A two-way adaptive multi-resolution guided sea surface temperature data fusion system

Abstract

The application discloses a kind of bidirectional adaptive multi-resolution guide sea surface temperature data fusion systems, belong to sea surface temperature fusion technical field, specifically include: data receiving layer receives infrared and microwave sea surface temperature data, mark different time phase;Characteristic decoupling layer handles data according to different time phase, separates the spatial detail feature of infrared and the semantic association feature of microwave;Bidirectional guide layer handles feature according to different time phase, with semantic association feature guide spatial detail feature to supplement reconstruction, while with spatial detail feature guide semantic association feature to enhance spatial structure;Dynamic fusion layer handles guided feature according to different time phase, dynamically calculates fusion weight according to real-time quality evaluation;Fusion output layer is based on weight integration feature, to target resolution, output day and night sea surface temperature fusion dataset;The application provides reliable data support for high-precision marine environment monitoring and climate research.

Description

Technical Field

[0001] This invention relates to the field of sea surface temperature fusion technology, specifically to a bidirectional adaptive multi-resolution guided sea surface temperature data fusion system. Background Technology

[0002] Sea surface temperature (SST), a key physical parameter of the ocean system, is fundamentally important for understanding global climate change, driving ocean circulation models, optimizing fisheries resource development, and providing early warning of extreme weather events. With the rapid development of satellite remote sensing technology, infrared and microwave sensors have become the main means of acquiring large-scale SST data. Infrared sensors, with their micrometer-level wavelength detection advantage, can capture fine spatial structures at the sub-kilometer level, clearly revealing small- to medium-scale ocean phenomena such as ocean current boundaries and eddies. Microwave sensors, relying on the ability of centimeter waves to penetrate clouds, enable continuous day and night observations, effectively compensating for the observation blind spots of infrared data during cloudy or rainy weather. The complementarity of these two types of data makes it possible to construct high-precision, high-coverage SST datasets; therefore, data fusion technology has gradually become a research hotspot in the field of marine remote sensing.

[0003] Current sea surface temperature data fusion methods have developed in parallel across multiple technological paths. Traditional statistical methods, by establishing linear regression models for infrared and microwave data and using spatiotemporal interpolation algorithms to fill data gaps, have demonstrated stable performance in routine observations in open sea areas. Physical modeling, based on the thermal radiation transfer equation and incorporating atmospheric correction parameters to optimize the matching relationship between the two types of data, has significantly improved the fusion accuracy in nearshore areas. Deep learning methods, which have emerged in recent years, have demonstrated particularly outstanding adaptability under complex sea conditions by constructing end-to-end feature mapping networks, providing a new paradigm for the nonlinear fusion of multi-source data. The application of these technologies has initially improved the temporal continuity and spatial integrity of sea surface temperature data, supporting breakthroughs in several marine scientific research projects.

[0004] However, existing technologies still have significant limitations. Regarding temporal adaptability, most methods employ a unified processing framework for day and night observation data, failing to adequately differentiate sensor response differences caused by factors such as solar radiation and atmospheric stability. This results in systematic biases in the fusion results during the twilight transition. In terms of feature interaction mechanisms, mainstream schemes often use unidirectional feature transfer modes, either focusing on global trend guidance from microwave data or relying on local detail correction from infrared data, making it difficult to achieve deep collaborative optimization of the two types of features. Regarding dynamic adjustment capabilities, the calculation of fusion weights largely depends on preset rules or static training parameters, failing to adapt in real-time to data quality fluctuations in different sea areas. The problem of fusion accuracy degradation is particularly pronounced during extreme weather events such as typhoons and sea fog. These shortcomings limit the effectiveness of sea surface temperature fusion data in high-precision marine environmental analysis. Summary of the Invention

[0005] The purpose of this invention is to provide a bidirectional adaptive multi-resolution guided sea surface temperature data fusion system to solve the following technical problems:

[0006] Existing technologies suffer from insufficient temporal adaptability, unidirectional feature interaction, and weak dynamic adjustment capabilities, which limit the high-precision application of sea surface temperature fusion data.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A bidirectional adaptive multi-resolution guided sea surface temperature data fusion system includes:

[0009] The data receiving layer is used to receive infrared and microwave sea surface temperature data transmitted by satellite sensors and to mark the observation time phase category for data recording, including daytime and nighttime phases.

[0010] The feature decoupling layer is used to process infrared sea surface temperature data and microwave sea surface temperature data according to the observation time phase category, and to separate the spatial detail features of infrared sea surface temperature data and the semantic association features of microwave sea surface temperature data.

[0011] A bidirectional guidance layer is used to process spatial detail features and semantic association features according to the observation time phase category. The semantic association features are used to guide the spatial detail features to supplement and reconstruct information, while the spatial detail features are used to guide the semantic association features to enhance the spatial structure.

[0012] The dynamic fusion layer is used to process the guided features according to the observation time phase category. Based on the real-time quality assessment results of the input data in spatial location, it dynamically calculates the fusion weights of spatial detail features and semantic association features.

