A remote sensing image ground surface temperature outlier repair method and related device

Abstract

The present application belongs to the technical field of image processing, and discloses a remote sensing image ground surface temperature abnormal value repairing method and related device to solve the technical problems of the existing ground surface temperature repairing technology, such as the non-adaptation of convolution network to per-pixel regression modeling, the insufficient fusion of multi-dimensional heterogeneous features, and the discontinuity of the repairing boundary of a large-area abnormal region. In the technical solution disclosed by the present application, three types of standardized multi-dimensional feature data sets of spectrum, environment and ground surface cover are constructed after spatial registration and normalization processing; an improved residual multi-layer perception machine containing an input projection layer, an adaptive channel attention layer, a residual block stack and a regression head is used to learn the nonlinear mapping relationship between the multi-dimensional features and the ground surface temperature, a high-precision temperature estimation model is trained, and the model prediction value is used to fill and repair the abnormal region. The technical solution disclosed by the present application significantly improves the temperature abnormality repairing accuracy and spatial continuity in a complex ground feature scene.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and specifically relates to a method and related apparatus for repairing surface temperature anomalies in remote sensing images. Background Technology

[0002] Land surface temperature (LST) is a comprehensive determinant of the Earth's surface thermal conditions and a key parameter in studies of surface energy balance, urban heat island effect, and climate change. The Landsat series of satellites provides long-term, continuous thermal infrared remote sensing data. Its Collection 2 Level-2 product includes atmospherically corrected LST data, representing a high-resolution product in the field of thermal environment research. However, due to factors such as cloud cover and sensor noise, Landsat LST products often contain missing pixels and localized anomalous pixels, resulting in spatial discontinuities that significantly impact subsequent processing and research.

[0003] Existing methods for restoring missing and anomalous pixels in land surface temperature data mainly include spatial interpolation methods, time series reconstruction methods, and machine learning-based restoration methods. Among them, spatial interpolation methods generally include inverse distance weighting (IDW), regular spline function (SPLINE), ordinary kriging (OK), and trend surface interpolation (TREND). These methods only utilize the spatial information of a single image and are not effective in restoring large areas of missing data. Time series reconstruction methods require the collaboration of multiple images to complete the restoration. Machine learning-based restoration methods are currently one of the mainstream restoration algorithms, but most existing models are shallow models with shortcomings in indicator selection and model architecture, and have limited ability to learn complex nonlinear relationships between multidimensional heterogeneous features.

[0004] In recent years, deep learning technology has demonstrated great potential in remote sensing image restoration. However, most existing deep learning-based remote sensing image restoration methods use convolutional neural networks to process image structural data, which is incompatible with the inherent characteristics of land surface temperature data. The relationship between land surface temperature and various auxiliary features is a pixel-by-pixel correspondence, essentially a multivariate regression modeling task, and the data format belongs to tabular data. The modeling approach of extracting spatial features from images using convolutional networks cannot fully explore multidimensional feature correlations and is not the optimal technical route. Therefore, existing technologies struggle to simultaneously achieve deep fusion of multidimensional heterogeneous auxiliary features and accurate fitting of the complex nonlinear relationship between features and land surface temperature. Furthermore, the restoration results are prone to problems such as abrupt changes in regional boundaries and spatial discontinuities. There is an urgent need for a novel Landsat land surface temperature anomaly restoration method that can balance multi-source heterogeneous feature fusion capabilities, high-precision nonlinear fitting capabilities, and smooth, seamless spatial continuity of the restored area. Summary of the Invention

