A power plant type detection method based on deep learning

Abstract

This application provides a deep learning-based method for power plant type detection, comprising: acquiring remote sensing images of power plants; inputting the remote sensing images of power plants into a power plant type detection model to obtain power plant type detection results, wherein the power plant type detection model includes a feature extraction module, a multi-scale fusion feature module, and a classification and prediction module connected in sequence; the multi-scale fusion feature module includes a local-global attention module and a multi-level feature fusion module. The method proposed in this application can detect different types of power plants, adapts to situations where the target power plants have vastly different scales, maintains accurate detection performance in multi-scale power plant detection, and retains accurate detection results even when the background is highly similar to the power plant.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to a deep learning-based method for detecting power plant types. Background Technology

[0002] The rapid construction of renewable energy power plants is causing significant changes in the power structure of certain regions, posing a major challenge to balancing power supply and demand. Accurate reporting and monitoring of power plant distribution maps are crucial to addressing this challenge. Such maps, combined with supplementary information (such as fossil fuel consumption and meteorological and hydrological conditions), can comprehensively assess power supply capacity, thereby supporting the analysis of power supply and demand relationships and the assessment of renewable energy potential.

[0003] The core task of power plant distribution map monitoring is to obtain power plant location information. Traditional power plant surveys rely on manual statistics, on-site inspections, or local remote sensing interpretation, which suffers from limited coverage, long processing times, and high labor costs, making it difficult to quickly achieve synchronous monitoring at a "whole-area" scale. Currently, with the rapid development of satellite imaging and processing technologies, satellite imagery combined with deep learning methods is widely considered a potential tool for efficiently monitoring power plant locations. Research on the detection of other types of power plants remains relatively scarce, mainly facing two challenges: First, the scale of power plants varies significantly. For example, large mega-power plants (such as the Three Gorges Dam) can reach 2300 meters in length, while small power plants are only tens of meters long, a difference of tens of times. Second, the relationship between power plants and their backgrounds is complex: power plants and their backgrounds may be highly similar or significantly different. For example, wind power stations may be distributed in different terrains such as hills, grasslands, and oceans, while coal-fired power plants are often built in urban areas, closely resembling the appearance of surrounding industrial buildings. These two major challenges require the proposed method to possess extremely high stability and be insensitive to the size and background of the power plants. However, existing methods cannot meet this requirement, and research on such methods is still in the exploratory stage. Summary of the Invention

[0004] This application provides a deep learning-based method for detecting power plant types. To solve the aforementioned technical problems, this application employs the following technical methods:

[0005] Firstly, this application provides a deep learning-based method for detecting power plant types, including:

[0006] Acquire remote sensing images of the power plant;

[0007] The remote sensing image of the power plant is input into the power plant type detection model to obtain the power plant type detection result. The power plant type detection model includes a feature extraction module, a multi-scale fusion feature module, and a classification and prediction module connected in sequence. The multi-scale fusion feature module includes a local-global attention module and a multi-level feature fusion module.

[0008] Optionally, inputting the remote sensing image of the power plant into the power plant type detection model to obtain the power plant type detection result includes:

[0009] The remote sensing image of the power plant is input into the feature extraction module to obtain multi-scale semantic features of the power plant.

[0010] The multi-scale power plant semantic features are input into the multi-scale fusion feature module to obtain multi-scale fusion features;

[0011] The multi-scale fusion features are input into the classification and prediction module to obtain the power plant type detection result.

[0012] Optionally, the step of inputting the multi-scale power plant semantic features into the multi-scale fusion feature module to obtain multi-scale fusion features includes:

[0013] The multi-scale power plant semantic features are input into the local-global attention module to obtain multi-scale local-global aggregated features.

[0014] The multi-scale local-global aggregated features are input into the multi-level feature fusion module to obtain multi-scale fused features.