[0013] The fusion output layer is used to generate fused features by integrating bidirectional guided spatial detail features and semantic association features based on fusion weights. The fused features are then improved to the target spatial resolution, and the sea surface temperature fusion datasets for daytime and nighttime phases are output separately.

[0014] As a further aspect of the present invention: in the data receiving layer, marking the observation phase category for data records specifically includes:

[0015] The system analyzes the precise observation timestamps attached to satellite data to obtain the absolute time information for each observation. Based on the observation timestamps and the astronomical information of the data acquisition location, it calculates the local solar altitude angle and marks observations with a solar altitude angle greater than or equal to zero as daytime phases and observations with a solar altitude angle less than zero as nighttime phases.

[0016] The observation phase category information is associated with the corresponding infrared sea surface temperature data and microwave sea surface temperature data as metadata, and this category information is passed to the subsequent processing level along with the data.

[0017] As a further aspect of the present invention: in the feature decoupling layer, the process of processing infrared sea surface temperature data and microwave sea surface temperature data according to the observation time phase category is as follows:

[0018] The received infrared sea surface temperature data and microwave sea surface temperature data are separated into daytime time phase data subsets and nighttime time phase data subsets based on the observation time phase category metadata carried in the data;

[0019] For a subset of daytime temporal data, a set of parameter configurations trained on daytime temporal data is used to separate the spatial detail features of infrared sea surface temperature data and extract the semantic association features of microwave sea surface temperature data.

[0020] For a subset of nighttime temporal data, another set of parameter configurations trained on the nighttime temporal data is used to separate the spatial detail features of infrared sea surface temperature data and extract the semantic association features of microwave sea surface temperature data.

[0021] The parameters are obtained through independent training on labeled data at different time phases, and the training process uses a time-sensitive objective function.

[0022] As a further aspect of the present invention: the process of separating the spatial detail features of infrared sea surface temperature data and the semantic association features of microwave sea surface temperature data in the feature decoupling layer is as follows:

[0023] Multi-scale analysis of infrared sea surface temperature data is performed to extract spatial gradient information and texture change information at different scales. Based on the extracted spatial gradient information and texture change information, a feature tensor representing local detail changes is constructed.

[0024] Time-frequency transformation is performed on microwave sea surface temperature data to extract its evolution pattern in the time dimension and its distribution pattern in the spatial dimension. Based on the extracted time-frequency features, a semantic feature tensor characterizing macroscopic physical processes is constructed.

[0025] Through the convolutional operation of the feature decoupling layer, spatial detail features and semantically related features are separated into different feature channels.

[0026] As a further aspect of the present invention: in the bidirectional guidance layer, the process of processing spatial detail features and semantic association features according to the observation time phase category is as follows:

[0027] Spatial detail features and semantic association features from the feature decoupling layer are routed to different processing branches based on the observation time phase of the data source.

[0028] In the daytime phase processing branch, the semantic association features of the daytime phase are used to guide the spatial detail features of the daytime phase for information supplementation and reconstruction. This process is achieved through a cross-modal information transmission mechanism.

[0029] In the nighttime phase processing branch, the semantic association features of the nighttime phase are used to guide the spatial detail features of the nighttime phase for information supplementation and reconstruction. This process is also achieved through the cross-modal information transmission mechanism.

[0030] In the two processing branches of the time phase, spatial detail features are used simultaneously to guide semantic association features to enhance spatial structure.

[0031] As a further aspect of the present invention: in the bidirectional guidance layer, the specific process of using semantic association features to guide spatial detail features for information supplementation and reconstruction, and simultaneously using spatial detail features to guide semantic association features for spatial structure enhancement, is as follows:

[0032] An attention mapping mechanism is established to transform semantic association features into spatial detail features. The guiding weight of semantic features on spatial features is calculated. Based on the guiding weight, semantic association features are converted into spatial detail supplementary information to reconstruct missing regions.

[0033] Establish a structural constraint mechanism from spatial detail features to semantic association features, extract edge and texture information of spatial features, and adjust the spatial distribution of semantic association features according to structural constraints to enhance their correspondence with geographic features;

[0034] Through bidirectional iterative loops, the synergistic optimization of spatial detail features and semantic association features is achieved.

[0035] As a further aspect of the present invention: the specific process of processing the guided features according to the observation time phase category in the dynamic fusion layer is as follows:

[0036] Maintain separate quality assessment rule bases and dynamic weight calculation models for daytime and nighttime phases;

[0037] For daytime data, when calculating the fusion weights of spatial detail features and semantic association features, the quality assessment rule base includes assessment criteria for solar flare intensity and sea fog coverage.

[0038] For nighttime data, when calculating the fusion weights of spatial detail features and semantic association features, the quality assessment rule base includes assessment criteria for radiative cooling uniformity and nighttime cloud characteristics.

[0039] Based on the quality assessment results of each time phase, the dynamic weight calculation model of the corresponding time phase is invoked to generate feature fusion weight values ​​for the daytime and nighttime phases respectively.

[0040] As a further aspect of the present invention: in the dynamic fusion layer, the specific process of dynamically calculating the fusion weights of spatial detail features and semantic association features based on the real-time quality assessment results of the input data in spatial location is as follows:

[0041] Real-time monitoring of cloud cover index, signal-to-noise ratio, and deviation from background field in infrared sea surface temperature data; real-time monitoring of brightness temperature consistency, precipitation impact index, and land pollution level in microwave sea surface temperature data.

