Building height estimation method and system based on single satellite image

Abstract

The invention discloses a building height estimation method and system based on a single satellite image, and belongs to the technical field of remote sensing image processing. The problem that fine-grained spatial information in shallow-layer features and advanced semantic related information in deep-layer features cannot be effectively utilized when an existing neural network is used for predicting the height is solved. Multilayer height features of a single satellite image are extracted by using an encoder, then processing is performed by using a decoder based on an AMHF multi-scale feature fusion module, the AMHF multi-scale feature fusion module integrates a DAFH attention mechanism of a double-branch structure, a branch 1 uses depth separable convolution to extract fine-grained spatial features, a branch 2 uses depth separable convolution to extract fine-grained spatial features, and a branch 3 uses depth separable convolution to extract fine-grained spatial features; and the branch 2 captures global spatial features by using spatial attention and facilitates extraction of macroscopic advanced semantic related information at the same time, a normalized height map is obtained after the two branches are fused, and the normalized height map is multiplied by maxheight to obtain a final prediction height map.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and in particular relates to a method and system for estimating the height of a building. Background Art

[0002] With the advancement of technology and the growing needs of the public, the possibility of building realistic 3D city models has gradually become a necessity. Realistic 3D city models are crucial for urban planning and navigation. They can further assist drones and small manned aircraft in urban navigation, contributing to the development of the low-altitude economy. A high-precision but costly method is used to fine-tune the reconstruction of a small number of key areas, while a low-cost method with a reasonable level of accuracy (LOD 1.3) is used to reconstruct the majority of areas.

[0003] Existing methods for generating 3D city models primarily use 3D sensors and reconstruction methods based on drone aerial photography. 3D sensors such as LiDAR and SAR can provide high-quality 3D measurement information, but they face challenges such as high data acquisition costs and difficulties. Reconstruction methods based on drone aerial photography primarily include multi-view image reconstruction and single-image height estimation. Multi-view drone image reconstruction methods require high data volumes, leading to the same challenges as 3D sensors—high economic and time costs. Single-image height estimation based on drone remote sensing imagery does reduce costs, but it cannot reduce the time cost of data collection, nor can it address the variability in drone poses caused by the vast coverage of urban-scale scenes, a critical issue for single-image height estimation. In summary, these methods cannot simultaneously achieve low cost and time costs while still meeting LOD 1.3 reconstruction requirements. Therefore, height estimation from single satellite images and reconstruction of 3D models based on these images has become the final piece in the puzzle of realistic 3D city construction technology.

[0004] The process of extracting height information from a single remote sensing image is called monocular height estimation, and the estimated result is a digital surface model (DSM) or a normalized digital surface model (nDSM). Estimating depth from a single image is an ill-posed problem because the only explicit depth features in a single image are variations and contrasts in size and texture, or occlusion and perspective relationships between different objects. Before the advancement of deep learning methods, traditional depth estimation methods mostly used random field methods to extract context and geometric features from the image to infer depth values. With the continuous development of neural networks and deep learning, monocular height estimation methods using convolutional neural networks (CNNs) and generative adversarial networks (GNNs) have emerged. IM2HEIGHT is the first to use a convolutional network based on residual learning to predict the DSM. To address the mismatch between DSM data and satellite imagery caused by real-world errors, the IM2ELEVATION network proposes a mutual information-based registration method and uses a Hough transform-based verification step to adjust mismatched image patches. To address the difficulty of extracting ground feature information from remote sensing imagery, IMG2HEIGHT proposes a spatially enhanced, multi-scale aggregated encoder-decoder height estimation network (SM-EDNet). Generative adversarial networks (GANs) have great potential in simulating data structures, which helps address image-to-image translation. IMG2DSM pioneered this approach for height estimation, including a generator and discriminator trained using DSMs and optical images. To reduce the time and computational costs of height estimation, U-IMG2DSM combines variational autoencoders (VAEs) with GANs for image-to-image translation, thereby obtaining digital surface model (DSM) data from a single remote sensing image. However, most of these methods neglect the ability to distinguish height features between adjacent and intersecting ground features and different building height layers, making them difficult to perform effectively when faced with the large number of objects and complex height distributions in satellite imagery. Summary of the Invention

[0005] The present invention aims to solve the problem that the existing neural network cannot effectively utilize the fine-grained spatial information in shallow features and the high-level semantic related information in deep features when predicting height.

