Sea surface methane remote sensing monitoring method and system based on sensitive factors and machine learning

Abstract

The invention discloses a sea surface methane remote sensing monitoring method and system based on sensitive factors and machine learning. The method comprises the following steps: acquiring multi-source remote sensing data in a target sea surface area; performing multi-modal data processing on data from different sensors in the multi-source remote sensing data to obtain spectral data; determining a weight value of a target sensitive factor according to a preset sensitive factor dynamic model, and establishing environment time sequence data according to the weight value of the target sensitive factor; and splicing the spectrum data and the environment time sequence data, and inputting the spliced data into a methane concentration prediction model to obtain a methane concentration prediction value. According to the method and the system provided by the invention, remote sensing data with different time-space and resolution ratios are fused to overcome the defect of data heterogeneity, and the sensitive factor weight value is screened and determined by the method provided by the embodiment of the invention to realize full fusion of the marine environment sensitive factors, so that the accuracy of a methane concentration monitoring result is improved.

Description

Technical Field

[0001] The present invention relates to the field of ocean remote sensing monitoring technology, and in particular to a sea surface methane remote sensing monitoring method and system based on sensitive factors and machine learning. Background Art

[0002] Methane is a potent greenhouse gas with a warming effect 80 times that of carbon dioxide. The ocean is an important natural source of methane emissions. Seabed sediments, hydrate decomposition, and permafrost melting may all release methane. Therefore, it is important to monitor the emission sources and changes of sea surface methane to assess environmental safety.

[0003] Existing remote sensing technologies for methane monitoring primarily focus on terrestrial environments, employing satellite remote sensing combined with physical inversion algorithms. However, these existing monitoring methods suffer from poor data heterogeneity, poor adaptability to marine environments, and insufficient temporal and spatial resolution, making it difficult to obtain dynamic data on marine methane changes.

[0004] Therefore, the prediction methods in the prior art need to be further improved. Summary of the Invention

[0005] In view of the shortcomings of the above-mentioned related technologies, the purpose of the present invention is to provide a sea surface methane remote sensing monitoring method and system based on sensitive factors and machine learning, so as to overcome the defects of the monitoring methods in the existing technology, such as poor data heterogeneity, poor adaptability to the marine environment, and insufficient temporal and spatial resolution.

[0006] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0007] In a first aspect, the present application provides a method for remote sensing monitoring of sea surface methane based on sensitive factors and machine learning, which includes:

[0008] Acquire multi-source remote sensing data in the target sea surface area;

[0009] Perform multimodal data processing on the data from different sensors in the multi-source remote sensing data to obtain spectral data;

[0010] Determine the weight value of the target sensitive factor according to the preset sensitive factor dynamic model, and establish environmental time series data according to the weight value of the target sensitive factor;

[0011] The spectral data and the environmental time series data are spliced and input into a trained methane concentration prediction model to obtain a methane concentration prediction value output by the methane concentration prediction model.

[0012] Optionally, the step of performing multimodal data processing on data from different sensors in the multi-source remote sensing data to obtain spectral data includes:

[0013] The satellite remote sensing data acquired from different satellite sensors are temporally and spatially aligned to obtain multi-source spatiotemporal-spectral data after temporal and spatial alignment;

[0014] Fusing low-resolution spectral data and high-resolution spatial data in multi-source spatiotemporal-spectral data to obtain fused multispectral data;

[0015] Perform atmospheric correction on the multispectral data to obtain atmospherically corrected multi-source remote sensing data;

[0016] Multi-source remote sensing data are classified into aerosol types to obtain classified and fused spectral data.

[0017] Optionally, the step of performing spatiotemporal alignment on satellite remote sensing data acquired from different satellite sensors to obtain spatiotemporal aligned multi-source spatiotemporal-spectral data includes:

[0018] Perform multi-scale detection on satellite remote sensing data from different sensors to screen out ocean feature points in satellite remote sensing data;

[0019] Perform feature point matching on each of the screened ocean feature points to obtain matched ocean feature points;

[0020] Calculating a sub-pixel offset of the ocean feature point, and correcting the ocean feature point according to the sub-pixel offset to obtain a corrected ocean feature point;

[0021] Cross-resolution fusion is performed based on the corrected ocean feature points to obtain fused multi-source spatiotemporal-spectral data.

[0022] Optionally, the step of fusing low-resolution spectral data and high-resolution spatial data in the multi-source spatiotemporal-spectral data to obtain fused multispectral data includes:

[0023] The input low-resolution spectral data and high-resolution spatial data are input into a preset spectral fusion model to obtain reconstructed multispectral data; wherein, the preset spectral fusion model is constructed based on a U-Net network model.

[0024] Optionally, the step of determining the weight value of the target sensitive factor according to the preset sensitive factor dynamic model and establishing the environmental time series data according to the weight value of the target sensitive factor includes:

[0025] The forward radiation transfer model is used to simulate the radiation characteristics under different methane concentrations and screen out a preset number of target sensitive factors;

[0026] Use the dynamic model of sensitive factors to calculate the weight value of each target sensitive factor;

[0027] Construct environmental time series data based on the weight values of target sensitivity factors.

