A monocular remote sensing image height estimation method and device based on semantic distribution and regional modulation

Abstract

The application discloses a monocular remote sensing image height estimation method and device based on semantic distribution and region modulation, and belongs to the technical field of computer vision and remote sensing image processing. The method first acquires a multi-scale feature map through bidirectional multi-scale feature extraction; then, a height interval semantic modeling module is used to introduce global height semantic information to modulate the height interval representation, so that preliminary height prediction is obtained; finally, a region-aware pixel-level height modulation module is used to explicitly model the height region to which the pixels belong and generate modulation parameters, and the preliminary prediction is differentially corrected. The application can significantly improve the precision, spatial consistency and robustness of building height estimation, and is suitable for application scenarios such as city modeling and remote sensing surveying and mapping.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and remote sensing image processing technology, specifically relating to a method and apparatus for estimating the height of monocular remote sensing images based on semantic distribution and region modulation. Background Technology

[0002] Monocular remote sensing image height estimation is a core task in the fields of remote sensing image processing and geographic information acquisition, and it is of great significance for urban planning, environmental monitoring, and 3D modeling. Existing methods are mainly divided into two categories: height regression modeling and height binning modeling.

[0003] Height regression modeling utilizes neural networks to directly regress a normalized digital surface model from a single RGB image. Existing research primarily employs convolutional neural networks or encoder-decoder structures, improving height regression accuracy through multi-scale feature extraction and fusion. Some methods combine convolutional neural networks with Transformers to balance local details with global context awareness. Furthermore, incorporating prior geometric and physical information such as shadows and off-axis views, or jointly learning height estimation with tasks like semantic segmentation and building extraction, also enhances model performance to some extent.

[0004] Height binning modeling divides a continuous height space into several discrete intervals and employs classification or a combined classification-regression approach to enhance the model's ability to represent the height distribution structure. Methods such as ordinal regression, DORN, and AdaBins transform depth estimation into a discrete interval classification task, improving prediction stability. Inspired by this, methods like HTC-DC Net and HeightFormer apply the height binning concept to remote sensing building height estimation, combining global context modeling with discrete height interval discrimination to achieve significant performance improvements.

[0005] However, existing technologies still have many shortcomings. Height regression modeling methods are prone to prediction bias under conditions of large height distribution spans and uneven sample distribution, have limited global context modeling capabilities, and are insufficient in characterizing large-scale building structures and overall height distribution. Methods incorporating physical priors are highly dependent on lighting conditions and imaging parameters, resulting in poor applicability in high-density urban areas. Multi-task learning methods are insufficient in explicitly modeling semantic relationships and spatial regional differences within height intervals, making it difficult to achieve regionally adaptive height correction.

[0006] The height binning modeling method has a relatively weak modeling of the relationship between height interval semantics and global height semantics, making it difficult to effectively guide the height interval discrimination process using global context information, and easily leading to interval confusion. In addition, the modeling of the relationship between height interval semantics and pixel-level spatial response is relatively coarse, failing to explicitly characterize the significant differences in height structure between different spatial regions, and making it difficult to achieve fine-grained, region-adaptive height correction.

[0007] Therefore, how to improve the accuracy, consistency and robustness of monocular remote sensing image height estimation in complex urban scenarios is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method and apparatus for estimating the height of monocular remote sensing images based on semantic distribution and regional modulation. It integrates bidirectional multi-scale feature extraction, global height interval semantic modeling, and regionally perceptive pixel-level height modulation. By explicitly modeling the semantic distribution of height and adaptively correcting spatial heterogeneity, it improves the accuracy and consistency of building height estimation in complex urban scenarios.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A method for estimating the height of monocular remote sensing images based on semantic distribution and region modulation, the method comprising:

[0011] Step 1: Preprocess the input monocular remote sensing image to obtain a processed image of uniform size;

[0012] Step 2: Perform bidirectional multi-scale feature extraction on the processed image to obtain a multi-scale feature map;

[0013] Step 3: Process the multi-scale feature map to transform continuous height estimation into a height interval probability inference problem. Adaptively modulate the height interval representation by introducing global height semantic information to obtain a preliminary pixel-level height prediction map.