[0006] A building height estimation method based on a single satellite image is proposed. The collected single satellite image is fed into an AMFHENet height estimation network for estimation to obtain a height estimation result. The processing process of the AMFHENet height estimation network includes the following steps: First, an encoder is used to extract N layers of height features from a single satellite image. The encoder depths are recorded as Feature 1, ..., Feature N from deep to shallow respectively. Then, a decoder is used for processing. The decoder uses N AMHF multi-scale feature fusion modules for processing, which are recorded as AMHF multi-scale feature fusion module 1, ..., AMHF multi-scale feature fusion module N respectively. The processing process of the decoder includes: The AMHF multi-scale feature fusion module has 2 inputs and 1 output. The 2 inputs correspond to input port 1 and input port 2 respectively. The input of input port 1 is the output of the previous module. The input port 2 of the AMHF multi-scale feature fusion module 1 has no input. The input of the input port 2 of other AMHF multi-scale feature fusion modules 1 is the feature map corresponding to the encoder; input port 1 is directly connected to the residual convolution unit. The output of the residual convolution unit of input port 1 and the input of input port 2 are fused and then pass through a residual convolution unit, and then processed by the DAFH attention mechanism, and then up-sampled; The DAFH attention processing process is as follows: Among them, X is the input feature, Indicates the average grouping operation by channel, Represents the kth feature after grouping; for The results after spatial attention modulation, represents a 3×3 convolutional unit, Indicates the operation of Hadamard product. Represents the Sigmoid activation function; For The extracted spatial features, represents depth-wise separable convolution; is the feature after adaptive fusion, is the CA channel attention mechanism, represents the learnable fusion parameters, Represents the final output, Indicates that the grouped features are spliced ​​by channel; The input of the AMHF multi-scale feature fusion module N is processed by a 3×3 convolution layer, and then restored to the original image size through bilinear interpolation. Finally, the normalized height map is obtained through the Sigmoid activation function. Multiply the normalized height map by max_height to get the final predicted height map, where max_height represents the set maximum height.

[0007] Furthermore, the CA channel attention mechanism is as follows: in, Represents the input features of CA.

[0008] Furthermore, after the DAFH attention mechanism is processed, the upsampling process includes: Features processed based on DAFH attention mechanism Perform upsampling: in, Express The result after upsampling; represents transposed convolution, Indicates bilinear interpolation, and either of the two sampling methods is selected for upsampling; Then make channel adjustments: in, For the final output, It is a 1×1 convolution unit.

[0009] Preferably, N is 4, that is, the encoder extracts 4 layers of height features of a single satellite image, and the decoder uses 4 AMHF multi-scale feature fusion modules for processing.

[0010] Preferably, the encoder adopts DINO v2 encoder.

[0011] Furthermore, the height estimation network AMFHENet is pre-trained, and the loss function of the height estimation network AMFHENet during training is as follows: in represents the height value predicted at point (i, j), Represents the true height value at point (i, j); Indicates the image height, Indicates the image width, is the balance parameter.

