Gobi desert surface temperature prediction method based on deep learning

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a gobi desert surface temperature prediction method based on deep learning. According to the method, a system architecture in which an acquisition end, a remote sensing end, an edge computing terminal and a central server cooperatively operate is constructed, and fusion and real-time transmission of multi-source observation data are realized. By designing a space-time neural network model fused with physical constraints, combining an angle-guided convolutional neural network and a gated cycle unit structure, and introducing surface energy balance and Stefan-Boltzmann radiation consistency constraints, high-precision prediction and physical consistency modeling of surface temperature are realized. And a cross-scale reconstruction and missing detection complementation mechanism is adopted, so that the spatial continuity and resolution of a prediction result are improved. According to the method, high-precision, stable and self-adaptive prediction of the surface temperature of the gobi desert area is realized, and intelligent support is provided for meteorological monitoring, ecological evaluation and early warning decision making of the arid area.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for predicting surface temperature in the Gobi Desert based on deep learning. Background Technology

[0002] The Gobi Desert region is located in an arid continental climate zone with sparse vegetation, high albedo, and large diurnal temperature variations. Surface thermal processes are influenced by multiple factors, including solar radiation, energy balance, and soil heat flux. Land surface temperature (LST) is a crucial parameter reflecting energy exchange between the Earth and atmosphere and changes in the ecological environment, playing a vital role in climate monitoring, drought early warning, and energy balance research. However, meteorological stations are sparse in the Gobi Desert, making it difficult for traditional ground-based observations to achieve large-scale, high spatiotemporal resolution coverage. Although the development of remote sensing technology has made it possible to retrieve LST using satellite or UAV thermal infrared imagery, the results are limited by long observation periods, frequent cloud and dust interference, and limited spatial resolution, resulting in insufficient continuity and accuracy in single-source remote sensing retrieval. Existing LST prediction methods largely rely on experience or physical models (such as single-channel methods and split-window methods), which exhibit significant errors under conditions of high albedo and strong thermal convection, and struggle to characterize nonlinear spatiotemporal variations. While deep learning methods have been introduced for LST estimation, they primarily focus on pixel-level predictions from single-source images, lacking multi-source fusion and spatiotemporal correlation modeling. In addition, communication conditions in the Gobi Desert are limited, and traditional cloud computing models are unable to support high-frequency prediction and early warning, resulting in insufficient system real-time performance and stability. Summary of the Invention

[0003] This invention addresses the problems of low accuracy, poor spatiotemporal continuity, and insufficient real-time performance in surface temperature prediction under the extreme conditions of aridity, sparse vegetation, high albedo, and large diurnal temperature range in the Gobi Desert. It proposes a deep learning-based method for predicting surface temperature in the Gobi Desert. This method constructs a collaborative architecture involving a data acquisition unit, a remote sensing unit, an edge computing terminal, and a central server. This architecture enables the fusion processing of ground observations and multi-source remote sensing data, and achieves real-time inference and online adaptive updates of the model under communication-constrained conditions through an edge-cloud collaborative mechanism. The core of the method lies in designing a spatiotemporal neural network model that incorporates physical constraints. This model combines angle-guided convolutional neural networks with gated recurrent unit structures, introducing surface energy balance and Stefan-Boltzmann radiation consistency constraints to ensure that the model's prediction results conform to the natural law of energy conservation. Simultaneously, a cross-scale reconstruction and missing data completion mechanism is proposed. This mechanism utilizes the joint reconstruction of coarse-resolution thermal infrared data with high-resolution albedo, topography, and texture features to achieve high spatiotemporal resolution surface temperature field recovery. Furthermore, an angle-guided activation module is introduced into the convolutional layer to enhance the model's responsiveness to surface directional features (such as slope and albedo gradient). This method achieves high-precision, physically consistent, and adaptive prediction of surface temperature in the Gobi Desert region, providing reliable intelligent support for monitoring surface thermal processes and meteorological early warning under extreme environments.

