Training sample construction method and recognition method for land cover recognition in remote sensing images

Abstract

This invention belongs to the field of geographic surveying and mapping technology, and provides a method for constructing training samples and a recognition method for land cover identification in remote sensing images. This method exhibits better generalization performance and higher recognition accuracy, making it suitable for practical engineering applications, especially power transmission line route selection. The sample construction method first labels the types and boundaries of land covers based on remote sensing images, obtaining mask images corresponding to the remote sensing images. Then, based on DSM and DEM data, it obtains nDSM data corresponding to each remote sensing image. Next, the nDSM data is appended as additional bands to the corresponding remote sensing images to obtain fused data. Finally, the fused data and mask images are cropped to obtain training samples for various land cover types. The recognition method employs the UCTransNet model and is trained using the training samples constructed by the training sample construction method.

Description

Technical Field

[0001] This invention belongs to the field of geographic surveying and mapping technology, and particularly relates to a method for constructing training samples and a method for recognizing ground features in remote sensing images. Background Technology

[0002] With the development of Earth observation technology, the means and types of remote sensing image acquisition have become increasingly diversified, and the timeliness and effectiveness of image data have also been greatly improved. In particular, the application of unmanned aerial survey technology in recent years has enabled the acquisition of different types of high-resolution remote sensing data, which has increased significantly. Therefore, high-resolution remote sensing images are now widely used in land use, urban planning, environmental monitoring and other fields.

[0003] The prerequisite and foundation for extracting information from high-resolution remote sensing images is remote sensing image classification. Based on different classification units, remote sensing image classification methods can be divided into pixel-based and object-oriented classification. Traditionally, object-oriented methods are mainly used, utilizing the spatial relationships between pixels. However, in the traditional process of power transmission line selection, the identification and extraction of ground features from remote sensing images relies primarily on manual labor. This is inefficient, costly, and time-consuming, and cannot maintain the timeliness and effectiveness of the data through regular updates, thus failing to enhance its application value. On the other hand, classifying ground features through visual interpretation is subjective; different operators may use different standards, potentially leading to different results and inaccuracies.

[0004] Since the emergence of semantic segmentation models, represented by fully convolutional neural networks (FCNs), these models have been able to perform pixel-based dense prediction tasks and accurately classify images by incorporating contextual information. Currently, this method is widely used in image classification tasks with good results and has seen significant development. Common existing semantic segmentation models include U-NET, SegNet, DeepLab, Dilated Convolutions, RefineNet, and PSPNet. Therefore, current research has also explored applying semantic segmentation models to remote sensing image feature recognition.

[0005] However, current semantic segmentation models are mainly designed for small-scale images, such as medical images and machine vision images. Therefore, when applied to ground object recognition in remote sensing images, they have the following shortcomings: On the one hand, although high-resolution remote sensing images can show clear details of ground targets and reduce inter-class differences of target ground objects, the spectral information is more complex, which also increases intra-class differences. On the other hand, high-resolution remote sensing images generally have fewer bands and relatively lack spectral information, which to some extent limits the richness of model feature learning and makes it difficult to handle misclassification phenomena caused by similar spectral information within ground objects.

[0006] In particular, most existing studies are experimental in nature, primarily utilizing publicly available data samples or small amounts of remote sensing data from a specific area. The data formats are standardized, the areas involved and the data volume are relatively small, and the morphological features to be identified are simple. However, in actual engineering projects, such as power transmission line route selection, the areas involved are vast, and the remote sensing images are typically aerial data taken by drones. This data volume is large, and due to the randomness of drone trajectories and differences in optical systems, the resolution, tone, and quality of the acquired remote sensing images vary considerably. The morphological features of the ground features to be identified differ significantly between different projects. Therefore, based on these differences, the ground feature identification models obtained from existing research often lack representativeness and a certain degree of generalization. Consequently, the accuracy of ground feature identification in practical engineering applications is low, making it difficult to meet the needs of actual engineering applications. Summary of the Invention

[0007] The technical problem to be solved by this invention is to propose a training sample construction method and a recognition method for remote sensing image ground object recognition, which has better generalization performance, higher recognition accuracy, and can meet the needs of practical engineering applications.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0009] The method for constructing training samples for remote sensing image ground feature recognition includes the following steps:

[0010] S1. Data Preparation

[0011] For each land feature in each remote sensing image, the type and boundary of each land feature are labeled based on a preset land feature classification. Then, for the image area selected by the boundary of each land feature, a preset gray value corresponding to each land feature classification is assigned to obtain a mask image corresponding to each remote sensing image.

