National village building identification method and device based on deep learning

Abstract

The invention discloses a national village building identification method and device based on deep learning, and the method comprises the steps: obtaining low-altitude unmanned plane remote sensing image data of a national building in a research region, and carrying out the preprocessing, and obtaining a digital orthoimage; constructing a ethnic building sample library based on the digital orthoimage and a sliding window method; constructing a national village building image segmentation model based on a channel attention mechanism, and performing training through a national building sample library; and obtaining segmented images through the trained national village building image segmentation model, and identifying national village buildings based on the segmented images. According to the method, the precision and efficiency of ethnic building recognition are effectively improved, the problems of low efficiency and difficult data acquisition of a traditional method are solved, the extraction capability of the model for ethnic building features is remarkably enhanced, the accuracy and robustness of a recognition result are improved, and efficient technical support is provided for ethnic building heritage protection.

Description

Technical Field

[0001] This invention belongs to the field of ethnic village building recognition technology, and particularly relates to a method and device for ethnic village building recognition based on deep learning. Background Technology

[0002] As an important carrier of the cultural heritage of China's ethnic minorities, ethnic architecture possesses rich architectural heritage, containing abundant historical data and ethnic cultural connotations. With the advancement of urbanization and modernization, as well as the impact of natural disasters, ethnic minority buildings have suffered varying degrees of damage. Quickly and accurately extracting these buildings from remote sensing imagery is of great significance for promoting the application of remote sensing information in urban management and architectural heritage protection.

[0003] Currently, the main methods for detailed investigation and identification of ethnic minority architecture are field surveys and traditional mapping. For example, a combination of 3D laser image scanning and digital close-range photogrammetry is used to protect the cultural heritage of ethnic minority architecture. However, due to the large number and wide distribution of ethnic minority buildings, manual field investigation and identification requires significant manpower and resources. Relying solely on existing methods such as field surveys makes rapid and accurate identification difficult. Therefore, some experts use medium-resolution remote sensing data, such as Landsat, to map remote rural settlements. However, compared to the scale of ethnic minority settlements, the identification results are relatively coarse, requiring higher-resolution imagery. Furthermore, remote sensing images of ethnic minority buildings possess numerous and detailed features, while traditional methods relying on simple remote sensing indices and manual feature descriptions cannot accurately and quickly identify ethnic minority buildings. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method and apparatus for identifying ethnic village architecture based on deep learning, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, this invention provides a method for identifying ethnic village architecture based on deep learning, comprising:

[0006] Low-altitude UAV remote sensing image data of ethnic buildings in the study area were acquired and preprocessed to obtain digital orthophotos;

[0007] A sample database of ethnic architecture was constructed based on the aforementioned digital orthophotos and the sliding window method.

[0008] A segmentation model for ethnic village architecture images was constructed based on the channel attention mechanism and trained using the aforementioned ethnic architecture sample database;

[0009] Segmented images are obtained by training an image segmentation model of ethnic village buildings, and ethnic village buildings are identified based on the segmented images.

[0010] Optionally, the process of acquiring low-altitude UAV remote sensing image data of ethnic buildings in the study area and preprocessing it to obtain digital orthophotos includes:

[0011] Based on environmental factors and field surveys to determine the data acquisition time and photography conditions, the low-altitude airborne measurement system is geometrically and radiometrically calibrated. Remote sensing image data is then obtained by combining the data acquisition time and photography conditions. Image processing is performed on the remote sensing image data to obtain digital orthophotos. The low-altitude airborne measurement system includes a multi-rotor UAV and an oblique photography camera.

[0012] Optionally, the process of constructing a sample database of ethnic architecture includes:

[0013] The digital orthophoto is visually interpreted to mark ethnic architectural areas, and binarized label images are generated through element transformation. The digital orthophoto and the corresponding binarized label images are cropped into windows of fixed pixel size using a sliding window method to obtain sub-samples. Data augmentation operations are then performed on the sub-samples to obtain an ethnic architectural sample library. The data augmentation operations include rotation, flipping, brightness and contrast adjustment, and Gaussian blur enhancement.

