviously erroneous results are eliminated through methods such as candidate region generation, connected component segmentation, shape constraint filtering, and thresholding.

[0045] The following uses visible light images and laser point cloud data as examples to detect and identify archaeological sites and remains in a certain area. The specific operation steps are as follows: Step 1: Acquisition and preprocessing of multimodal remote sensing data under vegetation shading conditions.

[0046] Taking an airborne lidar system platform as an example, we acquire multimodal remote sensing data such as optical images and laser point clouds of the target area under vegetation obstruction conditions, and preprocess each data to establish a dataset for subsequent archaeological site feature extraction and identification analysis.

[0047] First, based on the characteristics of vegetation cover and terrain undulation in the target area, an adaptive flight parameter planning algorithm is adopted. The main difference between this invention and existing data acquisition methods is that the flight altitude, overlap, and scanning angle of the airborne system platform are dynamically adjusted according to vegetation index grid and terrain elevation data. This difference can improve the echo ratio of ground point clouds in dense areas such as forest canopy and crops, thereby enhancing the data acquisition density of archaeological site point clouds under vegetation obstruction conditions.

[0048] Then, the acquired optical images, laser point clouds, and other data are preprocessed. For the preprocessing of lidar point cloud data, existing methods mainly retain ground points through filtering algorithms, which may mistakenly remove archaeological sites and relics. The main difference between this invention and existing methods is that it adopts a vegetation classification algorithm, which only separates and removes vegetation canopy points, while retaining all artificial structures such as buildings and archaeological sites and relics, as shown in formula (1). This difference can ensure the retention rate of micro-topographic point clouds such as archaeological sites and relics, thereby improving the ability to detect archaeological sites under vegetation cover; at the same time, this invention constructs a vegetation mask to guide the elevation interpolation algorithm of the point cloud, generating a high-precision digital elevation model, which effectively improves the fine description of the micro-topography of archaeological site areas under vegetation cover.

[0049] (1) in, Let i be the i-th point in the laser point cloud data; The height relative to the local ground level. This represents the minimum height threshold for the vegetation canopy. For local curvature or other geometric feature indices, The threshold value for vegetation canopy curvature; The set of points identified as vegetation canopy points.

[0050] Finally, based on existing archaeological data or known site distribution information, and combined with surface type and vegetation cover characteristics, several typical sample areas were selected as training and validation samples. Labels were assigned to the relics and background areas within the sample areas using annotation tools, and the multimodal remote sensing data were cropped and segmented to form a multimodal dataset with clear supervision information.

[0051] Step 2: Extraction and enhancement of multimodal remote sensing features of archaeological sites and remains.

[0052] Based on the respective advantages of remote sensing imagery and laser point cloud, as well as the characteristics of the archaeological site itself, weak signal features of the archaeological site are extracted and enhanced in both two-dimensional plane and three-dimensional vertical aspects.

[0053] First, based on the existing neural network backbone weights, basic features are extracted from remote sensing data, transferring existing knowledge. For image data, a ResNet network is used as the backbone to extract multi-scale basic two-dimensional features from multiple different stages; for laser point cloud data, a RandLA-Net network is used as the backbone to extract multi-scale basic three-dimensional features from multiple different stages and map them to two-dimensional space.

[0054] For example, the multiple different stages specifically refer to four different stages, corresponding to the four hierarchical stages in the ResNet (or RandLA-Net) backbone network, and the basic features of the image and point cloud are shown in formula (2). Of course, the present invention is not limited to four stages, and features can also be extracted from two, three or more stages.

[0055] (2) in, For different stages of the ResNet or RandLA-Net backbone network, To extract the basic two-dimensional features of the image for the ResNet backbone network in the s-stage, To extract the basic 3D features of the image for the RandLA-Net backbone network in the s-stage; Indicates the dimensions of the basic feature, where B, , , The batch size of the data, the number of channels in the feature map at stage s, and the height and width.

[0056] For example, the feature map sizes C×H×W for the four different stages are 64×128×128, 128×64×64, 256×32×32, and 512×16×16, respectively. However, this invention is not limited to the specific dimensions mentioned above; the corresponding feature map sizes can be adjusted when the input resolution or network changes.

[0057] Then, addressing the issue of fragmented and discontinuous remains due to surrounding vegetation obstruction and localized human activity, resulting in incomplete planar dimension representation and extraction, this invention creatively proposes a multi-scale feature extraction method for planar strips of archaeological remains based on optical image data. This method uses convolution operations with strip shapes of different receptive field sizes to constrain the process of extracting planar features of archaeological remains, as shown in formula (3). Through multi-stage, multi-level adaptive extraction of local, global, and planar strip features, the representation capability of weak signals of remains at the two-dimensional plane level is enhanced, improving the completeness of archaeological remains detection and identification. The flowchart is as follows: Figure 2 As shown.