[0015] Optionally, the local-global attention module includes a first channel separation module, a global information branch extraction module, a local information branch extraction module, and a concatenation module. The local information branch extraction module is a 3×3 depthwise separable convolution module. The step of inputting the multi-scale power plant semantic features into the local-global attention module to obtain multi-scale local-global aggregated features includes:

[0016] The multi-scale power plant semantic features are input into the first channel separation module to obtain global information branches and local information branches.

[0017] The global information branch is input into the global information branch extraction module to obtain global context features;

[0018] The local information branch is input into the local information branch extraction module to obtain local spatial features;

[0019] The global context features and the local spatial features are input into the splicing module to obtain multi-scale local-global aggregated features.

[0020] Optionally, the global information branch extraction module includes a first convolution module, an average pooling module, a max pooling module, and a transposed attention module; the step of inputting the global information branch into the global information branch extraction module to obtain global context features includes:

[0021] The global information branch is processed by a 1×1 convolution through the first convolution module to obtain multi-channel auxiliary features, queries, keys, and values;

[0022] The query is processed by the average pooling module to perform average pooling, resulting in robust mean features;

[0023] The key is subjected to max pooling by the max pooling module to obtain the advantageous features;

[0024] The robust mean feature, the dominance feature, and the value are processed by the transposed attention module to obtain the global context feature.

[0025] The step of inputting the global context features and the local spatial features into the concatenation module to obtain multi-scale local-global aggregated features includes:

[0026] The global context features, the local spatial features, and the multi-channel auxiliary features are input into the splicing module to obtain multi-scale local-global aggregated features.

[0027] Optionally, the multi-level feature fusion module includes a multi-level strong and weak feature enhancement module and a feature enhancement module; the multi-scale local-global aggregated feature includes at least a first-scale local-global aggregated feature and a second-scale local-global aggregated feature; the step of inputting the multi-scale local-global aggregated feature into the multi-level feature fusion module to obtain the multi-scale fused feature includes:

[0028] After processing the first-scale local-global aggregated features through a 1×1 convolution, the first-scale convolutional local-global aggregated features are obtained.

[0029] After upsampling and 1×1 convolution, the second-scale local-global aggregated features are processed to obtain the upsampled second-scale convolutional local-global aggregated features.

[0030] The first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features are input into the feature enhancement module to obtain the aggregated features enhanced by the first-scale features.

[0031] The first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features are input into the strong and weak feature enhancement module to obtain the aggregated features after the first-scale strong and weak feature enhancement.

[0032] The aggregated features enhanced by the first scale strength feature and the aggregated features enhanced by the first scale feature are added pixel-level to obtain the multi-scale fusion feature.

[0033] Optionally, the step of inputting the first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features into the strong-weak feature enhancement module to obtain the aggregated features enhanced by the first-scale strong-weak features includes:

[0034] The first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features are added pixel-level to obtain preliminary fused features;

[0035] The first-scale convolutional local-global aggregated features, the upsampled second-scale convolutional local-global aggregated features, and the preliminary fused features are aggregated along the spatial dimension using average pooling. Then, the channel dimension weights of the first-scale convolution, the second-scale convolution, and the channel dimension weights of the preliminary fused features are calculated using the sigmoid function.

[0036] The first scale convolution channel dimension weights and the preliminary fusion feature channel dimension weights are compared, and the weights with higher weights are taken as the first strong feature weights, and the weights with lower weights are taken as the first weak feature weights.

[0037] The weights of the second-scale convolutional channel dimension are compared with the weights of the preliminary fused feature channel dimension. The weights with higher weights are used as the second strong feature weights, and the weights with lower weights are used as the second weak feature weights.

[0038] The first strong feature weight and the second strong feature weight are used to modulate the preliminary fused features to obtain clustered strong features;

[0039] The preliminary fused features are modulated using the first weak feature weight and the second weak feature weight to obtain clustered weak features;

[0040] The strong clustering features are refined by 1×1 convolution and then added to the weak clustering features that have undergone channel attention processing to obtain the aggregated features enhanced by the first-scale strong and weak features.

[0041] Optionally, the step of inputting the first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features into the feature enhancement module to obtain the aggregated features enhanced at the first scale includes:

[0042] The first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features are fused using a 3×3 convolutional layer to obtain fused coarse features.