[0014] Optionally, when using the sliding window method for cropping, the overlap rate between adjacent windows is set to 10%, and the cropping area outside each image boundary is filled with zero pixel values.

[0015] Optionally, the ethnic village building image segmentation model includes an encoder, a decoder, and a channel attention mechanism SENet module; the encoder includes a backbone network and a spatial pyramid pooling module; and the channel attention mechanism SENet module is embedded at the output of the encoder.

[0016] Optionally, the SENet channel attention mechanism module uses the encoder output as the input feature map, performs global average pooling and global max pooling on the input feature map in parallel to generate two independent feature vectors; the two independent feature vectors are concatenated to obtain a fusion descriptor; the fusion descriptor is processed through two fully connected layers to obtain a normalized channel weight vector; a weighted feature map is obtained based on the channel weight vector and the input feature map.

[0017] Optionally, the process of obtaining segmented images through the trained ethnic village architectural image segmentation model includes:

[0018] A shallow feature map is obtained through the backbone network, and a weighted feature map is obtained through the encoder and the SENet channel attention mechanism module. The shallow feature map and the weighted feature map are input into the encoder and decoder to output a segmented image.

[0019] Optionally, an ethnic village identification model is constructed, and ethnic village buildings are identified based on the segmented image and the ethnic village identification model; when training the ethnic village building image segmentation model and the ethnic village identification model, a joint function of BCE loss and Dice loss is used as the loss function.

[0020] This invention also provides a deep learning-based device for identifying ethnic village architecture, comprising:

[0021] The data acquisition module is used to acquire high-resolution remote sensing images of ethnic village buildings using an oblique photography camera mounted on a drone.

[0022] The data processing module is used to perform geometric calibration, radiometric calibration and image preprocessing on the remote sensing image to generate a digital orthophoto.

[0023] The sample construction module is used to construct a sample library of ethnic architecture based on the digital orthophotos through sliding window cropping and data augmentation.

[0024] The model processing module constructs an image segmentation model of ethnic village buildings based on the channel attention mechanism and trains it using the ethnic building sample library.

[0025] The recognition output unit obtains segmented images through the trained ethnic village building image segmentation model, and identifies ethnic village buildings based on the segmented images.

[0026] Compared with the prior art, the present invention has the following advantages and technical effects:

[0027] This invention can obtain high-resolution multispectral remote sensing images without damaging ethnic architecture, clearly showing the geomorphological features of the study area, and also fully containing RGB and visible light three-channel information, providing high-quality data support for subsequent analysis and research. The sample library construction strategy not only effectively reduces the workload of deep learning semantic annotation, but also successfully expands the sample dataset. Based on high-performance computing, the channel attention mechanism SENet is introduced into the DeepLabv3+ network model, which improves the recognition model's attention to important features and recognition ability. The SE_DeepLabv3+ ethnic architecture extraction model shows excellent accuracy, with an accuracy of 0.9831, a recall of 0.9743, an F1 score of 0.9787, and a mean intersection-over-union ratio of 0.9582. Attached Figure Description

[0028] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0029] Figure 1This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0032] Example 1

[0033] like Figure 1 As shown, this embodiment provides a method for identifying ethnic village architecture based on deep learning, including:

[0034] Low-altitude UAV remote sensing image data of ethnic buildings in the study area were acquired and preprocessed to obtain digital orthophotos;

[0035] As a specific implementation method, the process of acquiring low-altitude UAV remote sensing image data of ethnic architecture in the study area and preprocessing it to obtain digital orthophotos includes:

[0036] Based on environmental factors and field surveys to determine data acquisition time and photography conditions, geometric and radiometric calibrations are performed on the low-altitude airborne measurement system. Remote sensing image data is then obtained by combining the data acquisition time and photography conditions. Image processing is performed on the remote sensing image data to obtain digital orthophotos. The low-altitude airborne measurement system includes a multi-rotor UAV and an oblique photography camera.