[0028] Optionally, the sensitive factor dynamic model is constructed based on a random forest model;

[0029] The step of calculating the weight value of the target sensitivity factor using the sensitivity factor dynamic model includes:

[0030] The contribution value of each target sensitivity factor is calculated using the random forest model;

[0031] The contribution value of each target sensitive factor is verified through SHAP value analysis, and according to the verification results, the contribution value of each target sensitive factor is dynamically adjusted, and the optimized contribution value is used as the weight value of each target sensitive factor.

[0032] Optionally, the methane concentration prediction model includes: a feature extraction layer, a time series modeling layer, a feature fusion layer, a fully connected layer and an output layer;

[0033] The steps of splicing the spectral data and the environmental time series data and inputting them into a trained methane concentration prediction model to obtain a methane concentration prediction value output by the methane concentration prediction model include:

[0034] Input the spectral data into the feature extraction layer to obtain the spectral feature vector, and input the environmental time series data into the time series modeling layer to obtain the environmental time series feature vector;

[0035] Inputting the spectral feature vector and the environmental time series feature vector into the feature fusion layer to obtain feature splicing data output by the feature fusion layer;

[0036] The feature splicing data is input into the fully connected layer to obtain the methane concentration prediction result output by the fully connected layer.

[0037] Optionally, the training step of the methane concentration prediction model further includes:

[0038] The Monte Carlo Dropout method is used to calculate the predicted mean output of the methane concentration prediction model, and the prediction variance is calculated based on the preset mean;

[0039] The level of the graded warning is determined according to the prediction variance.

[0040] Optionally, the training step of the methane concentration prediction model further includes:

[0041] The model error is calculated using a spatiotemporal cross-validation method, and the model parameters of the methane concentration prediction model are adjusted according to the model error.

[0042] The second aspect is a sea surface methane remote sensing monitoring system based on sensitive factors and machine learning, which includes:

[0043] A data acquisition module, used to acquire multi-source remote sensing data in the target sea surface area;

[0044] The data fusion module is used to perform multimodal data processing on the data from different sensors in the multi-source remote sensing data to obtain spectral data;

[0045] A sensitive factor construction module is used to determine the weight value of the target sensitive factor according to a preset sensitive factor dynamic model, and to establish environmental time series data according to the weight value of the target sensitive factor;

[0046] The prediction output module is used to splice the spectral data and the environmental time series data and input them into the trained methane concentration prediction model to obtain the methane concentration prediction value output by the methane concentration prediction model.

[0047] Beneficial effects:

[0048] This embodiment discloses a sea surface methane remote sensing monitoring method and system based on sensitive factors and machine learning. The method and system obtain multi-source remote sensing data within a target sea surface area; perform multimodal data processing on data from different sensors in the multi-source remote sensing data to obtain spectral data; determine the weight value of the target sensitive factor according to a preset sensitive factor dynamic model, and establish environmental time series data based on the weight value of the target sensitive factor; splice the spectral data and environmental time series data and input them into a trained methane concentration prediction model to obtain the methane concentration prediction value output by the methane concentration prediction model. The method and system provided by the present invention fuse remote sensing data of different time, space and resolution to overcome the data heterogeneity defects caused by different time, space and spectral resolutions. The method of this embodiment screens out sensitive factors and determines the weight values of the sensitive factors to achieve full fusion of marine environmental sensitive factors, thereby achieving methane concentration inversion results and achieving high-precision methane concentration monitoring results. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flowchart of the steps of a method for remote sensing monitoring of sea surface methane based on sensitivity factors and machine learning according to an embodiment of the present invention;

[0050] Figure 2 Flowchart of the processing steps of multi-source remote sensing data in an embodiment of the present invention;

[0051] Figure 3 Schematic diagram of the principle of obtaining methane concentration by inversion based on the methane concentration prediction model in an embodiment of the present invention;

[0052] Figure 4 This is a principle structural block diagram of the sea surface methane remote sensing monitoring system based on sensitive factors and machine learning disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0054] Since methane is a powerful greenhouse gas, its emissions have a significant impact on global climate change. Remote sensing monitoring can timely detect and locate methane emission sources, and take effective measures to reduce emissions, thereby mitigating global warming.

[0055] Existing technologies for methane remote sensing mainly include satellite remote sensing, airborne remote sensing (such as sensors carried by aircraft), and ground-based remote sensing (such as ground measurement stations). Each of these technologies has its own characteristics and applicable scenarios.

[0056] Satellite remote sensing technology uses sensors on satellites to monitor atmospheric methane concentrations. It primarily relies on shortwave infrared (SWIR) wavelengths of 1.65μm and 2.3μm, and thermal infrared (TIR) wavelengths around 8μm. Thermal infrared sensors measure atmospheric thermal radiation, while shortwave infrared sensors obtain information by measuring atmospheric backscatter. The latter is more sensitive to changes in near-ground methane (CH4) concentrations.

[0057] Airborne remote sensing technology primarily relies on sensors carried by aircraft to monitor methane emissions. Aerial imaging spectroscopy can identify methane emissions at a regional scale, particularly from high-velocity "super emitters" such as fossil fuel and waste treatment facilities.

