Satellite image time sequence carbon emission prediction method, equipment and medium

Abstract

The invention discloses a satellite image time sequence carbon emission prediction method, equipment and a medium, and belongs to the technical field of remote sensing image intelligent analysis and carbon emission estimation. The method comprises the following steps: acquiring a continuous multispectral satellite image sequence and denoising to output a cleaned image sequence; performing semantic segmentation processing on the cleaned image sequence to generate a probability graph of a plurality of surface feature categories; based on the output of the semantic segmentation processing, calculating the area of each type of surface features, and estimating a carbon emission sequence in combination with a preset emission coefficient and a molar mass conversion factor from carbon to carbon dioxide; performing fine tuning processing on the estimated carbon emission sequence to absorb imaging errors or emission coefficient errors, and outputting an adjusted carbon emission sequence; based on the historical carbon emission sequence, a time sequence prediction model is used for processing so as to predict future carbon emission; and outputting a carbon emission prediction result. According to the method, the accuracy and application universality of carbon emission prediction can be improved to a certain extent.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent analysis of remote sensing images and estimation of carbon emissions, and in particular to a satellite image time sequence carbon emission prediction method, device and medium. BACKGROUND

[0002] At present, carbon emission monitoring mainly relies on ground statistical list, fixed site sensor or airborne section survey. These methods generally have the following problems: insufficient spatial coverage - it is difficult to balance the multi-scale fine monitoring of "national - regional - urban - point source"; poor timeliness - data publication often lags more than one year, which is not conducive to real-time verification (MRV); high cost and poor scalability; although pure satellite remote sensing has the advantage of wide range of observation, traditional CNN, decision tree, mass balance and other algorithms are difficult to maintain segmentation granularity and emission estimation accuracy at the same time when processing high-dimensional time sequence + multi-scale + noise interference images, especially in the urban edge area where buildings, roads and vegetation are interlaced. For long sequence time prediction tasks, traditional LSTM and TFT have the problems of difficulty in processing long sequence and insufficient prediction accuracy.

[0003] Therefore, how to improve the accuracy and application universality of carbon emission prediction has become a technical problem to be solved. SUMMARY

[0004] The satellite image time sequence carbon emission prediction method, device and medium provided by the embodiments of the present application solve the technical problem of how to improve the accuracy and application universality of carbon emission prediction.

[0005] In a first aspect, the embodiments of the present application provide a satellite image time sequence carbon emission prediction method, characterized in that the method comprises: acquiring a continuous multi-spectral satellite image sequence of a target area, wherein the sequence contains image data at multiple time points, and the image data includes spatial dimension and spectral channel information; performing denoising processing on the satellite image sequence to remove noise interference in the imaging process, and outputting a cleaned image sequence; performing semantic segmentation processing on the cleaned image sequence to generate probability maps of multiple ground object categories, the ground object categories including buildings, roads, bare land, vegetation and water bodies; based on the output of the semantic segmentation processing, calculating the areas of each ground object category, and combining a preset emission coefficient and a carbon to carbon dioxide molar mass conversion factor to estimate a carbon emission amount sequence; performing fine tuning processing on the estimated carbon emission amount sequence to absorb imaging errors or emission coefficient errors, and outputting an adjusted carbon emission amount sequence; using a time sequence prediction model to process based on a historical carbon emission amount sequence to predict future carbon emission amount; and outputting a carbon emission prediction result, including a rasterized emission heat map, a regional total amount curve and a future prediction value.

[0006] In an implementation manner of the present application, the continuous multispectral satellite image sequence of the target area is obtained, specifically comprising: determining the geographical range and time period of the target area, wherein the target area includes city, regional or point source multi-scale monitoring area; obtaining continuous multispectral image data of the target area based on a satellite remote sensing platform, wherein the image data contains spatial dimensions and spectral channels; timestamp labeling the obtained image data to form a time sequence input to obtain the continuous multispectral satellite image sequence.

[0007] In an implementation manner of the present application, the satellite image sequence is subjected to denoising processing to remove noise interference in the imaging process, and the cleaned image sequence is output, specifically comprising: inputting the satellite image sequence into a Kalman filtering algorithm, wherein the Kalman filtering algorithm includes two stages of state prediction and state update; in the state prediction stage, the current image state is predicted based on a preset state transition matrix to generate a predicted state vector; in the state update stage, the predicted state vector is corrected based on an observation matrix and an actual image vector to generate a posterior state vector; and the posterior state vector is output as the cleaned image sequence.

[0008] In an implementation manner of the present application, the cleaned image sequence is subjected to semantic segmentation processing to generate probability maps of multiple ground object categories, specifically comprising: inputting the cleaned image sequence into a Mask2Former-Lite segmentation network, wherein the segmentation network includes a feature encoder and a segmentation decoder; in the feature encoder, multi-scale spatial features of the image are extracted using a MixVisionTransformer; in the segmentation decoder, a class query vector is generated by dynamically matching feature maps based on a multi-head self-attention mechanism; and a sigmoid activation function is applied to process the class query vector to output a ground object category probability map for each pixel.

[0009] In an implementation manner of the present application, based on the output of the semantic segmentation processing, the areas of various ground objects are calculated, and the preliminary carbon element emission amount is converted into a carbon emission sequence by combining a preset emission coefficient and a molar mass conversion factor from carbon to carbon dioxide, specifically comprising: based on the probability map of the semantic segmentation processing, an area vector of various ground objects is calculated, wherein the area calculation includes summing the probability of each pixel and multiplying by the actual area of a single pixel; the area vector is converted into a preliminary carbon element emission amount by combining a preset emission coefficient vector; and the preliminary carbon element emission amount is converted into a carbon emission sequence by applying a molar mass conversion factor from carbon to carbon dioxide.

