A ground-based carbon emission monitoring method based on a multi-source data interaction network
By designing and training a multi-source data interaction network, the problems of data consistency and standardization in carbon emission detection using deep learning methods were solved, achieving higher accuracy and robustness in carbon emission monitoring and improving the model's accuracy and generalization ability.
Patent Information
- Application Number
- CN202311456651.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-03
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-11-03
AI Technical Summary
Existing deep learning-based methods suffer from data consistency and standardization issues in carbon emission detection, leading to information loss, distortion, bias, and difficulty in establishing data correlations, which affects the accuracy and comprehensive understanding of the model.
A multi-source data interaction network is adopted, including a modular backbone network, an attention interaction module, and a cross-scale interaction module. Through data augmentation and training, the attention interaction module is used to enhance the attention distribution of multi-source data, and the cross-scale interaction module is used to improve the correlation between data at different scales, thereby achieving effective data fusion and prediction.
It improves the accuracy and robustness of carbon emission monitoring, enhances the generalization ability of the model, can more accurately capture carbon emission trends and changes, and reduces interference caused by data inconsistency.
Smart Images

Figure CN117454116B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of carbon emission detection and analysis technology, specifically relating to a ground-based carbon emission monitoring method based on a multi-source data interaction network. Background Technology
[0002] Carbon emission monitoring and analysis technologies play a crucial role in environmental science and meteorology today. With rapid global industrialization and urbanization, carbon emissions, as one of the main drivers of climate change, have attracted widespread concern and attention. Accurate monitoring, assessment, and in-depth understanding of the scale, distribution, and trends of carbon emissions are essential for better addressing climate challenges. However, carbon emission monitoring is not easy because emission sources are diverse, widely distributed, and influenced by weather and geographical conditions. To overcome these challenges, researchers and experts are constantly developing various technologies and methods to improve the accuracy of carbon emission monitoring and analysis. These methods include using satellite remote sensing data, meteorological monitoring data, ground sensors, and advanced data analysis techniques to achieve more accurate, comprehensive, and real-time collection and analysis of carbon emission information.
[0003] Carbon satellite remote sensing technology plays a crucial role in carbon emission monitoring. Carbon satellites provide high-resolution Earth observation images, helping to identify and locate carbon emission sources and accurately estimate emissions. Simultaneously, remote sensing data can also be used to monitor land use and cover changes. Activities such as deforestation, land clearing, and urban expansion affect carbon storage and emissions; remote sensing images can capture these changes, providing insights into the carbon cycle. However, simple remote sensing images typically cannot directly measure atmospheric carbon concentration. Instead, they need to be combined with ground-based specialized sensors and equipment such as Continuous Emissions Monitoring Systems (CEMS), vegetation monitoring sensors, and atmospheric samplers. These ground sensors provide critical data for monitoring and assessing the sources, distribution, and trends of carbon emissions. Furthermore, because carbon emission data is often very large, big data processing and storage technologies are required for effective management and analysis. Cloud computing technology enables scientists to build carbon emission simulation and prediction models on cloud platforms, improving computational efficiency and real-time monitoring capabilities.
[0004] Deep learning technology has made significant progress in carbon emission monitoring and environmental science. First, deep learning models can better process high-resolution remote sensing images, providing more detailed and accurate surface information. This is crucial for monitoring small-scale carbon emission sources. Second, deep learning models can better process high-resolution remote sensing images, providing more detailed and accurate surface information. This is crucial for monitoring small-scale carbon emission sources and land use change. Finally, deep learning models can be used to process time-series data, including time-series remote sensing data and meteorological data. This allows for better capture and analysis of seasonal and annual variations in carbon emissions. Satellites such as OCO-2 have begun using deep learning technology to process large-scale carbon emission data, demonstrating the enormous potential of deep learning technology in carbon emission monitoring and environmental science. However, existing deep learning-based methods in carbon emission monitoring still face issues of data consistency and standardization. First, information loss and distortion occur when inconsistent data types from different sources are fused into a unified data representation, potentially leading to information loss and distortion. Some data types may contain specific information, but this may not be fully preserved during the fusion process, thus reducing data quality and usability. Second, data inconsistency can lead to biases. If the data types of different data sources are inconsistent, the model may be more influenced by certain data sources during data fusion, while paying insufficient attention to other data sources. This can cause bias, making the model favor information from certain data sources and neglect the importance of other data sources, thus affecting the accuracy of the results. Third, it can be difficult to establish data relationships. Consistent data types may make it difficult to establish effective correlations and connections between data. This means that when fusing multi-source data, the model may fail to capture the potential relationships between different data sources, thus limiting a comprehensive understanding of the system or phenomenon. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a ground-based carbon emission monitoring method based on a multi-source data interaction network.