[0004] This invention provides a deep learning-based method for predicting land surface temperature (LST) in the Gobi Desert. This method is applied to the spatiotemporal prediction of LST in arid, sparsely vegetated, high-albedo, and diurnal temperature range environments in the Gobi Desert. The method operates within an architecture that integrates data acquisition, remote sensing, edge computing terminals, and a central server. The architecture includes:

[0005] The acquisition terminal is used to collect ground observation data in the Gobi Desert area;

[0006] The remote sensing terminal is used to acquire multi-source remote sensing data covering the target area.

[0007] Edge computing terminals are used to perform real-time preprocessing and quality control on data from the acquisition and remote sensing ends locally, and establish a two-way connection with the central server through the LoRaWAN wireless communication network.

[0008] The edge computing terminal is an industrial-grade box-type host with a built-in quad-core central processing unit (CPU), 8GB of memory and 512GB of solid-state storage. It integrates an embedded GPU module with CUDA acceleration capability, a GNSS high-precision timing module, and is equipped with an industrial Ethernet interface and a cellular communication module. It meets the operating temperature range of −20℃ to +55℃ and IP65 protection requirements. It is equipped with a photovoltaic-battery integrated power supply module and can operate stably for a long time in the extreme environment of high temperature, strong wind and sand in the Gobi Desert.

[0009] The central server receives multi-source data from edge computing terminals after real-time preprocessing and quality control. It performs radiometric correction, atmospheric correction, and surface emissivity estimation on the multi-source data to reconstruct a high-precision surface temperature field. The central server uses a GPU cluster to perform multi-source feature fusion, spatiotemporal neural network model training and inference, cross-scale reconstruction, bias correction, and parameter optimization. The optimized model parameters are then distributed to the edge computing terminals in real time via LoRaWAN, forming a closed-loop temperature prediction system that is collaborative between the edge and the cloud.

[0010] The method includes the following steps:

[0011] Step S1: Process ground observation data and multi-source remote sensing data to obtain multi-source fused input data;

[0012] Step S2: The central server performs atmospheric correction and surface emissivity estimation on the thermal infrared band image in the multi-source fusion input data to obtain the initial surface temperature data;

[0013] Step S3: The central server merges the initial surface temperature data and multi-source fusion input data to construct multi-source spatiotemporal feature data with unified resolution;

[0014] Step S4: Construct a spatiotemporal neural network model for land surface temperature prediction on the central server; use multi-source spatiotemporal feature data as input to the spatiotemporal neural network model to simultaneously learn the spatial distribution and temporal evolution of land surface temperature; introduce a coupling loss function to train the spatiotemporal neural network model to constrain the predicted values ​​to satisfy the laws of energy conservation and radiation, resulting in a trained spatiotemporal neural network model; perform inference on the trained spatiotemporal neural network model on the central server to output preliminary land surface temperature prediction results and pixel-level uncertainty assessment; the spatiotemporal neural network model includes an angle-guided CNN model and a GRU model;

[0015] Step S5: In the central server, determine the availability of low-resolution thermal infrared data in the preliminary surface temperature prediction results; jointly reconstruct the temperature dynamics information of coarse-resolution thermal infrared data in the preliminary surface temperature prediction results with high-resolution albedo, topography and texture features to achieve surface temperature overlay within the 30–100m resolution range, and obtain the reconstructed surface temperature prediction results; for the missing areas in the reconstructed surface temperature prediction results caused by cloud cover, dust obstruction and observation gaps, use mask convolution and time series graph regular interpolation methods to complete the data and obtain high-precision surface temperature prediction data.

[0016] Step S6: In the central server, using the measured surface temperature retrieved from the ground infrared radiometer deployed in the Gobi Desert as a reference, and combining it with high-precision surface temperature prediction data, the output of the spatiotemporal neural network model is subjected to bias correction and small sample migration fine-tuning to obtain the corrected spatiotemporal neural network model; the edge computing terminal performs incremental updates based on the corrected spatiotemporal neural network model, adaptively adjusts the model parameters online, and periodically uploads the data back to the central server; finally, the surface temperature prediction result is obtained.