[0042] Based on the monitoring results, the confidence score of spatial detail features and the reliability score of semantic association features are calculated; based on the confidence score and reliability score, the fusion weights of spatial detail features and semantic association features are obtained by normalization using the sigmoid function.

[0043] Establish a weight smoothing adjustment mechanism. When a quality indicator undergoes a sudden change, the weight value is gradually adjusted using an exponential decay method.

[0044] As a further aspect of the present invention: In the fusion output layer, based on the fusion weights, the spatial detail features and semantic association features after bidirectional guidance are integrated to generate fusion features, and the specific process of improving the fusion features to the target spatial resolution is as follows:

[0045] The fusion weights of spatial detail features and semantic association features output by the dynamic fusion layer are weighted and multiplied by the corresponding bidirectional guided spatial detail features and semantic association features, respectively.

[0046] The weighted spatial detail features and semantic association features are added element by element to generate initial fused features. Multi-scale feature extraction is then performed on the initial fused features to obtain a feature pyramid containing information at different scales.

[0047] Subpixel convolution operations are used to upsample the features at each level of the feature pyramid, unifying all features to the target spatial resolution.

[0048] The multi-scale features after unifying the resolution are concatenated by channels, and then compressed and reconstructed by convolutional layers to generate the final high-resolution fused features.

[0049] The beneficial effects of this invention are:

[0050] This invention effectively addresses the limitations of existing technologies through multi-dimensional technological innovation. Addressing the issue of insufficient temporal adaptability, it precisely labels daytime and nighttime phases, employing independently trained parameter configurations and sensitive objective functions for separate processing. This allows the fusion process to fully adapt to the sensor response characteristics caused by differences in solar radiation and atmospheric stability between day and night, fundamentally avoiding systematic biases in the twilight transition zone and significantly improving data accuracy in time-specific scenarios. Facing the limitations of unidirectional feature interaction, an innovative bidirectional guidance mechanism achieves synergistic optimization of semantic association features and spatial detail features. Semantic features provide macroscopic guidance for supplementing and reconstructing spatial details, while spatial features enhance the geographical structural correspondence of semantic features. This bidirectional iterative cycle ensures that the fusion result retains both the fine texture of infrared data and reflects the macroscopic physical laws of microwave data, overcoming the information limitations of single-direction guidance. To address the issue of weak dynamic adjustment capabilities, the dynamic fusion layer constructs a dedicated quality assessment rule base for different time phases, generates dynamic weights by combining real-time data quality monitoring, and responds to quality fluctuations through a smoothing mechanism. This enables the fusion process to accurately adapt to data quality fluctuations under different sea areas and extreme weather conditions, ensuring that the output high-resolution fusion dataset has more reliable practical value in high-precision applications such as climate simulation and disaster early warning. Attached Figure Description

[0051] The invention will now be further described with reference to the accompanying drawings.

[0052] Figure 1 This is a system schematic diagram of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 As shown, this invention is a bidirectional adaptive multi-resolution guided sea surface temperature data fusion system, comprising:

[0055] The data receiving layer is responsible for data access and basic classification, receiving infrared and microwave sea surface temperature data transmitted from satellite sensors. This layer parses the precise observation timestamps accompanying the satellite data to obtain the absolute time information for each observation. It then calculates the local solar altitude angle by combining this with astronomical information about the data acquisition location. Based on this, it labels the observation time phase category for the data record. Observations with a solar altitude angle greater than or equal to zero are labeled as daytime phases, and observations with a solar altitude angle less than zero are labeled as nighttime phases. The time phase category information is then passed to the next layer as metadata associated with the corresponding data.

[0056] The feature decoupling layer performs targeted processing based on the temporal classification results, processing infrared sea surface temperature (SST) data and microwave SST data separately according to the observation temporal category. During processing, multi-scale analysis is performed on the infrared SST data to extract spatial gradient and texture change information, constructing a feature tensor representing local detail changes. Time-frequency transformation is performed on the microwave SST data to extract evolution patterns in the temporal dimension and distribution patterns in the spatial dimension, constructing a semantic feature tensor representing macroscopic physical processes. Then, convolution operations are used to separate the two types of features into different channels, achieving decoupling of spatial detail features and semantically related features.

[0057] The bidirectional guidance layer continues the temporal classification processing logic, processing the decoupled features according to the observation temporal category. This layer achieves bidirectional guidance simultaneously in the daytime and nighttime processing branches through a cross-modal information transfer mechanism: on the one hand, it uses semantic association features to establish an attention mapping mechanism, converting them into spatial detail supplementary information to reconstruct missing regions; on the other hand, it uses the edge and texture information of spatial detail features to establish a structural constraint mechanism, adjusting the spatial distribution of semantic association features to enhance geographic correspondence, and completing the collaborative optimization of the two types of features through bidirectional iterative loops.