[0005] The purpose of this invention is to provide a method and related apparatus for repairing land temperature anomalies in remote sensing images, addressing the technical shortcomings of existing solutions such as the incompatibility of convolutional networks with pixel-by-pixel regression modeling of land temperature, insufficient fusion of multidimensional heterogeneous features, and discontinuous boundaries in repairing large-area missing data. Specifically, the technical solution disclosed in this invention is a land temperature anomaly repair scheme based on residual networks. It utilizes an improved residual multilayer perceptron network to learn the nonlinear mapping from features to temperature and employs a spatial smoothing algorithm to achieve seamless filling of anomaly regions, thereby improving the accuracy and continuity of land temperature repair.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for repairing surface temperature anomalies in remote sensing images, comprising: Acquire standard land satellite remote sensing data covering the target area, and extract multi-band surface reflectance data and surface temperature data to be repaired for the target area based on the standard land satellite remote sensing data. Obtain land cover classification data of the target area, and construct a standardized multidimensional feature dataset based on the multi-band land reflectance data of the target area, the standard data of the land satellite remote sensing and the land cover classification data, which includes spectral features, environmental features and land cover features, and has undergone spatial registration and normalization. Based on the surface temperature data to be restored, abnormal temperature pixels and high-confidence normal pixels are identified and marked. A sample dataset is constructed based on each high-confidence normal pixel and the standardized multidimensional feature dataset. The high-confidence normal pixels are the pixels other than the identified abnormal temperature pixels. Each sample data in the sample dataset includes: the standardized multidimensional feature vector corresponding to the high-confidence normal pixel and the ground truth label of the surface temperature. Based on the sample dataset, the pre-constructed surface temperature estimation model is iteratively trained until the model accuracy converges or the preset training termination condition is met, thus obtaining the trained surface temperature estimation model. Based on the standardized multidimensional feature dataset, the standardized multidimensional feature vector of each pixel is input into the trained surface temperature estimation model to estimate the temperature, thus obtaining the predicted surface temperature value of each pixel. The abnormal areas to be repaired are located based on the temperature anomaly pixels identified by the identification markers, and the abnormal areas are filled and repaired based on the predicted surface temperature values ​​of each pixel, so as to obtain the surface temperature data of the target area after repair and reconstruction. The pre-built land surface temperature estimation model is an improved residual multilayer perceptron, which sequentially includes an input projection layer, an adaptive channel attention layer, a residual block stack, and a regression head. The input projection layer is used to input standardized multidimensional feature vectors, map features from the original input dimensions to preset hidden dimensions, and output hidden dimension feature vectors. The adaptive channel attention layer is used to input hidden dimension feature vectors, calculate element-wise gating weights for each dimension of features, and adjust the original features element-wise based on the gating weights, outputting a filtered and enhanced weighted feature sequence. The residual block stack is used to input the weighted feature sequence. Each residual block is configured with an independent fully connected transformation main path and a residual cross-layer direct connection shortcut path. The transformation results of the main path are superimposed and fused with the original input features transmitted by the shortcut path to output deep fused features. The regression head is used to input deep fused features, and after multi-level dimensional compression linear mapping operations, outputs a single-dimensional predicted land surface temperature value.

[0007] A further improvement to the technical solution of this invention lies in constructing a standardized multidimensional feature dataset based on the multi-band surface reflectance data of the target area, the standard data of the Landsat remote sensing, and the classification data of the land cover, which includes spectral features, environmental features, and land cover features, and has undergone spatial registration and normalization processing. This dataset includes: Spectral feature indices are calculated based on multi-band surface reflectance data of the target area, and spectral features are constructed based on the spectral feature indices and the multi-band surface reflectance data of the target area. Extract the header information and atmospheric parameters from the Landsat remote sensing standard data, calculate the solar angle parameters based on the header information, and construct environmental features based on the solar angle parameters and atmospheric parameters. Based on the surface feature classification data, the statistical characteristics of the feature classification of each pixel neighborhood are calculated, and surface cover characteristics are constructed together with the surface feature classification data. By integrating the spectral features, environmental features, and land cover features, a preliminary multidimensional heterogeneous feature dataset is obtained. Spatial registration is then uniformly performed on the preliminary multidimensional heterogeneous feature dataset, and normalization is applied to all features to obtain a standardized multidimensional feature dataset.