[0017] Furthermore, step S4 specifically includes the following steps:

[0018] Step S41: Use an angle-guided CNN model to extract features from multi-source spatiotemporal feature data to obtain spatial feature data; the construction method of the angle-guided CNN model is as follows: build a CNN model, introduce an angle-guided module to optimize the activation of the convolutional layer of the CNN model, so that the model can capture the target spatial features more accurately, and obtain the angle-guided CNN model.

[0019] Step S42: Use the GRU model to extract features from multi-source spatiotemporal feature data to obtain temporal feature data;

[0020] Step S43: Input spatial and temporal feature data into the feature fusion layer for joint modeling, and perform prediction through a fully connected layer; introduce a dual physical constraint term based on the surface energy balance equation and the Stefan-Boltzmann law, and construct a coupling loss function by combining it with a time consistency regularization term, and train the angle-guided CNN model and GRU model to obtain the trained spatiotemporal neural network model; the trained spatiotemporal neural network model performs inference on the central server and outputs preliminary surface temperature prediction results and pixel-level uncertainty assessment.

[0021] Furthermore, step S41 specifically includes the following steps:

[0022] Step S411: Perform time axis registration, spatial resolution unification and normalization on the multi-source spatiotemporal feature data to obtain standardized multi-source feature data;

[0023] Step S412: Adjust the activation direction of the standardized multi-source feature data through the angle guidance module to obtain rotation-enhanced convolution activation features;

[0024] Step S413: Input the rotation-enhanced convolution activation features into the spatial pooling layer of the angle-guided CNN model, aggregate the local responses, and generate spatial feature data through feature integration processing.

[0025] Furthermore, step S412 specifically includes the following steps:

[0026] Step S4121: Extract the normalized activation vectors of the convolutional layer from the labeled samples (such as abnormal and normal regions) in the standardized multi-source feature data, calculate the candidate directions using the mean difference method, perform principal component analysis on the candidate direction set, extract the principal components, and define a two-dimensional rotation subspace.

[0027] Step S4122: Determine the rotation matrix, perform geometric rotation in the two-dimensional rotation subspace, adjust the directionality of the normalized activation vector of the convolutional layer, and obtain the feature direction of the activation vector;

[0028] Step S4123: Introduce an adaptive angle guidance mechanism to calculate a conditional mask based on the projection of the activation vector in the feature direction, and perform selective rotation based on the mask; this mechanism ensures that only activations consistent with the target direction are controlled, thereby improving the robustness and controllability of the network in complex scenes; and obtain rotation-enhanced convolution activation features.

[0029] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0030] This invention, through the construction of a collaborative architecture involving a data acquisition terminal, a remote sensing terminal, an edge computing terminal, and a central server, achieves the fusion acquisition and real-time transmission of multi-source observation data in the Gobi Desert region, improving the spatiotemporal integrity and data reliability of surface temperature prediction. The edge computing terminal possesses local preprocessing and inference capabilities, enabling independent operation under communication constraints, and achieving data quality control and preliminary prediction. The central server is responsible for model training, reconstruction, and parameter optimization, and uses the LoRaWAN communication network to achieve incremental synchronization and backhaul updates of model parameters. This edge-cloud collaborative mechanism effectively solves problems such as unstable communication, high network latency, and difficulties in data upload in the Gobi Desert region, enhancing the system's long-term continuous operation and real-time response capabilities in extremely arid and windy environments.

[0031] This invention designs a spatiotemporal neural network model that integrates physical constraints, achieving energy conservation and radiation consistency constraints in the land surface temperature prediction process, significantly improving the physical rationality and numerical stability of the model's predictions. By introducing surface energy balance constraints and Stefan-Boltzmann radiation constraints during model training, the prediction results not only numerically match the observed data but also maintain consistency with physical laws, thus solving the problem of non-physical biases easily generated by traditional deep learning methods. Combining an angle-guided convolutional neural network with a gated recurrent unit joint structure, this invention achieves simultaneous modeling of the spatial distribution and temporal evolution of land surface temperature, significantly improving the ability to characterize the complex terrain and thermal processes of the Gobi Desert, and providing accurate support for dynamic prediction of the regional thermal environment.