[0058] Ground-based remote sensing relies on ground-based measurement stations, analyzing data through complex models to identify emission points and track their changes over time. This method directly measures methane concentrations in the atmosphere, providing highly accurate data suitable for detailed monitoring of localized areas. For example, high-resolution satellites, sensors carried by drones, and ground-based measurement stations are combined with complex models to analyze data to identify and track methane emission sources.

[0059] Existing remote sensing technologies for methane monitoring primarily focus on terrestrial environments, employing satellite remote sensing (e.g., TROPOMI and GOSAT) combined with physical inversion algorithms (e.g., WFM-DOAS and OEM). Ocean methane (CH4) monitoring is crucial for global climate change and marine resource development. However, these methods have the following drawbacks:

[0060] 1. In situ measurement: Data is directly obtained through ship-borne or buoy sensors, but the coverage is limited and the cost is high.

[0061] 2. Physical inversion algorithms: Algorithms such as WFM-DOAS and OEM based on spaceborne sensors (such as TROPOMI and GOSAT) rely on atmospheric parameters (temperature, humidity, and aerosol distribution) and have high computational complexity.

[0062] 3. Traditional remote sensing methods: Severely affected by cloud cover and insufficient temporal and spatial resolution (usually >1km / week), it is difficult to capture the dynamic changes of ocean methane.

[0063] Furthermore, satellite remote sensing data also has problems such as data heterogeneity, poor adaptability to the marine environment, insufficient spatiotemporal resolution, and insufficient coupling of sensitive factors. The differences in spatiotemporal and spectral resolution of multi-source remote sensing data (Landsat, MODIS, Sentinel-2) make fusion difficult; traditional algorithms rely on prior information on atmospheric parameters, and their accuracy drops significantly in complex sea-air interaction scenarios (such as high humidity and aerosol interference). Marine methane monitoring lacks a dynamic database with high spatiotemporal resolution and cannot effectively capture sudden methane leaks. Existing models do not fully integrate marine environmental sensitive factors (such as sea surface temperature, chlorophyll concentration, and wind speed), resulting in poor physical interpretability of the inversion results.

[0064] In order to overcome the above problems, the present application provides a sea surface methane remote sensing monitoring method and system based on sensitive factors and machine learning. First, multi-source remote sensing data in the target sea surface area is obtained, the multi-source remote sensing data is fused to obtain spectral data, and then the sensitive factors are screened out. The weight values of each screened sensitive factor are confirmed, and environmental time series data are established based on the determined weight values. The concentration prediction value of sea surface methane is obtained based on the spectral data and the environmental time series data based on machine learning inversion. The method and system provided by the present application reduce the problem of data specificity by fusing the collected remote sensing data from multiple sources, and dynamically determine the weight values of sensitive factors through a multi-physics coupling model, thereby increasing the full fusion of sensitive factors and improving the interpretability of the inversion results. The spectral features and environmental time series data are combined based on machine learning to invert the methane concentration, thereby improving the prediction accuracy.

[0065] The following further describes the sea surface methane remote sensing monitoring method and system based on sensitive factors and machine learning disclosed in this embodiment with reference to the accompanying drawings.

[0066] In the first aspect, the present application provides a method for remote sensing monitoring of sea surface methane based on sensitive factors and machine learning, such as Figure 1 Shown, including:

[0067] Step S1: Acquire multi-source remote sensing data within the target sea surface area.

[0068] In this step, multi-source remote sensing data for the target sea surface area is obtained from satellites or aircraft. When the multi-source remote sensing data comes from one or more satellites, remote sensing data of the target sea surface area captured by various sensors onboard the satellites is obtained. When the multi-source remote sensing data comes from one or more aircraft, flight plans are developed for the aircraft over the target sea surface area, and sensors are deployed to image the sea surface, collecting high-resolution remote sensing image data. The collected data is transmitted to the ground via radio or other means. The ground system preprocesses the received data to obtain the multi-source remote sensing data.

[0069] Specifically, data is downloaded from the remote sensing satellite data platform and preprocessed. When downloading data from the remote sensing WeChat data platform, data is searched and downloaded based on the area corresponding to the target sea surface area, the selected data type (e.g., total carbon dioxide, total sulfur dioxide, or total methane), and the selected time range, thereby obtaining relevant remote sensing data.

[0070] In one implementation, multi-source remote sensing data includes primary data and auxiliary data. The primary data is remote sensing data acquired using Sentinel-5P and / or Landsat-8 / 9. The auxiliary data includes MODIS and HY-1C / D water color data. Sentinel-5P is a global atmospheric pollution monitoring satellite equipped with a tropospheric observation instrument, capable of providing high-temporal and spatial resolution atmospheric composition data, including carbon dioxide and methane. Landsat-8 / 9 are two satellites designed to provide high-resolution remote sensing imagery of the Earth's surface. MODIS is a spaceborne remote sensing instrument. MODIS data features high temporal resolution and wide spectral coverage, making it suitable for dynamic global environmental monitoring and multi-disciplinary scientific research, including monitoring of sea surface areas. MODIS data can be downloaded from NASA's official platform or a geospatial data cloud platform. Downloading remote sensing data also requires defining the target area, selecting a time range, selecting a data type, and setting data parameters (such as spatial resolution and spectral bands). MODIS data has different spatial resolutions and spectral bands, which can be selected based on needs. The HY-1C / D ocean color data comes from a dual-satellite network of high-resolution ocean satellites. These satellites carry a variety of high-resolution imaging sensors, including the Ocean Color Scanner (COCTS), the Ultraviolet Imager (UVI), and the Coastal Zone Imager (CZI). These sensors can be used to obtain ocean environmental parameters such as sea surface height, sea surface temperature, sea surface salinity, sea surface wind speed, wave height, sea ice distribution, and seawater pollution. The ocean color data is primarily used to monitor the optical properties of ocean water, such as chlorophyll concentration and suspended matter concentration.