[0037] Specifically, a low-altitude airborne measurement system was constructed using a DJI Matrice M300 RTK multi-rotor drone equipped with a Ruipu D2M oblique photography camera (104.5 x 104.587 mm in size, with a total pixel count of 130 million). Rigorous geometric and radiometric calibrations were performed on the low-altitude airborne measurement system to ensure the quality of the acquired data. When planning the aerial photography process, this embodiment fully considered the potential impact of sunlight and wind on image quality, avoiding the loss of some image texture features due to cloud cover. Data acquisition was conducted between 9:00 AM and 11:00 AM, when wind speeds were low and lighting conditions were suitable. Field surveys determined a flight altitude of 120 meters above the terrain, and a waypoint hovering photography method was adopted. The heading and lateral overlap rates were carefully set to 80% and 70%, respectively, to ensure image integrity and redundancy. All flight parameters were set using DJIPilot 2 software, and detailed flight mission planning was executed accordingly, achieving fully automated waypoint flight operations, improving operational efficiency and flight safety. Utilizing the built-in application software on the controller and the DJI Maps platform, the acquired data was efficiently transmitted and meticulously processed. The processing workflow encompassed key stages such as image preprocessing and accuracy assessment. In the data preprocessing stage, PSO data from the lens was first exported, followed by the export of raw photos from the camera. The corresponding positional information was then calculated using CGO (CHC Geomatics Office) (a professional GNSS data processing software developed by Shanghai Huace, primarily used for post-processing of static and dynamic data, supporting multiple data formats and calculation functions), and uniformly imported into aerial triangulation software for aerial triangulation. This yielded the precise spatial coordinates of each photo. Accuracy was then verified using checkpoints obtained during the preparation stage, and an accuracy report was exported. In the image preprocessing stage, the main focus was on addressing issues such as deformation, distortion, blurring, and noise caused by drone flight vibrations. Image enhancement, retouching, cropping, and reconstruction techniques were employed to acquire clear, centimeter-high resolution images of the study area. To verify the image accuracy, visible light imagery was used as a base map, and data acquisition was conducted through a human-computer interaction approach, followed by on-site verification within the study area. Finally, a digital orthophoto (DOM) of the study area was successfully generated. This image not only clearly shows the topographic features of the study area, but also fully contains RGB visible light three-channel information, providing high-quality data support for subsequent analysis and research.

[0038] A sample database of ethnic architecture was constructed based on digital orthophotos and the sliding window method.

[0039] As a specific implementation method, the process of constructing a sample database of ethnic architecture includes:

[0040] The digital orthophotos were visually interpreted to identify ethnic architectural areas, and binarized label images were generated through feature transformation. The digital orthophotos and their corresponding binarized label images were cropped into fixed-pixel windows using a sliding window method to obtain subsamples. Data augmentation operations were then performed on the subsamples to obtain an ethnic architectural sample library. The data augmentation operations included rotation, flipping, brightness and contrast adjustment, and Gaussian blur enhancement.

[0041] As a specific implementation method, when using the sliding window method for cropping, the overlap rate between adjacent windows is set to 10%, and the cropping area outside each image boundary is filled with zero pixel values.