[0010] In an implementation form of the present application, fine-tuning is performed on the estimated carbon emission sequence to absorb imaging errors or emission coefficient errors, and an adjusted carbon emission sequence is output, specifically comprising: inputting the carbon emission sequence into a lightweight multi-layer perception regression module, wherein the regression module includes a linear fine-tuning layer; in the fine-tuning layer, applying weight and bias parameters to perform linear transformation on the carbon emission sequence; outputting the transformed carbon emission sequence as the adjusted carbon emission sequence.

[0011] In an implementation form of the present application, based on the historical carbon emission sequence, a time series prediction model is used for processing to predict future carbon emissions, specifically comprising: inputting the historical carbon emission sequence into a PatchTST model, wherein the PatchTST model includes sequence segmentation, linear projection and multi-head self-attention mechanism; in the sequence segmentation stage, the input sequence is cut into multiple local time windows; in the linear projection stage, a linear transformation is applied to each time window to generate an embedding vector; in the multi-head self-attention stage, the embedding vector is modeled for global and local patterns; outputting the carbon emission prediction value at the future time point.

[0012] In an implementation form of the present application, the carbon emission prediction result is output, specifically comprising: based on the adjusted carbon emission sequence, a rasterized emission heat map is generated, wherein the heat map resolution matches the original image spatial dimension; aggregating the carbon emission sequence data to generate a regional total amount time curve; based on the output of the time series prediction model, the carbon emission prediction value in the future period of time is generated; outputting the rasterized emission heat map, the regional total amount curve and the future prediction value to a preset visualization system.

[0013] In a second aspect, the embodiments of the present application also provide a satellite image time sequence carbon emission prediction device, characterized in that the device comprises: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: acquire a continuous multispectral satellite image sequence of a target region, wherein the sequence contains image data at multiple time points, and the image data includes spatial dimensions and spectral channel information; perform denoising processing on the satellite image sequence to remove noise interference in the imaging process, and output a cleaned image sequence; perform semantic segmentation processing on the cleaned image sequence to generate probability maps of multiple land cover categories, and the land cover categories include buildings, roads, bare land, vegetation and water bodies; based on the output of the semantic segmentation processing, calculate the areas of various land covers, and combine a preset emission coefficient and a carbon-to-carbon dioxide molar mass conversion factor to estimate a carbon emission sequence; perform fine-tuning processing on the estimated carbon emission sequence to absorb imaging errors or emission coefficient errors, and output an adjusted carbon emission sequence; based on a historical carbon emission sequence, use a time sequence prediction model for processing to predict future carbon emissions; and output a carbon emission prediction result, including a rasterized emission heat map, a regional total quantity curve and a future prediction value.

[0014] In a third aspect, the embodiments of the present application also provide a nonvolatile computer storage medium for satellite image time sequence carbon emission prediction, which stores computer executable instructions, characterized in that the computer executable instructions are configured to: acquire a continuous multispectral satellite image sequence of a target region, wherein the sequence contains image data at multiple time points, and the image data includes spatial dimensions and spectral channel information; perform denoising processing on the satellite image sequence to remove noise interference in the imaging process, and output a cleaned image sequence; perform semantic segmentation processing on the cleaned image sequence to generate probability maps of multiple land cover categories, and the land cover categories include buildings, roads, bare land, vegetation and water bodies; based on the output of the semantic segmentation processing, calculate the areas of various land covers, and combine a preset emission coefficient and a carbon-to-carbon dioxide molar mass conversion factor to estimate a carbon emission sequence; perform fine-tuning processing on the estimated carbon emission sequence to absorb imaging errors or emission coefficient errors, and output an adjusted carbon emission sequence; based on a historical carbon emission sequence, use a time sequence prediction model for processing to predict future carbon emissions; and output a carbon emission prediction result, including a rasterized emission heat map, a regional total quantity curve and a future prediction value.

[0015] The satellite image time sequence carbon emission prediction method, device and medium provided by the embodiments of the present application at least have the following technical effects: The accuracy of multi-scale carbon emission estimation and prediction is significantly improved by improving the end-to-end fusion Transformer segmentation model, the area-weighted carbon emission regression module, and the PatchTST prediction network. The semantic segmentation module uses a multi-head self-attention mechanism to capture the global spatial relationship of remote sensing images, effectively improving the fine-grained recognition ability in complex scenes such as building-vegetation interlaced areas, and reducing the misjudgment rate of roads; the area-weighted module explicitly embeds physical constraint rules to ensure the mass conservation of carbon elements to carbon dioxide, enhancing the credibility of emission level estimation; the regression fine-tuning module compensates for imaging or coefficient errors through a lightweight MLP, further improving the accuracy of annual cumulative emission estimation.

[0016] The time series prediction mechanism based on PatchTST breaks through the long-term memory decay bottleneck of traditional LSTM / TFT models. By cutting the historical emission sequence into local time windows and applying the Transformer multi-head self-attention mechanism, the ability to capture long-term sequence dependencies is significantly improved, effectively reducing the mean square error of predictions over 72 hours. At the same time, the AWGT mechanism provides traceable intermediate results (feature area vectors, emission factor lists), realizing the full-chain traceable mapping from "pixel → area → carbon → carbon dioxide", enhancing the explainability of the prediction results, and meeting the audit needs of regulatory agencies for emission source tracing.