[0071] Step S2: performing multimodal data processing on the data from different sensors in the multi-source remote sensing data to obtain spectral data.

[0072] After receiving the multi-source remote sensing data in the above step S1, since the multi-source remote sensing data comes from different sensors, in order to overcome the heterogeneity of the data, this step fuses the collected multi-source remote sensing data to obtain spectral data.

[0073] Further, combined with Figure 2 As shown, the step of performing multimodal data processing on the data from different sensors in the multi-source remote sensing data to obtain spectral data includes:

[0074] Step S11: performing spatiotemporal alignment on satellite remote sensing data acquired from different satellite sensors to obtain spatiotemporal aligned multi-source spatiotemporal-spectral data.

[0075] This step uses a modified SIFT (Scale-Invariant Feature Transform) algorithm for multi-source data registration. By extracting local feature points from Sentinel-5P (10m resolution) and MODIS (1km resolution), and combining them with a bicubic interpolation algorithm, we achieve sub-pixel alignment (error <0.2 pixels), eliminating geometric offsets caused by differences in satellite orbits and enabling precise registration across resolutions.

[0076] Specifically, this step involves performing spatiotemporal alignment on satellite remote sensing data acquired from different satellite sensors to obtain the spatiotemporal aligned multi-source spatiotemporal-spectral data. The steps include:

[0077] Step S211: Perform multi-scale detection on satellite remote sensing data from different sensors to filter out ocean feature points of the satellite remote sensing data.

[0078] Multi-scale detection: Key points are extracted through the Difference of Gaussian Pyramid (DoG) to screen stable ocean features (such as sea-land boundaries).

[0079] Because satellite remote sensing data from different sensors may correspond to different image scales, and a single scale cannot effectively capture all targets, this step uses multi-scale detection to extract key points in order to screen out stable ocean features. Multi-scale detection in this step processes images or features at different scales to ensure that targets of all sizes in the image can be identified.

[0080] Specifically, satellite remote sensing data from different sensors is first registered to ensure alignment within the same geographic coordinate system. The aligned data is then converted to the same scale. Feature extraction is performed on remote sensing images using deep learning models such as convolutional neural networks. For ocean feature points of different scales, features are extracted at different convolutional layers of the CNN to generate feature maps of different scales. Feature maps of different scales are then fused. Fusion of multi-scale feature maps can be achieved using a feature pyramid network (FPN). Object detection algorithms, such as YOLO, are applied to the fused feature maps to detect ocean feature points. Detected ocean feature points are then screened. This screening can be based on pre-set rules, such as shape, texture, or shape. Alternatively, machine learning algorithms can be used to distinguish true ocean feature points from false positives using trained classifiers.

[0081] Step S212: performing feature point matching on each of the screened ocean feature points to obtain matched ocean feature points.

[0082] Feature point matching is performed on each selected ocean feature point, which specifically includes two important steps: feature point description and feature point matching. The first feature point description involves extracting local features from each ocean feature point. Feature vectorization converts the extracted features into feature vectors. Feature point matching involves selecting a matching algorithm, setting a matching threshold, and determining whether two feature points match. If the distance between the local features (i.e., feature descriptors) corresponding to two ocean feature points is less than the threshold, they are considered to match, thus obtaining the matched ocean feature points.

[0083] Furthermore, to achieve more accurate matching of ocean feature points, a bidirectional matching method is used in this step. The bidirectional matching method combines the kd-tree nearest neighbor search with the RANSAC algorithm to eliminate false matches and ensure the robustness of feature point pairs.

[0084] A kd-tree is constructed for the feature point descriptor set. For each ocean feature point, a nearest neighbor search is performed using the kd-tree within the feature point set of the other image to find the most similar feature point. Bidirectional matching involves forward matching and reverse matching. This involves searching for the nearest neighbor feature point in image B for each feature point in image A to form a matching pair, and searching for the nearest neighbor feature point in image A for each feature point in image B to form a matching pair. A matching pair is considered valid only when both forward and reverse matching point to the same pair of feature points. The RANSAC algorithm is then used to eliminate mismatched pairs.

[0085] Step S213: Calculate the sub-pixel offset of the ocean feature point, and correct the ocean feature point according to the sub-pixel offset to obtain a corrected ocean feature point.

[0086] Taylor expansion correction: Using the second-order Taylor expansion of the DoG function (Difference of Gaussians), the sub-pixel offset of the feature points is calculated to improve the positioning accuracy to sub-pixel. The sub-pixel offset is:

[0087]

[0088] Step S214: performing cross-resolution fusion based on the corrected ocean feature points to obtain fused multi-source spatiotemporal-spectral data.