[0008] A further improvement of the technical solution of the present invention is that the spectral characteristic indices include vegetation-related indices, water-related indices, soil-related indices, and building-related indices; the atmospheric-related parameters include upward and downward propagating thermal radiation-related parameters and atmospheric transmittance-related parameters; the solar angle parameters include solar zenith angle and solar azimuth angle; the surface cover characteristics include surface classification categories and spatial neighborhood statistical characteristics, wherein the surface classification categories take values ​​from 1 to M integers, corresponding to M types of land cover; and the spatial neighborhood statistical characteristics are the number of pixels of each land cover category within a preset-size neighborhood window around each pixel.

[0009] A further improvement to the technical solution of this invention lies in that, based on the surface temperature data to be repaired, during the process of identifying and marking temperature anomaly pixels and high-confidence normal pixels, the surface temperature anomaly values ​​are identified from two dimensions: data value range and pixel geospatial correlation, including: Pixels with null or non-finite surface temperatures are marked as temperature anomalous pixels; Pixels with surface temperatures below a preset low temperature threshold and above a preset high temperature threshold are marked as temperature anomalous pixels; Obtain the central tendency statistics of surface temperature of different land types in the target area. And the dispersion statistic D will deviate Pixels within a certain range are labeled as temperature anomaly pixels; among them, 1 represents the preset sensitivity coefficient; The remote sensing image is traversed pixel by pixel using two preset sliding windows. The temperature central tendency and dispersion statistics within the sliding windows of the two scales are calculated respectively. The local normal temperature range is constructed by combining the preset sensitivity coefficients of each scale. Pixels that exceed the local normal temperature range are marked as temperature anomalous pixels.

[0010] A further improvement to the technical solution of the present invention lies in that, in the pre-constructed surface temperature estimation model, The input projection layer consists of a single linear transformation layer; The adaptive channel attention layer consists of a linear transformation layer, a ReLU nonlinear activation layer, another linear transformation layer, and a Sigmoid activation layer arranged in sequence. The residual block stack consists of L residual learning blocks with the same structure stacked together. Each residual learning block contains a fully connected transformation main path and a residual cross-layer direct connection shortcut path set in parallel. The regression head is composed of multiple residual learning blocks with progressively decreasing dimensions, strung together.

[0011] A further improvement of the technical solution of the present invention is that the fully connected transformation main path is composed of a first linear transformation layer, a batch normalization layer, a nonlinear activation layer, a regularization discard layer, a second linear transformation layer, and a batch normalization layer in sequence, which is used to learn the residual mapping; the residual cross-layer direct connection shortcut path is composed of a linear transformation layer and a batch normalization layer to perform linear projection when the input and output dimensions are different, and is an identity connection when the input and output dimensions are the same.

[0012] A further improvement to the technical solution of the present invention lies in locating the abnormal region to be repaired based on the temperature anomaly pixels identified by the identification marker, including: The temperature anomaly pixels identified by the merging and identification are expanded using a morphological dilation operator to absorb isolated anomaly pixels in the neighborhood, resulting in an outlier binary image; wherein, in the outlier binary image, the abnormal region is 0 and the normal region is 1; Based on the binary image of the outlier, the standard land satellite remote sensing data of the target area is masked to complete the location of the outlier area to be repaired.