[0014] Step 4: Obtain the height region probability distribution of each pixel based on the multi-scale feature map, and generate pixel-level modulation parameters according to the height region probability distribution;

[0015] Step 5: Modulate the preliminary pixel-level height prediction map pixel by pixel using the pixel-level modulation parameters to generate and output the final height estimation map.

[0016] Furthermore, in step 2: the backbone feature encoder constructed using a deep neural network performs multi-level feature encoding on the processed image, and through a bidirectional multi-scale feature aggregation structure, it fuses low-level spatial detail features while maintaining the integrity of high-level semantic information, and outputs the multi-scale feature map.

[0017] Furthermore, in step 3:

[0018] The Patch Transformer encoder is used to generate a token sequence. The first query token is extracted as a regression head to capture global scene semantics and global features are mapped to the probability distribution of the height interval. At the same time, multiple subsequent query tokens are extracted for foreground modeling and background modeling, respectively.

[0019] Two convolutional layers are used to process multi-scale feature maps, generating foreground and background feature maps respectively.

[0020] The probability distribution of height intervals is predicted using a regression head. The interval width is normalized and linearly mapped to a predefined range from minimum to maximum height. Continuous height interval boundaries are constructed through accumulation operations, and the center value of adjacent intervals is calculated as a representative of discrete height.

[0021] Furthermore, in step 3, the SEQT (semantic enhancement query transformation) module is used to adaptively modulate the height interval query vector using global height semantics:

[0022] First, semantic features are obtained by semantically encoding the regression head using a lightweight multilayer perceptron, from which modulation parameters specific to the query vector are generated. Then, feature-level linear modulation is applied to adaptively transform the query vector. Next, a multi-head attention mechanism is used to enhance the consistency between query vectors. Finally, the original information is preserved through residual connections. The foreground query vector and background query vector are processed by independent SEQT modules to generate enhanced foreground query vector and background query vector, and pixel-level attention maps are calculated through dot product operations.

[0023] Furthermore, in step 3, the preliminary pixel-level height prediction map is generated by weighted summation based on height intervals, including: weighting and accumulating the corresponding interval responses using the center value of each height interval to obtain continuous height prediction results.

[0024] Furthermore, in step 4: a lightweight convolution operation is used, with the multi-scale feature map as input, to output a multi-channel pixel-level region probability map, wherein each channel corresponds to a predefined height region, and the predefined height region includes ground region, low region, medium height region and high altitude region.

[0025] For each pixel, the modulation factor is calculated by weighting and aggregating the region probability map with learnable modulation weights, so that each pixel receives a spatial variation correction that reflects its high region affiliation distribution.

[0026] Furthermore, in step 5: after weighted modulation of the preliminary pixel-level height prediction map and the pixel-level modulation parameters, the pixel-level building height map corresponding to the image is output.

[0027] On the other hand, the present invention provides a monocular remote sensing image height estimation device based on semantic distribution and region modulation, comprising:

[0028] The image input module is used to preprocess the input monocular remote sensing image to obtain the processed image;

[0029] A multi-scale feature extraction module is used to extract multi-scale features from the processed image to obtain a multi-scale feature map.

[0030] The height interval semantic modeling module is used to process the multi-scale feature map, transforming continuous height estimation into a height interval probability inference problem. By introducing global height semantic information, the height interval representation is adaptively modulated to obtain a preliminary pixel-level height prediction map.

[0031] A region-aware pixel-level height modulation module is used to obtain the height region probability distribution of each pixel based on the multi-scale feature map, and generate pixel-level modulation parameters according to the height region probability distribution.

[0032] The result output module is used to perform pixel-by-pixel modulation on the preliminary pixel-level height prediction map using the pixel-level modulation parameters, and generate and output the final height estimation map.

[0033] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned monocular remote sensing image height estimation method based on semantic distribution and regional modulation.

[0034] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned monocular remote sensing image height estimation method based on semantic distribution and region modulation.

[0035] The beneficial effects of this invention are as follows:

[0036] Significantly improved prediction accuracy: Through a collaborative mechanism of bidirectional multi-scale feature aggregation and semantically guided interval modeling, state-of-the-art performance was achieved on the DFC2023 dataset, with an RMSE of 4.61m, a MAE of 1.59m, and an absRel reduced to 0.2180. Compared with HTC-DC Net, the RMSE was reduced by 0.9m, the MAE by 0.58m, and the absRel by 0.027.