[0058] (3) in, and This refers to performing n×1 and 1×n strip convolution operations on the image feature tensor, where n takes the values ​​7, 11, and 21 in the above formula; The corresponding planar strip features with a receptive field size of n, where n takes the values ​​7, 11, or 21 in the above formula; The two-dimensional planar features of the ruins plane; BN refers to the Batch Normalization (BN) operation. This refers to the addition operation of the feature tensor.

[0059] For example, the strip convolution sizes are 7×1 and 1×7, 11×1 and 1×11, and 21×1 and 1×21. However, the present invention is not limited to the specific sizes mentioned above, and can be adjusted accordingly based on the input image resolution, the scale of the target relic, and the actual computing power configuration.

[0060] Finally, addressing the technical challenges of representing and extracting incomplete vertical features of archaeological sites due to localized erosion and collapse caused by natural weathering, this invention creatively proposes a method for extracting and enhancing vertical features of archaeological sites that takes into account elevation difference attention mechanisms. Based on the advantages of three-dimensional laser point clouds in representing archaeological sites in a three-dimensional manner, a novel elevation difference attention mechanism is proposed. This mechanism calculates the elevation differences in the point cloud and generates attention weights, which are then embedded into a self-attention mechanism and a convolutional neural network. This achieves adaptive representation of weak signals in the vertical structure of the archaeological site, as shown in formula (4), thereby improving the completeness of archaeological site detection and identification.

[0061] (4) in, Based on three-dimensional features, MLP refers to Multilayer Perceptron. The classification results of elevation information, i.e., the category index. The category index is expanded into a one-dimensional sequence. PE represents the number of elevation categories, and PE represents the location index. This is an elevation index matrix. For the flattened elevation index, For parameter weights, Pay attention to elevation differences.

[0062] The Softmax and Argmax functions transform the laser point cloud feature map into a probability distribution and a class index, respectively; the flatten function expands the class index into a one-dimensional feature, and the reshape function transforms it into a vertical one-dimensional feature. The two are then subtracted. Operations, and elevation categories Adding the position index PE to the elevation index matrix yields the elevation index matrix. Finally, the elevation index is used after flattening. Assign different parameter weights Constructing elevation difference attention The flowchart is as follows Figure 3 As shown.

[0063] Step 3: Archaeological relic detection and identification and anomaly handling based on cross-fusion of planar-vertical multimodal features.

[0064] Based on the characteristics of archaeological sites and relics from two dimensions—image plane and point cloud elevation difference—this invention designs a network model for the detection and identification of archaeological sites and relics by cross-fusion of image and point cloud plane-elevation features, and processes the abnormal identification results.

[0065] First, an encoder-decoder architecture is adopted to design the network model, using high-resolution imagery and 3D laser point cloud data as input. In the encoder stage, modules for extracting 2D image planar strip features and 3D point cloud elevation difference features are designed, rather than the simple fusion of existing general CNN / RandLA-Net technologies. Figure 4 As shown in the figure. Based on the different modal features of the image and point cloud, a combined loss function specifically designed for modal differences is presented, as shown in formula (5). This fully utilizes the complementary advantages of high-resolution texture in the image and stereoscopic elevation in the point cloud to accurately capture the planar distribution and vertical structure of archaeological remains, thereby improving the accuracy of remains detection under complex vegetation cover.

[0066] (5) in, The combined loss function, the main loss It is implemented by combining Dice loss and cross-entropy loss; two-dimensional image scale-assisted loss (i.e., image-assisted loss). And 3D point cloud auxiliary loss (i.e., point cloud auxiliary loss) It is mainly based on the Cross Entropy Loss construction; the hyperparameters α and β are set to 0.4 in this invention, mainly because the loss terms corresponding to images and point clouds are equivalent, but it is not limited to this value and can be adjusted according to the loss distribution of the specific task and the training effect.

[0067] First, existing fusion methods often employ simple concatenation or additive fusion, neglecting the scale and semantic differences between image and point cloud modalities. This invention innovatively designs a feature alignment module based on channel and spatial attention mechanisms, as shown in formula (6). It adaptively selects key feature channels through channel attention and focuses on significant areas of the archaeological site through spatial attention, thereby achieving optimal selection of multimodal features in the local area of ​​the archaeological site.