[0017] The lightweight network architecture combined with structure optimization strategies enables the system to achieve sub-second real-time inference on single-GPU or edge computing platforms, supporting continuous monitoring of urban areas at all times. The modular design supports cross-regional migration: by adjusting the emission coefficient vector and fine-tuning layer parameters, different spectral characteristics, feature distribution, and imaging frequencies of satellite data can be adapted without retraining, significantly enhancing the engineering practicality of carbon monitoring in multiple geographic regions. BRIEF DESCRIPTION OF DRAWINGS

[0018] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions serve to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A satellite image time series carbon emission prediction method flowchart is provided for the embodiments of the present application; Figure 2 A satellite image time series carbon emission prediction device internal structure schematic diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0019] In order to make the purposes, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below in combination with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0020] The embodiments of the present application provide a satellite image time sequence carbon emission prediction method, device and medium, to solve the technical problems of how to improve the accuracy and application universality of carbon emission prediction.

[0021] The technical solutions provided by the embodiments of the present application will be described in detail below with reference to the drawings.

[0022] Figure 1 A satellite image time sequence carbon emission prediction flowchart is provided for the embodiments of the present application. As shown in the figure, the satellite image time sequence carbon emission prediction method provided by the embodiments of the present application specifically includes the following steps: Figure 1 Step 1, acquiring a continuous multispectral satellite image sequence of a target area, wherein the sequence contains image data at multiple time points, and the image data includes spatial dimension and spectral channel information.

[0023] This step aims to build an initial data source for carbon emission prediction. The "continuous multispectral satellite image sequence" refers to an image set formed by repeated observation of a target area by a satellite remote sensing platform within a fixed time period, and each image contains spatial dimension (height H x width W pixels) and spectral channel (λ bands, such as visible light, infrared bands, etc.).

[0024] Step 1.1, determining the geographical range and time period of the target area, wherein the target area includes city, regional or point source multi-scale monitoring area.

[0025] The target area refers to the geographical range that needs to be monitored and predicted for carbon emissions. The area usually has clear boundaries, such as a city administrative district, an industrial park, a river basin or a national park. The size of the target area needs to be determined according to the resolution of the satellite image and the monitoring requirements.

[0026] Step 1.2, acquiring continuous multispectral image data of the target area based on a satellite remote sensing platform, wherein the image data contains spatial dimension and spectral channel.

[0027] ​Select a satellite platform with multispectral imaging capability (e.g. Landsat series, Sentinel-2 series, etc.) as the data source. Determine the density of the time series (e.g. one per month, one per quarter) based on the monitoring target (e.g. monthly monitoring, quarterly monitoring). The time series should cover a sufficiently long historical period (e.g. the past 3-5 years) to support subsequent time series analysis and prediction. The time points refer to the specific dates when the satellite passes over and acquires the imagery.

[0028] Step 1.3, Timestamp annotation of the acquired image data to form a time series input to obtain a continuous multispectral satellite image sequence.

[0029] Download multispectral satellite imagery covering the target area at the selected time points through a satellite data distribution platform (e.g. USGS Earth Explorer, ESA Copernicus Open Access Hub, etc.). Each image data contains spatial dimension information (i.e. geographic coordinates and surface coverage corresponding to each pixel) and spectral channel information (i.e. spectral band data of multiple specific wavelength ranges, such as visible light, near-infrared, short-wave infrared, etc.).

[0030] Perform necessary preprocessing on the acquired raw images, including but not limited to: radiometric calibration (converts raw DN values to surface reflectance or radiance), atmospheric correction (eliminates the effects of atmospheric scattering and absorption), image registration (ensures accurate spatial alignment of images at different time points). This step aims to improve the data quality basis for subsequent processing.

[0031] In a specific example: a user needs to conduct quarterly carbon emission monitoring for B industrial park located in A region. The target area is set as the administrative boundary range of the industrial park. Choose Sentinel-2 satellite as the data source due to its high spatial resolution (10 meters) and free access policy. Determine the time series density as one image per quarter (January, April, July, October), and the time range covers the past 4 years (e.g. January 2020 to October 2023). Through the Copernicus Open Access Hub platform, download all selected time point Sentinel-2 Level-1C products (containing apparent reflectance) covering B industrial park. After downloading, perform atmospheric correction (e.g. using Sen2Cor processor) to convert all images to surface reflectance data, and perform accurate image registration to ensure spatial alignment of all image pixels.

[0032] Step 2, Perform denoising processing on the satellite image sequence to remove noise interference during imaging, output the cleaned image sequence.

[0033] Step 2.1, input the satellite image sequence into the Kalman filter algorithm, wherein the Kalman filter algorithm comprises two stages of state prediction and state update.

[0034] The noise interference in the imaging process mainly includes: random noise (such as sensor thermal noise), strip noise (inconsistent response of sensor detection elements), cloud cover and its shadow shielding, thin fog influence, etc. These noises will reduce the image quality and affect the accuracy of subsequent ground object identification.

[0035] Step 2.2, in the state prediction stage, the current image state is predicted based on the preset state transition matrix to generate a predicted state vector.

[0036] Step 2.2.1, spatial filtering: for random noise and slight strip noise, spatial filtering methods such as mean filtering, median filtering or Gaussian filtering can be used. These methods use the statistical information in the neighborhood of the pixel to smooth the noise.