[0012] Using interpolation, the DSM data and DEM data of the corresponding regions of each remote sensing image are interpolated to obtain DSM data and DEM data with the same pixel size as each remote sensing image. Then, based on the interpolated DSM data and DEM data, nDSM data corresponding to each remote sensing image are calculated.

[0013] S2, Data Overlay

[0014] The corresponding nDSM data and remote sensing images are registered separately. Then, the nDSM data is added as an additional band to the corresponding remote sensing image to obtain the fused data corresponding to each remote sensing image.

[0015] S3. Sample construction

[0016] Crop the fused data and the mask image according to the data requirements of the deep neural network adopted by the remote sensing image ground object recognition model to obtain training samples of various ground object types.

[0017] Specifically, in step S1, the remote sensing image is a remote sensing image obtained by unmanned aerial vehicle (UAV) aerial photography; and based on the point cloud data obtained by UAV aerial photography, the progressive irregular encrypted triangular network filtering algorithm is used to generate DSM data and DEM data corresponding to each remote sensing image.

[0018] Furthermore, in step S1, before annotating the types and boundaries of each ground object, first convert each obtained remote sensing image to the same resolution through the three - time convolution interpolation method.

[0019] Specifically, in step S1, the inverse distance weighting method is used to interpolate the DSM data and DEM data of the area corresponding to the remote sensing image.

[0020] Furthermore, in step S3, before cropping the fused data and the mask image, normalize the data of each channel of the fused data respectively.

[0021] Furthermore, in step S3, divide the sample data into a training set, a test set and a validation set according to a preset ratio, and perform data augmentation on the training set data by adopting random horizontal flipping, vertical flipping, diagonal flipping and linear ratio stretching, perform data augmentation on the validation set data by adopting random linear stretching, and the percentage of linear stretching of the training set data is less than the percentage of linear stretching of the validation set data.

[0022] Furthermore, the preset ground object classification includes small - target independent ground objects and other independent ground objects. Among them, the small - target independent ground objects are independent ground objects with a resolution less than 32×32, and the independent ground objects other than the small - target independent ground objects are other independent ground objects.

[0023] Furthermore, according to the ratio value M between the number of pixels in the image area selected by the boundary of each ground object and the total number of pixels of the remote sensing image corresponding to this ground object, classify each ground object. The smaller the ratio value M, the higher the level; in step S3, for the training samples corresponding to each ground object obtained by cropping, perform up - sampling processing different times according to the level of each ground object. The higher the level, the more times of up - sampling.

[0024] Specifically, classify the ground objects and perform up - sampling according to the following requirements:

[0025] Sixth level: M≤0.05%, the number of up - sampling times is 5 times;

[0026] Fifth level: 0.05% < M≤0.1%, the number of up - sampling times is 4 times;

[0027] Level 4: 0.1% < M ≤ 1%, with the upsampling times being 3 times;

[0028] Level 3: 1% < M ≤ 5%, with the upsampling times being 2 times;

[0029] Level 2: 5% < M ≤ 10%, with the upsampling times being 1 time;

[0030] Level 1: M > 10%, with the upsampling times being 0 times.

[0031] The present invention also provides a method for identifying ground objects in remote sensing images, including:

[0032] Using the training sample construction method for identifying ground objects in remote sensing images as described above, constructing training samples covering preset ground object classifications, and forming a training sample set;

[0033] Based on the UCTransNet model, constructing a remote sensing image ground object recognition model, and training it using the training sample set to obtain a completed remote sensing image ground object recognition model;

[0034] Based on the remote sensing image, DSM data, and DEM data of the area to be classified, constructing the fusion data of the area to be classified and inputting it into the completed remote sensing image ground object recognition model to complete the recognition of ground objects in the area to be classified.

[0035] Further, the preset ground object classifications include small target independent ground objects and other independent ground objects, where the small target independent ground objects are independent ground objects with a resolution less than 32×32, and the independent ground objects other than the small target independent ground objects are other independent ground objects;

[0036] The remote sensing image ground object recognition model constructed based on the UCTransNet model includes the UCTransNet model and a shallow feature extraction channel; the shallow feature extraction channel is a skip channel established between the first layer convolution and the last layer deconvolution of the UCTransNet model;

[0037] The shallow feature extraction channel takes the output E1 of the first layer convolution of the UCTransNet model as the input. First, feature extraction is performed on E1 through convolution layers of different scales, and the features of different scales obtained by extraction are spliced to obtain a spliced feature; then, through convolution operations, the spliced feature is converted into a shallow feature map A1, and the size and number of channels of the shallow feature map A1 are the same as those of the output D1 of the last layer deconvolution of the UCTransNet model; afterwards, A1 and D1 are fused to obtain a fused feature map and used as the output of the decoding module of the UCTransNet model.