[0068] (6) in, Based on two-dimensional features, and These are the image and point cloud output features after multimodal feature alignment, respectively. and These are image and point cloud modal features that apply channel attention weights. and These are image and point cloud modal features that apply spatial attention weights, respectively.

[0069] Then, in response to the technical problem of the lack of global feature fusion in existing methods, this invention proposes a feature fusion module based on cross-attention mechanism, as shown in formula (7), which performs global-level cross-fusion and complementary perception of features from different sources of images and point clouds, solves the problem of modal heterogeneity of images and point clouds, and thus achieves the ability to densely and accurately detect and identify the boundaries of relics in a vegetation-covered environment.

[0070] (7) in, and These are the image and point cloud output features after global multimodal feature fusion, respectively; , , and , , These refer to the Query, Key, and Value vectors in the image and point cloud modal self-attention mechanisms, respectively, where T represents transpose and d... k This refers to the size of the feature dimension; This refers to the multiplication operation of tensors.

[0071] Finally, morphological operations and multi-scale filtering can be used to remove isolated small regions and noisy false detection points, while retaining candidate areas of relics with continuity and typical morphological features.

[0072] In summary, the site and relic detection and identification method provided in Example 1 is based on multimodal remote sensing data and deep learning methods to detect and identify archaeological relics. It can overcome the challenge of the current difficulty in detecting and identifying archaeological sites and relics using a single band or a single data source. In areas with dense vegetation or severe surface disturbance, it can improve the completeness and automation of archaeological site and relic detection and identification.

[0073] Example 2: Example 2 provides a method for detecting and identifying archaeological sites and remains. The difference between Example 1 and Example 2 is that, after obtaining the detection and identification results, Example 2 further includes: verifying and evaluating the detection and identification results. (See [link to example 1]). Figure 5 .

[0074] This involves establishing verification benchmarks based on existing archaeological survey results, ground exploration data, and high-precision geographic information data to verify and evaluate the identification results of archaeological remains, including indicators such as accuracy and intersection-over-union ratio. Based on quantitative indicators, the results of archaeological site and remains detection and identification can be visualized and analyzed for interpretability.

[0075] Specifically, the results of intelligent identification of archaeological sites are verified by combining interpretation by archaeological experts with on-site investigations, and the accuracy and completeness are evaluated to provide support for further promotion and application.

[0076] In terms of evaluation metrics for the accuracy and integrity of archaeological site detection, this invention uses pixel accuracy (PA) and intersection over union (IoU).

[0077] Where PA represents the ratio of the number of correctly extracted pixels to the total number of pixels in the image, as shown in formula (8).

[0078] (8) IoU represents the ratio of the intersection of the predicted map set and the ground truth map set to the union of the two sets, as shown in formula (9).

[0079] (9) In the formula, k represents the number of pixel categories; This represents the number of pixels whose actual category is i and whose predicted category is also i. This represents the total number of pixels of category i; This represents the number of pixels whose actual category is i and whose predicted category is j.

[0080] Furthermore, by overlaying multimodal remote sensing images with actual site distribution maps, the spatial distribution consistency and identification completeness of site targets can be intuitively assessed. Additionally, by combining the candidate region confidence scores output by the post-processing module, a site probability heatmap can be generated to analyze the model's identification sensitivity under different vegetation densities.

[0081] Example 3: Example 3 provides a site and relic detection and identification system, comprising: The model building unit is used to build a detection and identification network model and train the detection and identification network model using sample data labeled with site and relic tags; The data acquisition unit is used to acquire image data and laser point cloud data; The detection and identification unit includes a trained detection and identification network model. The detection and identification unit is used to detect and identify the site and relics based on the input image data and laser point cloud data using the trained detection and identification network model, and output the detection and identification result information. The site and relic detection and identification system is used to perform the steps in the site and relic detection and identification method as described in Example 1.

[0082] Since the functions of each unit in the site and relic detection and identification system provided in Embodiment 3 correspond to the steps in the site and relic detection and identification method provided in Embodiment 1, Embodiment 3 can be understood by referring to the description of Embodiment 1, and will not be repeated here.

[0083] Example 4: Example 4 provides a site and relic detection and identification system. The difference between Example 4 and Example 3 is that Example 4 further includes a verification and evaluation unit; the verification and evaluation unit is used to verify and evaluate the detection and identification result information.

[0084] The verification and evaluation unit performs the verification and evaluation operation as described in Example 2, and therefore can be understood by referring to the description of Example 2, which will not be repeated here.