[0037] Step 2.2.2, time domain filtering: for time series data, the information of images at different time points can be used for denoising. For example, time domain median filtering or mean filtering of spectral values of the same pixel at different time points can effectively suppress transient noise (such as cloud shadow). For cloud cover, a cloud detection algorithm (such as using short-wave infrared or a specific cloud index) can be used to identify pixels contaminated by clouds, and the values of cloud-free images at adjacent time points (such as the previous and next days) can be used for replacement or interpolation.

[0038] Step 2.2.3, frequency domain filtering: for periodic strip noise, the image can be converted to the frequency domain by Fourier transform, and the frequency components corresponding to the strip noise can be filtered out in the frequency domain, and then converted back to the spatial domain Step 2.3, in the state update stage, the predicted state vector is corrected based on the observation matrix and the actual image vector to generate a posterior state vector.

[0039] According to the identified noise type and image characteristics, one or a combination of multiple denoising methods is selected for processing. The processed image sequence should significantly reduce noise interference, with clear ground object boundaries and reliable spectral information. The cleaned image sequence is output for the next step.

[0040] Step 2.4, output the posterior state vector as the cleaned image sequence.

[0041] In a specific example, the Sentinel-2 image sequence of Industrial Park B acquired in Step 1 is denoised. First, pixels covered by clouds and cloud shadows are identified using the image's QA60 band (quality assurance band) or cloud detection algorithms based on shortwave infrared bands (e.g., B11 and B12) (e.g., s2cloudless). For these pixels, a temporal median filter is applied: the cloud-free image values ​​for the pixel at adjacent time points (e.g., the previous and next quarters) are found and the median value is used to replace the image. In areas not covered by clouds, slight banding noise (manifested as light and dark stripes along the track) is detected and suppressed using frequency-domain filtering. The final output is a cleaned image sequence that removes significant cloud cover, cloud shadows, and banding noise.

[0042] Step 3: Perform semantic segmentation processing on the cleaned image sequence to generate probability maps of multiple object categories, including buildings, roads, bare land, vegetation, and water bodies.

[0043] Step 3.1: Input the cleaned image sequence into the Mask2Former-Lite segmentation network, where the segmentation network includes a feature encoder and a segmentation decoder.

[0044] Identify the specific land feature categories to be identified. In this implementation, core categories include: buildings (including factories, residences, and commercial buildings), roads (including highways, main roads, secondary roads, and dirt roads), bare land (uncovered soil, sand, rock, and construction areas), vegetation (including forests, grasslands, farmland, and parks), and water bodies (including rivers, lakes, reservoirs, and ponds). Additional categories can be added or refined based on the characteristics of the target area (e.g., distinguishing between farmland and forestland).

[0045] Step 3.2: In the feature encoder, use MixVisionTransformer to extract the multi-scale spatial features of the image.

[0046] Use deep learning-based semantic segmentation models, such as U-Net, DeepLabv3+, and PSPNet. These models learn spectral, textural, and contextual features of images to classify each pixel. If using a supervised learning model, a training dataset is required. This dataset contains cleaned image samples and their corresponding pixel-level object category labels (GroundTruth). These labels must be accurately annotated manually or semi-automatically. Use the training dataset to train the selected semantic segmentation model and optimize the model parameters to accurately identify various ground objects.

[0047] Step 3.3: In the segmentation decoder, the feature maps are dynamically matched based on the multi-head self-attention mechanism to generate the category query vector.

[0048] Each scene image in the cleaned image sequence is input into the trained (or pre-trained) semantic segmentation model. The model processes each pixel of the input image and outputs a probability value (usually between 0 and 1) for each predefined land cover class that the pixel belongs to.

[0049] Step 3.4, apply the sigmoid activation function to process the class query vector, output the land cover class probability map of each pixel.

[0050] For each land cover class, the model outputs a probability map (ProbabilityMap) corresponding to the class. The probability map has the same spatial dimensions as the input image, and each pixel value in the map represents the probability of the location belonging to the class. For example, if a pixel value in the "building probability map" is 0.95, it means that the pixel has a 95% chance of belonging to a building.

[0051] In a specific example: for the carbon emission monitoring needs of B industrial park, the land cover classes are defined as: buildings, roads (including hardened surfaces), bare land (including construction sites and unhardened ground), vegetation (including green spaces and farmland), and water bodies. The U-Net model architecture is selected. Historical Sentinel-2 images and manually annotated label maps of A region (containing similar industrial area landscapes) are collected as the training set and validation set to train the U-Net model. The cleaned image sequence (a total of 16 scenes, one per quarter) of B industrial park output in step 2 is input into the trained U-Net model. The model processes each scene image and outputs five probability maps corresponding to the five classes of buildings, roads, bare land, vegetation, and water bodies. For example, for the October 2023 image, the model generates a probability map for each pixel in the B industrial park area belonging to the above five land cover classes on that date.

[0052] Step 4, based on the output of the semantic segmentation processing, calculate the area of each land cover class, and combine the pre-set emission coefficient and the molar mass conversion factor from carbon to carbon dioxide to estimate the carbon emission sequence.

[0053] Step 4.1, based on the probability map of semantic segmentation processing, calculate the area vector of each land cover class, where the area calculation includes summing the probabilities of each pixel and multiplying by the actual area of a single pixel.

[0054] Step 4.1.1, determine the dominant class: for each pixel, compare its values on all land cover class probability maps. Assign the pixel to the class with the highest probability value (i.e., the dominant class). For example, if a pixel has a value of 0.8 in the building probability map, 0.1 in the road, 0.05 in the bare land, 0.05 in the vegetation, and 0.0 in the water body, it is classified as a building.