[0043] The fused coarse feature is averaged along the channel dimension to obtain the first segmented fused coarse feature and the second segmented fused coarse feature.

[0044] The first segmented coarse features are subjected to spatial attention processing to obtain spatial focused features;

[0045] After the spatial focusing feature and the second segmentation fusion coarse feature are concatenated in the channel dimension, and then processed by a 1×1 convolutional layer, the aggregated feature after the first scale feature enhancement is obtained.

[0046] In a second aspect, this application also provides a computer-readable storage medium, characterized in that it stores computer program code, which, when executed by a processor, implements the method as described in any of the first aspects.

[0047] This application has the following beneficial effects:

[0048] The method proposed in this application can detect different types of power plants and is also suitable for situations where the target power plants vary greatly in size. It can also achieve accurate detection performance in multi-scale power plant detection and maintain accurate detection results when facing situations where the background is highly similar to the power plant. Attached Figure Description

[0049] Figure 1 A flowchart illustrating a deep learning-based power plant type detection method provided in this application embodiment;

[0050] Figure 2 This is a schematic diagram of the structure of the power plant type detection model provided in the embodiments of this application;

[0051] Figure 3 This is a schematic diagram of the structure of the feature extraction module provided in an embodiment of this application;

[0052] Figure 4 This is a schematic diagram of the structure of the multi-scale fusion feature module provided in the embodiments of this application;

[0053] Figure 5 This is a schematic diagram of the structure of the local-global attention module provided in an embodiment of this application;

[0054] Figure 6 This is a schematic diagram of the structure of the multi-level feature fusion module provided in the embodiments of this application.

[0055] Explanation of reference numerals in the attached figures

[0056] The system includes: a power plant type detection model 200; a feature extraction module 210; a multi-scale fusion feature module 220; a classification and prediction module 230; a local-global attention module 221; a multi-level feature fusion module 222; a first channel separation module 510; a global information branch extraction module 520; a local information branch extraction module 530; and a splicing module 540; a first convolution module 521; an average pooling module 522; a max pooling module 523; and a transposed attention module 524; a multi-level strong and weak feature enhancement module 610; a feature enhancement module 620; a preliminary fusion module 611; a weight calculation module 612; a first weight comparison module 613; a second weight comparison module 614; a clustered strong feature modulation module 615; a clustered weak feature modulation module 616; a second convolution module 617; and a channel attention module 618; a third convolution module 621; a second channel separation module 622; a spatial attention module 623; and a fourth convolution module 624. Detailed Implementation

[0057] To facilitate understanding by those skilled in the art, the present application will be further described below in conjunction with embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present application.

[0058] To solve the above technical problems, such as Figure 1 As shown, this application proposes a deep learning-based method for power plant type detection, including:

[0059] Step S101: Acquire remote sensing images of the power plant;

[0060] The acquisition of remote sensing images of power plants in this step can be achieved using any existing technical means, and this application does not impose any specific limitations on this method.

[0061] Step S102: Input the remote sensing image of the power plant into the power plant type detection model to obtain the power plant type detection result. The power plant type detection model includes a feature extraction module, a multi-scale fusion feature module, and a classification and prediction module connected in sequence. The multi-scale fusion feature module includes a local-global attention module and a multi-level feature fusion module.

[0062] By inputting the remote sensing images obtained in the previous step into the power plant type detection model, the power plant type detection results can be obtained.

[0063] This section provides a detailed explanation of this step, taking into account the specific structure of the power plant type detection model:

[0064] like Figure 2 As shown, the power plant type detection model 200 of this application includes a feature extraction module 210, a multi-scale fusion feature module 220, and a classification and prediction module 230 connected in sequence.