[0042] Specifically, ArcGIS software was used to visually interpret and annotate the entire UAV imagery. The resulting vector files were then converted into raster images as label images. These label images represent the actual surface conditions of the ethnic minority buildings in the UAV imagery. Since ethnic minority villages often contain various land types, such as roads and trees, to ensure data quality, all features except ethnic minority buildings were treated as background during the sample annotation process. Furthermore, during the conversion of vector features to raster images, the raster pixel value for ethnic minority buildings was specifically set to 255 (representing maximum brightness or white), while the background raster pixel value was set to 0 (representing black or minimum brightness). This binarization of feature categories facilitates the training process of the deep learning model. A sliding window method was used to crop the low-altitude remote sensing and annotation files of ethnic minority buildings into 512×512 pixel windows, ensuring a 10% overlap between adjacent windows. During the cropping process, to maintain consistency between training and testing samples, the cropped area outside the boundaries of each image was filled with zero pixel values. A sample set of ethnic architectural features was constructed by randomly cropping images of the study area and their corresponding labeled images. In addition, sample augmentation techniques were employed to increase the diversity of samples used for training on ethnic architectural features. Specifically, geometric transformations were achieved through rotation (including 90° and 270°) and flipping (vertical and horizontal flipping), color transformations were performed using color adjustments (such as adjusting brightness, contrast, and chroma), and blurring (Gaussian blur) was used for pixel transformation. Finally, 70% of the samples were selected as training data to train and update the deep learning model parameters, 20% as validation data to evaluate issues such as model overfitting during training, and 10% as test data to evaluate the final model. Random sampling was used to divide the dataset to ensure fairness and balance in sample quality.

[0043] A segmentation model for ethnic village architecture images was constructed based on the channel attention mechanism and trained using an ethnic architecture sample database.

[0044] Segmented images are obtained by training an image segmentation model of ethnic village buildings, and ethnic village buildings are identified based on the segmented images.

[0045] As a specific implementation method, the ethnic village building image segmentation model includes an encoder, a decoder, and a channel attention mechanism SENet module; the encoder includes a backbone network and a spatial pyramid pooling module; and the channel attention mechanism SENet module is embedded at the output of the encoder.

[0046] As a specific implementation, the SENet channel attention mechanism module takes the encoder output as the input feature map, performs global average pooling and global max pooling on the input feature map in parallel to generate two independent feature vectors; the two independent feature vectors are concatenated to obtain a fusion descriptor; the fusion descriptor is processed through two fully connected layers to obtain a normalized channel weight vector; a weighted feature map is obtained based on the channel weight vector and the input feature map.

[0047] As a specific implementation method, the process of obtaining segmented images through a trained ethnic village architectural image segmentation model includes:

[0048] Shallow feature maps are obtained through the backbone network, and weighted feature maps are obtained through the encoder and channel attention mechanism SENet module. The shallow feature maps and weighted feature maps are input into the encoder and decoder to output the segmented image.

[0049] As a specific implementation method, an ethnic village identification model is constructed, and ethnic village buildings are identified based on the segmented image and the ethnic village identification model; when training the ethnic village building image segmentation model and the ethnic village identification model, a joint function of BCE loss and Dice loss is used as the loss function.

[0050] Specifically, the experimental hardware configuration was as follows: the training environment consisted of an Ubuntu 18.04 operating system, using Python 3.7, and TensorFlow-GPU version 2.4 was installed. The CPU was an Intel Xeon E5 series, and the GPU was an NVIDIA GeForce RTX 3090.

[0051] The experimental training parameters are shown in Table 1. The input image resolution is 512 pixels × 512 pixels.

[0052] Table 1

[0053]

[0054] The DeepLabv3+ model works by fusing shallow image features with deep upsampled features, creating an encoder-decoder mechanism. This mechanism effectively improves segmentation while paying attention to image boundary information. In the encoding stage, the model mainly consists of a feature extraction backbone network and an Atrous Spatial Pyramid Pooling (ASPP) module. The backbone network is responsible for feature learning optimization and generating coarse feature maps, while the ASPP module uses convolutional kernels of different sizes to obtain high-level abstract feature maps, thus achieving deeper feature sampling. In the decoding stage, the model combines pixel location information with spatial information using high-dimensional and low-dimensional features, and finally uses bilinear interpolation upsampling to complete pixel-level segmentation. Specifically, after UAV imagery is input into the DeepLabv3+ network, the encoder module encodes multi-scale contextual information by applying zigzag convolutions at different scales, thereby extracting image features. Subsequently, the decoder module refines the segmentation results along object boundaries, restoring image features and compensating for missing edge information.