[0038] The beneficial effects of the present invention are:

[0039] First, this invention proposes a standardized and streamlined sample creation method. For image recognition of actual power transmission projects, a deep learning sample database is established according to ground features, which facilitates sample accumulation and provides the possibility for the diversity of sample data. This results in better generalization performance and higher recognition accuracy of remote sensing image ground feature recognition, and can meet the needs of large-area and long-span remote sensing image ground feature recognition in actual power transmission line route selection projects.

[0040] Secondly, high-resolution image data lacks sufficient spectral information, making it difficult to effectively handle misclassification of features due to similar spectral information. This invention introduces nDSM data as an additional band superimposed onto the high-resolution image. The geometric information of nDSM can serve as supplementary information for spectral features, playing a significant role in identifying woodlands and farmlands, as well as some houses and roads, with similar spectral characteristics, thus effectively improving the accuracy of feature identification.

[0041] Third, the UCTransNet model is introduced. In the model encoding stage, the important features in each convolutional layer are automatically assigned greater weights, while the unimportant features are suppressed. Multi-scale features are selectively fused to construct more fully expressive features. In the decoding stage, the model can autonomously and effectively connect the multi-scale features from the encoding stage with the features from the decoding stage, providing better assurance for practical engineering applications, especially power transmission line route selection projects. Attached Figure Description

[0042] Figure 1 This is a flowchart of the remote sensing image ground feature identification method of the present invention;

[0043] Figure 2 Here is a diagram of the UCTransNet model structure;

[0044] Figure 3 This is a road extraction map based on the remote sensing image feature recognition method of the present invention;

[0045] Figure 4 This is a building extraction map based on the remote sensing image feature recognition method of the present invention. Detailed Implementation

[0046] This invention aims to propose a training sample construction method and a recognition method for remote sensing image ground feature recognition. It has better generalization performance and higher recognition accuracy, and can be adapted to practical engineering applications, especially the needs of power transmission line route selection projects, and can meet the requirements of remote sensing image ground feature recognition in scenarios such as large spans and large data differences.

[0047] The following description, in conjunction with the embodiments and accompanying drawings, provides further details.

[0048] Example 1:

[0049] The method for constructing training samples for remote sensing image ground feature recognition includes the following steps:

[0050] S1. Data Preparation

[0051] For each land feature in each remote sensing image, the type and boundary of each land feature are labeled based on a preset land feature classification. Then, for the image area selected by the boundary of each land feature, a preset gray value corresponding to each land feature classification is assigned to obtain a mask image corresponding to each remote sensing image.

[0052] For the DSM (Digital Surface Model) and DEM (Digital Elevation Model) data of the regions corresponding to each remote sensing image, interpolation is performed using the inverse distance weighting method to obtain DSM and DEM data with the same pixel size as each remote sensing image. Then, based on the interpolated DSM and DEM data, nDSM (Normalized Digital Surface Model) data corresponding to each remote sensing image are calculated. Besides the inverse distance weighting method, other existing interpolation methods can also be used to interpolate the DSM and DEM data of the regions corresponding to the remote sensing images.

[0053] In this embodiment, the remote sensing images are obtained from aerial photography by a drone. Based on the point cloud data obtained from the drone aerial photography, a progressive irregular encrypted triangular mesh filtering algorithm is used to generate DSM data and DEM data corresponding to each remote sensing image. Since aerial photography data is used, the resolution of each remote sensing image is different. Therefore, before labeling the types and boundaries of each landform, the acquired remote sensing images are first converted to the same resolution using a cubic convolution interpolation method.

[0054] S2, Data Overlay

[0055] Linear registration was used to register the corresponding nDSM data and remote sensing images separately. Then, the nDSM data was added as an additional band to the corresponding remote sensing images to obtain fused data corresponding to each remote sensing image.