[0065] The feature extraction module 210 of this application is a multi-scale power plant remote sensing image feature extraction architecture mainly constructed with residual blocks (ResBlocks). It can extract multi-scale power plant semantic features, and its structure is as follows: Figure 3 As shown, the remote sensing image, as input, first enters the "Scale 1" feature extraction block, which consists of two ResBlocks, outputting "Scale 1" features. These features are then downsampled and used as input to the "Scale 2" feature extraction block and as output to the "Scale 1" feature extraction block. The operations for the "Scale 2," "Scale 3," and "Scale 4" feature extraction blocks are similar to those for the "Scale 1" block. Ultimately, the module achieves hierarchical extraction of multi-scale features from the remote sensing image through the cascading of feature extraction blocks at four scales. This helps the model meet the recognition needs of targets at different scales, allowing the model to capture hierarchical information from small-scale details and local morphology to large-scale overall structure and spatial layout during feature learning.

[0066] like Figure 4 As shown, the multi-scale fusion feature module 220 of this application includes a local-global attention module 221 and a multi-level feature fusion module 222. The features at four scales (multi-scale power plant semantic features) output by the feature extraction module are used as input. After passing through the local-global attention module, the local and global features are concatenated to form local-global aggregated features at four scales. Finally, the local-global aggregated features are used as input to the multi-level feature fusion module. This multi-level feature fusion module replaces the Feature Pyramid (FPN) commonly used in traditional object detection tasks, enabling more effective interaction and fusion of features at different scales. The multi-level feature fusion module includes a multi-level strong / weak feature selection module and a feature enhancement module. In general, the local-global aggregated features of two adjacent scales are used as input to the multi-level feature fusion module, and finally, the fused features at four scales, i.e., the multi-scale fusion features, are output.

[0067] like Figure 5 As shown, the local-global attention module 221 of this application includes a first channel separation module 510, a global information branch extraction module 520, a local information branch extraction module 530, and a splicing module 540. The local information branch extraction module 530 is a 3×3 depthwise separable convolution module. The global branch extraction module 520 includes a first convolution module 521, an average pooling module 522, a max pooling module 523, and a transposed attention module 524. The specific processing procedure of multi-scale power plant semantic features in the local-global attention module is as follows:

[0068] Multi-scale power plant semantic features are generally represented by feature maps, let them be... ,in , and These represent the height, width, and number of channels of the feature map, respectively. First, the multi-scale power plant semantic features are... The input to the first channel separation module 510 is divided into a global information branch and a local information branch according to the channel dimensions, which are used for subsequent extraction of local and global information, respectively.

[0069] (1)

[0070] In the formula, and Representing the global information branch and the local information branch, the two branches can focus on feature extraction of the global context and local details respectively, achieving task decoupling and information complementarity.

[0071] In the local information branch, As input, local spatial features are then obtained through a 3×3 depthwise separable convolution. Depth convolution performs convolution operations within a single channel, eliminating the need for cross-channel computation. It can accurately capture spatial local dependencies between pixels with very few parameters (far fewer than standard convolution) and minimal computational cost, such as the edges and detailed textures of objects in remote sensing images. At the same time, it avoids the need for additional location encoding (because local spatial relationships are directly modeled through the convolution kernel).

[0072] In the global information branch, As input, the features are first split into four features by a 1×1 convolution in the first convolution module 521, which are multi-channel auxiliary features. Query ,key ,value ,in , , As features for obtaining global context, their number of channels is all 1. Secondly, regarding After the average pooling module 522 performs 3×3 average pooling (with a step size of 2), robust mean features are obtained. ;right After performing 2×2 max pooling (with a step size of 2) using the max pooling module, the dominant features are obtained. Then based on and Calculating attention weights is... Multiply, and then perform transpose attention processing (524) to obtain global context features. :

[0073] (2)

[0074] in This represents matrix multiplication. This represents the Softmax activation function (used to normalize attention weights). express The transpose of the matrix is ​​used to obtain the final global context features. and The latitude is consistent. This module uses two steps—pooling dimensionality reduction and transposed attention—to capture long-range global dependencies (such as the overall spatial layout of a power plant and its association with surrounding features) while reducing computational cost. Pooling dimensionality reduction lowers feature resolution, reducing the computational cost of subsequent matrix multiplications; while based on the dimensionality-reduced... , Calculate attention weights, and then... The features are multiplied to obtain the global context features. The computational complexity of this process is linearly related to the number of feature tokens, which is significantly better than traditional self-attention.