[0006] The objective of this invention is achieved as follows: a ground-based carbon emission monitoring method based on a multi-source data interaction network, comprising the following steps:
[0007] S1. Design of a multi-source data interaction network: The multi-source data interaction network includes a modular backbone network, an attention interaction module (AIM), and a cross-scale interaction module (CIM).
[0008] S2. Training the multi-source data interaction network: Perform data analysis and comparison on satellite data and auxiliary data; perform data augmentation processing to transform these numerical data into tensors with image shapes, and then input them into the multi-source data interaction network model for training;
[0009] S3. Perform carbon emission trend prediction: Use a trained multi-source data interaction network model for prediction, input satellite carbon remote sensing data and auxiliary data, and use optimized weights for encoding and decoding operations to finally obtain a high-level carbon emission trend map of a certain time and place.
[0010] Furthermore, the attention interaction module utilizes features from multi-source data to construct a shared global attention label sequence, which is then fed back to both input parties to strengthen their attention distribution. This interacts with key features in satellite and auxiliary data to generate a probability-weighted map of all pixels, guiding the model to assign high-probability weights to pixels in the carbon emission region. This module adopts a multi-source data fusion approach based on attention sharing, shifting the model's learning focus from the representational attributes of multi-source data to the attention distribution of multi-source data features. This reduces the interference of specific information contained in different data types on the fused data, minimizing information loss and distortion caused by data inconsistency.
[0011] When two features at the same scale interact, satellite data and auxiliary data first undergo channel transformation through three convolutional layers. The input features are then mapped to the feature spaces of query (Q), key (K), and value (V) for subsequent self-attention computation. Specifically, the three features after channel mapping are... and The compression factor r of the first two channels remains consistent; each pixel of the triple feature is used as a label, and then flattened into a query sequence, key sequence, and value sequence of satellite data or auxiliary data that the transformer layer can understand, which are respectively Where H represents the vertical pixel count of the satellite data or auxiliary data feature map, W represents the horizontal pixel count of the satellite data or auxiliary data feature map, C represents the number of channels of the satellite data or auxiliary data feature map, and H×W represents the total pixel count obtained after the vertical and horizontal pixels are transformed into a coherent label sequence in the transformer layer; the triple sequence can be represented as:
[0012] S q =Trans(Conv) 1×1 (Q))
[0013] S k =Trans(Conv) 1×1 (K))
[0014] S v=Trans(Conv) 1×1 (V))
[0015] Where Trans(·) represents transforming the matrix into a sequence of labels, Conv 1×1 (·) indicates a convolutional layer;
[0016] The input multi-source data features are concatenated by channel, and adaptive average pooling is used to compress the feature size to 1×1. Channel compression is also used to obtain globally adaptive carbon channel attention. The carbon channel feature matrix of a single pixel is used as a single label and replicated to make the total number of pixels in the label sequence reach H×W. This sequence is the attention interaction sequence. Represented as:
[0017] S i =Copy(Trans(Conv) 1×1 (Pool(I1+I2))))
[0018] Where Copy(·) represents cell reuse, Conv 1×1 (·) represents a 1×1 convolution, and Pool(·) represents a pooling layer;
[0019] S i Same as S k The sequence will also be similar to S after transposition. q The similarity between the two types of data is obtained through matrix multiplication. The two similarities are then fused at the pixel level to obtain the attention interaction matrix. This matrix is then normalized using the softmax function to obtain the attention weights. Finally, matrix multiplication is performed on the sum of these weights, and the matrix is transformed back to the input feature size. The attention interaction is calculated using the following dot product formula:
[0020] AttInteraction(S q S k S v S I ) = Softmax(S q S k T +S q S I T )S v
[0021] Finally, residual connections are used before output to prevent feature distortion.