[0013] In a second aspect, the present invention provides a system for repairing surface temperature anomalies in remote sensing images, comprising: The data acquisition module is used to acquire standard land satellite remote sensing data covering the target area, and extract multi-band surface reflectance data and surface temperature data to be repaired in the target area based on the standard land satellite remote sensing data. The multidimensional feature construction module is used to acquire the land cover classification data of the target area. Based on the multi-band land reflectance data of the target area, the standard data of land satellite remote sensing and the land cover classification data, a standardized multidimensional feature dataset containing spectral features, environmental features and land cover features is constructed and spatially registered and normalized. The anomaly identification and sample construction module is used to identify and mark temperature anomaly pixels and high-confidence normal pixels based on the surface temperature data to be repaired, and to construct a sample dataset based on each high-confidence normal pixel and the standardized multidimensional feature dataset; wherein, the high-confidence normal pixels are the remaining pixels other than the identified and marked temperature anomaly pixels; each sample data in the sample dataset includes: the standardized multidimensional feature vector corresponding to the high-confidence normal pixel and the ground truth label of the surface temperature; The model training and temperature prediction module is used to iteratively train the pre-built surface temperature estimation model based on the sample dataset until the model accuracy converges or the preset training termination condition is met, so as to obtain the trained surface temperature estimation model; based on the standardized multidimensional feature dataset, the standardized multidimensional feature vector of each pixel is input into the trained surface temperature estimation model to perform temperature estimation, so as to obtain the predicted surface temperature value of each pixel. The abnormal area filling and repair module is used to locate the abnormal area to be repaired based on the temperature abnormality pixels identified by the identification marker, and fill and repair the abnormal area based on the predicted surface temperature values ​​of each pixel to obtain the surface temperature data of the target area after repair and reconstruction. The pre-built land surface temperature estimation model is an improved residual multilayer perceptron, which sequentially includes an input projection layer, an adaptive channel attention layer, a residual block stack, and a regression head. The input projection layer is used to input standardized multidimensional feature vectors, map features from the original input dimensions to preset hidden dimensions, and output hidden dimension feature vectors. The adaptive channel attention layer is used to input hidden dimension feature vectors, calculate element-wise gating weights for each dimension of features, and adjust the original features element-wise based on the gating weights, outputting a filtered and enhanced weighted feature sequence. The residual block stack is used to input the weighted feature sequence. Each residual block is configured with an independent fully connected transformation main path and a residual cross-layer direct connection shortcut path. The transformation results of the main path are superimposed and fused with the original input features transmitted by the shortcut path to output deep fused features. The regression head is used to input deep fused features, and after multi-level dimensional compression linear mapping operations, outputs a single-dimensional predicted land surface temperature value.

[0014] In a third aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the remote sensing image surface temperature anomaly repair method as described in any one of the first aspects of the present invention.

[0015] In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the remote sensing image surface temperature anomaly repair method as described in any one of the first aspects of the present invention.

[0016] Compared with the prior art, the present invention has the following beneficial effects: Existing methods for restoring land surface temperature anomalies in remote sensing images often employ local spatial interpolation or single statistical threshold methods, relying solely on neighboring pixel values ​​to fill in the anomalous areas. This fails to adequately correlate the nonlinear interactions between surface temperature and multiple dimensional factors, including spectral, atmospheric, and surface cover dimensions. Furthermore, in complex terrain scenarios, the restored results are prone to temperature discontinuities and regional distortions. To address these issues, this invention discloses a novel method for restoring land surface temperature anomalies in remote sensing images. This method first integrates multi-source remote sensing data to construct a standardized multi-dimensional feature set. Then, it identifies and distinguishes between anomalous temperature pixels and high-confidence normal pixels, constructing a normal sample set based on the high-confidence normal pixels. The training process, based on these normal samples, includes an input projection layer, an adaptive channel attention layer, a residual block stack, and a return layer. The improved residual multilayer perceptron model, developed by Guitou, uses the predicted temperature values ​​of each pixel output from the trained model to fill in abnormal areas and complete the repair. Specifically, the adaptive channel attention layer automatically calculates the feature weights of each dimension, amplifying the effective feature weights strongly correlated with surface temperature from the spectrum, environment, and land cover, while suppressing the weights of interfering and redundant features. This addresses the problem of mixed heterogeneous features and invalid features interfering with prediction accuracy. The residual connection mechanism of the residual block stack adds cross-layer direct connection branches in each fully connected transformation layer, directly superimposing the original features from the upper layer to the transformation output of the current layer. This avoids the gradient vanishing defect during deep network training, fully exploits the complex nonlinear mapping relationship between multidimensional features and surface temperature, and improves the accuracy of global surface temperature estimation. This invention can achieve high-precision automated repair of temperature anomaly areas in large-scale complex remote sensing images. After repair, the temperature field is continuous and smooth, without obvious halo or fault distortion, and the entire process requires no manual zoning intervention, making it suitable for batch remote sensing image processing.