[0055] Step 4.1.2, Generate Classification Map: Execute Step 4.1.1 for all pixels to generate a discrete Classification Map where each pixel belongs to only one determined land cover class.

[0056] Step 4.1.3, Count Pixels: On the Classification Map, count the total number of pixels belonging to each land cover class.

[0057] Step 4.1.4, Calculate Actual Area: Based on the spatial resolution of the satellite imagery (e.g., 10 meters for Sentinel-2), calculate the actual area represented by a single pixel (e.g., 10m x 10m = 100 square meters). Multiply the total number of pixels for each class by the area of a single pixel to obtain the actual covered area for that class (e.g., total area of buildings, total area of vegetation, etc.). Repeat this process for each image in the time series to obtain the area sequence for each land cover class at each time point.

[0058] Step 4.2, Convert Area Vector to Preliminary Carbon Emissions using Preset Emission Coefficient Vector.

[0059] Preset carbon emission (or absorption) coefficients for each land cover class. These coefficients are typically determined based on industry standards, scientific research, or local measurements. For example: Buildings: Annual carbon emission coefficient per unit area (considering building energy consumption, unit: tons of carbon / square meter / year).

[0060] Roads: Annual carbon emission coefficient per unit area (considering transportation emissions and road materials, unit: tons of carbon / square meter / year).

[0061] Bare land: Generally considered negligible or set a very small background value (or consider specific activities such as construction).

[0062] Vegetation: Annual carbon absorption coefficient per unit area (negative value, indicating carbon sink, unit: tons of carbon / square meter / year).

[0063] Water bodies: Generally considered negligible or set a very small background value.

[0064] Step 4.3, Convert Preliminary Carbon Element Emissions to Carbon Dioxide Equivalent Emissions using the Molar Mass Conversion Factor.

[0065] Carbon emission coefficients are usually given in terms of the mass of carbon (C). To obtain the more commonly used carbon dioxide equivalent emissions, the mass of carbon needs to be converted to the mass of carbon dioxide. The conversion factor is the ratio of the molecular weight of carbon dioxide to the atomic weight of carbon, which is 44 / 12 ≈ 3.67. This means that 1 ton of carbon is equivalent to 3.67 tons of carbon dioxide.

[0066] For each time point t in the time series: For each land cover type k (e.g. building, road), calculate its carbon emission (in carbon): Carbon emission Ck(t) = Area k(t) * Emission coefficient k.

[0067] Add up the carbon emissions Ck(t) of all land cover types to get the total carbon emission (in carbon) of the target area at this time point: Total carbon emission C(t) = Σ (Carbon emission Ck(t)).

[0068] Multiply the total carbon emission C(t) by the conversion factor 3.67 to get the total carbon dioxide equivalent emission CO2e(t) of the target area at this time point: Total carbon emission CO2e(t) = Total carbon emission C(t) * 3.67.

[0069] Repeat the above calculation for all time points in the time series to get the carbon emission sequence (usually in CO2e) of the target area.

[0070] In one specific example: Area calculation on the B Industrial Park 2023 Oct classification map generated in Step 3. Image resolution 10m, pixel area 100 sqm. Statistics: Nbuilding pixel number of building type, then building total area = Nbuilding * 100 sqm. Similarly calculate road, bare land, vegetation, water area. Preset emission coefficients (annual): building 0.05 tons C / sqm / year, road 0.02 tons C / sqm / year, bare land 0.001 tons C / sqm / year, vegetation -0.01 tons C / sqm / year (absorption), water 0.000 tons C / sqm / year. Since the image is quarterly data, the annual coefficient needs to be converted to quarterly coefficient (divide by 4). Calculate for 2023 Oct (representing this quarter): Building carbon emission C = (Building total area) * (0.05 / 4); Road carbon emission C = (Road total area) * (0.02 / 4); Bare land carbon emission C = (Bare land total area) * (0.001 / 4); Vegetation carbon absorption C = (Vegetation total area) * (-0.01 / 4) (negative value); Water carbon emission C = 0; Quarterly total carbon emission C = Building C + Road C + Bare land C + Vegetation C + Water C; Quarterly total carbon emission CO2e = Quarterly total carbon emission C * 3.67; Repeat this calculation process for each image of the past 4 years (16 quarters) to get the quarterly carbon emission (CO2e) sequence of B Industrial Park for the past 16 time points.

[0071] Step 5. Fine-tuning the estimated carbon emission sequence to absorb imaging errors or emission factor errors, output the adjusted carbon emission sequence.

[0072] Step 5.1. Input the carbon emission sequence into a lightweight multi-layer perceptron regression module, where the regression module includes a linear fine-tuning layer.

[0073] The estimated carbon emission sequence may have systematic biases, mainly from two aspects: Imaging errors: errors introduced in satellite image acquisition, preprocessing, denoising, segmentation, etc., ultimately leading to inaccurate estimation of feature area.

[0074] Emission factor errors: the pre-set emission factor is based on statistical or typical values, which may deviate from the actual emission intensity of the target area.

[0075] Step 5.2. In the fine-tuning layer, apply weight and bias parameters to linearly transform the carbon emission sequence.

[0076] Step 5.2.1. Correction based on reference data: If relatively reliable reference carbon emission data for part of the time points in the target area is available (such as emission inventory based on energy statistics, interpolation data from high-precision ground monitoring stations), the estimated sequence can be compared with the reference data. Calculate the difference (residual) between the two at the same time point, analyze the pattern of the residual (such as constant bias, proportional bias, time-varying bias). According to the residual pattern, establish a correction model (such as a linear regression model: adjustment value = a*estimated value + b), and use the reference data to fit the model parameters (a, b). Then apply the model to correct the entire estimated sequence.