[0056] S3, Sample Construction

[0057] According to the data requirements of the deep neural network used in the remote sensing image land cover recognition model, the fused data and masked images are cropped to obtain training samples for various land cover types. The pixel values ​​of remote sensing image data are typically in the range [0, 255]. When the data distribution range is large, the deep learning model's solution process is slow. In particular, the remote sensing images in this embodiment are aerial images, and the tone and quality of each image data are different. Therefore, to facilitate training, before cropping the fused data and masked images, the data of each channel of the fused data are normalized. Through normalization, the data range of each channel is mapped from the [0, 255] range to the [0, 1] range.

[0058] In remote sensing image feature recognition, conventional feature classification typically includes roads, water bodies, farmland, forest land, and buildings. However, in power transmission line route selection projects, individual features such as graves, wells, and towers, while having a very small pixel footprint, have a significant impact. Therefore, this invention introduces individual features and further subdivides them into small-target individual features and other individual features. The preset feature classifications include roads, water bodies, farmland, forest land, buildings, small-target individual features, and other individual features.

[0059] Independent features are those features that cannot be represented to scale on a map. Specifically, as map scales decrease, the amount of content represented also decreases. On small-scale topographic maps, independent features cannot be displayed in their true form, so they are usually represented by geometric symbols or side-view pictographic symbols, primarily indicating historical and cultural independent features. In this invention, small-target independent features are defined as independent features with a resolution less than 32×32. All other independent features are categorized as "other independent features."

[0060] Since this embodiment is applied to the identification of ground features in remote sensing images for power transmission line route selection projects, the size of ground features in remote sensing images varies greatly. For example, larger ones include contiguous buildings and farmland, while smaller ones include independent buildings and small, independent ground features such as wells, iron towers, and graves. Among these, ground features with small pixels, such as small, independent ground features, have problems such as few usable features, small proportion of data concentration, imbalanced samples, and susceptibility to interference from various adverse factors.

[0061] Therefore, first, the sample data is divided into a training set, a test set, and a validation set according to a preset ratio, and data augmentation is performed on the training set data by means of random horizontal flipping, vertical flipping, diagonal flipping, and linear ratio stretching. Random linear stretching is used for data augmentation of the validation set data, and the percentage of linear stretching of the training set data is less than that of the validation set data. Through sample augmentation, the problems of low proportion and sample imbalance in the dataset are solved. Specifically, in this embodiment, the training set, the test set, and the validation set are divided in the ratio of 6:3:1. The percentage of linear stretching of the training set data is 0.5%, and the percentages of linear stretching of the validation set data are 0.8%, 1%, and 2%.

[0062] Secondly, according to the ratio value M between the number of pixels in the image region selected by the boundaries of each ground object and the total number of pixels in the corresponding remote sensing image of the ground object, each ground object is classified. The smaller the ratio value M, the higher the level; in step S3, for the training samples corresponding to each ground object obtained by cropping, upsampling is performed different times according to the level of each ground object. The higher the level, the more times of upsampling. Data augmentation is performed by means of upsampling, so as to solve the problem of few available features for small targets. Specifically, in this embodiment, the ground objects are classified and upsampled according to the following requirements:

[0063] Sixth level: M ≤ 0.05%, the number of upsampling times is 5 times;

[0064] Fifth level: 0.05% < M ≤ 0.1%, the number of upsampling times is 4 times;

[0065] Fourth level: 0.1% < M ≤ 1%, the number of upsampling times is 3 times;

[0066] Third level: 1% < M ≤ 5%, the number of upsampling times is 2 times;

[0067] Second level: 5% < M ≤ 10%, the number of upsampling times is 1 time;

[0068] First level: M > 10%, the number of upsampling times is 0 times.

[0069] Embodiment 2:

[0070] The training samples constructed by the training sample construction method for ground object recognition of remote sensing images of the present invention are applicable to the training of various semantic segmentation models. However, in order to further improve the recognition rate, the present invention also provides an aspect of ground object recognition based on the UCTransNet model architecture. This will be described below in conjunction with Embodiment 2.

[0071] A method for ground object recognition of remote sensing images, as Figure 1 shown, includes:

[0072] Using the training sample construction method for remote sensing image land cover recognition in Example 1, training samples covering preset land cover classifications are constructed to form a training sample set; the preset land cover classifications include roads, water bodies, cultivated land, forest land, buildings, small target independent land cover and other independent land cover.

[0073] Then, a remote sensing image ground feature recognition model is constructed based on the UCTransNet model, and trained using the training sample set to obtain a trained remote sensing image ground feature recognition model.