[0075] Finally, cascading operations are used to combine local features. Global context features and multi-channel auxiliary features The features are concatenated along the channel dimension to obtain the final local-global aggregated features. Specifically, this can be expressed using a relational formula:

[0076] (3)

[0077] This architecture, through a split-branch design, retains the detailed representation of target details by local features while incorporating the ability of global features to understand scene context, resulting in fused features with richer semantic information and discriminative capabilities. By efficiently capturing the local spatial information and global contextual dependencies of remote sensing images, it addresses the limitation of traditional self-attention mechanisms where "computational complexity has a quadratic relationship with image resolution." Ultimately, it achieves more accurate remote sensing change detection feature extraction with fewer model parameters and less computation, providing high-quality feature support for subsequent change area identification.

[0078] The specific structure of the multi-level feature fusion module 222 is as follows: Figure 6 As shown, the multi-level feature fusion module 222 includes a multi-level strong and weak feature enhancement module 610 and a feature enhancement module 620. The multi-scale local-global aggregated features output above generally include at least four different scale local-global aggregated features. For ease of subsequent understanding, the local-global aggregated features of two adjacent scales are respectively named the first-scale local-global aggregated features. and the second-scale local-global aggregated features This leads to the aggregation of local-global features at the first scale. and the second-scale local-global aggregated features The following describes the specific processing procedure of the multi-level feature fusion module on the input features, which serves as the input:

[0079] First Upsampling with At the same scale, after passing through a 1×1 convolution to unify the channel dimension, we obtain upsampled second-scale convolutional local-global aggregated features, which will... After passing through a 1×1 convolution with a unified channel, the first-scale convolutional local-global aggregated features are obtained. The upsampled second-scale convolutional local-global aggregated features and the first-scale convolutional local-global aggregated features are then input into the feature enhancement module to obtain the aggregated features enhanced from the first-scale features. The first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features are input into the strong and weak feature enhancement module to obtain the aggregated features enhanced by the first-scale strong and weak features. Finally, the aggregated features after enhancing the first-scale features are... Aggregated features enhanced with first-scale strength features Pixel-level addition is performed to obtain multi-scale fused features. .

[0080] The multi-level strong and weak feature enhancement module 610 includes a preliminary fusion module 611, a weight calculation module 612, a first weight comparison module 613, a second weight comparison module 614, a strong feature aggregation modulation module 615, a weak feature aggregation modulation module 616, a second convolution module 617, and a channel attention module 618. Combined with... Figure 6 The following is a detailed explanation of the specific processing procedures of the first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features in the multi-level strong and weak feature enhancement module 610:

[0081] For ease of understanding, the variable representations of the first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features are compared with those of the first-scale local-global aggregated features. Second-scale local-global aggregation features The variable representations must remain consistent. First, let... and In the initial fusion module 611, pixel-level addition is performed in the channel dimension to achieve initial fusion of cross-level features and obtain the initial fused features. Then , and The input is fed into the weight calculation module 612, where it is aggregated along the spatial dimension using average pooling. Then, the sigmoid function is used to calculate the weights for each channel dimension, resulting in the first-scale convolutional channel dimension weights. Second-scale convolution channel dimension weights and preliminary fusion of feature channel dimension weights :

[0082] (4)

[0083] in, sig represents the sigmoid function. This indicates global average pooling.

[0084] Weights based on the first-scale convolutional channel dimension Weights of preliminary fusion feature channel dimensions Input the first weight comparison module 613. This module compares the weight values ​​of each channel and defines the channel weight with the higher weight value as the first strong feature weight, and the channel weight with the lower weight value as the first weak feature weight. Then, input the second-scale convolution channel dimension weights. Weights of preliminary fusion feature channel dimensions Input the second weight comparison module 614. This module compares the weight values ​​of each channel and defines the channel with the higher weight value as the second strong feature weight, and the channel with the lower weight value as the second weak feature weight. This can be formulated as follows:

[0085] (5)

[0086] in, and These represent the strong feature weights and weak feature weights obtained by comparing their magnitudes, respectively.