[0017] In the preferred embodiment of the present invention, three types of features—spectral, environmental, and land cover—are constructed independently in stages. First, all features are integrated to form a preliminary multidimensional heterogeneous dataset. Then, spatial registration is uniformly implemented to eliminate band spatial offset. Finally, global normalization is used to eliminate feature dimension differences. This technique can fully cover the multidimensional influence information of land cover spectral response, atmospheric radiation interference, and land surface spatial distribution, unify the feature spatial scale and coordinate reference, and reduce model training bias caused by heterogeneous features.

[0018] In the preferred embodiment of this invention, a multi-stage progressive identification logic is adopted, sequentially filtering null / non-finite values ​​and global high and low temperature extreme values. Temperature statistical intervals are constructed according to land cover categories, and the global sensitivity coefficient Se1 is used to screen for deviations and anomalies. A dual-scale sliding window of different sizes is used in conjunction with their respective local sensitivity coefficients to identify local scale anomalies. The small window identifies isolated small clusters of anomalies, and the large window identifies regional overall shifts. This technique covers all anomaly types in a layered manner, sets differentiated judgment intervals for different land cover types to reduce the false detection rate, and uses dual-scale windows to identify distortions at different spatial scales. By adjusting the sensitivity coefficient to adapt to complex surface scenes, the completeness of temperature anomaly pixel identification is significantly improved, while simultaneously reducing the probability of false positives and false negatives.

[0019] In the preferred embodiment of this invention, the model sequentially connects an input projection layer, an adaptive channel attention layer, a multi-layer residual block stack, and a multi-level dimensionality reduction regression head. Specific implementation methods are provided: the input projection layer uniformly maps features to a fixed hidden dimension; the attention layer generates gating weights of the same dimension to achieve dynamic feature recalibration; the residual block stack stacks multiple layers of residual learning blocks containing independent main paths and shortcut paths to mine deep nonlinear correlations; and the regression head progressively compresses feature dimensions to output single-dimensional temperature. This method unifies feature input specifications, autonomously filters redundant interference features, relies on residual direct connection branches to solve the gradient vanishing problem in deep networks, and smoothly compresses features to avoid information loss, thereby improving the model's ability to extract and fit high-dimensional remote sensing features, and ensuring that the predicted temperature values ​​closely match the actual thermal radiation patterns of ground objects. Furthermore, the implementation methods for the main path and shortcut paths are provided to ensure stable fitting of complex surface temperature mapping relationships. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a method for repairing anomalies in surface temperature from remote sensing images, as described in Embodiment 1 of the present invention. Figure 2 This is a flowchart illustrating a Landsat surface temperature anomaly repair method based on residual networks in Embodiment 2 of the present invention. Figure 3 This is a schematic diagram of the improved residual multilayer perceptron network in Embodiment 2 of the present invention; Figure 4 This is a comparative schematic diagram of surface temperature images before and after restoration in Embodiment 3 of the present invention; Figure 5 This is a schematic diagram of a remote sensing image surface temperature anomaly repair system in Embodiment 4 of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of protection of the present invention, but rather to illustrate specific selected embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0023] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. It should be noted that, unless otherwise specified, the following embodiments and features can be combined with each other.

[0024] Example 1 Please see Figure 1 The present invention provides a method for repairing surface temperature anomalies in remote sensing images, comprising the following steps: Step 1: Obtain standard land satellite remote sensing data covering the target area. Based on the standard land satellite remote sensing data, obtain multi-band surface reflectance data of the target area within the target area, as well as surface temperature data to be repaired for missing or abnormal pixels, as the basic raw data for constructing features.

[0025] Step 2: Obtain surface cover classification data of the target area. Based on the multi-band surface reflectance data of the target area, the standard data of land satellite remote sensing, and the surface cover classification data, construct a standardized multidimensional feature dataset that includes spectral features, environmental features, and surface cover features, and has undergone spatial registration and normalization.