[0058] The dynamic fusion layer continues to process the guided features by category, maintaining independent quality assessment rule bases and dynamic weight calculation models for daytime and nighttime phases respectively. During the assessment process, the cloud coverage index, signal-to-noise ratio, and background field deviation of infrared data, as well as the brightness-temperature consistency, precipitation impact index, and land pollution level of microwave data are monitored in real time. Based on these, the confidence and reliability scores of the two types of features are calculated, and fusion weights are generated after normalization. An exponential decay mechanism is used to cope with sudden changes in quality indicators, ensuring smooth and stable weight adjustment.

[0059] The fusion output layer integrates features based on the aforementioned dynamic weights, multiplying the weights by the corresponding guided spatial detail features and semantic association features, and generating initial fusion features through element-wise summation. Subsequently, a feature pyramid is constructed through multi-scale feature extraction, and sub-pixel convolution operations are used to unify features at each level to the target spatial resolution. After channel concatenation and convolutional layer compression optimization, high-resolution fusion features are finally generated, outputting sea surface temperature fusion datasets for daytime and nighttime phases respectively.

[0060] Specifically, in the data receiving layer, marking the observation phase category for data records includes:

[0061] First, the infrared and microwave sea surface temperature data transmitted by satellite sensors are preprocessed to analyze the precise observation timestamp attached to each data point. This timestamp uses Coordinated Universal Time (UTC) format, accurate to the millisecond level, and directly reflects the absolute time information of data acquisition, providing a basis for subsequent time phase determination.

[0062] Subsequently, calculations were performed using astronomical information from the data acquisition location. This astronomical information included the region's latitude and longitude coordinates, altitude, and Earth's rotation parameters, which were acquired synchronously via satellite ephemeris data or the positioning system of ground receiving stations. Based on the observation timestamp and the latitude and longitude of the data acquisition location, astronomical algorithms were used to calculate the local solar altitude angle. This calculation process comprehensively considered the effects of solar declination, hour angle, Earth's curvature, and atmospheric refraction to ensure the accuracy of the results. Atmospheric refraction correction reduced errors during low-altitude observations, keeping the accuracy of the solar altitude angle calculation within a preset range.

[0063] The time phase is classified based on the calculated solar altitude angle. When the solar altitude angle is greater than or equal to zero degrees, it indicates that the sun is above the local horizon, and the observation is marked as the daytime phase, covering the entire period from sunrise to sunset, including the dawn and twilight phases. When the solar altitude angle is less than zero degrees, the sun is below the local horizon, and the observation is marked as the nighttime phase, corresponding to the period from sunset to sunrise the next day.

[0064] After completing the time phase labeling, the observation time phase category information is integrated into metadata. In addition to the time phase category, the metadata also includes auxiliary information such as the observation timestamp, latitude and longitude of the data acquisition location, and sensor type, forming a complete dataset description system. Through a data encapsulation mechanism, the time phase category metadata is bound to the corresponding infrared and microwave sea surface temperature data, ensuring that the two remain correlated during data flow. Finally, the data receiving layer transmits the infrared and microwave sea surface temperature data, along with the time phase category metadata, to the subsequent feature decoupling layer via an internal data bus, providing a clear basis for the feature decoupling layer to process the data according to different time phase categories.

[0065] The process of processing infrared sea surface temperature data and microwave sea surface temperature data according to the observation time phase category in the feature decoupling layer is as follows:

[0066] The feature decoupling layer first receives infrared and microwave sea surface temperature data from the data receiving layer, both of which carry metadata about the observation time phase. The system extracts the time phase marker for each data point through the metadata parsing module. Based on the criteria for dividing daytime and nighttime phases, the two types of sensor data are separated into daytime and nighttime data subsets, respectively. During the separation process, a data grouping and indexing mechanism is employed to ensure that infrared and microwave data from the same time phase maintain a spatiotemporal correspondence, providing a collaborative foundation for subsequent feature extraction.

[0067] For the separated daytime temporal data subset, the feature decoupling layer uses the first set of parameter configurations for processing. This parameter configuration, generated through training on large-scale daytime temporal sample data, includes key parameters such as the hierarchical structure of the feature extraction network, convolutional kernel size, and activation function type, accurately adapting to the feature distribution patterns of infrared and microwave data during the daytime. In practice, the system uses this parameter configuration to separate spatial detail features from infrared sea surface temperature data and simultaneously extract semantic association features from microwave sea surface temperature data. For the nighttime temporal data subset, a second independent set of parameter configurations is used. This configuration is trained based on nighttime observation data, and its parameter settings fully consider factors such as nighttime sea surface radiation characteristics and atmospheric interference patterns, specifically capturing infrared spatial details and microwave semantic associations in nighttime scenes.

[0068] The generation of the parameter configurations relies on a dual-temporal independent training mechanism. During training, labeled datasets for daytime and nighttime are constructed separately, containing measured sea surface temperature values ​​and corresponding sensor observation data under different sea areas and weather conditions. By inputting the labeled data into a feature decoupling network, a temporally sensitive objective function is used for iterative optimization: the objective function introduces a temporal weight factor in the loss calculation, emphasizing spatial detail fidelity constraints for daytime data and strengthening temporal continuity constraints for semantic features for nighttime data, ultimately enabling the two sets of parameter configurations to achieve optimal adaptation to their respective temporal data.