[0022] Furthermore, the cross-scale interaction module utilizes existing cross-scale features, including high-resolution low-channel low-order data feature maps and low-resolution high-channel high-order data feature maps, to extract semantically rich carbon channel data from the high-order data feature maps. It then guides the low-order data according to probability distribution, thereby fully communicating carbon channel features at various levels and enhancing the perception of semantic information about carbon emissions. This module strengthens the correlation between homogeneous multi-scale features with lower computational cost, which helps the model capture potential relationships between homogeneous data at different scales, uncover subtle carbon emission changes in high-resolution carbon remote sensing images, and improve the accuracy of carbon emission detection.
[0023] The input high-order data feature map guides the low-order data feature map in two ways: First, keeping the high-order carbon features unchanged, the low-order carbon features are reduced in size by channel reduction and pooling compression to the same size as the high-order carbon features. Then, the low-order carbon features are weighted using a weight matrix obtained by the Sigmoid activation function, and finally, the size is restored by upsampling. The formula is expressed as:
[0024] M1 = Upsample(Pool(Conv) 3×3 (I1)⊙σ(Conv 3×3 (I2)))
[0025] Second, the low-order carbon features are kept unchanged, while the high-order carbon features are upsampled to a size consistent with the low-order carbon features. Then, a weighted matrix is obtained using the Sigmoid activation function to weight the low-order carbon features. The formula is expressed as:
[0026] M2 = Conv 3×3 (I1)⊙σ(Upsample(Conv 3×3 (I2)))
[0027] Where I1 and I2 represent low-order carbon features and high-order carbon features, Conv 3×3 (·) represents a 3×3 convolution operation, Pool(·) represents a pooling layer, Upsample(·) represents upsampling, σ(·) represents the Sigmoid activation function, and ⊙ represents pixel-level multiplication; the outputs of the last two methods are added together and then output.
[0028] Beneficial effects of the present invention: The present invention provides a ground carbon emission monitoring method based on a multi-source data interaction network, comprising the following steps: S1, designing a multi-source data interaction network: the multi-source data interaction network includes a modular backbone network, an attention interaction module (AIM), and a cross-scale interaction module (CIM);
[0029] S2. Training the multi-source data interaction network: Analyze and compare satellite data and auxiliary data; perform data augmentation processing, converting these numerical data into tensors with image shapes, and then inputting them into the multi-source data interaction network model for training; S3. Predicting carbon emission trends: Use the trained multi-source data interaction network model for prediction, inputting satellite carbon remote sensing data and auxiliary data, and performing encoding and decoding operations using optimized weights to finally obtain a high-level carbon emission trend map for a specific time and location; Through the above steps, this invention proposes two types of interaction methods: same-scale and cross-scale; same-scale interaction includes an attention interaction module, pixel-level multiplication, and pixel-level addition, with attention... The interaction module utilizes features from multi-source data to construct a common global attention label sequence, which feeds back to the multi-source data features to strengthen their attention distribution; pixel-level multiplication can emphasize the interaction between multi-source data; pixel-level addition can help the model capture common information and shared features among multi-source data; cross-scale interaction includes a cross-scale interaction module and a channel-cascaded decoder. The former can use the attention of high-order carbon features to guide low-order carbon features, improving their attention intensity; the latter can balance the fusion and upsampling of interaction features; the overall architecture implemented in this invention can be composed of a fully convolutional neural network consisting of an encoder and a decoder, which improves the generalization ability of the model while improving accuracy. Attached Figure Description
[0030] Figure 1 This is a structural diagram of the multi-source data interaction network based on the present invention;
[0031] Figure 2 This is a schematic diagram of the attention interaction module of the present invention;
[0032] Figure 3 This is a schematic diagram of the cross-scale interaction module of the present invention. Detailed Implementation
[0033] The present invention will now be further described with reference to the accompanying drawings.
[0034] To better understand the technical solution of the present invention, detailed descriptions are provided below through specific embodiments.
[0035] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.
[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0037] like Figure 1 As shown, the present invention provides a ground-based carbon emission monitoring method based on a multi-source data interaction network, which includes the following steps:
[0038] S1. Design of a multi-source data interaction network: The multi-source data interaction network includes a modular backbone network, an attention interaction module (AIM), and a cross-scale interaction module (CIM).