[0012] A building height estimation system based on a single satellite image, comprising: An input unit, used for inputting a single satellite image for which height estimation is to be performed; The height estimation unit uses the height estimation network AMFHENet to estimate the height of the input single satellite image and obtain the height estimation result. The processing process of the height estimation network AMFHENet includes the following steps: First, an encoder is used to extract N layers of height features from a single satellite image. The encoder depths are recorded as Feature 1, ..., Feature N from deep to shallow respectively. Then, a decoder is used for processing. The decoder uses N AMHF multi-scale feature fusion modules for processing, which are recorded as AMHF multi-scale feature fusion module 1, ..., AMHF multi-scale feature fusion module N respectively. The processing process of the decoder includes: The AMHF multi-scale feature fusion module has 2 inputs and 1 output. The 2 inputs correspond to input port 1 and input port 2 respectively. The input of input port 1 is the output of the previous module. The input port 2 of the AMHF multi-scale feature fusion module 1 has no input. The input of the input port 2 of other AMHF multi-scale feature fusion modules 1 is the feature map corresponding to the encoder; input port 1 is directly connected to the residual convolution unit. The output of the residual convolution unit of input port 1 and the input of input port 2 are fused and then pass through a residual convolution unit, and then processed by the DAFH attention mechanism, and then up-sampled; The DAFH attention processing process is as follows: Among them, X is the input feature, Indicates the average grouping operation by channel, Represents the kth feature after grouping; for The results after spatial attention modulation, represents a 3×3 convolutional unit, Indicates the operation of Hadamard product. Represents the Sigmoid activation function; For The extracted spatial features, represents depth-wise separable convolution; is the feature after adaptive fusion, is the CA channel attention mechanism, represents the learnable fusion parameters, Represents the final output, Indicates that the grouped features are spliced ​​by channel; The input of the AMHF multi-scale feature fusion module N is processed by a 3×3 convolution layer, and then restored to the original image size through bilinear interpolation. Finally, the normalized height map is obtained through the Sigmoid activation function. Multiply the normalized height map by max_height to get the final predicted height map, where max_height represents the set maximum height.

[0013] Furthermore, the CA channel attention mechanism is as follows: in, Represents the input features of CA.

[0014] Furthermore, after the DAFH attention mechanism is processed, the upsampling process includes: Features processed based on DAFH attention mechanism Perform upsampling: in, Express The result after upsampling; represents transposed convolution, Indicates bilinear interpolation, and either of the two sampling methods is selected for upsampling; Then make channel adjustments: in, For the final output, It is a 1×1 convolution unit.

[0015] Preferably, N is 4, that is, the encoder extracts 4 layers of height features of a single satellite image, and the decoder uses 4 AMHF multi-scale feature fusion modules for processing.

[0016] The beneficial effects of the present invention are: The DAFH attention proposed in the present invention and the AMHF multi-scale feature fusion module based on this attention can help the network distinguish between different building features and different land features in the feature fusion step, and enhance the network's attention to building height features. Specifically, the AMHF multi-scale feature fusion module fuses features of each layer at different scales, focusing on fusing fine-grained spatial information in shallow features and high-level semantic related information in deep features. This mainly relies on the dual-branch structure of DAFH. Branch 1 uses depth-separable convolution to extract fine-grained spatial features, and branch 2 uses spatial attention to capture global spatial features, while helping to extract macroscopic high-level semantic related information. Furthermore, branch 1 utilizes the characteristics of depth-separable convolution to avoid cross-channel interference, which is more suitable for the extraction of fine-grained spatial features. Branch 2 is not a simple series connection of the proposed spatial attention and channel attention, but first uses spatial attention to modulate the features, allowing high-level semantic information to be concentrated in large quantities, and then uses the proposed channel attention to capture channel weights. The two branches adopt a fusion strategy to flexibly adjust the fusion information focused on by AMFH. Therefore, the fusion ratio of branch 2 can be increased when fusing deep features to retain more high-level semantic information, and the fusion ratio of branch 1 can be increased when fusing shallow features to retain more fine-grained spatial information. The feature fusion strategy proposed in the present invention is more flexible and can adapt to the feature fusion operation at each stage, which is better than the single fusion strategy in the existing height estimation method. Therefore, the height estimation network AMFHENet proposed in the present invention can effectively mine deep height information in satellite images. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flowchart of a building height estimation method based on a single satellite image; Figure 2 This is a structural diagram of AMFHENet, a building height estimation network based on a single satellite image; Figure 3 This is the structural diagram of the AMHF multi-scale feature fusion module; Figure 4 This is the structural diagram of DAFH attention; Figure 5 Schematic diagram of the results of a building height estimation method based on a single satellite image. DETAILED DESCRIPTION

[0018] The description of the embodiments of this specification should be combined with the corresponding drawings, which should be considered as part of the complete specification. In the drawings, the shape or thickness of the embodiments may be enlarged and indicated for simplification or convenience. Furthermore, the parts of the structures in the drawings will be described separately. It is worth noting that the elements not shown in the drawings or not described in words are in a form known to those of ordinary skill in the art. The description of the embodiments herein, any reference to direction and orientation, are for ease of description only and should not be construed as any limitation on the scope of protection of the invention. The following description of the preferred embodiments will involve a combination of features, which may exist independently or in combination, and the present invention is not particularly limited to the preferred embodiments. The scope of the invention is defined by the claims.