[0069] In the feature decoupling layer, the process of separating the spatial detail features of infrared sea surface temperature data and the semantic association features of microwave sea surface temperature data is as follows:

[0070] The feature decoupling layer employs a differentiated processing strategy. For infrared sea surface temperature data, the system executes a multi-scale analysis process: first, the data is mapped to different spatial scales through Gaussian pyramid decomposition; at each scale, gradient operators are used to calculate spatial gradient information in the horizontal and vertical directions; simultaneously, a local binary mode algorithm is used to extract texture change features, including the directional distribution of temperature gradients and the density of texture units. Based on these extraction results, the system integrates parameters such as gradient magnitude and texture entropy into a multi-dimensional array according to spatial location, constructing a feature tensor representing local detail changes. This tensor can accurately reflect the small- and medium-scale structures in the sea surface temperature field, such as eddy boundaries and temperature fronts.

[0071] For microwave sea surface temperature data, a time-frequency transformation process is implemented in the feature decoupling layer. The system first segments the time-series microwave observation data into a sliding window, then uses a short-time Fourier transform to convert the data within each window to the time-frequency domain, extracting the energy intensity and phase information corresponding to different frequency components. This captures the evolution patterns of sea surface temperature over time, such as diurnal variation cycles and temperature trends. Spatially, by calculating the mean, variance, and spatial autocorrelation coefficient of temperature in different regions, the distribution patterns of the temperature field are analyzed, identifying macroscopic features such as large-scale ocean currents and temperature anomalies. Based on the time-frequency domain features and spatial statistical characteristics, the system constructs a semantic feature tensor representing macroscopic physical processes. This tensor reflects the overall evolution and physical correlation of the sea surface temperature field.

[0072] After constructing the feature tensors, the feature decoupling layer separates the channels of the two types of features through multiple convolutional operations. The convolutional layers employ a grouped convolution mechanism, allocating spatial detail features and semantically related features to different feature channels for independent storage. Specifically, the spatial detail features of infrared data are mapped to high-dimensional channels, preserving rich local texture information; while the semantically related features of microwave data occupy another set of channels, focusing on the representation of macroscopic physical laws. Through this channel separation design, the two types of features are effectively decoupled, laying the data foundation for the subsequent bidirectional guidance process.

[0073] In the bidirectional guidance layer, the process of processing spatial detail features and semantic association features according to the observation time phase category is as follows:

[0074] This layer first receives two types of feature data from the feature decoupling layer. By parsing the observation time-phase metadata carried in the data, a feature routing mechanism is established. Specifically, the system performs time-phase label recognition on each set of input features, automatically assigning spatial detail features and semantic association features labeled as daytime phases to the daytime processing branch, and assigning the corresponding features labeled as nighttime phases to the nighttime processing branch. This branched processing architecture ensures that features from different time phases can be optimized in computational paths adapted to their physical characteristics, avoiding interference caused by differences in day and night features.

[0075] In the daytime temporal processing branch, the core operation is to use the semantic association features of the daytime temporal phase to guide the information supplementation and reconstruction of spatial detail features. This process is achieved through a cross-modal information transfer mechanism: the semantic association features are first adjusted in dimension to match the scale of the spatial detail features, and then a correlation mapping between the two is established through a feature interaction matrix. With the help of prior physical laws in the daytime scene (such as the influence pattern of solar radiation on sea surface temperature distribution), the large-scale temperature gradient trend contained in the semantic features is transformed into the basis for repairing spatial details, focusing on filling in local information loss areas in infrared data caused by cloud obstruction or sensor noise.

[0076] The nighttime temporal processing branch employs a similar cross-modal information transmission mechanism, but adapts to the evolution characteristics of sea surface temperature at night. Nighttime semantic association features reflect the uniformity of temperature distribution under radiative cooling effects. When the system uses this characteristic to guide the reconstruction of spatial detail features, it focuses on repairing infrared data anomalies caused by atmospheric inversion layers, ensuring that the supplemented detail information conforms to the heat conduction patterns of the ocean at night.

[0077] In the two-phase processing branches, spatial structure enhancement of semantically related features is performed simultaneously using spatial detail features. Fine geographic markers such as coastlines and ocean current boundaries contained in the spatial detail features are extracted as structural reference signals to calibrate the ambiguity of the spatial distribution of semantically related features, making the representation of macroscopic physical processes more consistent with the actual geographic scene.

[0078] In the bidirectional guidance layer, the specific process of using semantic association features to guide spatial detail features for information supplementation and reconstruction, and using spatial detail features to guide semantic association features for spatial structure enhancement, is as follows:

[0079] In guiding spatial detail features with semantic association features, the system first establishes an attention mapping mechanism: by calculating the similarity between semantic features and spatial features in high-dimensional space, point-by-point guiding weights are generated. The higher the weight value, the more the spatial details at that location need to rely on semantic information for correction. Based on this weight distribution, semantic association features are mapped into supplementary information of the same dimension as spatial detail features through a feature transformation network. This information will specifically fill in the missing areas in the spatial features, while retaining the original reliable detail data, forming a "macro-guided micro" repair logic.