[0026] Explained, the standardized multidimensional feature dataset contains standardized multidimensional feature vectors corresponding to all pixels in the target area. Each set of standardized multidimensional features is a fusion of spectral features, environmental features, and land cover features matched to that pixel.

[0027] In a specific exemplary technical solution, step 2 specifically includes the following steps: Step 2.1: Calculate the spectral feature index based on the multi-band surface reflectance data of the target area; construct the spectral features required for surface temperature anomaly repair based on the spectral feature index and the multi-band surface reflectance data of the target area. Step 2.2: Extract the header information and atmospheric parameters of the Landsat remote sensing standard data, calculate the solar angle parameters based on the header information, and construct environmental features based on the solar angle parameters and the atmospheric parameters. Step 2.3: Obtain the land cover classification data of the target area, calculate the land cover classification statistical characteristics of each pixel's neighborhood based on the land cover classification data, and jointly construct the land cover characteristics; Step 2.4: Integrate the spectral features, environmental features, and land cover features to obtain a complete preliminary multidimensional heterogeneous feature dataset; perform unified spatial registration on all features in the preliminary multidimensional heterogeneous feature dataset, and then normalize all features to eliminate dimensional differences, outputting a standardized multidimensional feature dataset to provide unified feature support for subsequent anomaly identification, sample construction, and model input.

[0028] Step 3: Based on the surface temperature data to be repaired, identify and mark temperature anomalous pixels and high-confidence normal pixels, and construct a sample dataset based on each high-confidence normal pixel and the standardized multidimensional feature dataset; wherein, the high-confidence normal pixels are pixels other than the identified and marked temperature anomalous pixels; each sample data in the sample dataset includes: the standardized multidimensional feature vector corresponding to the high-confidence normal pixel and the ground truth label of surface temperature (i.e., the actual surface temperature value).

[0029] Step 4: Based on the sample dataset, iteratively train the pre-built surface temperature estimation model until the model accuracy converges or the preset training termination condition is met, and obtain the trained surface temperature estimation model; based on the standardized multidimensional feature dataset, input the standardized multidimensional feature vector of each pixel into the trained surface temperature estimation model to estimate the temperature, and obtain the predicted surface temperature value of each pixel.

[0030] Specifically, the pre-built land surface temperature estimation model is an improved residual multilayer perceptron, which sequentially includes an input projection layer, an adaptive channel attention layer, a residual block stack, and a regression head. The adaptive channel attention layer calculates the importance weights of features in each dimension, adjusts the original features based on these importance weights, and outputs a filtered and enhanced weighted feature sequence. This part enhances the input features by highlighting effective features strongly correlated with land surface temperature and suppressing redundant interference features. The output weighted feature sequence is then transmitted to the multilayer residual fully connected layer to address the issues of mixed multi-source heterogeneous features and interference from irrelevant features. To address the issue of prediction accuracy, a residual block stack is used to input weighted feature sequences. Each layer has an independent fully connected transformation main path and a residual cross-layer direct connection shortcut path. The transformation results of the main path and the original input features transmitted by the shortcut path are superimposed and fused before being passed to the next layer, alleviating the gradient vanishing problem in deep networks. The complex nonlinear relationship between multidimensional features and land surface temperature is explored layer by layer, and finally, deep fused features are output and transmitted to the regression head (i.e., the output layer). The output layer receives the deep fused features output from multiple residual fully connected layers, and outputs a single-channel land surface temperature prediction value after linear mapping operation.

[0031] Step 5: Locate the abnormal area to be repaired based on the temperature anomaly pixels identified by the identification markers, and fill and repair the abnormal area based on the predicted surface temperature values ​​of each pixel to obtain the repaired and reconstructed surface temperature data of the target area.