[0019] Specific implementation method 1: refer to Figure 1 Specifically describe this embodiment.

[0020] The method for estimating building height based on a single satellite image described in this embodiment includes: Step 1: Set the initial parameters, including the image processing size image_size and the maximum height max_height, and build a satellite image monocular height estimation dataset; Step 2: Use the encoder and decoder to build a single image height estimation network AMFHENet, as shown in Figure 2 As shown in Figure 2, the processing of AMFHENet includes: S21. The encoder in the height estimation network AMFHENet uses the DINO v2 encoder as the height encoder to extract the four-layer height features of a single satellite image, and then converts the number of channels of the four-layer height features to 256. The height feature extraction process is expressed as: in, Represents the height coordinate of the image, Represents the DINO v2 encoder, A tensor representing the transformation of the input image.

[0021] The four-layer height feature sizes are 1 / 28, 1 / 14, 1 / 7, and 2 / 7 of image_size from small to large (corresponding to the encoder depth from deep to shallow), and are named Feature 1, Feature 2, Feature 3, and Feature 4 respectively; S22: The decoder in the height estimation network AMFHENet uses the AMHF multi-scale feature fusion module for encoding. In this embodiment, the number of AMHF multi-scale feature fusion modules is set to 4, respectively denoted as AMHF multi-scale feature fusion module 1 to AMHF multi-scale feature fusion module 4. The processing process of the decoder includes: S221: The AMHF multi-scale feature fusion module has two inputs and one output. The two inputs correspond to input port 1 and input port 2 respectively. Input port 1 must have input, and input port 2 can have no input. When input port 2 has no input, the AMHF multi-scale feature fusion module plays the role of feature enhancement; when input port 2 has input, the AMHF multi-scale feature fusion module plays the role of feature fusion. Input port 1 is directly connected to the residual convolution unit. The output of the residual convolution unit of input port 1 and the input of input port 2 are fused and then passed through a residual convolution unit, and then processed by the DAFH attention mechanism, and then up-sampled. Figure 3 As shown in FIG, the process of the AMHF multi-scale feature fusion module is expressed as follows: in, is the input of input port 1, For the input of entrance 2, is the preliminary fusion result, RCU is the residual convolution unit, and 、 Indicates two RCU processes; is a 3×3 convolution unit, is the activation function.

[0022] in, The result of feature refinement.

[0023] DAFH is DAFH attention, which is the enhanced feature.

[0024] DAFH attention Figure 4 As shown, DAFH attention is expressed as: Among them, X is the input feature, Indicates the average grouping operation by channel, represents the kth feature after grouping. for The results after spatial attention modulation, represents a 3×3 convolutional unit, Indicates the operation of Hadamard product. Represents the Sigmoid activation function. For The extracted spatial features, represents depth-wise separable convolution; is the feature after adaptive fusion, It is a channel attention improved based on SE channel attention. represents the learnable fusion parameters, . Represents the final output, Indicates channel-wise concatenation of grouped features.

[0025] described Attention is expressed as: in, is the input feature of CA here, and A represents the intermediate variable after dimensionality reduction.

[0026] based on Perform upsampling: in, Express The result after upsampling; the sampling method can be selected, represents transposed convolution, Represents bilinear interpolation.

[0027] in, For The output of channel adjustment is the final output of the AMHF multi-scale feature fusion module; It is a 1×1 convolution unit.

[0028] For AMHF multi-scale feature fusion module 1, feature 1 is input into input port 1, and input port 2 has no input; the output of AMHF fusion module 1 is fusion feature 1; the output size of the AMHF multi-scale feature fusion module can be freely set. Here, for the feature fusion operation of the next layer, the size is set to the size of feature 2; S222: Input feature 2 and fused feature 1 into input port 1 and input port 2 of AMHF multi-scale feature fusion module 2 respectively, and the output size is 1 / 7 of image_size, and the output is fused feature 2; S223: Input feature 3 and fused feature 2 into input port 1 and input port 2 of AMHF multi-scale feature fusion module 3 respectively, and the output size is 2 / 7 of image_size, and the output is fused feature 3; S224: Input feature 4 and fused feature 3 into input port 1 and input port 2 of AMHF multi-scale feature fusion module 4 respectively, and the output size is 2 / 7 of image_size, and the output is fused feature 4; S23: Use 3×3 convolution to convert the number of channels of fusion feature 4 to 128, and then restore it to the original image size through bilinear interpolation. Then use 3×3 convolution and 1×1 convolution to convert the number of channels from 128 to 32, and then from 32 to 1. Finally, use the Sigmoid activation function to obtain the normalized height map; S24: Multiply the normalized height map by max_height to obtain the final predicted height map.