[0087] Then, the first and second strong feature weights are input into the clustered strong feature modulation module 615. Modulation is performed to obtain strong clustering features. The first weak feature weight and the second weak feature weight are input into the aggregation weak feature modulation module 616 for processing. Modulation is performed to obtain clustered weak features. :

[0088] (6)

[0089] (7)

[0090] In the formula, This indicates pixel-by-pixel multiplication.

[0091] The strong features are aggregated and then refined through a 1×1 convolution in the second convolution module 617 to further preserve their detailed information. Meanwhile, to extract useful information hidden within the weak features, the weak features... Then, a channel attention module 618 is added to enhance the semantic features by improving the information interaction between channels and reducing the computational cost. Finally, pixel-level summation is performed to obtain the aggregated features after the first-scale strong and weak feature enhancement. :

[0092] (8)

[0093] In the formula, Represents a 1×1 convolution. This indicates the channel attention module.

[0094] The multi-level strong and weak feature selection module addresses the scale and semantic heterogeneity issues of features at different levels by filtering strong and weak features across multiple levels. It first separates strong information (effective features for small targets) from weak information (redundancy and noise) within the features. Then, it refines the strong information separately, while channel attention is used to highlight the channel specificity of the weak features, improving the robustness of recognition for multi-scale targets (such as power plants of different sizes), especially small targets. Finally, the two types of features are fused to achieve a balance between "preserving details" and "enhancing semantics."

[0095] The feature enhancement module 620 includes a third convolution module 621, a second channel separation module 622, a spatial attention module 623, and a fourth convolution module 624. Combined with... Figure 6 The following is a detailed explanation of the specific processing procedures for the first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features in the feature enhancement module 620:

[0096] First, and After pixel-level addition along the channel dimension, the result is fed into the third convolutional layer module for 3×3 convolutional fusion. and To obtain coarse features Then the coarse features The input to the second channel separation module 622 is averaged along the channel dimension to form the first segmentation fusion coarse feature. Second segmentation fusion coarse features :

[0097] (9)

[0098] in, and The length, width and number of channels are , , .

[0099] Then the first segmentation is merged with coarse features The input attention module 623 enhances the feature responses of key regions (such as the spatial location of the core area of ​​a power plant) in the spatial dimension. First, spatial attention weights are generated using a sigmoid function. Then to Weighted spatial focusing features :

[0100] (10)

[0101] Then, the second segmentation fusion coarse features and spatial focusing features are concatenated along the channel dimension, and the enhanced features and original features are fused. Finally, the data is input into the fourth convolutional module 624 for 1×1 convolutional layer processing to obtain the aggregated features after the first scale strong and weak feature enhancement. :

[0102] (11)

[0103] This design achieves complementary features such as multi-scale feature fusion, spatial attention enhancement, and original feature preservation without increasing computational burden, thereby improving feature representation capabilities.

[0104] As the core functional unit responsible for outputting results in the technical solution of this application, the classification and prediction module 230 takes the feature data (i.e., multi-scale fusion features) after multi-scale fusion processing as the core input. By performing a series of targeted processing on the feature data, such as feature parsing, dimension mapping and category discrimination, it finally outputs accurate power plant type detection results, providing direct data support for subsequent power plant monitoring, management and other applications.