[0055] Meanwhile, the DeepLabv3+ network model introduces the channel attention mechanism SENet. The SENet module, by assigning different weights to each channel, effectively enhances information-rich channel features while suppressing less information-rich channel features, thus focusing on improving the network's ability to extract ethnic architectural features. This SE module mainly consists of two key operations: Squeeze and Excitation. The Squeeze operation performs global average pooling and global max pooling based on the height and width of the feature map, compressing the feature map within each channel into a global value. This value reflects a certain response characteristic of the channel's global receptive field. The Squeeze operation is defined as follows:

[0056]

[0057] Among them, f c Let h be the feature map with feature channel c, h be the height of the feature map, and w be the width of the feature map.

[0058] The excitation operation aims to implement the functionality of an attention mechanism. It uses a parameter w to generate corresponding weights for each feature channel. These weights are then normalized using a sigmoid activation function to ensure they fall within the range (0,1). The detailed definition of this process is as follows:

[0059] s = F ex (z,ω)=σ(h(z,ω1))=σ(ω2δ(ω1*z))

[0060] In the formula, ω (including ω1 and ω2) are the parameters of the fully connected layer, δ is the ReLU activation function, and σ is the sigmoid activation function. The normalized weights obtained using the σ (i.e., sigmoid) activation function are used to weight the features of the original channels. That is, feature extraction and weighting are achieved through two fully connected layers. The first fully connected layer typically has a smaller output dimension to reduce computational complexity and the number of parameters, and uses the ReLU activation function for non-linear transformation. The second fully connected layer restores the feature dimension to the same number of channels as the original. Subsequently, a weight recalibration operation is performed, where the feature maps obtained from the two pooling operations (average pooling and max pooling) are weighted and summed, and the result is normalized to between 0 and 1 using the sigmoid function, thus obtaining the attention weight for each channel. Finally, these attention weights are multiplied by the original feature maps to achieve channel-level feature reweighting, resulting in the final weighted feature map.

[0061] In image segmentation tasks, a loss function is typically used for backpropagation to guide the optimization of network parameters during each learning iteration, ultimately classifying pixels in the sample. The loss function is mainly used to evaluate the degree of inconsistency between the model's predicted values ​​and the true values ​​(labeled values); the smaller the loss function value, the higher the model's classification performance is considered. The problem of segmenting images of ethnic buildings and non-ethnic buildings from drones is a binary classification problem. A joint function of BCE Loss and Dice Loss is used as the loss function in the experiment. This allows the loss function to consider both the classification error of each pixel (i.e., the role of the BCE part), handling class imbalance, and the overall structural similarity between the predicted and true labels (i.e., the role of the Dice part), stabilizing the training process and improving model performance.

[0062] Images from ethnic architectural training samples are input into the depplabv3+ network incorporating channel attention mechanism and trained to obtain semantic-level segmented images. The semantic-level segmented images are then evaluated for semantic segmentation accuracy.

[0063] The deep learning framework Keras extracts global features from segmented images that meet accuracy requirements and trains them to obtain an automatic classification model for ethnic architecture based on deep learning.

[0064] This invention uses the depplabv3+ network to preprocess training samples and performs accuracy evaluation. Image sets whose accuracy meets the requirements are then trained using the Keras deep learning framework to obtain an automatic ethnic architecture recognition model. Images meeting certain accuracy evaluation requirements are used to establish a feature model of ethnic architecture—the automatic ethnic architecture recognition model—through global feature learning. This enables large-scale automatic recognition and classification of ethnic architecture. When remote sensing images of ethnic architecture need to be identified, simply input the remote sensing images into the trained automatic ethnic architecture recognition model to obtain the recognition and classification results.