[0039] S2. Training the multi-source data interaction network: Perform data analysis and comparison on satellite data and auxiliary data to clean the data and improve its quality; perform data augmentation processing to transform these numerical data into tensors with image shapes, and then input them into the multi-source data interaction network model for training.
[0040] S3. Perform carbon emission trend prediction: Use a trained multi-source data interaction network model for prediction, input satellite carbon remote sensing data and auxiliary data, and use optimized weights for encoding and decoding operations to finally obtain a high-level carbon emission trend map of a certain time and place.
[0041] This invention constructs a multi-source data interaction network. In the encoding stage, a modular hierarchical backbone network is adopted to initially extract multi-scale features from the multi-source data. Then, an attention interaction module is used to connect key features of the multi-source data, thereby constructing a shared global attention to guide the attention distribution among them. In the decoding stage, the cross-scale interaction module uses the attention of higher-order features to guide lower-order features, enhancing their attention intensity. The parallel combination of pixel-level multiplication and pixel-level addition helps the model capture common information and shared features among the multi-source data while emphasizing the interaction between them. The complete architecture of the model is as follows: Figure 1 As shown.
[0042] Furthermore, such as Figure 2 As shown, the attention interaction module utilizes features from multi-source data to construct a common global attention label sequence, which is then fed back to both input parties to strengthen their attention distribution. This interacts with key features in satellite and auxiliary data to generate a probability-weighted map of all pixels, guiding the model to assign high-probability weights to pixels in the carbon emission region. This module adopts a multi-source data fusion approach based on attention sharing, shifting the model's learning focus from the representational attributes of multi-source data to the attention distribution of multi-source data features. This reduces the interference of specific information contained in different data types on the fused data, minimizing information loss and distortion caused by data inconsistency.
[0043] When two features at the same scale interact, satellite data and auxiliary data first undergo channel transformation through three convolutional layers. The input features are then mapped to the feature spaces of query (Q), key (K), and value (V) for subsequent self-attention computation. Specifically, the three features after channel mapping are... and To conform to transformer computation, the compression factor *r* of the first two channels remains consistent. This operation also facilitates the fusion of multi-channel carbon remote sensing data, thereby improving the model's global perception while reducing computational load. Each pixel of the triple feature is used as a label, and then flattened into query sequences, key sequences, and value sequences of satellite or auxiliary data that the transformer layer can understand. Where H represents the vertical pixel count of the satellite data or auxiliary data feature map, W represents the horizontal pixel count of the satellite data or auxiliary data feature map, C represents the number of channels of the satellite data or auxiliary data feature map, and H×W represents the total pixel count obtained after the vertical and horizontal pixels are transformed into a coherent label sequence in the transformer layer; the triple sequence can be represented as:
[0044] S q =Trans(Conv) 1×1 (Q))
[0045] S k =Trans(Conv) 1×1 (K))
[0046] S v =Trans(Conv) 1×1 (V))
[0047] Where Trans(·) represents transforming the matrix into a sequence of labels, Conv 1×1 (·) indicates a convolutional layer;
[0048] To enhance the interaction between the transformer layers of satellite data and auxiliary data, the input multi-source data features are concatenated by channel, and adaptive average pooling is used to compress the feature size to 1×1. Channel compression is also employed to obtain globally adaptable carbon channel attention. To achieve information mapping capabilities similar to the key sequence, the carbon channel feature matrix of a single pixel is used as a single label and replicated to make the total number of pixels in the label sequence reach H×W. This sequence is the attention interaction sequence. Represented as:
[0049] S i =Copy(Trans(Conv) 1×1 (Pool(I1+I2))))
[0050] Where Copy(·) represents cell reuse, Conv 1×1 (·) represents a 1×1 convolution, and Pool(·) represents a pooling layer;
[0051] S i Same as S k The sequence will also be similar to S after transposition. q The similarity between the two types of data is obtained through matrix multiplication. The two similarities are then fused at the pixel level to obtain the attention interaction matrix. This matrix is then normalized using the softmax function to obtain the attention weights. Finally, matrix multiplication is performed on the sum of these weights, and the matrix is transformed back to the input feature size. The attention interaction is calculated using the following dot product formula:
[0052] AttInteraction(S q S k S v S I ) = Softmax(S q S k T +S q S I T )S v
[0053] Finally, residual connections are used before output to prevent feature distortion. Functionally, the attention interaction module fully learns the global information of each source data and effectively guides satellite and auxiliary data for complementary learning through attention interaction. Simultaneously, it mines the attention distribution of carbon emission-related features, suppresses irrelevant interference, and more accurately captures target information in the monitoring area. This interactive modeling avoids the model being biased towards information from certain data sources, balancing the influencing factors among data sources to a certain extent. This is crucial for improving the accuracy and robustness of ground-based carbon emission monitoring.