[0089] The bicubic interpolation method is used to achieve cross-resolution data fusion. The bicubic interpolation method first resamples the low-resolution remote sensing data (such as MODIS) to retain edge details, and then adapts it to the high-resolution data (such as Sentinel-5P). The bicubic interpolation kernel function is:

[0090] W(t)=-0.5|t| 3 +2.5|t| 2 -4|t|+2(|t|≤2);

[0091] After obtaining the fused multi-source spatiotemporal-spectral data, local feature points can be dynamically supplemented based on the registration residual map, and iterative optimization can be performed until the global error meets the standard.

[0092] Step S22: Fusing the low-resolution spectral data and the high-resolution spatial data in the multi-source spatiotemporal-spectral data to obtain fused multispectral data.

[0093] Since there are low-resolution spectral data and high-resolution spatial data in the fused multi-source spatiotemporal-spectral data, in order to avoid the low accuracy of the prediction results caused by data specificity, in this step, the high-resolution spatial data and the low-resolution spatial data are fused using the preset spectral fusion model to obtain the reconstructed multispectral data.

[0094] Specifically, the step of fusing low-resolution spectral data and high-resolution spatial data in the multi-source spatiotemporal-spectral data to obtain fused multispectral data includes:

[0095] The input low-resolution spectral data and high-resolution spatial data are input into a preset spectral fusion model to obtain reconstructed multispectral data; wherein, the preset spectral fusion model is constructed based on the U-Net network model. Since the U-Net network model has a symmetrical encoder-decoder structure, the low-resolution spectral data and the high-resolution spectral data are input into the encoder. The encoder gradually extracts high-level semantic features through convolution and pooling layers, reduces the spatial resolution, and then inputs the extracted high-level semantic features into the decoder. The decoder gradually restores the resolution of the data through upsampling and preserves details while encoding and decoding. Furthermore, the U-Net network model is also provided with skip connections, which can directly pass low-level details to the decoder through skip connections, solving the problem of information loss during the upsampling process and realizing efficient fusion of low-resolution data and high-resolution data.

[0096] Step S23: Perform atmospheric correction on the multispectral data to obtain atmospherically corrected multi-source remote sensing data.

[0097] After fusing data of varying resolutions, the multispectral data must undergo atmospheric correction. Atmospheric correction aims to eliminate atmospheric influences on remote sensing images, thereby obtaining more accurate ground reflectance data and improving the accuracy and usability of remote sensing data. Specifically, atmospheric influences on remote sensing data primarily include absorption, scattering, and reflection, which can lead to radiometric errors in the data.

[0098] In one embodiment, atmospheric correction is performed using a radiative transfer model, which can be the 6S (Second Simulation of the Satellite Signal in the Solar Spectrum) radiative transfer model or the MODTRANM model. The 6S model simulates the atmospheric radiative transfer process, inputting atmospheric parameters (e.g., aerosol type, visibility) and multispectral data, and inverting surface reflectance through iterative calculation. The MODTRANM model can provide atmospheric transfer calculations at different resolutions.

[0099] In this step, the 6S radiative transfer model is combined with the FLAASH (Fast Line-of-sight Atmospheric Analysis of Hypercubes) algorithm to eliminate aerosol scattering effects on a band-by-band basis. For marine environments, the HYBRID algorithm is introduced to optimize the inversion accuracy of the water vapor absorption channel (1.38μm), achieving a reflectivity error of ≤3% after correction.

[0100] Step S24: classify the multi-source remote sensing data into aerosol types to obtain classified and fused spectral data.

[0101] Aerosols (such as smoke and dust) contained in multi-source remote sensing data absorb and scatter electromagnetic waves, significantly affecting the radiation received by remote sensing sensors. Different aerosol types have different absorption and scattering characteristics of electromagnetic waves. By classifying aerosol types and applying corresponding correction layers to different aerosol types, the accuracy of multi-source remote sensing data can be improved.

[0102] Specifically, the steps for aerosol type classification include:

[0103] Water vapor inversion: Based on the ratio method of shortwave infrared (SWIR, 1.38μm) and near infrared (NIR, 0.94μm) bands, a water vapor content inversion model was constructed with an accuracy of 93% (compared with AERONET station data).

[0104] Aerosol classification: A k-means clustering algorithm (k=4) is used to classify aerosols into marine (dominated by salt particles), dust (mineral aerosols), pollution (industrial emissions), and mixed types. After classification, different atmospheric correction parameters are matched to each type to improve the robustness of the inversion in complex ocean-air interaction scenarios.

[0105] The classified aerosol types are refined and processed to obtain processed spectral data. In this step, multimodal data processing is performed on the data from different sensors in the multi-source remote sensing data to obtain spectral data to improve the prediction accuracy of methane concentration.

[0106] Step S3: Determine the weight value of the target sensitive factor according to the preset sensitive factor dynamic model, and establish environmental time series data according to the weight value of the target sensitive factor.

[0107] In this step, the target sensitive factors are first screened out, and then the weight values of the target sensitive factors are determined based on the constructed dynamic model of sensitive factors. Finally, the environmental time series data are established according to the weight values of the target sensitive factors.