[0065] Precision, recall, intersection-over-union (IoU), and F1 score are used to measure the deep learning model's ability to identify ethnic minority architecture. These metrics are calculated by constructing a confusion matrix, which is defined in Table 2 for binary classification tasks in deep learning.

[0066] Table 2

[0067]

[0068] True positive (TP) represents the number of samples correctly predicted as positive. False positive (FP) represents the number of samples incorrectly predicted as positive (i.e., actually negative). True negative (TN) represents the number of samples correctly predicted as negative. False negative (FN) represents the number of samples incorrectly predicted as negative (i.e., actually positive).

[0069] Category pixel accuracy, also known as precision, represents the probability that a certain category is correctly predicted in the prediction results.

[0070]

[0071] Recall, also known as sensitivity, is the ratio of correctly predicted positive observations to all observations in a class.

[0072]

[0073] The mean intersection-union ratio (MIRR) is one of the standard metrics for semantic segmentation. It represents the ratio of the intersection to the union of the ground truth and predicted segmentations.

[0074]

[0075] The F1 score is the harmonic mean of precision and recall. It comprehensively considers the precision and recall of a classification model, providing a balanced metric for evaluating model performance. The F1 score ranges from 0 to 1, with values ​​closer to 1 indicating higher prediction accuracy and values ​​closer to 0 indicating lower prediction accuracy.

[0076]

[0077] In addition, to verify the classification results, a ground-based verification method was used to visually interpret, compare, and analyze the identification results of ethnic buildings.

[0078] This invention also provides a deep learning-based device for identifying ethnic village architecture, comprising:

[0079] The data acquisition module is used to acquire high-resolution remote sensing images of ethnic village buildings using an oblique photography camera mounted on a drone.

[0080] The data processing module is used to perform geometric calibration, radiometric calibration, and image preprocessing on remote sensing images to generate digital orthophotos.

[0081] The sample construction module is used to build a sample library of ethnic architecture based on digital orthophotos through sliding window cropping and data augmentation.

[0082] The model processing module constructs an image segmentation model of ethnic village buildings based on the channel attention mechanism and trains it using an ethnic building sample database.

[0083] The recognition output unit obtains segmented images through the trained ethnic village building image segmentation model, and identifies ethnic village buildings based on the segmented images.

[0084] This invention proposes a method for identifying ethnic minority architectural heritage by combining low-altitude UAV remote sensing technology and improved deep learning algorithms. To overcome the limitations of satellite remote sensing in acquiring high spatial resolution images in a timely manner, and to enable precise detection in unique geographical environments and architectural types, obtaining data from areas difficult for humans to access (such as small gaps between buildings), this invention effectively addresses the limitations of manual surveys in providing accurate or comprehensive data. It utilizes UAVs to acquire ethnic architectural data and employs deep learning algorithms to extract identification features from massive amounts of data, overcoming the low efficiency of traditional identification methods and reducing the workload of data collection. The inclusion of an attention mechanism in the deep learning algorithm significantly reduces information loss and obtains higher-level features through weighted processing, thereby better preserving key information such as edges and textures in low-altitude remote sensing images of ethnic buildings, and improving the accuracy of edge segmentation.

[0085] Based on high-performance computing, the SENet channel attention mechanism was introduced into the DeepLabv3+ network model, which improved the recognition model's attention to and recognition ability of important features. The SE_DeepLabv3+ ethnic architecture extraction model showed excellent accuracy, with an accuracy of 0.9831, a recall of 0.9743, an F1 score of 0.9787, and a mean intersection-over-union ratio of 0.9582.

[0086] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying ethnic village buildings based on deep learning, characterized in that, The method comprises the following steps: acquiring low-altitude unmanned aerial vehicle remote sensing image data of ethnic buildings in a research area and preprocessing the data to obtain a digital orthographic image; constructing an ethnic building sample library based on the digital orthographic image and a sliding window method; constructing an ethnic village building image segmentation model based on a channel attention mechanism and training the model using the ethnic building sample library; obtaining a segmented image using the trained ethnic village building image segmentation model, and identifying ethnic village buildings based on the segmented image.