[0080] In guiding semantic association features with spatial detail features, the structural constraint mechanism plays a core role. The system extracts strong gradient regions (such as the boundaries between warm and cold water masses) from spatial detail features using edge detection algorithms, and uses these regions as spatial anchor points to construct a constraint matrix for semantic features. Based on the constraint matrix, the spatial distribution of semantic association features is fine-tuned: the steepness of the temperature gradient is enhanced near the anchor points, while the smoothness of the temperature field is maintained in uniform regions, thereby strengthening the correspondence between semantic features and actual geographical features.

[0081] Through bidirectional iterative loops, the two types of features are progressively optimized. In each iteration, the missing regions of spatial detail features are first updated based on the current semantic features, and then the spatial structure of the semantic features is adjusted according to the optimized spatial features. This process is repeated until the feature change is below a preset threshold. This collaborative optimization mode avoids information bias caused by a single feature dominance, while fully leveraging the detail advantages of infrared data and the global advantages of microwave data, providing high-quality feature input for subsequent dynamic fusion.

[0082] The specific process of processing the guided features according to the observation time phase category in the dynamic fusion layer is as follows:

[0083] Independent quality assessment rule bases and dynamic weight calculation models are built and maintained for daytime and nighttime phases, respectively. The two types of model systems are completely independent in parameter settings and assessment logic to adapt to the data characteristics of different time phases.

[0084] For daytime data, the quality assessment rule base primarily incorporates assessment criteria for solar flare intensity and sea fog coverage. Solar flare intensity assessment analyzes the distribution of high-brightness areas caused by direct sunlight in infrared data to determine the degree of interference with sea surface temperature observations. Sea fog coverage assessment relies on the collaborative verification of infrared and microwave data to determine the spatial proportion of cloud and fog obscured areas and their impact on data integrity. These criteria collectively constitute the basis for judging the quality of daytime data.

[0085] For nighttime data, the quality assessment rule base includes assessment criteria for radiative cooling uniformity and nighttime cloud characteristics. Radiative cooling uniformity assessment analyzes the spatial gradient of sea surface temperature at night to determine if there are localized temperature anomalies caused by differences in atmospheric stability. Nighttime cloud characteristics assessment combines the penetration capability of microwave data to identify potential interference from thin clouds, high clouds, etc., on infrared observations. These criteria provide a targeted framework for assessing the quality of nighttime data.

[0086] In practical processing, the dynamic fusion layer first calls the corresponding quality assessment rule base according to the temporal category of the data to perform a comprehensive quality assessment on the guided features. Then, based on the assessment results, it triggers the dynamic weight calculation model for the corresponding temporal phase to generate feature fusion weight values ​​for daytime and nighttime phases respectively, ensuring that the weight calculation matches the temporal characteristics of the data.

[0087] In the dynamic fusion layer, the specific process of dynamically calculating the fusion weights of spatial detail features and semantic association features based on the real-time quality assessment results of the input data in spatial location is as follows:

[0088] First, multi-dimensional real-time monitoring of infrared sea surface temperature data is performed, including cloud cover index, signal-to-noise ratio, and deviation from the background field. The cloud cover index reflects the proportion of pixels in the infrared data that are obscured by clouds, the signal-to-noise ratio reflects the relative intensity of the data signal and noise, and the deviation from the background field measures the degree of deviation of the observed value from the historical average field.

[0089] Simultaneously, the system conducts real-time monitoring of microwave sea surface temperature data, focusing on brightness-temperature consistency, precipitation impact index, and land pollution level. Brightness-temperature consistency is used to determine the inherent coordination between observations from different microwave channels, the precipitation impact index assesses the intensity of interference from precipitation processes on microwave detection, and land pollution level quantifies the extent of nearshore terrestrial radiation contamination of microwave data.

[0090] Based on the above monitoring results, the dynamic fusion layer calculates the confidence score of spatial detail features and the reliability score of semantic association features. The confidence score of spatial detail features integrates the monitoring indicators of infrared data to reflect the credibility of local detail information; the reliability score of semantic association features combines the monitoring results of microwave data to assess the reliability of the macroscopic physical process representation.

[0091] Subsequently, the confidence score and reliability score are normalized using the sigmoid function to obtain the fusion weights of spatial detail features and semantic association features, respectively, so that the two types of weight values ​​are in a unified numerical range, which facilitates subsequent fusion calculation.

[0092] To address sudden changes in data quality, the dynamic fusion layer establishes a weight smoothing adjustment mechanism. When a sudden change in quality indicators is detected, the system gradually adjusts the weight values ​​using an exponential decay method to avoid drastic fluctuations in weights from impacting the fusion results, thus ensuring the stability of the fusion process and the continuity of the output data.

[0093] In the fusion output layer, based on the fusion weights, spatial detail features and semantic association features after bidirectional guidance are integrated to generate fusion features. The specific process of improving the fusion features to the target spatial resolution is as follows:

[0094] First, the weight modulation mechanism is activated, calling the spatial detail feature fusion weights output by the dynamic fusion layer and performing position-wise weighted multiplication with the spatial detail features processed by bidirectional guidance; at the same time, the semantic association feature fusion weights are performed with the corresponding bidirectional guided semantic association features, thereby strengthening the dominant position of high reliability features in the fusion process through weight allocation.