[0032] Furthermore, the cross-scale reconstruction and missing data completion mechanism proposed in this invention effectively solves the common problems of cloud cover, dust obstruction, and observation gaps in multi-source remote sensing images of the Gobi Desert, improving the spatial continuity and prediction resolution of the surface temperature field. By jointly reconstructing coarse-resolution thermal infrared images with high-resolution albedo, topographic, and texture features, the temperature prediction results are pushed down from the hundred-meter scale to a finer scale. The masked convolution and temporal map regularized interpolation algorithms provide physically consistent temperature completion for missing areas, significantly enhancing the spatiotemporal consistency and reliability of the prediction results. Overall, this invention achieves high-precision, physically consistent, and adaptive prediction of surface temperature in the Gobi Desert, providing valuable data support and intelligent decision-making basis for regional meteorological early warning, ecological environment assessment, and energy development planning. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the architecture of a deep learning-based method for predicting surface temperature in the Gobi Desert proposed in this invention.

[0034] Figure 2 This is a flowchart illustrating a deep learning-based method for predicting surface temperature in the Gobi Desert, as proposed in this invention. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0036] Example 1, according to Figure 1 , Figure 2 This invention provides a deep learning-based method for predicting land surface temperature (LST) in the Gobi Desert. This method is applied to the spatiotemporal prediction of LST in arid, sparsely vegetated, high-albedo, and diurnal temperature range environments in the Gobi Desert. The method operates within an architecture that integrates a data acquisition unit, a remote sensing unit, an edge computing terminal, and a central server. The architecture includes:

[0037] The data acquisition unit is used to collect ground observation data in the Gobi Desert area. It consists of two parts: an automatic weather station and surface measurement stakes. The automatic weather station is equipped with air temperature and humidity sensors, wind speed and direction sensors, total radiation and net radiation sensors, and air pressure sensors to collect atmospheric environmental data. The surface measurement stakes are set up in typical landform sample areas and integrate downward-looking infrared radiometers and multi-layer (0–20cm) soil temperature and humidity probes to monitor the surface radiance and temperature and the thermal and moisture status of shallow soil in real time.

[0038] The remote sensing end is used to acquire multi-source remote sensing data covering the target area, including thermal infrared satellite remote sensing images and visible-near infrared satellite remote sensing images acquired by satellite payloads, as well as long-wave infrared thermal image data acquired by UAV aerial surveys. The UAV aerial survey thermal imager has imaging capabilities in the 8–14 μm band. It achieves high-precision registration with satellite images through POS positioning and image beam adjustment. By collaboratively fusing long-wave infrared thermal image data with satellite remote sensing images, it provides high spatiotemporal resolution supplementary information support for surface temperature inversion and subsequent spatiotemporal prediction models.

[0039] Edge computing terminals are used to perform real-time preprocessing and quality control on data from the acquisition and remote sensing ends locally, and establish a two-way connection with the central server through the LoRaWAN wireless communication network.

[0040] The edge computing terminal is an industrial-grade box-type host with a built-in quad-core central processing unit (CPU), 8GB of memory and 512GB of solid-state storage. It integrates an embedded GPU module with CUDA acceleration capability, a GNSS high-precision timing module, and is equipped with an industrial Ethernet interface and a cellular communication module. It meets the operating temperature range of −20℃ to +55℃ and IP65 protection requirements. It is equipped with a photovoltaic-battery integrated power supply module and can operate stably for a long time in the extreme environment of high temperature, strong wind and sand in the Gobi Desert.

[0041] The central server receives multi-source data from edge computing terminals after real-time preprocessing and quality control. It performs radiometric correction, atmospheric correction, and surface emissivity estimation on the multi-source data to reconstruct a high-precision surface temperature field. The central server uses a GPU cluster to perform multi-source feature fusion, spatiotemporal neural network model training and inference, cross-scale reconstruction, bias correction, and parameter optimization. The optimized model parameters are then distributed to the edge computing terminals in real time via LoRaWAN to achieve incremental updates and inference deployment of the model at the edge, forming a closed-loop temperature prediction system that is collaborative between the edge and the cloud.