[0054] Furthermore, the cross-scale interaction module utilizes existing cross-scale features, including high-resolution low-channel low-order data feature maps and low-resolution high-channel high-order data feature maps, to extract semantically rich carbon channel data from the high-order data feature maps. It then guides the low-order data according to probability distribution, thereby fully communicating carbon channel features at various levels and enhancing the perception of semantic information about carbon emissions. This module strengthens the correlation between homogeneous multi-scale features with lower computational cost, which helps the model capture potential relationships between homogeneous data at different scales, uncover subtle carbon emission changes in high-resolution carbon remote sensing images, and improve the accuracy of carbon emission detection.
[0055] The input high-order data feature map guides the low-order data feature map in two ways: First, keeping the high-order carbon features unchanged, the low-order carbon features are reduced in size by channel reduction and pooling compression to the same size as the high-order carbon features. Then, the low-order carbon features are weighted using a weight matrix obtained by the Sigmoid activation function, and finally, the size is restored by upsampling. The formula is expressed as:
[0056] M1 = Upsample(Pool(Conv) 3×3 (I1)⊙σ(Conv 3×3 (I2)))
[0057] Second, the low-order carbon features are kept unchanged, while the high-order carbon features are upsampled to a size consistent with the low-order carbon features. Then, a weighted matrix is obtained using the Sigmoid activation function to weight the low-order carbon features. The formula is expressed as:
[0058] M2 = Conv 3×3 (I1)⊙σ(Upsample(Conv 3×3 (I2)))
[0059] Where I1 and I2 represent low-order carbon features and high-order carbon features, Conv 3×3 (·) represents a 3×3 convolution operation, Pool(·) represents a pooling layer, Upsample(·) represents upsampling, σ(·) represents the Sigmoid activation function, and ⊙ represents pixel-level multiplication;
[0060] The outputs from the last two methods are then added and fused. Through this operation, the cross-scale interaction module guides high-order carbon feature information into low-order carbon feature maps, enhancing the semantic expressiveness of low-order carbon features. This allows low-order carbon features to retain richer semantic expression while retaining less channel information. This probability distribution-based interactive modeling reduces data bias caused by the fusion of features at different scales. When processing high-resolution carbon remote sensing images, it helps the model combine contextual information to enhance its global understanding of data features, thereby capturing subtle changes in carbon emissions.
[0061] In summary, this invention provides a ground-based carbon emission monitoring method based on a multi-source data interaction network, comprising the following steps: S1, designing a multi-source data interaction network: the multi-source data interaction network includes a modular backbone network, an attention interaction module (AIM), and a cross-scale interaction module (CIM); S2, training the multi-source data interaction network: performing data analysis and comparison on satellite data and auxiliary data; performing data augmentation processing to convert these numerical data into tensors with image shapes, and then inputting them into the multi-source data interaction network model for training; S3, predicting carbon emission trends: using the trained multi-source data interaction network model for prediction, inputting satellite carbon remote sensing data and auxiliary data, and performing encoding and decoding operations using optimized weights to finally obtain a high-level carbon emission trend map of a certain time and place; through the above steps, this invention... Two types of interaction methods are proposed: same-scale and cross-scale. Same-scale interaction includes an attention interaction module, pixel-level multiplication, and pixel-level addition. The attention interaction module uses features from multi-source data to construct a common global attention label sequence to feed back to the multi-source data features to strengthen their attention distribution. Pixel-level multiplication can emphasize the interaction between multi-source data. Pixel-level addition can help the model capture common information and shared features among multi-source data. Cross-scale interaction includes a cross-scale interaction module and a channel-cascaded decoder. The former can use the attention of high-order carbon features to guide low-order carbon features and improve their attention intensity. The latter can balance the fusion and upsampling of interaction features. The overall architecture of this invention can be constructed by a fully convolutional neural network consisting of an encoder and a decoder. While improving accuracy, the generalization ability of the model is also improved.