[0095] After weighting, the two types of features are added element-wise to form an initial fused feature that combines local texture details with macroscopic physical laws. Then, a multi-scale feature extraction process is initiated. By setting feature extraction windows of different sizes, multi-dimensional information ranging from centimeter-level micro-temperature gradients to kilometer-level regional temperature distribution is captured. A feature pyramid is constructed that includes bottom-level pixel-level details, mid-level structural trends, and high-level physical semantics. Each level of features corresponds to the variation features at different scales in the sea surface temperature field.

[0096] For feature data of different resolutions in the feature pyramid, subpixel convolution is used for upsampling. This operation accurately transforms low-resolution features to the target spatial resolution dimension by rearranging and mapping feature pixels, while maintaining the spatial topological relationships and physical correlations within the features. This ensures that feature data at all levels are unified in spatial scale, laying a spatial consistency foundation for subsequent feature integration.

[0097] The unified resolution multi-scale features encompass multi-level information, ranging from microscopic temperature gradients to macroscopic distribution patterns. The bottom-level features retain the fine texture details of infrared data, the mid-level features carry structural information such as ocean current boundaries, and the top-level features contain the semantics of physical processes reflected in microwave data. When these features are stitched together along the channel dimension, a feature alignment mechanism ensures precise spatial matching of features at different scales, forming a high-dimensional feature tensor containing rich hierarchical information. Each channel corresponds to a specific scale of feature representation, collectively constituting a multi-dimensional description of the sea surface temperature field.

[0098] After concatenation, the feature tensors are processed in convolutional layers. First, feature compression is achieved using multiple sets of convolutional kernels, reducing the number of channels and data redundancy without losing key information. Simultaneously, different receptive field sizes are used to enhance the correlation between features. Then, the reconstruction process begins, using adaptive weight adjustments to strengthen the expression intensity of effective features, such as assigning higher weights to features in key regions like the edges of ocean current eddies. At the same time, noise interference is suppressed, irrelevant information such as cloud reflections is smoothed, and minor deviations in local areas are corrected.

[0099] After compression and refinement optimization of the convolutional layers, the feature tensor achieves deep fusion of multi-scale information while maintaining high spatial resolution, ultimately generating fused features with both high spatial resolution and physical consistency, providing core data support for outputting sea surface temperature fusion datasets for daytime and nighttime phases respectively.

[0100] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A bidirectional adaptive multi-resolution guided sea surface temperature data fusion system, characterized in that, include: The data receiving layer is used to receive infrared and microwave sea surface temperature data transmitted by satellite sensors and to mark the observation time phase category for data recording, including daytime and nighttime phases. The feature decoupling layer is used to process infrared sea surface temperature data and microwave sea surface temperature data according to the observation time phase category, and to separate the spatial detail features of infrared sea surface temperature data and the semantic association features of microwave sea surface temperature data. A bidirectional guidance layer is used to process spatial detail features and semantic association features according to the observation time phase category. The semantic association features are used to guide the spatial detail features to supplement and reconstruct information, while the spatial detail features are used to guide the semantic association features to enhance the spatial structure. The dynamic fusion layer is used to process the guided features according to the observation time phase category. Based on the real-time quality assessment results of the input data in spatial location, it dynamically calculates the fusion weights of spatial detail features and semantic association features. The fusion output layer is used to generate fusion features by integrating bidirectional guided spatial detail features and semantic association features based on fusion weights, improve the fusion features to the target spatial resolution, and output sea surface temperature fusion datasets for daytime and nighttime phases respectively. In the bidirectional guidance layer, the process of processing spatial detail features and semantic association features according to the observation time phase category is as follows: Spatial detail features and semantic association features from the feature decoupling layer are routed to different processing branches based on the observation time phase of the data source. In the daytime phase processing branch, the semantic association features of the daytime phase are used to guide the spatial detail features of the daytime phase for information supplementation and reconstruction. This process is achieved through a cross-modal information transmission mechanism. In the nighttime phase processing branch, the semantic association features of the nighttime phase are used to guide the spatial detail features of the nighttime phase for information supplementation and reconstruction. This process is also achieved through the cross-modal information transmission mechanism. In the two-phase processing branches, spatial detail features are used simultaneously to guide semantic association features for spatial structure enhancement; In the bidirectional guidance layer, the specific process of using semantic association features to guide spatial detail features for information supplementation and reconstruction, and simultaneously using spatial detail features to guide semantic association features for spatial structure enhancement, is as follows: An attention mapping mechanism is established to transform semantic association features into spatial detail features. The guiding weight of semantic features on spatial features is calculated. Based on the guiding weight, semantic association features are converted into spatial detail supplementary information to reconstruct missing regions. Establish a structural constraint mechanism from spatial detail features to semantic association features, extract edge and texture information of spatial features, and adjust the spatial distribution of semantic association features according to structural constraints to enhance their correspondence with geographic features; Through bidirectional iterative loops, the synergistic optimization of spatial detail features and semantic association features is achieved.