[0042] The central server includes: GPU-accelerated training nodes with 24GB of video memory, distributed object storage and spatiotemporal database, and a containerized service environment for implementing deep learning model training, batch processing and reconstruction of remote sensing data, and service-oriented publishing of prediction results.

[0043] The method includes the following steps:

[0044] Step S1: Data Synchronization and Quality Control: Ground observation data and multi-source remote sensing data are time-axis aligned using GNSS timing signals to ensure that ground observations and remote sensing images match at the same time. Radiometric calibration, cloud and dust identification masking, geometric fine correction, and projection unification are performed on the multi-source remote sensing data to ensure consistency of radiometric physical quantities and spatial coordinates. Sensor anomaly detection, missing data interpolation, and noise removal are performed on the ground observation data to obtain continuous and effective ground observation time series. Finally, spatiotemporally synchronized and reliable multi-source fusion input data is generated.

[0045] Step S2: Atmospheric correction and emissivity estimation: The central server performs atmospheric correction and surface emissivity estimation on the thermal infrared band images in the multi-source fusion input data to obtain the initial surface temperature data;

[0046] Step S3: Spatiotemporal feature fusion and scale unification: The central server fuses the initial surface temperature data and multi-source fusion input data to construct multi-source spatiotemporal feature data with unified resolution;

[0047] Step S4: Spatiotemporal Neural Network Prediction: A spatiotemporal neural network model for land surface temperature prediction is constructed on the central server. Multi-source spatiotemporal feature data is used as input to the spatiotemporal neural network model to simultaneously learn the spatial distribution and temporal evolution of land surface temperature. A coupling loss function is introduced to train the spatiotemporal neural network model, which is used to constrain the predicted values ​​to satisfy the laws of energy conservation and radiation, resulting in a trained spatiotemporal neural network model. The trained spatiotemporal neural network model performs inference on the central server, outputting preliminary land surface temperature prediction results for the next 1–72 hours and pixel-level uncertainty assessment. The spatiotemporal neural network model includes an angle-guided CNN model and a GRU model.

[0048] Step S5: Cross-scale reconstruction and missing data completion: In the central server, the availability of low-resolution thermal infrared data in the preliminary surface temperature prediction results is determined; the temperature dynamic information of coarse-resolution thermal infrared data in the preliminary surface temperature prediction results is jointly reconstructed with high-resolution albedo, topography and texture features to achieve surface temperature overshoot within the 30–100m resolution range, resulting in reconstructed surface temperature prediction results; for the missing data areas in the reconstructed surface temperature prediction results caused by cloud cover, dust obstruction and observation gaps, mask convolution and time-series regular interpolation methods are used to complete the missing data, obtaining high-precision surface temperature prediction data;

[0049] Step S6: Calibration and Adaptive Update: In the central server, using the measured surface temperature retrieved from the ground infrared radiometers deployed in the Gobi Desert as a reference, and combined with high-precision surface temperature prediction data, the output of the spatiotemporal neural network model is subjected to bias correction and small-sample migration fine-tuning to obtain a calibrated spatiotemporal neural network model. The edge computing terminal performs incremental updates based on the calibrated spatiotemporal neural network model, adaptively adjusts the model parameters online, and periodically uploads the data back to the central server. This ensures the robustness and consistency of the model during long-term operation, ultimately obtaining the surface temperature prediction results for the next 1–72 hours, generating high-temperature risk warning information, and providing decision support for environmental monitoring and meteorological early warning in the Gobi Desert area.

[0050] Among them, steps S1–S6 are completed collaboratively by the edge computing terminal and the central server: when communication is restricted, the edge computing terminal performs local inference and caches the results in downsampling mode, and uploads them to the central server for high-precision reconstruction and prediction back transmission when communication is normal.

[0051] Example 2, this example is based on Example 1.