[0062] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0063] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.
Claims
1. A ground-based carbon emission monitoring method based on a multi-source data interaction network, characterized in that, Includes the following steps: S1. Design of a multi-source data interaction network: The multi-source data interaction network includes a modular backbone network, an attention interaction module (AIM), and a cross-scale interaction module (CIM). S2. Training a multi-source data interaction network: performing data analysis and comparison on satellite data and auxiliary data; Data augmentation is performed to transform these numerical data into tensors with image shapes, which are then fed into a multi-source data interaction network model for training. S3. Perform carbon emission trend prediction: Use a trained multi-source data interaction network model for prediction, input satellite carbon remote sensing data and auxiliary data, use optimized weights for encoding and decoding operations, and finally obtain a high-level carbon emission trend map of a certain time and place. The attention interaction module uses features from multi-source data to construct a common global attention label sequence to feed back to both inputs and outputs to strengthen their attention distribution. This allows for the interaction of key features in satellite data and auxiliary data to generate a probability weight map of all pixels, which then guides the model to assign high probability weights to pixels in the carbon emission area. When two features at the same scale interact, satellite data and auxiliary data first undergo channel transformation through three convolutional layers. The input features are then mapped to the feature space of query Q, key K, and value V for subsequent self-attention computation. Specifically, the three features after channel mapping are... , and The compression factor r of the first two channels remains consistent; each pixel of the triple feature is used as a label, and then flattened into a query sequence, key sequence, and value sequence of satellite data or auxiliary data that the transformer layer can understand, which are respectively Where H represents the vertical pixel count of the satellite data or auxiliary data feature map, W represents the horizontal pixel count of the satellite data or auxiliary data feature map, C represents the number of channels of the satellite data or auxiliary data feature map, and H×W represents the total pixel count obtained after the vertical and horizontal pixels are transformed into a coherent label sequence in the transformer layer; the triple sequence can be represented as: in This means transforming the matrix into a sequence of tags. Indicates a convolutional layer; The input multi-source data features are concatenated by channel, and adaptive average pooling is used to compress the feature size to 1×1, and channels are compressed to obtain globally adaptive carbon channel attention; The carbon channel feature matrix of a single pixel is used as a single label and replicated itself, making the total number of pixels in the label sequence reach H×W. This sequence is the attention interaction sequence. , is represented as: in, Indicates unit reuse, Represents a 1×1 convolution. Indicates the pooling layer; same Similar sequences will also be the same after transposition. The similarity between the two types of data is obtained through matrix multiplication. The two similarities are then fused at the pixel level to obtain the attention interaction matrix. This matrix is then normalized using the softmax function to obtain the attention weights. Finally, matrix multiplication is performed on the sum of these weights, and the matrix is transformed back to the input feature size. The attention interaction is calculated using the following dot product formula: Finally, residual connections are used before output to prevent feature distortion. The cross-scale interaction module utilizes existing cross-scale features, including high-resolution low-channel low-order data feature maps and low-resolution high-channel high-order data feature maps, to extract semantically rich carbon channel data from the high-order data feature maps, and guides the low-order data according to the probability distribution, thereby fully communicating the carbon channel features at each level and improving the perception of semantic information about carbon emissions. The input high-order data feature map guides the low-order data feature map in two ways: First, keeping the high-order carbon features unchanged, the low-order carbon features are reduced in size by channel reduction and pooling compression to the same size as the high-order carbon features. Then, the low-order carbon features are weighted using a weight matrix obtained by the Sigmoid activation function, and finally, the size is restored by upsampling. The formula is expressed as: Second, the low-order carbon features are kept unchanged, while the high-order carbon features are upsampled to a size consistent with the low-order carbon features. Then, a weighted matrix is obtained using the Sigmoid activation function to weight the low-order carbon features. The formula is expressed as: in, and Representing low-order carbon characteristics and high-order carbon characteristics, This represents a 3×3 convolution operation. Indicates the pooling layer. Indicates upsampling, This represents the Sigmoid activation function. This represents pixel-level multiplication; the outputs of the last two methods are added together and then merged.