[0108] Specifically, the steps of determining the weight value of the target sensitive factor according to the preset sensitive factor dynamic model and establishing the environmental time series data according to the weight value of the target sensitive factor include:

[0109] Step S31: Simulate the radiation characteristics under different methane concentrations through the forward radiation transfer model to screen out a preset number of target sensitivity factors.

[0110] The forward radiative transfer model (MODTRAN) was used to simulate the radiation characteristics under different CH4 concentrations and extract sensitive band combinations. Pearson correlation analysis was performed on 1000 sets of simulated data, and the top five sensitive factors (|γ|>0.75) were screened.

[0111] In this step, the MODTRAN (MODerate resolution atmospheric TRANsmission) radiation transfer model is used to simulate the methane absorption characteristics in the 1.65 μm band and calculate the absorption depth (Depth = 1-R observed / R background ) as the spectral sensitivity factor.

[0112] Spectral characteristics: Absorption depth Quantify the methane absorption intensity in the 1.65μm band;

[0113] Second derivative spectrum: Enhance the detection capability of weak absorption peaks.

[0114] Sea surface temperature (SST): methane solubility corrected by the Arrhenius equation S = κ H ·e -ΔH / (R·T) , quantifying the effect of temperature on methane release.

[0115] Wind speed (U 10 ): Build a gas transmission rate model Characterize the driving effect of wind speed on methane diffusion at the air-sea interface.

[0116] Chlorophyll concentration (Chl-a): CH 4bio =0.023·Chl-a 1.4 (Unit: nmol / L) Correlates plankton activity with methane release.

[0117] Step S32: Calculate the weight value of each target sensitivity factor using the sensitivity factor dynamic model.

[0118] This step uses a preset dynamic model of sensitive factors to dynamically assign weight values of target sensitive factors. In one embodiment, the dynamic model of sensitive factors adopts a random forest model, and the contribution of each factor (such as SST weight 40%, wind speed 30%, chlorophyll 20%) is calculated by the dynamic model of sensitive factors constructed based on the random forest model, and the rationality of the weight is verified by SHAP (SHapley Additive exPlanations) value analysis to ensure the physical interpretability of the model. By dynamically assigning weights to environmental factors, the consistency between the inversion results and the ocean methane generation-transport mechanism is improved.

[0119] Specifically, the step of calculating the weight value of the target sensitivity factor using the sensitivity factor dynamic model includes:

[0120] The contribution value of each target sensitive factor is calculated using the random forest model; the contribution value of each target sensitive factor is verified through SHAP value analysis, and according to the verification results, the contribution value of each target sensitive factor is dynamically adjusted, and the optimized contribution value is used as the weight value of each target sensitive factor.

[0121] Step S33: construct environmental time series data according to the weight value of the target sensitivity factor.

[0122] After determining the weights of each target sensitivity factor, the environmental time series data is constructed based on the weights of each target sensitivity factor. One implementation method is to construct time series data including SST, wind speed, Chl-a and other information in chronological order.

[0123] Step S4: splicing the spectral data and the environmental time series data and inputting them into the trained methane concentration prediction model to obtain the methane concentration prediction value output by the methane concentration prediction model.

[0124] After obtaining the spectral data through step S2 and the environmental time series data through step S3, the spectral data and the environmental time series data are spliced, and the spliced data are input into the trained methane concentration prediction model, and the methane concentration prediction model is used to obtain the methane concentration prediction value.

[0125] Combine Figure 3 As shown, the methane concentration prediction model includes: a feature extraction layer, a time series modeling layer, a feature fusion layer, a fully connected layer and an output layer.

[0126] The steps of splicing the spectral data and the environmental time series data and inputting them into the trained methane concentration prediction model to obtain the methane concentration prediction value output by the methane concentration prediction model include:

[0127] Step S41: input the spectral data into the feature extraction layer to obtain the spectral feature vector, and input the environmental time series data into the time series modeling layer to obtain the environmental time series feature vector.

[0128] Step S42: input the spectral feature vector and the environmental time series feature vector into the feature fusion layer to obtain feature splicing data output by the feature fusion layer.

[0129] Step S43: input the feature splicing data into the fully connected layer to obtain the methane concentration prediction result output by the fully connected layer.

[0130] In one embodiment, combined Figure 3As shown, the 1D-CNN module is used to analyze spectral absorption features (1.65μm band). The input layer performs band selection and normalization on the raw spectral data, and then the 1D-CNN module extracts features from the selected and normalized spectral bands. For example, 11 spectral bands (VIS-NIR-SWIR) are selected from the raw spectral data obtained from Landsat-8, and then the selected spectral band data is normalized to the [0,1] range.

[0131] Specifically, the 1D-CNN module contains three layers of 1D convolution (convolution kernel K = 3, number of channels 32→64→128), ReLU activation function, and maximum pooling (pooling window = 2), which are used to process the input raw spectral data in sequence to extract the local spectral features of the methane absorption band (1.65μm), thereby outputting a 64-dimensional spectral feature vector to characterize the methane absorption intensity and spectral morphology.