[0074] Finally, based on the remote sensing images, DSM data, and DEM data of the area to be classified, the fused data of the area to be classified is constructed and input into the trained remote sensing image land cover recognition model to complete the recognition of land cover in the area to be classified.

[0075] The aforementioned UCTransNet model, such as Figure 2 As shown, this is an improvement on the classic U-Net network. The skip connections of U-Net are replaced by the CTrans (Channel Converter) module. CTrans consists of a submodule that performs multi-scale channel cross-fusion (CCT) and a submodule that guides the fusion of multi-scale channel cross-information (CCA) to effectively connect to the decoder features, thereby eliminating ambiguity.

[0076] Because this invention utilizes small, independent ground features, which have few pixels and limited usable features, they are highly susceptible to environmental influences. After processing with the UCTransNet model, the feature information of these small, independent ground features is easily lost, resulting in incomplete representation and potentially affecting the final recognition accuracy. In contrast, shallow, high-resolution features are more suitable for representing the features of small, independent ground features. Therefore, to ensure the recognition of small, independent ground features, a remote sensing image ground feature recognition model based on the UCTransNet model is further constructed, including the UCTransNet model and a shallow feature extraction channel; the shallow feature extraction channel is a skip channel established between the first convolutional layer and the last deconvolutional layer of the UCTransNet model.

[0077] The shallow feature extraction channel takes the output E1 of the first convolutional layer of the UCTransNet model as input. First, features are extracted from E1 through 1×1, 3×3, and 5×5 convolutional layers, and the extracted features at different scales are concatenated to obtain concatenated features. Then, through a convolution operation, the concatenated features are converted into a shallow feature map A1, and the size and number of channels of the shallow feature map A1 are the same as the size and number of channels of the output D1 of the last deconvolutional layer of the UCTransNet model. Afterward, A1 and D1 are fused to obtain a fused feature map, which is used as the output of the UCTransNet model decoding module. The fusion of A1 and D1 can adopt any existing fusion method, including but not limited to concatenation, addition, etc. In this embodiment, concatenation is used to achieve the fusion of A1 and D1.

[0078] The UCTransNet model, based on the output of its decoding module, maps the fused feature map through a 1×1 convolution, and then activates it through the Sigmoid function to obtain the final prediction result of the land cover type.

[0079] To test the performance of the recognition method of this invention, the recognition method of this embodiment was implemented using the PyTorch deep learning framework. The joint function of soft cross-entropy and Dice Loss was selected as the loss function to measure the loss value between the predicted classification result and the actual land cover classification. The existing UCTransNet model training strategy was used for training.

[0080] A trained model was used to extract ground features within the selected area of ​​the power transmission line. The classified images were then stitched together using a dilation prediction method to create the final image. The dilation prediction steps included: filling the border with a factor of one to ensure the central region was preserved during dilation prediction; using a sliding window size of 256*256; retaining 128*128 pixels at the center each time; and setting the step size to 128. This dilation prediction resulted in seamless stitching of the aerial images, improving the segmentation effect of large-sized images. The results are as follows: Figure 3 , Figure 4 As shown, where, Figure 3 This is a road extraction map based on the remote sensing image feature recognition method in this embodiment; Figure 4 This is a building extraction map based on the remote sensing image feature recognition method of this embodiment.

[0081] It should be noted that the above embodiments are merely preferred embodiments and are not intended to limit the present invention. It should be pointed out that those skilled in the art can make various modifications, equivalent substitutions, and improvements without departing from the spirit and scope of the claims, and all such modifications, substitutions, and improvements should be included within the scope of protection of the present invention.