[0032] Existing technologies for land surface temperature restoration in remote sensing images suffer from several unavoidable drawbacks: the heterogeneous features derived from multi-source remote sensing, including spectral, environmental, and land cover characteristics, are complex and redundant, and the lack of a targeted screening mechanism can easily interfere with model building; conventional multilayer fully connected networks are prone to gradient vanishing after depth enhancement, failing to fully exploit the complex nonlinear relationship between features and land surface temperature; if missing or anomalous temperature pixels in the original image are directly used in training, they will contaminate the dataset with incorrect labels, significantly reducing the model's estimation accuracy; traditional restoration methods that rely solely on local interpolation lack global data constraints, resulting in abrupt temperature changes, boundary distortion, and poor spatial continuity in the restored area. This invention discloses a method for restoring surface temperature anomalies in remote sensing images. It employs a multi-layered progressive improvement scheme to address the aforementioned problems one by one: three independent feature classes are constructed step-by-step, and unified registration and normalization are used to eliminate spatial and dimensional biases, resulting in a standardized multidimensional feature dataset; high-confidence normal pixels are selected through anomaly identification, and only reliable ground truth labels are used to build the sample dataset, avoiding interference from anomalous data at the source; an adaptive channel attention layer is introduced into the model to autonomously allocate feature weights, strengthening effective features and suppressing ineffective interference features; a residual fully connected layer with cross-layer direct connection branches is used to solve the gradient vanishing problem in deep networks, accurately fitting the temperature mapping relationship; finally, based on the model's global prediction results, anomaly areas are uniformly restored, and the original effective ground truth temperatures are fused to generate a complete temperature field, effectively improving the spatial transition effect of the restored area and comprehensively enhancing the overall accuracy of surface temperature estimation and anomaly restoration.

[0033] Example 2 Please see Figure 2 In this embodiment of the invention, a method for repairing Landsat surface temperature anomalies based on residual networks is provided, comprising the following steps: S1: Data preprocessing: Obtain Landsat Collection 2 Level-2 data covering the study area, directly extract multi-band surface reflectance (SR) data and surface temperature (ST) data, and then crop to obtain multi-band surface reflectance data and surface temperature data to be restored for the study area.

[0034] S2: Construct a multidimensional feature dataset: Calculate spectral feature indices based on the multi-band surface reflectance data of the study area obtained in step S1, and construct spectral features together with the multi-band surface reflectance data of the study area obtained in step S1; calculate solar angle-related parameters based on product header file data, and construct environmental features together with atmospheric-related parameters in the L2C2 product; obtain surface cover classification data of the study area and statistical features of the cover classification of each pixel's neighborhood, and construct surface cover features together; integrate and construct a standardized multidimensional feature dataset containing spectral features, environmental features, and surface cover features, and after spatial registration and normalization.

[0035] In a specific exemplary technical solution, the content of the multidimensional feature dataset constructed in step S2 is as follows: Spectral features include spectral feature indices and multi-band surface reflectance; wherein, the spectral feature indices are calculated by multi-band combination of multispectral surface reflectance, and the calculable feature indices related to typical land features include vegetation-related indices, water-related indices, soil-related indices, and building-related indices.

[0036] Environmental features include solar angle parameters and atmospheric parameters; among them, atmospheric parameters include parameters related to upward and downward propagating thermal radiation and atmospheric transmittance; solar angle parameters include solar zenith angle and solar azimuth angle, which are calculated using imaging time and geographic latitude and longitude.

[0037] The land cover features include land cover categories and their spatial neighborhood statistical features; wherein, the land cover category takes the value of an integer from 1 to M, corresponding to M land cover types, forming 1 feature; the spatial neighborhood statistical feature data is the number of pixels of each land cover type within an X*X neighborhood window around each pixel, thus forming M features; X*X is the neighborhood window size.

[0038] The multidimensional feature dataset consists of N1 + N2 + N3 features. N1 is the total number of spectral features, which is equal to the sum of the number of preprocessed multispectral bands and the number of calculated spectral feature indices. N2 is the total number of environmental features, including 3 atmospheric related parameters and 2 solar angle parameters. N3 is the total number of land cover features, consisting of 1 land cover classification feature and M spatial neighborhood statistical features, i.e., N3 = M + 1.