[0105] From a technical implementation perspective, the classification and prediction module 230 here is not limited to a specific architecture; any existing technical solution that can meet the core requirement of "input feature data - output category judgment result" is applicable. Specifically, the selectable architecture types cover multiple technical directions: for example, in the field of traditional machine learning, mature models such as Support Vector Machine (SVM), Random Forest, and Gradient Boosting Decision Tree (GBDT) can be used. These models have the advantages of low training cost and fast inference speed in scenarios with relatively fixed feature dimensions, and can quickly complete the classification and discrimination of power plant types; in the field of deep learning, variant architectures of Convolutional Neural Networks (CNN) (such as ResNet, DenseNet, etc.) can be relied upon to achieve deep feature mining, which is especially suitable for scenarios where multi-scale fusion features contain complex spatial correlation information. Lightweight network architectures (such as MobileNet, EfficientNet) can also be used to meet real-time detection requirements, and attention mechanisms (such as SE module, CBAM module) can be combined to strengthen the weights of key features and further improve the accuracy of type detection.

[0106] The core innovation of this application focuses on the construction process of multi-scale fusion features, rather than the specific implementation architecture of classification and prediction. Therefore, this application maintains an open attitude towards the technical selection of the classification and prediction module 230, and does not impose mandatory requirements on the specific algorithm framework, number of network layers, parameter settings, and other details. As long as the architecture can effectively receive multi-scale fusion features and stably output accurate power plant type detection results (such as the distinction between specific types such as thermal power, hydropower, wind power, and photovoltaic power), it falls within the technical scope covered by this application.

[0107] In summary, the method proposed in this application can detect different types of power plants (coal-fired, hydropower, wind power, photovoltaic, and nuclear power), and is adaptable to situations where the scale of the target power plants varies greatly. It maintains accurate detection performance in multi-scale power plant detection and retains accurate detection results even when the background is highly similar to the power plant. This allows for the acquisition of spatiotemporal distribution information of power plants over a large area, providing support for cross-regional power supply and demand analysis and promoting power emission reduction.

[0108] This application also provides a computer-readable storage medium for storing a computer program. This computer-readable storage medium can be applied to a computer device, and the computer program causes the computer device to execute the corresponding processes in the methods described above in the embodiments of this application; for brevity, further details are omitted here.

[0109] The above embodiments are preferred implementations of this application. In addition, this application can be implemented in other ways. Any obvious substitutions without departing from the concept of this technical solution are within the protection scope of this application.

[0110] To facilitate understanding by those skilled in the art of the improvements made by this application compared to the prior art, some of the accompanying drawings and descriptions have been simplified, and for clarity, some other elements have been omitted from this application. Those skilled in the art should realize that these omitted elements may also constitute the content of this application.

Claims

1. A deep learning-based method for detecting power plant types, characterized in that, include: Acquire remote sensing images of the power plant; The remote sensing image of the power plant is input into the power plant type detection model to obtain the power plant type detection result. The power plant type detection model includes a feature extraction module, a multi-scale fusion feature module, and a classification and prediction module connected in sequence. The multi-scale fusion feature module includes a local-global attention module and a multi-level feature fusion module. The step of inputting the remote sensing image of the power plant into the power plant type detection model to obtain the power plant type detection result includes: The remote sensing image of the power plant is input into the feature extraction module to obtain multi-scale semantic features of the power plant. The multi-scale power plant semantic features are input into the multi-scale fusion feature module to obtain multi-scale fusion features; The multi-scale fusion features are input into the classification and prediction module to obtain the power plant type detection result; The step of inputting the multi-scale power plant semantic features into the multi-scale fusion feature module to obtain multi-scale fusion features includes: The multi-scale power plant semantic features are input into the local-global attention module to obtain multi-scale local-global aggregated features. The multi-scale local-global aggregated features are input into the multi-level feature fusion module to obtain multi-scale fused features; The multi-level feature fusion module includes a multi-level strong and weak feature enhancement module and a feature enhancement module; the multi-scale local-global aggregated feature includes at least a first-scale local-global aggregated feature and a second-scale local-global aggregated feature; the step of inputting the multi-scale local-global aggregated feature into the multi-level feature fusion module to obtain multi-scale fused features includes: After processing the first-scale local-global aggregated features through a 1×1 convolution, the first-scale convolutional local-global aggregated features are obtained. After upsampling and 1×1 convolution, the second-scale local-global aggregated features are processed to obtain the upsampled second-scale convolutional local-global aggregated features. The first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features are input into the feature enhancement module to obtain the aggregated features enhanced by the first-scale features. The first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features are input into the strong and weak feature enhancement module to obtain the aggregated features after the first-scale strong and weak feature enhancement. The aggregated features enhanced by the first scale strength feature and the aggregated features enhanced by the first scale feature are added pixel-level to obtain the multi-scale fusion feature. The step of inputting the first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features into the strong-weak feature enhancement module to obtain the aggregated features enhanced by the first-scale strong-weak features includes: The first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features are added pixel-level to obtain preliminary fused features; The first-scale convolutional local-global aggregated features, the upsampled second-scale convolutional local-global aggregated features, and the preliminary fused features are aggregated along the spatial dimension using average pooling. Then, the channel dimension weights of the first-scale convolution, the second-scale convolution, and the channel dimension weights of the preliminary fused features are calculated using the sigmoid function. The first scale convolution channel dimension weights and the preliminary fusion feature channel dimension weights are compared, and the weights with higher weights are taken as the first strong feature weights, and the weights with lower weights are taken as the first weak feature weights. The weights of the second-scale convolutional channel dimension are compared with the weights of the preliminary fused feature channel dimension. The weights with higher weights are used as the second strong feature weights, and the weights with lower weights are used as the second weak feature weights. The first strong feature weight and the second strong feature weight are used to modulate the preliminary fused features to obtain clustered strong features; The preliminary fused features are modulated using the first weak feature weight and the second weak feature weight to obtain clustered weak features; The strong clustering features are refined by 1×1 convolution and then added to the weak clustering features that have undergone channel attention processing to obtain the aggregated features enhanced by the first-scale strong and weak features.