[0052] In the conventional technical field, step S4 specifically includes the following steps:

[0053] Step R1: Use an angle-guided CNN model to extract features from multi-source spatiotemporal feature data to obtain spatial feature data. The construction method of the angle-guided CNN model is as follows: build a CNN model, introduce an angle-guided module to optimize the activation of the convolutional layer of the CNN model, so that the model can capture the target spatial features more accurately, and obtain the angle-guided CNN model.

[0054] Step R2: Use the GRU model to extract features from multi-source spatiotemporal feature data to obtain temporal feature data;

[0055] Step R3: Input spatial and temporal feature data into the feature fusion layer for joint modeling, and perform prediction through a fully connected layer; introduce the mean square error loss function to train the angle-guided CNN model and GRU model to obtain the trained spatiotemporal neural network model; the trained spatiotemporal neural network model performs inference on the central server and outputs the preliminary surface temperature prediction results for the next 1–72 hours and the pixel-level uncertainty assessment.

[0056] In this invention, step S4 specifically includes the following steps:

[0057] Step S41: Use an angle-guided CNN model to extract features from multi-source spatiotemporal feature data to obtain spatial feature data; the construction method of the angle-guided CNN model is as follows: build a CNN model, introduce an angle-guided module to optimize the activation of the convolutional layer of the CNN model, so that the model can capture the target spatial features more accurately, and obtain the angle-guided CNN model.

[0058] Step S42: Use the GRU model to extract features from multi-source spatiotemporal feature data to obtain temporal feature data;

[0059] Step S43: Spatial and temporal feature data are input into the feature fusion layer for joint modeling, and prediction is performed through a fully connected layer. A dual-physics constraint term based on the surface energy balance equation and the Stefan-Boltzmann law is introduced, and a coupling loss function is constructed by combining it with a time consistency regularization term. Angle-guided CNN and GRU models are trained to obtain the trained spatiotemporal neural network model. The trained spatiotemporal neural network model performs inference on the central server, outputting preliminary surface temperature predictions for the next 1–72 hours and a pixel-level uncertainty assessment. The formulas used are as follows:

[0060] Coupling loss function formula:

[0061] ;

[0062] in, Represents the coupling loss function. This indicates the loss of consistency in the Earth's surface energy balance; This indicates the loss of Stefan-Boltzmann radiation uniformity; This represents the time consistency regularization term; , and Indicates non-negative weighting coefficients;

[0063] SEB Physical Consistency Loss :

[0064] ;

[0065] in, Indicates spatial index, Indicates a point-in-time index. Represents a physical constraint availability mask. This represents the SEB residual, which is "net radiation − (sensible heat + latent heat + soil heat)"; Indicates the length of the time window. This represents the net radiation, a physical quantity. The scaling factor representing net radiation. Represents the numerical stability constant; Indicates the summation symbol;

[0066] This parameter ensures that the model's output temperature matches the conservation of surface energy, directly constraining the model's physical output and avoiding temperature predictions that do not conform to the laws of radiation and heat transfer in nature.

[0067] SB radiation uniformity loss :

[0068] 1) There is an upward long wave reference:

[0069] ;

[0070] in, Indicates the availability of the radiation term mask. It represents the surface emissivity, reflecting the radiation emission capacity of different surface types in the infrared band; This represents the Stefan-Boltzmann constant. This represents the predicted land surface temperature (LST) from the model, expressed in Kelvin (K). It represents its fourth power and is used to calculate the predicted upward longwave radiation; This indicates a high-confidence inversion value of the reference uplink longwave radiation, derived from the measured inversion results of a ground radiometer.

[0071] This section constrains the prediction of surface temperature by comparing the deviation between the model-predicted upwave radiation and the reference upwave radiation, ensuring that it conforms to physical consistency.

[0072] 2) No reference data:

[0073] ;

[0074] in, Indicates incident long-wave radiation, This represents the external baseline estimate of the net longwavelength;

[0075] This part constrains the surface temperature (LST) by estimating the net longwave (i.e., the difference between the incident longwave and the emitted longwave), ensuring that the model still conforms to the laws of radiation physics when there is a lack of uplink longwave reference data.