[0029] Step 3: Use the height of each pixel in the final predicted height map And the height of each pixel in the real height map Structural height prediction loss ; The estimated loss of height Expressed as: in represents the height value predicted at point (i, j), Indicates the true height value at point (i, j). Indicates the image height, Indicates the image width, is a hyperparameter to balance the two terms.

[0030] Utilizing high prediction loss The height estimation network AMFHENet is trained, and the above steps are repeated for multiple trainings. After the training is completed, the trained AMFHENet is obtained.

[0031] Step 4: Use the trained AMFHENet to estimate the height of the image to be predicted and generate a predicted height map.

[0032] This implementation uses only a single satellite image as input, utilizing a feature extraction network to fully extract the height features of buildings in the image. A decoding network driven by multi-scale feature fusion, built around a four-layer AMHF core, ultimately achieves an accurate height prediction map. This method can accurately and rapidly predict building height information in satellite images, significantly reducing manpower, time, and economic costs, and facilitating the construction of realistic three-dimensional cities. To address the difficulty in distinguishing height features between adjacent and intersecting features and between different building height layers in monocular height estimation tasks using satellite remote sensing images, the present invention designs an AMHF multi-scale feature fusion module and DAFH attention. Specifically, the AMHF multi-scale feature fusion module fuses features from various layers at different scales, focusing on integrating fine-grained spatial information from shallow features with high-level semantic information from deep features. This is primarily achieved through the dual-branch structure of DAFH: branch 1 uses depthwise separable convolution to extract fine-grained spatial features, while branch 2 uses spatial attention to capture global spatial features, while also facilitating the extraction of macroscopic, high-level semantic information. Furthermore, branch 1 leverages the properties of depthwise separable convolution to avoid cross-channel interference, making it more suitable for extracting fine-grained spatial features. Branch 2 is not a simple concatenation of the proposed spatial attention and channel attention, but first uses spatial attention to modulate the features, so that high-level semantic information is concentrated in large quantities, and then uses the proposed channel attention to capture the channel weights. The two branches adopt an adaptive fusion strategy to flexibly adjust the fusion information focused on by AMFH. Therefore, more high-level semantic related information can be retained when fusing deep features, and more fine-grained spatial information can be retained when fusing shallow features. Finally, residual connections are used to solve the problem of possible feature loss. Existing single-image height estimation methods face the problem of being unable to take into account the fusion of high-level semantic related information and fine-grained spatial information due to a single feature fusion strategy. Starting from the attention mechanism, the present invention designs two attention branches with completely opposite focuses, and lets them compete with each other during training, adaptively retaining a reasonable weight ratio, so as to achieve flexible feature fusion. Figure 5 The height map is colored in the image, and the color changes are used to intuitively represent the changes in height. It can be clearly seen that AMFHENet accurately extracts the building boundaries and can predict the heights between different building layers.