[0037] Enhanced modeling and high consistency of complex structures: Global multi-scale feature interaction and region-aware modulation can effectively maintain the high consistency of the building interior, reduce the height interference of adjacent buildings, fully restore complex geometric forms such as slender buildings, and ensure the smoothness of height distribution and structural stability.

[0038] Improved boundary delineation accuracy and anti-interference capability: The region-aware pixel-level modulation mechanism differentiates the height response of different spatial regions, making the height change at the building boundary steeper and more stable, effectively suppressing background error response, and significantly improving boundary alignment accuracy and contour clarity.

[0039] Breakthroughs in robustness and generalization in complex scenarios: Multi-module collaborative design explicitly models scale differences, interval semantic relationships, and regional characteristics, effectively alleviating scale ambiguity, adapting to urban scenarios with frequent height changes and complex geometric shapes, and significantly improving model robustness and generalization capabilities.

[0040] Expanding the adaptability and generalization performance of cross-border domestic data sources: The multi-module collaborative design takes into account both semantic distribution modeling and regional modulation mechanism, which can effectively adapt to different sensor imaging conditions and resolution differences, maintain stable performance on GF-7 domestic remote sensing data, and significantly improve the model's generalization ability and robustness in engineering applications. Attached Figure Description

[0041] Figure 1 This is a flowchart of a monocular remote sensing image height estimation method based on semantic distribution and region modulation according to the present invention.

[0042] Figure 2 This is a flowchart of the bidirectional multi-scale feature extraction process;

[0043] Figure 3 Flowchart for semantic modeling of height intervals;

[0044] Figure 4 Flowchart for pixel-level height modulation for region-aware applications;

[0045] Figure 5 This is a block diagram of a monocular remote sensing image height estimation device based on semantic distribution and regional modulation according to the present invention. Detailed Implementation

[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0047] This invention provides a method and apparatus for monocular remote sensing image height estimation based on semantic distribution and region modulation. It is designed collaboratively from three levels: feature representation, height modeling, and pixel-level modulation. The overall approach consists of three key stages: multi-scale feature extraction, height interval semantic modeling, and region-aware pixel-level height modulation. Figure 1 As shown, the method specifically includes:

[0048] Step 1: Preprocess the input monocular remote sensing image to obtain a processed image of uniform size;

[0049] The image input module receives monocular remote sensing images, scales and crops the input images to a uniform size (e.g., 512×512 pixels), and normalizes their heightmap labels to the range [0,1]. The image format is RGB three-channel.

[0050] Step 2: Perform bidirectional multi-scale feature extraction on the processed image to obtain a multi-scale feature map;

[0051] like Figure 2 As shown, a multi-scale feature extraction module is used to obtain multi-scale feature maps. This module is based on a backbone feature encoder constructed from deep neural networks (such as EfficientNet, Dinov3, or Vision Transformer (ViT) series models, lightweight backbone networks such as MobileNet, and multi-branch convolutional networks such as ResNeXt; the core of which is to extract multi-level, multi-scale feature representations from remote sensing images) to perform multi-level feature encoding on the input monocular remote sensing image. A multi-scale feature fusion module is introduced to effectively fuse low-level spatial detail features while maintaining the integrity of high-level semantic information. The multi-scale feature fusion module includes a bidirectional multi-scale feature aggregation structure (such as the FPN-PAN structure). Specifically, the encoder first extracts multi-level, multi-scale feature representations from the input single remote sensing image, and then the decoder performs step-by-step feature recovery and fusion to output a multi-scale feature map.

[0052] Step 3: Process the multi-scale feature map to transform continuous height estimation into a height interval probability inference problem. Adaptively modulate the height interval representation by introducing global height semantic information to obtain a preliminary pixel-level height prediction map.