[0076] Time Consistency Regular :

[0077] ;

[0078] in, Indicates a time-regular availability mask. This represents the second-order time difference, which measures the change between adjacent time steps. This indicates Huber's losses;

[0079] This loss function suppresses temporal jitter and non-physical abrupt changes in temperature prediction by imposing constraints on the second-order time difference of temperature, and ensures that the surface temperature output by the model has a smooth temporal evolution.

[0080] Example 3, this example is based on Example 2.

[0081] In the conventional technical field, step S41 specifically includes the following steps:

[0082] Step E1: Perform time axis registration, spatial resolution unification and normalization on the multi-source spatiotemporal feature data to obtain standardized multi-source feature data;

[0083] Step E2: Input standardized multi-source feature data into the convolutional layer of the convolutional neural network (CNN) model to extract local spatial response features;

[0084] Step E3: Input the local spatial response features into the spatial pooling layer of the CNN model, aggregate the local responses, and generate spatial feature data through feature integration processing.

[0085] In this invention, step S41 specifically includes the following steps:

[0086] Step S411: Perform time axis registration, spatial resolution unification and normalization on the multi-source spatiotemporal feature data to obtain standardized multi-source feature data;

[0087] Step S412: Adjust the activation direction of the standardized multi-source feature data through the angle guidance module to obtain rotation-enhanced convolution activation features;

[0088] Step S413: Input the rotation-enhanced convolution activation features into the spatial pooling layer of the angle-guided CNN model, aggregate the local responses, and generate spatial feature data through feature integration processing; the feature vector retains the spatial structure information of surface temperature distribution, albedo gradient, and terrain slope.

[0089] Example 4, this example is based on Example 3. In this example, step S412 specifically includes the following steps:

[0090] Step S4121: Extract the normalized activation vectors of the convolutional layer from the labeled samples (such as abnormal and normal regions) in the standardized multi-source feature data, calculate the candidate directions using the mean difference method, perform principal component analysis on the candidate direction set, extract the principal components, and define a two-dimensional rotation subspace.

[0091] Step S4122: Determine the rotation matrix and perform geometric rotation in the two-dimensional rotation subspace to adjust the directionality of the activation vectors after normalization of the convolutional layer, so that the model can focus more on spatial features with significant directionality (such as surface albedo gradient or terrain undulation changes) and obtain the feature direction of the activation vectors.

[0092] Applying geometric rotation formulas in a two-dimensional rotating subspace:

[0093] ;

[0094] in, , Let be a set of orthogonal unit bases for a two-dimensional rotational subspace. Indicates the transpose symbol. Represents the identity matrix. Indicates , It is a basis matrix composed of column vectors; Indicates the rotation angle. Represents the rotation matrix; Represents a two-dimensional rotating subspace Perform rotation angle The subsequent global rotation operator;

[0095] Step S4123: Introduce an adaptive angle guidance mechanism to calculate a conditional mask based on the projection of the activation vector in the feature direction, and perform selective rotation based on the mask; this mechanism ensures that only activations consistent with the target direction are controlled, thereby improving the robustness and controllability of the network in complex scenes; and obtain rotation-enhanced convolution activation features.

[0096] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A deep learning-based method for predicting surface temperature in the Gobi Desert, characterized in that: The method operates within an architecture that includes a data acquisition terminal, a remote sensing terminal, an edge computing terminal, and a central server working together. This architecture includes: The acquisition terminal collects ground observation data; Remote sensing terminal, acquire multi-source remote sensing data; The method includes the following steps: Step S1: Process ground observation data and multi-source remote sensing data to obtain multi-source fused input data; Step S2: The central server performs atmospheric correction and surface emissivity estimation on the thermal infrared band image in the multi-source fusion input data to obtain the initial surface temperature data; Step S3: The central server merges the initial surface temperature data and multi-source fusion input data to construct multi-source spatiotemporal feature data; Step S4: Use a spatiotemporal neural network model to process multi-source spatiotemporal feature data and output preliminary surface temperature prediction results; Step S5: In the central server, determine the availability of low-resolution thermal infrared data in the preliminary surface temperature prediction results to obtain the reconstructed surface temperature prediction results; for the missing areas in the reconstructed surface temperature prediction results, use mask convolution and time series graph regular interpolation methods to fill in the missing areas and obtain high-precision surface temperature prediction data. Step S6: In the central server, the output of the spatiotemporal neural network model is subjected to bias correction and small sample migration fine-tuning, based on the high-precision surface temperature prediction data, to finally obtain the surface temperature prediction result.