2. The bidirectional adaptive multi-resolution guided sea surface temperature data fusion system according to claim 1, characterized in that, In the data receiving layer, marking the observation phase category for data records specifically includes: The system analyzes the precise observation timestamps attached to satellite data to obtain the absolute time information for each observation. Based on the observation timestamps and the astronomical information of the data acquisition location, it calculates the local solar altitude angle and marks observations with a solar altitude angle greater than or equal to zero as daytime phases and observations with a solar altitude angle less than zero as nighttime phases. The observation phase category information is associated with the corresponding infrared sea surface temperature data and microwave sea surface temperature data as metadata, and this category information is passed to the subsequent processing level along with the data.

3. The bidirectional adaptive multi-resolution guided sea surface temperature data fusion system according to claim 1, characterized in that, In the feature decoupling layer, the process of processing infrared sea surface temperature data and microwave sea surface temperature data according to the observation time phase is as follows: The received infrared sea surface temperature data and microwave sea surface temperature data are separated into daytime time phase data subsets and nighttime time phase data subsets based on the observation time phase category metadata carried in the data; For a subset of daytime temporal data, a set of parameter configurations trained on daytime temporal data is used to separate the spatial detail features of infrared sea surface temperature data and extract the semantic association features of microwave sea surface temperature data. For a subset of nighttime temporal data, another set of parameter configurations trained on the nighttime temporal data is used to separate the spatial detail features of infrared sea surface temperature data and extract the semantic association features of microwave sea surface temperature data. The parameters are obtained through independent training on labeled data at different time phases, and the training process uses a time-sensitive objective function.

4. The bidirectional adaptive multi-resolution guided sea surface temperature data fusion system according to claim 3, characterized in that, In the feature decoupling layer, the process of separating the spatial detail features of infrared sea surface temperature data and the semantic association features of microwave sea surface temperature data is as follows: Multi-scale analysis of infrared sea surface temperature data is performed to extract spatial gradient information and texture change information at different scales. Based on the extracted spatial gradient information and texture change information, a feature tensor representing local detail changes is constructed. Time-frequency transformation is performed on microwave sea surface temperature data to extract its evolution pattern in the time dimension and its distribution pattern in the spatial dimension. Based on the extracted time-frequency features, a semantic feature tensor characterizing macroscopic physical processes is constructed. Through the convolutional operation of the feature decoupling layer, spatial detail features and semantically related features are separated into different feature channels.

5. The bidirectional adaptive multi-resolution guided sea surface temperature data fusion system according to claim 1, characterized in that, In the dynamic fusion layer, the specific process of processing the guided features according to the observation time phase category is as follows: Maintain separate quality assessment rule bases and dynamic weight calculation models for daytime and nighttime phases; For daytime data, when calculating the fusion weights of spatial detail features and semantic association features, the quality assessment rule base includes assessment criteria for solar flare intensity and sea fog coverage. For nighttime data, when calculating the fusion weights of spatial detail features and semantic association features, the quality assessment rule base includes assessment criteria for radiative cooling uniformity and nighttime cloud characteristics. Based on the quality assessment results of each time phase, the dynamic weight calculation model of the corresponding time phase is invoked to generate feature fusion weight values ​​for the daytime and nighttime phases respectively.

6. The bidirectional adaptive multi-resolution guided sea surface temperature data fusion system according to claim 5, characterized in that, In the dynamic fusion layer, the specific process of dynamically calculating the fusion weights of spatial detail features and semantic association features based on the real-time quality assessment results of the input data in spatial location is as follows: Real-time monitoring of cloud cover index, signal-to-noise ratio, and deviation from background field in infrared sea surface temperature data; real-time monitoring of brightness temperature consistency, precipitation impact index, and land pollution level in microwave sea surface temperature data. Based on the monitoring results, the confidence score of spatial detail features and the reliability score of semantic association features are calculated; based on the confidence score and reliability score, the fusion weights of spatial detail features and semantic association features are obtained by normalization using the sigmoid function. Establish a weight smoothing adjustment mechanism. When a quality indicator undergoes a sudden change, the weight value is gradually adjusted using an exponential decay method.

7. The bidirectional adaptive multi-resolution guided sea surface temperature data fusion system according to claim 1, characterized in that, In the fusion output layer, based on the fusion weights, spatial detail features and semantic association features after bidirectional guidance are integrated to generate fusion features. The specific process of improving the fusion features to the target spatial resolution is as follows: The fusion weights of spatial detail features and semantic association features output by the dynamic fusion layer are weighted and multiplied by the corresponding bidirectional guided spatial detail features and semantic association features, respectively. The weighted spatial detail features and semantic association features are added element by element to generate initial fused features. Multi-scale feature extraction is then performed on the initial fused features to obtain a feature pyramid containing information at different scales. Subpixel convolution operations are used to upsample the features at each level of the feature pyramid, unifying all features to the target spatial resolution. The multi-scale features after unifying the resolution are concatenated, and then the features are compressed and reconstructed through convolutional layers to generate the final high-resolution fused features.

Images