[0132] In addition, the LSTM module is used to process the environmental parameter time series data to obtain the environmental time series feature vector. Figure 3 As shown, the time series environmental parameters (SST, wind speed, Chl-a) are constructed with a time step of T = 30 days, and the environmental time series data is input into the input layer. The input layer interpolates and normalizes the environmental time series data and then inputs it into the LSTM module. The LSTM module performs time series modeling to capture the seasonal and mutation characteristics of environmental factors. In one embodiment, the LSTM module is a bidirectional LSTM model with a hidden layer dimension of 64. The environmental factor features are extracted jointly by the forward LSTM and the backward LSTM.

[0133] After obtaining local spectral features and environmental factor features, they are fused through a feature fusion layer. Specifically, the spectral features (64 dimensions) are concatenated with the environmental time series features (32 dimensions), and the methane concentration (ppb) is output through a fully connected layer (256→128→1). Huber Loss (δ=1.0) is used to balance the influence of outliers.

[0134] Furthermore, in order to obtain accurate prediction results and respond to the prediction results, the training step of the methane concentration prediction model also includes:

[0135] The Monte Carlo Dropout method is used to calculate the predicted mean output by the methane concentration prediction model, and the prediction variance is calculated based on the preset mean; the level of the graded warning is determined based on the prediction variance.

[0136] In one implementation, when using the Monte Carlo method to calculate the predicted mean, Dropout is activated (dropout rate = 0.5) during the inference phase, and T = 100 random forward propagations are performed to calculate the predicted mean and variance. A graded warning rule is used: low uncertainty (σ < 5 ppb): the predicted value is directly output; medium uncertainty (5 ≤ σ < 15 ppb): the prediction is flagged for manual review; high uncertainty (σ ≥ 15 ppb): the data re-collection process is triggered.

[0137] Furthermore, the training step of the methane concentration prediction model further includes:

[0138] The model error is calculated using a spatiotemporal cross-validation method, and the model parameters of the methane concentration prediction model are adjusted according to the model error.

[0139] In order to obtain a more accurate methane concentration prediction model, when training the methane concentration prediction model, the model error can be calculated by cross-validation and actual measurement comparison, thereby improving the prediction accuracy of the methane concentration prediction model.

[0140] In specific implementation, cross-validation: spatiotemporal cross-validation (STCV) showed RMSE = 2.8 ppb, R = 0.92; actual measurement comparison: compared with ship-borne Picarro G4301 data, the average error was <5%.

[0141] The detection method provided in this embodiment uses multimodal data fusion: the TSA-Fusion algorithm resolves the scale differences of multiple sensors, and uses a hybrid neural network architecture: CNN-LSTM-Attention to jointly model spectral and spatiotemporal features; and constructs dynamic sensitivity factor weights: a dynamic allocation model based on physical mechanisms and SHAP value analysis, thereby achieving data integrity and improving the accuracy of monitoring data.

[0142] Secondly, a sea surface methane remote sensing monitoring system based on sensitive factors and machine learning, such as Figure 4 Shown, including:

[0143] The data acquisition module 100 is used to acquire multi-source remote sensing data in the target sea surface area; its function is as described in step S1.

[0144] The data fusion module 200 is used to perform multimodal data processing on the data from different sensors in the multi-source remote sensing data to obtain spectral data; its function is as described in step S2.

[0145] The sensitive factor construction module 300 is used to determine the weight value of the target sensitive factor according to the preset sensitive factor dynamic model, and to establish the environmental time series data according to the weight value of the target sensitive factor; its function is as described in step S3.

[0146] The prediction output module 400 is used to splice the spectral data and the environmental time series data and input them into the trained methane concentration prediction model to obtain the methane concentration prediction value output by the methane concentration prediction model. Its function is as described in step S4.

[0147] This embodiment discloses a sea surface methane remote sensing monitoring method and system based on sensitive factors and machine learning. The method and system obtain multi-source remote sensing data within a target sea surface area; perform multimodal data processing on data from different sensors in the multi-source remote sensing data to obtain spectral data; determine the weight value of the target sensitive factor according to a preset sensitive factor dynamic model, and establish environmental time series data based on the weight value of the target sensitive factor; splice the spectral data and environmental time series data and input them into a trained methane concentration prediction model to obtain the methane concentration prediction value output by the methane concentration prediction model. The method and system provided by the present invention fuse remote sensing data of different time, space and resolution to overcome the data heterogeneity defects caused by different time, space and spectral resolutions. The method of this embodiment screens out sensitive factors and determines the weight values of the sensitive factors to achieve full fusion of marine environmental sensitive factors, thereby achieving methane concentration inversion results and achieving high-precision methane concentration monitoring results.

[0148] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0149] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A sea surface methane remote sensing monitoring method based on sensitive factors and machine learning, characterized in that: include: Acquire multi-source remote sensing data in the target sea surface area; Perform multimodal data processing on the data from different sensors in the multi-source remote sensing data to obtain spectral data; Determine the weight value of the target sensitive factor according to the preset sensitive factor dynamic model, and establish environmental time series data according to the weight value of the target sensitive factor; The spectral data and the environmental time series data are spliced and input into a trained methane concentration prediction model to obtain a methane concentration prediction value output by the methane concentration prediction model.