2. The method for predicting surface temperature in the Gobi Desert based on deep learning according to claim 1, characterized in that: Step S4 specifically includes: using a spatiotemporal neural network model to process multi-source spatiotemporal feature data, learning the laws of surface temperature in spatial distribution and temporal evolution, introducing a coupling loss function, training the spatiotemporal neural network model, and outputting preliminary surface temperature prediction results.

3. The method for predicting surface temperature in the Gobi Desert based on deep learning according to claim 2, characterized in that: Spatiotemporal neural network models include angle-guided CNN models and GRU models.

4. The method for predicting surface temperature in the Gobi Desert based on deep learning according to claim 3, characterized in that: Step S4 specifically includes the following steps: Step S41: Use an angle-guided CNN model to extract features from multi-source spatiotemporal feature data to obtain spatial feature data; Step S42: Use the GRU model to extract features from multi-source spatiotemporal feature data to obtain temporal feature data; Step S43: Jointly model spatial and temporal feature data and perform prediction through a fully connected layer; introduce a dual physical constraint term based on the surface energy balance equation and the Stefan-Boltzmann law, and construct a coupling loss function by combining it with a time consistency regularization term, train the angle-guided CNN model and GRU model to obtain the trained spatiotemporal neural network model; the trained spatiotemporal neural network model performs inference on the central server and outputs preliminary surface temperature prediction results.

5. The method for predicting surface temperature in the Gobi Desert based on deep learning according to claim 4, characterized in that: The construction method of the angle-guided CNN model is as follows: build a CNN model, introduce an angle-guided module to optimize the activation of the convolutional layers of the CNN model, and obtain the angle-guided CNN model.

6. The method for predicting surface temperature in the Gobi Desert based on deep learning according to claim 4, characterized in that: Step S41 specifically includes the following steps: Step S411: Perform time axis registration, spatial resolution unification and normalization on the multi-source spatiotemporal feature data to obtain standardized multi-source feature data; Step S412: Adjust the activation direction of the standardized multi-source feature data through the angle guidance module to obtain rotation-enhanced convolution activation features; Step S413: Input the rotation-enhanced convolution activation features into the spatial pooling layer of the angle-guided CNN model, aggregate the local responses, and generate spatial feature data through feature integration processing.

7. The method for predicting surface temperature in the Gobi Desert based on deep learning according to claim 6, characterized in that: Step S412 specifically includes the following steps: Step S4121: Extract the normalized activation vectors of the convolutional layer from the labeled samples in the standardized multi-source feature data, calculate the candidate directions using the mean difference method, perform principal component analysis on the candidate direction set, extract the principal components, and define a two-dimensional rotation subspace. Step S4122: Determine the rotation matrix, perform geometric rotation in the two-dimensional rotation subspace, adjust the directionality of the normalized activation vectors of the convolutional layer, focus on the spatial features with significant directionality, and obtain the feature direction of the activation vectors; Step S4123: Introduce an adaptive angle guidance mechanism, calculate a conditional mask based on the projection of the activation vector in the feature direction, and perform selective rotation based on the mask to obtain rotation-enhanced convolution activation features.

8. The method for predicting surface temperature in the Gobi Desert based on deep learning according to claim 1, characterized in that: The edge computing terminal is an industrial-grade box-type host with a built-in quad-core central processing unit, 8GB of memory and 512GB of solid-state storage. It integrates an embedded GPU module with CUDA acceleration capability, a GNSS high-precision timing module, and is equipped with an industrial Ethernet interface and a cellular communication module. It meets the operating temperature range of −20℃ to +55℃ and IP65 protection level requirements, and uses photovoltaic-battery integrated power supply.

Images