[0053] like Figure 3 As shown, the height interval semantic modeling module is used to perform preliminary height interval prediction on the feature map. This module specifically includes the following sub-modules:

[0054] (1) Global Context Feature Extraction Module: The Patch Transformer encoder processes the input features (i.e., multi-scale feature maps) to generate a token sequence. The first query token is extracted as a regression head to capture global scene semantics and map global features to the probability distribution of the height interval, thereby enabling continuous depth prediction through expectation calculation. The next S query tokens are used for foreground modeling, and the other S query tokens are used for background modeling. The regression head is output to the height interval segmentation module, and the foreground and background query tokens are output to the SEQT module.

[0055] (2) Local feature extraction module: Two 3×3 convolutional layers are used to process the input features (i.e., multi-scale feature maps) to generate foreground and background feature maps. The foreground and background feature maps are output as local features to the global context modeling module.

[0056] (3) Height Interval Division Module: The regression head predicts the probability distribution of height bins through a regressor composed of a multi-layer fully connected network, and normalizes the distribution to form an effective height interval probability representation. Based on the predicted height bin probability distribution, the interval width is normalized and linearly mapped to a predefined height range [H_min, H_max]. Subsequently, continuous height interval boundaries are constructed through accumulation operations, and the center value of adjacent intervals is calculated as a discrete height representative to obtain the divided height intervals. The divided height intervals are output to the global context modeling module.

[0057] (4) SEQT Module: Adaptively modulates the height interval query vector using global height semantics. First, semantic features are obtained by semantically encoding the regression head using a lightweight multilayer perceptron, from which query-specific modulation parameters are generated. Then, feature-level linear modulation is applied to adaptively transform the query tokens. Next, a multi-head attention mechanism is used to enhance the consistency between query tokens. Finally, the original information is preserved through residual connections. Foreground and background query tokens are processed by independent SEQT modules to generate enhanced foreground and background query tokens. The enhanced query tokens generate pixel-level attention maps through dot product operations and are output to the global context modeling module.

[0058] (5) Global Context Modeling Module: Receives the foreground and background feature maps output by the local feature extraction module, the pixel-level attention map output by the SEQT module, and the divided height intervals output by the height interval segmentation module. Based on the pixel-level attention map, the local features are weighted and fused, and combined with the center values ​​of each height bin in the divided height intervals, the corresponding interval responses are weighted and accumulated to obtain continuous height prediction results, i.e., a preliminary pixel-level height prediction map.

[0059] Step 4: Obtain the height region probability distribution of each pixel based on the multi-scale feature map, and generate pixel-level modulation parameters according to the height region probability distribution;

[0060] like Figure 4 As shown, a region-aware pixel-level height modulation module is used to obtain the modulation parameters of each pixel in each height range. This module includes:

[0061] (1) Region classification module: Using lightweight convolution operation, with multi-scale feature map as input, output a four-channel pixel-level region probability map, where each channel corresponds to one of four predefined height regions: Ground (1-2m), Low (2-20m), Mid (20-40m) and High (40m+).

[0062] Without loss of generality, the generation method of the four-channel height region probability map can be replaced by other region segmentation strategies, including but not limited to 2 / 3 / 5-channel predefined region classification, dynamic region segmentation based on semantic segmentation, and automatic mining of height region features through self-supervised learning. The convolution operation can be replaced by lightweight Transformer blocks and depthwise separable convolution. The core is to output pixel-level height region attribution information.

[0063] (2) Weighted aggregation module: For each pixel, the modulation factor is calculated by weighting and aggregating the pixel-level region probability map with the learnable modulation weight. Each pixel receives spatial variation correction, reflecting its distribution in the height region.

[0064] Step 5: Modulate the preliminary pixel-level height prediction map pixel by pixel using the pixel-level modulation parameters to generate and output the final height estimation map.

[0065] The modulation parameters output by the region-aware pixel-level height modulation module are used to perform pixel-by-pixel height modulation on the preliminary height map output by the height interval semantic modeling module, ultimately obtaining the predicted height map. The results output module then weights and modulates the results from both modules to output the corresponding height map of the image, representing the height value of each pixel in the image.