2. The method according to claim 1, characterized in that, The local-global attention module includes a first channel separation module, a global information branch extraction module, a local information branch extraction module, and a concatenation module. The local information branch extraction module is a 3×3 depthwise separable convolution module. The multi-scale power plant semantic features are input into the local-global attention module to obtain multi-scale local-global aggregated features, including: The multi-scale power plant semantic features are input into the first channel separation module to obtain global information branches and local information branches. The global information branch is input into the global information branch extraction module to obtain global context features; The local information branch is input into the local information branch extraction module to obtain local spatial features; The global context features and the local spatial features are input into the splicing module to obtain multi-scale local-global aggregated features.

3. The method according to claim 2, characterized in that, The global information branch extraction module includes a first convolution module, an average pooling module, a max pooling module, and a transposed attention module; The step of inputting the global information branch into the global information branch extraction module to obtain global context features includes: The global information branch is processed by a 1×1 convolution through the first convolution module to obtain multi-channel auxiliary features, queries, keys, and values; The query is processed by the average pooling module to perform average pooling, resulting in robust mean features; The key is subjected to max pooling by the max pooling module to obtain the advantageous features; The robust mean feature, the dominance feature, and the value are processed by the transposed attention module to obtain the global context feature. The step of inputting the global context features and the local spatial features into the concatenation module to obtain multi-scale local-global aggregated features includes: The global context features, the local spatial features, and the multi-channel auxiliary features are input into the splicing module to obtain multi-scale local-global aggregated features.

4. The method according to claim 1, characterized in that, The step of inputting the first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features into the feature enhancement module to obtain the aggregated features enhanced at the first scale includes: The first-scale convolutional local-global aggregated features and the upsampled second-scale convolutional local-global aggregated features are fused using a 3×3 convolutional layer to obtain fused coarse features. The fused coarse feature is averaged along the channel dimension to obtain the first segmented fused coarse feature and the second segmented fused coarse feature. The first segmented coarse features are subjected to spatial attention processing to obtain spatial focused features; After the spatial focusing feature and the second segmentation fusion coarse feature are concatenated in the channel dimension, and then processed by a 1×1 convolutional layer, the aggregated feature after the first scale feature enhancement is obtained.

5. A computer-readable storage medium, characterized in that, The system contains computer program code that, when executed by a processor, implements the method as described in any one of claims 1 to 4.

Images