[0033] In summary, this method can accurately and quickly predict building heights, greatly reducing manpower, time and economic costs, and is conducive to the construction of real-life three-dimensional cities, as well as remote sensing applications in urban planning, post-disaster reconstruction, military reconnaissance, and other aspects. Specific implementation method 2: This embodiment is a building height estimation system based on a single satellite image. It is actually a computer program system corresponding to a building height estimation method based on a single satellite image. The system specifically includes: An input unit, used for inputting a single satellite image for which height estimation is to be performed; The height estimation unit uses the height estimation network AMFHENet to estimate the height of the input single satellite image and obtain the height estimation result. The processing process of the height estimation network AMFHENet includes the following steps: First, an encoder is used to extract four layers of height features from a single satellite image. The encoder depths are recorded as features 1, ..., and 4 from deep to shallow respectively. Then, a decoder is used for processing. The decoder uses four AMHF multi-scale feature fusion modules for processing, which are recorded as AMHF multi-scale feature fusion module 1, ..., and AMHF multi-scale feature fusion module 4 respectively. The processing process of the decoder includes: The AMHF multi-scale feature fusion module has 2 inputs and 1 output. The 2 inputs correspond to input port 1 and input port 2 respectively. The input of input port 1 is the output of the previous module. The input port 2 of the AMHF multi-scale feature fusion module 1 has no input. The input of the input port 2 of other AMHF multi-scale feature fusion modules 1 is the feature map corresponding to the encoder; input port 1 is directly connected to the residual convolution unit. The output of the residual convolution unit of input port 1 and the input of input port 2 are fused and then pass through a residual convolution unit, and then processed by the DAFH attention mechanism, and then up-sampled; The DAFH attention processing process is as follows: Among them, X is the input feature, Indicates the average grouping operation by channel, Represents the kth feature after grouping; for The results after spatial attention modulation, represents a 3×3 convolutional unit, Indicates the operation of Hadamard product. Represents the Sigmoid activation function; For The extracted spatial features, represents depth-wise separable convolution; is the feature after adaptive fusion, is the CA channel attention mechanism, represents the learnable fusion parameters, Represents the final output, Indicates that the grouped features are spliced ​​by channel; The CA channel attention mechanism is as follows: in, Represents the input features of CA.

[0035] After the DAFH attention mechanism is processed, the upsampling process is as follows: Features processed based on DAFH attention mechanism Perform upsampling: in, Express The result after upsampling; represents transposed convolution, Indicates bilinear interpolation, and either of the two sampling methods is selected for upsampling; Then make channel adjustments: in, For the final output, It is a 1×1 convolution unit.

[0036] The input of the AMHF multi-scale feature fusion module 4 is then processed by a 3×3 convolution layer, and then restored to the original image size through bilinear interpolation. Finally, the normalized height map is obtained through the Sigmoid activation function; Multiply the normalized height map by max_height to get the final predicted height map, where max_height represents the set maximum height.

[0037] The output unit is used to output the segmented image.

[0038] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A building height estimation method based on a single satellite image, characterized in that: The collected single satellite image is sent to the height estimation network AMFHENet for estimation to obtain the height estimation result. The processing process of the height estimation network AMFHENet includes the following steps: First, an encoder is used to extract N layers of height features from a single satellite image. The encoder depths are recorded as Feature 1, ..., Feature N from deep to shallow respectively. Then, a decoder is used for processing. The decoder uses N AMHF multi-scale feature fusion modules for processing, which are recorded as AMHF multi-scale feature fusion module 1, ..., AMHF multi-scale feature fusion module N respectively. The processing process of the decoder includes: The AMHF multi-scale feature fusion module has 2 inputs and 1 output. The 2 inputs correspond to input port 1 and input port 2 respectively. The input of input port 1 is the output of the previous module. The input port 2 of the AMHF multi-scale feature fusion module 1 has no input. The input of the input port 2 of other AMHF multi-scale feature fusion modules 1 is the feature map corresponding to the encoder; input port 1 is directly connected to the residual convolution unit. The output of the residual convolution unit of input port 1 and the input of input port 2 are fused and then pass through a residual convolution unit, and then processed by the DAFH attention mechanism, and then up-sampled; The DAFH attention processing process is as follows: Among them, X is the input feature, Indicates the average grouping operation by channel, Represents the kth feature after grouping; for The results after spatial attention modulation, represents a 3×3 convolutional unit, Indicates the operation of Hadamard product. Represents the Sigmoid activation function; For The extracted spatial features, represents depth-wise separable convolution; is the feature after adaptive fusion, is the CA channel attention mechanism, represents the learnable fusion parameters, Represents the final output, Indicates that the grouped features are spliced ​​by channel; The input of the AMHF multi-scale feature fusion module N is processed by a 3×3 convolution layer, and then restored to the original image size through bilinear interpolation. Finally, the normalized height map is obtained through the Sigmoid activation function. Multiply the normalized height map by max_height to get the final predicted height map, where max_height represents the set maximum height.