[0066] In one specific embodiment, the DFC2023 dataset was used to validate the method of the present invention. The dataset was divided into training and testing sets. The input monocular remote sensing images were preprocessed and then input into the model for training and inference. Experimental results show that the method of the present invention achieves high height estimation accuracy on the testing set, for example, RMSE of 4.61m and MAE of 1.59m, which is significantly better than existing methods. In addition, the US3D dataset was also used for method validation. Experimental results show that the method of the present invention achieves high height estimation accuracy on the testing set, for example, RMSE of 4.18m and MAE of 1.49m, which is significantly better than existing methods.

[0067] Furthermore, when tested on GF-7 remote sensing image data, the method of this invention can also obtain stable height estimation results, such as RMSE of 4.11m and MAE of 2.18m, indicating that it has good generalization ability across data sources.

[0068] On the other hand, such as Figure 5 As shown, this invention provides a monocular remote sensing image height estimation device based on semantic distribution and region modulation, which includes modules capable of implementing the steps of the aforementioned method, specifically including:

[0069] Image input module: Used to receive input remote sensing image data in RGB three-channel format, and perform preprocessing to obtain the processed image.

[0070] Multi-scale feature extraction module: The backbone feature encoder, built based on deep neural networks (such as EfficientNet, Dinov3, etc.), performs multi-level feature encoding on the preprocessed monocular remote sensing image. By introducing a bidirectional multi-scale feature aggregation structure, it effectively integrates low-level spatial detail features while maintaining the integrity of high-level semantic information, and extracts multi-scale information of the remote sensing image to obtain a multi-scale feature map.

[0071] Height Interval Semantic Modeling Module: This module is suitable for explicitly modeling continuous height estimation tasks as a probabilistic inference problem related to height intervals. By constructing intermediate representations related to height intervals and adaptively modulating them with global height semantic information, the module enables the height interval discrimination process to simultaneously perceive global scene height distribution features and local structural information. This improves the consistency and stability of modeling different height intervals and yields preliminary pixel-level height prediction results.

[0072] Region-aware pixel-level height modulation module: This module obtains the height region probability distribution of each pixel by explicitly modeling the region to which the pixel belongs, and generates pixel-level modulation parameters based on the height region probability distribution.

[0073] The result output module uses the pixel-level modulation parameters to perform pixel-by-pixel differential modulation on the preliminary pixel-level height prediction map, generating and outputting the final height estimation map. This effectively characterizes the spatial heterogeneity of pixel-level height behavior while maintaining the continuity of overall height estimation, further improving the local accuracy and overall robustness of height prediction.

[0074] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for estimating the height of monocular remote sensing images based on semantic distribution and regional modulation.

[0075] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for estimating the height of monocular remote sensing images based on semantic distribution and regional modulation.

[0076] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for estimating the height of monocular remote sensing images based on semantic distribution and region modulation, characterized in that, The method includes: Step 1: Preprocess the input monocular remote sensing image to obtain a processed image of uniform size; Step 2: Perform bidirectional multi-scale feature extraction on the processed image to obtain a multi-scale feature map; Step 3: Process the multi-scale feature map, transforming continuous height estimation into a height interval probability inference problem. Adaptive modulation of the height interval representation is achieved by introducing global height semantic information to obtain a preliminary pixel-level height prediction map; including: The Patch Transformer encoder is used to generate a token sequence. The first query token is extracted as a regression head to capture global scene semantics and global features are mapped to the probability distribution of the height interval. At the same time, multiple subsequent query tokens are extracted for foreground modeling and background modeling, respectively. Two convolutional layers are used to process multi-scale feature maps, generating foreground and background feature maps respectively. The probability distribution of height intervals is predicted using a regression head. The interval width is normalized and linearly mapped to a predefined range from minimum to maximum height. Continuous height interval boundaries are constructed through accumulation operations, and the center value of adjacent intervals is calculated as a representative of discrete height. The semantic enhancement query transformation module adaptively modulates the height interval query vector using global height semantics: First, semantic features are obtained by semantically encoding the regression head using a lightweight multilayer perceptron, from which modulation parameters specific to the query vector are generated. Then, feature-level linear modulation is applied to adaptively transform the query vector. Next, a multi-head attention mechanism is used to enhance the consistency between query vectors. Finally, the original information is preserved through residual connections. The foreground query vector and background query vector are processed by independent SEQT modules to generate enhanced foreground query vector and background query vector, and pixel-level attention maps are calculated through dot product operations. Based on the pixel-level attention map, local features are weighted and fused, and combined with the center value of each height bin in the divided height interval, the corresponding interval response is weighted and accumulated to obtain continuous height prediction results, which serve as a preliminary pixel-level height prediction map. Step 4: Obtain the height region probability distribution of each pixel based on the multi-scale feature map, and generate pixel-level modulation parameters according to the height region probability distribution; Step 5: Modulate the preliminary pixel-level height prediction map pixel by pixel using the pixel-level modulation parameters to generate and output the final height estimation map.