2. The deep learning-based ethnic village building identification method according to claim 1, wherein the process of acquiring low-altitude unmanned aerial vehicle remote sensing image data of ethnic buildings in a research area and preprocessing the data to obtain a digital orthographic image comprises: based on environmental factors and field surveys, determining data acquisition time and photography conditions, after geometric calibration and radiation calibration of a low-altitude airborne measurement system, combining the data acquisition time and photography conditions to obtain remote sensing image data, and performing image processing on the remote sensing image data to obtain a digital orthographic image; wherein the low-altitude airborne measurement system comprises a multi-rotor unmanned aerial vehicle and a tilt photography camera.

3. The deep learning-based ethnic village building identification method according to claim 1, wherein the process of constructing an ethnic building sample library comprises: performing visual interpretation on the digital orthographic image, marking ethnic building areas, and generating a binary label image through feature conversion; using a sliding window method to crop the digital orthographic image and the corresponding binary label image into windows of a fixed pixel size to obtain sub-samples, and performing data enhancement operations on the sub-samples to obtain an ethnic building sample library; wherein the data enhancement operations include rotation, flipping, brightness and contrast adjustment, and Gaussian blur enhancement.

4. The deep learning-based ethnic village building identification method according to claim 3, wherein when the sliding window method is used for cropping, the overlap rate between adjacent windows is set to 10%, and the cropped areas outside each image boundary are filled with zero pixel values.

5. The deep learning-based ethnic village building identification method according to claim 1, wherein the ethnic village building image segmentation model comprises an encoder, a decoder, and a channel attention mechanism SENet module; the encoder comprises a backbone network and a spatial pyramid pooling module; and the channel attention mechanism SENet module is embedded at the output end of the encoder.

6. The deep learning-based ethnic village building identification method according to claim 5, wherein the channel attention mechanism SENet module takes the output of the encoder as an input feature map, performs global average pooling and global maximum pooling on the input feature map in parallel to generate two independent feature vectors; the two independent feature vectors are spliced to obtain a fusion descriptor; the fusion descriptor is processed through two fully connected layers to obtain a normalized channel weight vector; and a weighted feature map is obtained based on the channel weight vector and the input feature map.

7. The deep learning-based ethnic village building identification method according to claim 6, wherein ​ ​ ​ ​ ​ The process of obtaining a segmentation image by the trained ethnic village building image segmentation model comprises: obtaining a shallow feature map through the backbone network, obtaining a weighted feature map through the encoder and the channel attention mechanism SENet module, inputting the shallow feature map and the weighted feature map into the add-decoder, and outputting a segmentation image.

8. The deep learning-based ethnic village building identification method according to claim 1, characterized in that, an ethnic village identification model is constructed, and ethnic village buildings are identified based on the segmentation image and the ethnic village identification model; during training of the ethnic village building image segmentation model and the ethnic village identification model, a joint function of a BCE loss and a Dice loss is used as a loss function. 9.A deep learning based ethnic village building recognition device for implementing the method of any one of claims 1-8, characterized in that, comprises: a data acquisition module configured to acquire high-resolution remote sensing images of ethnic village buildings by a tilt photography camera carried by a UAV; a data processing module configured to perform geometric calibration, radiation calibration and image preprocessing on the remote sensing images to generate digital orthographic images; a sample construction module configured to construct an ethnic building sample library based on the digital orthographic images by sliding window cropping and data enhancement; a model processing module configured to construct an ethnic village building image segmentation model based on a channel attention mechanism and train the model by using the ethnic building sample library; an identification output unit configured to obtain a segmentation image by the trained ethnic village building image segmentation model and identify ethnic village buildings based on the segmentation image.

Images