[0077] Step 5.2.2. Smoothing / filtering based on time series characteristics: If there is a lack of reliable reference data, but it is believed that the noise (random error) in the estimated sequence is greater than the systematic error, or the systematic error changes slowly over time, time series smoothing or filtering techniques can be used to absorb high-frequency noise and small systematic drift. Common methods include moving average (Moving Average), exponential smoothing (Exponential Smoothing), or Kalman filter (Kalman Filter). These methods use the historical information of the sequence itself to estimate the current value, smoothing out unreasonable fluctuations.

[0078] Step 5.3. Output the transformed carbon emission sequence as the adjusted carbon emission sequence.

[0079] According to the data situation and error characteristics, select the method (or combination) in step 5.2.1 or 5.2.2 to process the original carbon emission sequence. The processed sequence should be closer to the true situation, reducing the influence of systematic bias and random noise. Output the adjusted carbon emission sequence for prediction.

[0080] In a specific case: for the quarterly carbon emission sequence of B industrial park (16 time points), it is found that the estimated value is generally higher than the annual emission data of the park released by the local environmental protection department based on the annual energy consumption reported by enterprises (assuming that 3 years of data are available). Analysis shows that the main error comes from the high emission coefficient (especially for buildings and roads). Use the method of step 5.2.1: distribute the annual reference data quarterly (assuming uniform emission within the year) to get the quarterly reference value. Compare the estimated sequence with the quarterly reference value at the corresponding time points, and find that there is an approximately linear relationship: estimated value ≈ 1.2*reference value + 500 tons CO2e (constant term). Use the available reference data points to fit the linear model parameters (a = 1.2, b = 500). Then, apply this correction model to the entire estimated sequence (including time points with reference data and no reference data), adjust all estimated values: adjusted value = (estimated value - 500) / 1.2. Output the adjusted quarterly carbon emission sequence.

[0081] Step 6, based on the historical carbon emission sequence, use time series prediction model processing to predict future carbon emissions.

[0082] Step 6.1, input the historical carbon emission sequence into the PatchTST model, where the PatchTST model includes sequence segmentation, linear projection and multi-head self-attention mechanism.

[0083] Step 6.2, in the sequence segmentation stage, cut the input sequence into multiple local time windows.

[0084] Step 6.3, in the linear projection stage, apply linear transformation to each time window to generate embedding vectors.

[0085] Step 6.4, in the multi-head self-attention stage, model global and local patterns for embedding vectors.

[0086] Step 6.5, output the carbon emission prediction value of future time points.

[0087] In one specific example: Adjust the carbon emission (CO2e) sequence of the past 16 quarters (4 years) of B Industrial Park using Step 5. The sequence shows clear annual seasonality (higher emissions in winter, lower in summer) and a slow upward trend. The Facebook Prophet model is chosen for prediction because it can better handle seasonality and trend changes. The first 14 quarters of data (2019Q4-2022Q3) are used as the training set, and the last 2 quarters (2022Q4, 2023Q1) are used as the validation set (actually 2023Q1, the validation set is used for tuning). Fit the Prophet model on the training set, and optimize the prediction effect on the validation set by adjusting the seasonality and trend parameters. After determining the optimal model, retrain the optimal Prophet model using all 16 quarters of data (2019Q4-2023Q3). Then, use this final model to predict carbon emissions for the next 4 quarters (2023Q4, 2024Q1, 2024Q2, 2024Q3).

[0088] Step 7, output carbon emission prediction results, including rasterized emission heat map, regional total curve and future prediction value.

[0089] Step 7.1, based on the adjusted carbon emission sequence, generate a rasterized emission heat map, where the heat map resolution matches the original image spatial dimensions.

[0090] Step 7.1.1, spatialize emission intensity: Carbon emissions mainly come from buildings and roads and other land features. Use the building probability map and road probability map (or classification map) generated in Step 3 at the most recent time point to approximate the current spatial distribution. Weighted fusion of building probability map and road probability map (for example, give higher weight to buildings), generate a spatial distribution map (intensity map) representing "potential emission source intensity".

[0091] Step 7.1.2, allocate regional total: allocate the regional total carbon emissions (CO2e) at a certain time point in the future predicted in Step 6 to each pixel according to the proportion of the intensity map generated in Step 7.1.1. For example, high-intensity pixels are allocated more emissions. In this way, each pixel has a predicted emission value (units such as tons CO2e / pixel / quarter).

[0092] Step 7.1.3, visualize heat map: map the predicted emission value at the pixel level to different colors according to the numerical value (such as low emission with blue / green, high emission with yellow / red), generate a visual rasterized emission heat map (Heatmap). The map clearly shows the spatial distribution of carbon emissions in the target area during the predicted period.

[0093] Step 7.2, aggregate carbon emission sequence data to generate a regional total time curve.

[0094] Step 7.2.1, Plotting Historical Curve: Plot the adjusted historical carbon emission sequence (CO2e) output from Step 5 in chronological order as a line or curve graph.

[0095] Step 7.2.2, Plotting Forecast Curve: Plot the future carbon emission sequence (CO2e) forecasted in Step 6, continuing from the historical curve. Usually, historical data and forecast data are distinguished by different line types (such as dashed lines) or colors.

[0096] Step 7.2.3, Adding Confidence Interval: If the prediction model provides uncertainty estimates for the prediction (such as the Prophet model will give prediction intervals), confidence intervals (such as 80% or 95% confidence intervals) can be plotted around the prediction curve to represent the credible range of the prediction.