Claims

1. A method for constructing training samples for remote sensing image feature recognition, characterized in that, Includes the following steps: S1. Data Preparation For each land feature in each remote sensing image, the type and boundary of each land feature are labeled based on a preset land feature classification. Then, for the image area selected by the boundary of each land feature, a preset gray value corresponding to each land feature classification is assigned to obtain a mask image corresponding to each remote sensing image. Using interpolation, the DSM data and DEM data of the corresponding regions of each remote sensing image are interpolated to obtain DSM data and DEM data with the same pixel size as each remote sensing image. Then, based on the interpolated DSM data and DEM data, nDSM data corresponding to each remote sensing image are calculated. S2, Data Overlay The corresponding nDSM data and remote sensing images are registered separately. Then, the nDSM data is added as an additional band to the corresponding remote sensing image to obtain the fused data corresponding to each remote sensing image. S3, Sample Construction According to the data requirements of the deep neural network used in the remote sensing image ground feature recognition model, the fused data and masked images are cropped to obtain training samples for various ground feature types. The preset land feature classification includes small target independent land features and other independent land features. Among them, small target independent land features are independent land features with a resolution of less than 32×32. Other independent land features are those other than small target independent land features. The remote sensing image land cover recognition model based on the UCTransNet model includes the UCTransNet model and a shallow feature extraction channel; the shallow feature extraction channel is a skip channel established between the first convolutional layer and the last deconvolutional layer of the UCTransNet model. The shallow feature extraction channel takes the output E1 of the first convolutional layer of the UCTransNet model as input. First, features are extracted from E1 through convolutional layers of different scales, and the extracted features of different scales are concatenated to obtain concatenated features. Then, through convolution operations, the concatenated features are converted into a shallow feature map A1, and the size and number of channels of the shallow feature map A1 are the same as the size and number of channels of the output D1 of the last deconvolutional layer of the UCTransNet model. After that, A1 and D1 are fused to obtain a fused feature map, which is used as the output of the decoding module of the UCTransNet model.

2. The method for constructing training samples for remote sensing image feature recognition as described in claim 1, characterized in that, In step S1, the remote sensing images are obtained by drone aerial photography; and based on the point cloud data obtained by drone aerial photography, a progressive irregular encrypted triangular mesh filtering algorithm is used to generate DSM data and DEM data corresponding to each remote sensing image.

3. The method for constructing training samples for remote sensing image feature recognition as described in claim 2, characterized in that, In step S1, before labeling the types and boundaries of features in each location, the acquired remote sensing images are first converted to the same resolution using a cubic convolution interpolation method.

4. The method for constructing training samples for remote sensing image feature recognition as described in claim 1, characterized in that, In step S1, the inverse distance weighting method is used to interpolate the DSM and DEM data of the region corresponding to the remote sensing image.

5. The method for constructing training samples for remote sensing image feature recognition as described in claim 1, characterized in that, In step S3, before cropping the fused data and the mask image, the data of each channel of the fused data is normalized.

6. The method for constructing training samples for remote sensing image feature recognition as described in any one of claims 1 to 5, characterized in that, In step S3, the sample data is divided into a training set, a test set, and a validation set according to a preset ratio, and the training set data is augmented by randomly horizontal flipping, vertical flipping, diagonal flipping, and linear ratio stretching, the validation set data is augmented by random linear stretching, and the percentage of linear stretching of the training set data is less than that of the validation set data.

7. The method for constructing training samples for remote sensing image feature recognition as described in claim 1, characterized in that, The preset ground object classification includes small target independent ground objects and other independent ground objects. Among them, small target independent ground objects are independent ground objects with a resolution less than 32×32, and other independent ground objects are those other than small target independent ground objects among the independent ground objects.

8. The method for constructing training samples for remote sensing image feature recognition as described in any one of claims 1, 2, 3, 4, 5, or 7, characterized in that, According to the ratio value M between the number of pixels in the image area selected by the boundary of each ground object and the total number of pixels in the corresponding remote sensing image of the ground object, each ground object is classified, and the smaller the ratio value M, the higher the level; in the step S3, for the training samples corresponding to each ground object obtained by cropping, upsampling is performed different times according to the level of each ground object, and the higher the level, the more times of upsampling.

9. The method for constructing training samples for remote sensing image feature recognition as described in claim 8, characterized in that, The ground objects are classified and upsampled according to the following requirements: Level six: M≤0.05%, and the number of upsampling times is 5 times; Level five: 0.05% < M≤0.1%, and the number of upsampling times is 4 times; Level four: 0.1% < M≤1%, and the number of upsampling times is 3 times; Level three: 1% < M≤5%, and the number of upsampling times is 2 times; Level two: 5% < M≤10%, and the number of upsampling times is 1 time; Level one: M > 10%, and the number of upsampling times is 0 times.

10. A method for identifying ground features in remote sensing images, characterized in that, It includes: Adopt the training sample construction method for remote sensing image ground object recognition described in any one of claims 1 to 9 to construct training samples covering the preset ground object classifications and form a training sample set; Construct a remote sensing image ground object recognition model based on the UCTransNet model, and use the training sample set to train it to obtain a trained remote sensing image ground object recognition model; Based on the remote sensing image, DSM data, and DEM data of the area to be classified, construct the fusion data of the area to be classified and input it into the trained remote sensing image ground object recognition model to complete the recognition of the ground objects in the area to be classified.

Images