2. The method for estimating building height based on a single satellite image according to claim 1, wherein: The CA channel attention mechanism is as follows: in, Represents the input features of CA.

3. The method for estimating building height based on a single satellite image according to claim 2, wherein: After the DAFH attention mechanism is processed, the upsampling process includes: Features processed based on DAFH attention mechanism Perform upsampling: in, Express The result after upsampling; represents transposed convolution, Indicates bilinear interpolation, and either of the two sampling methods is selected for upsampling; Then make channel adjustments: in, For the final output, It is a 1×1 convolution unit.

4. The method for estimating building height based on a single satellite image according to claim 1, wherein: The N is 4, that is, the encoder extracts 4 layers of height features of a single satellite image, and the decoder uses 4 AMHF multi-scale feature fusion modules for processing.

5. The method for estimating building height based on a single satellite image according to claim 4, wherein: The encoder adopts DINO v2 encoder.

6. A building height estimation method based on a single satellite image according to any one of claims 1 to 5, characterized in that: The height estimation network AMFHENet is pre-trained. The loss function of the height estimation network AMFHENet during training is as follows: in represents the height value predicted at point (i, j), Represents the true height value at point (i, j); Indicates the image height, Indicates the image width, is the balance parameter.

7. A building height estimation system based on a single satellite image, characterized in that: include: An input unit, used for inputting a single satellite image for which height estimation is to be performed; The height estimation unit uses the height estimation network AMFHENet to estimate the height of the input single satellite image and obtain the height estimation result. The processing process of the height estimation network AMFHENet includes the following steps: First, an encoder is used to extract N layers of height features from a single satellite image. The encoder depths are recorded as Feature 1, ..., Feature N from deep to shallow respectively. Then, a decoder is used for processing. The decoder uses N AMHF multi-scale feature fusion modules for processing, which are recorded as AMHF multi-scale feature fusion module 1, ..., AMHF multi-scale feature fusion module N respectively. The processing process of the decoder includes: The AMHF multi-scale feature fusion module has 2 inputs and 1 output. The 2 inputs correspond to input port 1 and input port 2 respectively. The input of input port 1 is the output of the previous module. The input port 2 of the AMHF multi-scale feature fusion module 1 has no input. The input of the input port 2 of other AMHF multi-scale feature fusion modules 1 is the feature map corresponding to the encoder; input port 1 is directly connected to the residual convolution unit. The output of the residual convolution unit of input port 1 and the input of input port 2 are fused and then pass through a residual convolution unit, and then processed by the DAFH attention mechanism, and then up-sampled; The DAFH attention processing process is as follows: Among them, X is the input feature, Indicates the average grouping operation by channel, Represents the kth feature after grouping; for The results after spatial attention modulation, represents a 3×3 convolutional unit, Indicates the operation of Hadamard product. Represents the Sigmoid activation function; For The extracted spatial features, represents depth-wise separable convolution; is the feature after adaptive fusion, is the CA channel attention mechanism, represents the learnable fusion parameters, Represents the final output, Indicates that the grouped features are spliced ​​by channel; The input of the AMHF multi-scale feature fusion module N is processed by a 3×3 convolution layer, and then restored to the original image size through bilinear interpolation. Finally, the normalized height map is obtained through the Sigmoid activation function. Multiply the normalized height map by max_height to get the final predicted height map, where max_height represents the set maximum height.

8. The building height estimation system based on a single satellite image according to claim 7, characterized in that: The CA channel attention mechanism is as follows: in, Represents the input features of CA.

9. The building height estimation system based on a single satellite image according to claim 8, characterized in that: After the DAFH attention mechanism is processed, the upsampling process includes: Features processed based on DAFH attention mechanism Perform upsampling: in, Express The result after upsampling; represents transposed convolution, Indicates bilinear interpolation, and either of the two sampling methods is selected for upsampling; Then make channel adjustments: in, For the final output, It is a 1×1 convolution unit.

10. A building height estimation system based on a single satellite image according to any one of claims 7 to 9, characterized in that: The N is 4, that is, the encoder extracts 4 layers of height features of a single satellite image, and the decoder uses 4 AMHF multi-scale feature fusion modules for processing.

Images