2. The method for remote sensing monitoring of sea surface methane based on sensitive factors and machine learning according to claim 1, characterized in that: The step of performing multimodal data processing on data from different sensors in the multi-source remote sensing data to obtain spectral data includes: The satellite remote sensing data acquired from different satellite sensors are temporally and spatially aligned to obtain multi-source spatiotemporal-spectral data after temporal and spatial alignment; Fusing low-resolution spectral data and high-resolution spatial data in multi-source spatiotemporal-spectral data to obtain fused multispectral data; Perform atmospheric correction on the multispectral data to obtain atmospherically corrected multi-source remote sensing data; Multi-source remote sensing data are classified into aerosol types to obtain classified and fused spectral data.

3. The method for remote sensing monitoring of sea surface methane based on sensitive factors and machine learning according to claim 2, characterized in that: The step of performing spatiotemporal alignment on satellite remote sensing data acquired from different satellite sensors to obtain spatiotemporal aligned multi-source spatiotemporal-spectral data comprises: Perform multi-scale detection on satellite remote sensing data from different sensors to screen out ocean feature points in satellite remote sensing data; Perform feature point matching on each of the screened ocean feature points to obtain matched ocean feature points; Calculating a sub-pixel offset of the ocean feature point, and correcting the ocean feature point according to the sub-pixel offset to obtain a corrected ocean feature point; Cross-resolution fusion is performed based on the corrected ocean feature points to obtain fused multi-source spatiotemporal-spectral data.

4. The method for remote sensing monitoring of sea surface methane based on sensitive factors and machine learning according to claim 2, characterized in that: The step of fusing low-resolution spectral data and high-resolution spatial data in the multi-source spatiotemporal-spectral data to obtain fused multispectral data comprises: The input low-resolution spectral data and high-resolution spatial data are input into a preset spectral fusion model to obtain reconstructed multispectral data; wherein, the preset spectral fusion model is constructed based on a U-Net network model.

5. The method for remote sensing monitoring of sea surface methane based on sensitive factors and machine learning according to claim 1, characterized in that: The steps of determining the weight value of the target sensitive factor according to the preset sensitive factor dynamic model and establishing the environmental time series data according to the weight value of the target sensitive factor include: The forward radiation transfer model is used to simulate the radiation characteristics under different methane concentrations and screen out a preset number of target sensitive factors; Use the dynamic model of sensitive factors to calculate the weight value of each target sensitive factor; Construct environmental time series data based on the weight values of target sensitivity factors.

6. The method for remote sensing monitoring of sea surface methane based on sensitive factors and machine learning according to claim 5, characterized in that: The dynamic model of sensitive factors is constructed based on the random forest model; The step of calculating the weight value of the target sensitivity factor using the sensitivity factor dynamic model includes: The contribution value of each target sensitivity factor is calculated using the random forest model; The contribution value of each target sensitive factor is verified through SHAP value analysis, and according to the verification results, the contribution value of each target sensitive factor is dynamically adjusted, and the optimized contribution value is used as the weight value of each target sensitive factor.

7. The method for remote sensing monitoring of sea surface methane based on sensitive factors and machine learning according to claim 1, characterized in that: The methane concentration prediction model includes: a feature extraction layer, a time series modeling layer, a feature fusion layer, a fully connected layer and an output layer; The steps of splicing the spectral data and the environmental time series data and inputting them into a trained methane concentration prediction model to obtain a methane concentration prediction value output by the methane concentration prediction model include: Input the spectral data into the feature extraction layer to obtain the spectral feature vector, and input the environmental time series data into the time series modeling layer to obtain the environmental time series feature vector; Inputting the spectral feature vector and the environmental time series feature vector into the feature fusion layer to obtain feature splicing data output by the feature fusion layer; The feature splicing data is input into the fully connected layer to obtain the methane concentration prediction result output by the fully connected layer.

8. The method for remote sensing monitoring of sea surface methane based on sensitive factors and machine learning according to claim 7, characterized in that: The training step of the methane concentration prediction model further includes: The Monte Carlo Dropout method is used to calculate the predicted mean output of the methane concentration prediction model, and the prediction variance is calculated based on the preset mean; The level of the graded warning is determined according to the prediction variance.

9. The method for remote sensing monitoring of sea surface methane based on sensitive factors and machine learning according to claim 8, characterized in that: The training step of the methane concentration prediction model further includes: The model error is calculated using a spatiotemporal cross-validation method, and the model parameters of the methane concentration prediction model are adjusted according to the model error.

10. A sea surface methane remote sensing monitoring system based on sensitive factors and machine learning, characterized in that: include: A data acquisition module, used to acquire multi-source remote sensing data in the target sea surface area; The data fusion module is used to perform multimodal data processing on the data from different sensors in the multi-source remote sensing data to obtain spectral data; A sensitive factor construction module is used to determine the weight value of the target sensitive factor according to a preset sensitive factor dynamic model, and to establish environmental time series data according to the weight value of the target sensitive factor; The prediction output module is used to splice the spectral data and the environmental time series data and input them into the trained methane concentration prediction model to obtain the methane concentration prediction value output by the methane concentration prediction model.

Images