[0097] Step 7.3, Based on the output of the time series prediction model, generate carbon emission prediction values for a future period of time.

[0098] Output the specific carbon emission values (CO2e) at each future time point predicted in Step 6 in a structured data form (such as a table, list), including point estimates (most likely values) and possible interval estimates.

[0099] Step 7.4, Output the rasterized emission heat map, regional total curve, and future prediction value to the preset visualization system.

[0100] Integrate the rasterized emission heat map, regional total curve, and future prediction value table together to form the final carbon emission prediction result report or visualization interface for users to view and analyze.

[0101] The above is the method embodiment of the present application. Based on the same inventive concept, the present application also provides a satellite image time series carbon emission prediction device, the structure of which is shown in Figure 2 .

[0102] Figure 2 A satellite image time series carbon emission prediction device provided by an embodiment of the present application is shown in Figure 2 . The device comprises: at least one processor 201; and a memory 202 in communication connection with the at least one processor; wherein the memory 202 stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor 201 to enable the at least one processor 201 to: Obtain a continuous multispectral satellite image sequence of a target area, wherein the sequence contains image data at multiple time points, and the image data includes spatial dimension and spectral channel information; perform denoising processing on the satellite image sequence to remove noise interference in the imaging process, and output a cleaned image sequence; perform semantic segmentation processing on the cleaned image sequence to generate probability maps of multiple ground object categories, including buildings, roads, bare land, vegetation, and water bodies; based on the output of the semantic segmentation processing, calculate the area of each ground object category, and combine a preset emission coefficient and a molar mass conversion factor of carbon to carbon dioxide to estimate a carbon emission sequence; perform fine-tuning processing on the estimated carbon emission sequence to absorb imaging errors or emission coefficient errors, and output an adjusted carbon emission sequence; based on the historical carbon emission sequence, use a time series prediction model for processing to predict future carbon emissions; and output carbon emission prediction results, including a rasterized emission heat map, a regional total quantity curve, and a future prediction value.

[0103] Some embodiments of the present application provide a non-volatile computer storage medium for satellite image time series carbon emission prediction corresponding to Figure 1 The computer executable instructions are set to: Obtain a continuous multispectral satellite image sequence of a target area, wherein the sequence contains image data at multiple time points, and the image data includes spatial dimension and spectral channel information; perform denoising processing on the satellite image sequence to remove noise interference in the imaging process, and output a cleaned image sequence; perform semantic segmentation processing on the cleaned image sequence to generate probability maps of multiple ground object categories, including buildings, roads, bare land, vegetation, and water bodies; based on the output of the semantic segmentation processing, calculate the area of each ground object category, and combine a preset emission coefficient and a molar mass conversion factor of carbon to carbon dioxide to estimate a carbon emission sequence; perform fine-tuning processing on the estimated carbon emission sequence to absorb imaging errors or emission coefficient errors, and output an adjusted carbon emission sequence; based on the historical carbon emission sequence, use a time series prediction model for processing to predict future carbon emissions; and output carbon emission prediction results, including a rasterized emission heat map, a regional total quantity curve, and a future prediction value.

[0104] Each of the embodiments in the present application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the IoT device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiments.

[0105] The system and medium provided by the embodiments of the present application are one-to-one corresponding, and therefore the system and medium also have similar beneficial technical effects to the method corresponding thereto. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the system and medium will not be described here again.

[0106] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. In addition, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0107] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0108] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0109] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0110] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memories.

[0111] Memory can include non-persistent memory, Random Access Memory (RAM), and / or non-volatile memory, such as Read Only Memory (ROM) or flash memory, in computer readable media. Memory is an example of computer readable media.

[0112] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include transitory media, such as modulated data signals and carrier waves.

[0113] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or apparatus that comprises a list of elements does not include only those elements in the list, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without further limitation, an element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0114] The above only is an embodiment of the present application, and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. A satellite image time series carbon emission prediction method, characterized in that: The method comprises: Acquire a continuous multispectral satellite image sequence of a target area, wherein the sequence includes image data at multiple time points, and the image data includes spatial dimension and spectral channel information; performing denoising processing on the satellite image sequence to remove noise interference in the imaging process and outputting a cleaned image sequence; performing semantic segmentation processing on the cleaned image sequence to generate probability maps of multiple object categories, wherein the object categories include buildings, roads, bare land, vegetation, and water bodies; Based on the output of the semantic segmentation process, the area of ​​each type of land feature is calculated, and the carbon emission series is estimated by combining a preset emission coefficient and a carbon to carbon dioxide molar mass conversion factor; Performing fine-tuning processing on the estimated carbon emission series to absorb imaging errors or emission coefficient errors, and outputting an adjusted carbon emission series; Based on the historical carbon emission series, a time series forecasting model is used to predict future carbon emissions; Output carbon emission prediction results, including rasterized emission heat map, regional total amount curve and future prediction values.

2. The satellite image time series carbon emission prediction method according to claim 1, characterized in that: Acquire a continuous multispectral satellite image sequence of the target area, including: Determine the geographic scope and time period of the target area, which may include a city, region, or point source multi-scale monitoring area; Acquire continuous multispectral image data of the target area based on satellite remote sensing platforms, where the image data includes spatial dimensions and spectral channels; The acquired image data are timestamped to form a time series input to obtain a continuous multispectral satellite image sequence.