2. The method for estimating the height of monocular remote sensing images based on semantic distribution and region modulation according to claim 1, characterized in that, In step 2: the backbone feature encoder constructed using a deep neural network performs multi-level feature encoding on the processed image, and through a bidirectional multi-scale feature aggregation structure, it fuses low-level spatial detail features while maintaining the integrity of high-level semantic information, and outputs the multi-scale feature map.

3. The method for estimating the height of monocular remote sensing images based on semantic distribution and region modulation according to claim 1, characterized in that, In step 4: a lightweight convolution operation is used, with the multi-scale feature map as input, to output a multi-channel pixel-level region probability map, where each channel corresponds to a predefined height region, including ground region, low region, medium height region and high altitude region. For each pixel, the modulation factor is calculated by weighting and aggregating the region probability map with learnable modulation weights, so that each pixel receives a spatial variation correction that reflects its high region affiliation distribution.

4. The method for estimating the height of monocular remote sensing images based on semantic distribution and region modulation according to claim 1, characterized in that, In step 5: the preliminary pixel-level height prediction map is weighted and modulated with the pixel-level modulation parameters, and the pixel-level building height map corresponding to the image is output.

5. A monocular remote sensing image height estimation device based on semantic distribution and region modulation, characterized in that, include: The image input module is used to preprocess the input monocular remote sensing image to obtain the processed image; A multi-scale feature extraction module is used to extract multi-scale features from the processed image to obtain a multi-scale feature map. A height interval semantic modeling module is used to process the multi-scale feature map, transforming continuous height estimation into a height interval probability inference problem. It adaptively modulates the height interval representation by introducing global height semantic information to obtain a preliminary pixel-level height prediction map; including: The Patch Transformer encoder is used to generate a token sequence. The first query token is extracted as a regression head to capture global scene semantics and global features are mapped to the probability distribution of the height interval. At the same time, multiple subsequent query tokens are extracted for foreground modeling and background modeling, respectively. Two convolutional layers are used to process multi-scale feature maps, generating foreground and background feature maps respectively. The probability distribution of height intervals is predicted using a regression head. The interval width is normalized and linearly mapped to a predefined range from minimum to maximum height. Continuous height interval boundaries are constructed through accumulation operations, and the center value of adjacent intervals is calculated as a representative of discrete height. The semantic enhancement query transformation module adaptively modulates the height interval query vector using global height semantics: First, semantic features are obtained by semantically encoding the regression head using a lightweight multilayer perceptron, from which modulation parameters specific to the query vector are generated. Then, feature-level linear modulation is applied to adaptively transform the query vector. Next, a multi-head attention mechanism is used to enhance the consistency between query vectors. Finally, the original information is preserved through residual connections. The foreground query vector and background query vector are processed by independent SEQT modules to generate enhanced foreground query vector and background query vector, and pixel-level attention maps are calculated through dot product operations. Based on the pixel-level attention map, local features are weighted and fused, and combined with the center value of each height bin in the divided height interval, the corresponding interval response is weighted and accumulated to obtain continuous height prediction results, which serve as a preliminary pixel-level height prediction map. A region-aware pixel-level height modulation module is used to obtain the height region probability distribution of each pixel based on the multi-scale feature map, and generate pixel-level modulation parameters according to the height region probability distribution. The result output module is used to perform pixel-by-pixel modulation on the preliminary pixel-level height prediction map using the pixel-level modulation parameters, and generate and output the final height estimation map.

6. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the monocular remote sensing image height estimation method based on semantic distribution and regional modulation as described in any one of claims 1-3.

7. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the monocular remote sensing image height estimation method based on semantic distribution and region modulation as described in any one of claims 1-3.

Images