[0085] In summary, this invention provides a multimodal remote sensing data scheme for detecting and identifying archaeological sites and relics under vegetation cover. By comprehensively utilizing different modalities of data and employing data registration, feature enhancement, and fusion modeling techniques, this invention significantly improves the salience of relics in vegetation-covered areas. Furthermore, through deep feature representation and multi-scale classification strategies, this invention achieves accurate identification and automated discrimination against different surface backgrounds, thereby enhancing the accuracy and stability of relic identification. Finally, this invention provides efficient and reliable technical support for archaeological surveys and cultural heritage protection.

[0086] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting and identifying archaeological sites and remains, characterized in that, Includes the following steps: Construct and train the detection and recognition network model; The detection and recognition network model includes a preprocessing unit, a basic feature extraction unit, a feature enhancement unit, and a feature fusion unit connected in sequence. The feature enhancement unit is used to enhance the basic features in both planar and vertical aspects. The acquired image data and laser point cloud data are input into the trained detection and recognition network model, which outputs the detection and recognition results of the site and relics.

2. The method for detecting and identifying archaeological sites and remains according to claim 1, characterized in that, The preprocessing unit is used to preprocess image data and laser point cloud data; The preprocessing of laser point cloud data includes: classifying vegetation, separating and removing vegetation canopy points, and preserving the structure of the site and the micro-topography of the surface; constructing a vegetation mask based on the vegetation classification results; and using the vegetation mask to guide the elevation interpolation algorithm of the point cloud to generate a digital elevation model, thereby achieving a detailed description of the structure of the site and the micro-topography of the surface under vegetation shading conditions.

3. The method for detecting and identifying archaeological sites and remains according to claim 1, characterized in that, The basic feature extraction unit is used to extract basic features; Specifically, for image data, a convolutional neural network is used as the backbone to extract multi-scale basic two-dimensional features from multiple different stages; for laser point cloud data, a three-dimensional point cloud deep learning network is used as the backbone network to extract multi-scale basic three-dimensional features from multiple different stages.

4. The method for detecting and identifying archaeological sites and remains according to claim 1, characterized in that, The feature enhancement unit includes a planar feature enhancement module and a vertical feature enhancement module; The planar feature enhancement module uses strip convolutions of different widths to obtain different receptive field features and performs multi-scale feature fusion to obtain enhanced image planar features. The vertical feature enhancement module calculates the elevation difference of the point cloud, generates elevation difference attention weights, and embeds the elevation difference attention weights into the self-attention mechanism and the convolutional neural network to obtain the enhanced vertical features of the point cloud.

5. The method for detecting and identifying archaeological sites and remains according to claim 1, characterized in that, The combined loss function of the detection and recognition network model includes the main loss, image-assisted loss, and point cloud-assisted loss; the main loss consists of Dice loss and cross-entropy loss.

6. The method for detecting and identifying archaeological sites and remains according to claim 4, characterized in that, The feature fusion unit includes a feature alignment module and a feature fusion module; The feature alignment module is used to align the enhanced planar features of the image with the enhanced vertical features of the point cloud in terms of spatial location, scale, and channel dimension; wherein, key feature channels are selected through channel attention, and significant areas of the site and remains are aligned through spatial attention. The feature fusion module is used to perform a global-level cross-fusion operation of aligned image planar features and point cloud vertical features to obtain fused features.

7. The method for detecting and identifying archaeological sites and remains according to claim 1, characterized in that, The image data includes one or more of the following: visible light and near-infrared optical image data, synthetic aperture radar data, and multispectral / hyperspectral image data.

8. The method for detecting and identifying archaeological sites and remains according to claim 1, characterized in that, After obtaining the detection and identification result information, the method further includes: verifying and evaluating the detection and identification result information.

9. A system for detecting and identifying archaeological sites and remains, characterized in that, include: The model building unit is used to build a detection and identification network model and train the detection and identification network model using sample data labeled with site and relic tags; The data acquisition unit is used to acquire image data and laser point cloud data; The detection and identification unit includes a trained detection and identification network model. The detection and identification unit is used to detect and identify the site and relics based on the input image data and laser point cloud data using the trained detection and identification network model, and output the detection and identification result information. The site and relic detection and identification system is used to perform the steps in the site and relic detection and identification method as described in any one of claims 1 to 8.

10. The archaeological site and relic detection and identification system according to claim 9, characterized in that, Also includes: Verification and evaluation unit; The verification and evaluation unit is used to verify and evaluate the detection and identification result information.

Images