3. The satellite image time series carbon emission prediction method according to claim 1, characterized in that: Performing denoising processing on the satellite image sequence to remove noise interference in the imaging process and outputting a cleaned image sequence, specifically including: The satellite image sequence is input into the Kalman filter algorithm, where the Kalman filter algorithm includes two stages: state prediction and state update; In the state prediction stage, the current image state is predicted based on the preset state transfer matrix to generate a predicted state vector; In the state update phase, the predicted state vector is corrected based on the observation matrix and the actual image vector to generate the posterior state vector; The output posterior state vector is used as the cleaned image sequence.

4. The satellite image time series carbon emission prediction method according to claim 1, characterized in that: Perform semantic segmentation on the cleaned image sequence to generate probability maps for multiple object categories, including: The cleaned image sequence is input into the Mask2Former-Lite segmentation network, where the segmentation network includes a feature encoder and a segmentation decoder; In the feature encoder, MixVisionTransformer is used to extract multi-scale spatial features of the image; In the segmentation decoder, the feature maps are dynamically matched based on the multi-head self-attention mechanism to generate the category query vector; The sigmoid activation function is applied to process the category query vector and output the ground object category probability map for each pixel.

5. The satellite image time series carbon emission prediction method according to claim 1, characterized in that: Based on the output of the semantic segmentation process, the area of ​​each type of land feature is calculated, and the carbon emission series is estimated by combining the preset emission coefficient and the carbon to carbon dioxide molar mass conversion factor, specifically including: Based on the probability map processed by semantic segmentation, the area vector of each type of ground object is calculated. The area calculation includes summing the probability of each pixel and multiplying it by the actual area of ​​the single pixel. Combined with the preset emission coefficient vector, the area vector is converted into preliminary carbon emission; The carbon to carbon dioxide molar mass conversion factor was applied to convert the preliminary carbon element emissions to generate a carbon emission series.

6. The satellite image time series carbon emission prediction method according to claim 1, characterized in that: Fine-tune the estimated carbon emission series to absorb imaging errors or emission coefficient errors, and output the adjusted carbon emission series, including: Input the carbon emission sequence into a lightweight multi-layer perceptron regression module, where the regression module includes a linear fine-tuning layer; In the fine-tuning layer, the weight and bias parameters are applied to perform a linear transformation on the carbon emission series; The transformed carbon emission series is output as the adjusted carbon emission series.

7. The satellite image time series carbon emission prediction method according to claim 1, characterized in that: Based on the historical carbon emissions series, a time series forecasting model is used to predict future carbon emissions, including: The historical carbon emission sequence is input into the PatchTST model, where the PatchTST model includes sequence segmentation, linear projection, and multi-head self-attention mechanism; In the sequence segmentation stage, the input sequence is divided into multiple local time windows; In the linear projection stage, a linear transformation is applied to each time window to generate an embedding vector; In the multi-head self-attention stage, global and local patterns are modeled on the embedding vectors; Output the predicted carbon emissions at a future point in time.

8. The satellite image time series carbon emission prediction method according to claim 1, characterized in that: Output carbon emission prediction results, including: Generate a rasterized emission heat map based on the adjusted carbon emission series, where the heat map resolution matches the original image spatial dimensions; Aggregate carbon emission series data to generate regional total time curves; Generate carbon emission forecasts for a period of time in the future based on the output of the time series forecasting model; Output the gridded emission heat map, regional total amount curve and future forecast value to a preset visualization system.

9. A satellite image time series carbon emission prediction device, characterized in that: The device comprises: at least one processor; and, a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: Acquire a continuous multispectral satellite image sequence of a target area, wherein the sequence includes image data at multiple time points, and the image data includes spatial dimension and spectral channel information; performing denoising processing on the satellite image sequence to remove noise interference in the imaging process and outputting a cleaned image sequence; performing semantic segmentation processing on the cleaned image sequence to generate probability maps of multiple object categories, wherein the object categories include buildings, roads, bare land, vegetation, and water bodies; Based on the output of the semantic segmentation process, the area of ​​each type of land feature is calculated, and the carbon emission series is estimated by combining a preset emission coefficient and a carbon to carbon dioxide molar mass conversion factor; Performing fine-tuning processing on the estimated carbon emission series to absorb imaging errors or emission coefficient errors, and outputting an adjusted carbon emission series; Based on the historical carbon emission series, a time series forecasting model is used to predict future carbon emissions; Output carbon emission prediction results, including rasterized emission heat map, regional total amount curve and future prediction values.

10. A non-volatile computer storage medium for satellite image time series carbon emission prediction, storing computer executable instructions, characterized in that: The computer executable instructions are configured to: Acquire a continuous multispectral satellite image sequence of a target area, wherein the sequence includes image data at multiple time points, and the image data includes spatial dimension and spectral channel information; performing denoising processing on the satellite image sequence to remove noise interference in the imaging process and outputting a cleaned image sequence; performing semantic segmentation processing on the cleaned image sequence to generate probability maps of multiple object categories, wherein the object categories include buildings, roads, bare land, vegetation, and water bodies; Based on the output of the semantic segmentation process, the area of ​​each type of land feature is calculated, and the carbon emission series is estimated by combining a preset emission coefficient and a carbon to carbon dioxide molar mass conversion factor; Performing fine-tuning processing on the estimated carbon emission series to absorb imaging errors or emission coefficient errors, and outputting an adjusted carbon emission series; Based on the historical carbon emission series, a time series forecasting model is used to predict future carbon emissions; Output carbon emission prediction results, including rasterized emission heat map, regional total amount curve and future prediction values.

Images