[0039] S3: Outlier Screening and Model Input Sample Construction: Based on the surface temperature data to be repaired, a multi-stage outlier identification algorithm is used to extract temperature outlier pixels and generate an outlier mask. The remaining pixels that are not identified by the algorithm as temperature outlier pixels are identified as high-confidence normal pixels. The model input sample dataset is constructed based on the high-confidence normal pixels. Then, the sample dataset is randomly divided into training set, validation set and test set according to a set ratio.

[0040] In a specific exemplary technical solution, the multi-stage anomaly recognition algorithm in step S3 specifically includes: S3.1: Identify surface temperature anomalies from the data itself and the geospatial relationships between pixels, including: S3.1.1: Pixels with null or non-finite surface temperature data are directly marked as anomalies; S3.1.2: Mark pixels with surface temperatures below a preset low temperature threshold or above a preset high temperature threshold as anomalies; S3.1.3: Calculate the basic statistics of surface temperature within different land types in the study area. and discrete case statistics D , will deviate Cells within a certain range are marked as abnormal. 1 represents the preset sensitivity coefficient; S3.1.4: Determine a smaller window size w and a larger window size W, perform a pixel-by-pixel traversal of the image of the study area, and calculate the basic statistical measures of the local window of each pixel during the traversal process. , and discrete case statistics D , , will deviate , Pixels within a range are marked as anomalies; among them, a smaller window is used to mark isolated points and small clusters of anomalies, and a larger window is used to mark regional overall shift anomalies. Se2 and Se3 are the preset sensitivity coefficients for the smaller window and the larger window, respectively. S3.2: Merge the identified outlier regions and use a morphological dilation method to absorb isolated pixels in the neighborhood to obtain the final outlier binary image, where outlier regions are 0 and normal value regions are 1.

[0041] S3.3: Use outlier binary images to mask the remote sensing image of the study area to obtain image data of the study area that retains only normal values.

[0042] S3.4: Extract the feature vectors and surface temperature values ​​corresponding to the image data of the study area with normal values, and randomly divide them into training set, validation set and test set according to the ratio of a1:a2:a3; where a1, a2 and a3 are the proportions of the total number of samples in the training set, validation set and test set, respectively.

[0043] S4: Model Construction and Temperature Estimation: Construct a fully connected neural network model that incorporates residual connection mechanism and feature attention mechanism. Input the pre-divided sample dataset into the model and train it through multiple rounds until the validation accuracy converges. Then, take the model with the best validation performance as the final trained surface temperature estimation model. Use the trained surface temperature estimation model to estimate the temperature of all pixels in the study area to obtain the predicted surface temperature value of each pixel.

[0044] Please see Figure 3 In a specific exemplary technical solution, the fully connected neural network model with residual connections in step S4 is an improved residual multilayer perceptron (ResMLP), whose structure consists of an input projection layer, an adaptive channel attention layer, a residual block stack, and a regression head in sequence.

[0045] The specific training process includes: S4.1: The training set is input into the neural network model, passing through an input projection layer consisting of a single linear transformation layer, and then the input feature vector is transformed from the input dimension D. in Mapping to hidden dimension D h ; S4.2: D h The feature data of the dimension is input into the adaptive channel attention layer, which consists of a linear transformation layer (Linear), a non-linear activation layer (ReLU), a linear transformation layer (Linear), and a sigmoid activation layer. After passing through the adaptive channel attention layer, element-wise gated weights of the same dimension are generated for each sample. The input features are multiplied element-wise with the corresponding gated weights to complete the dynamic recalibration of multidimensional features, automatically amplify effective features, and suppress redundant interference features. S4.3: The output of S4.2 is input into the residual block layer, which consists of L residual learning blocks with identical structures stacked sequentially. Each residual learning block contains a main path and a shortcut path. The output of the main path and the output of the shortcut path are added together and then processed by a nonlinear activation layer (ReLU) to obtain the output of the residual learning block. In the main path, the input features undergo multiple fully connected nonlinear transformations to perform deep fitting of the features and extract complex nonlinear correlations, outputting the transformed feature map. The shortcut path does not undergo any weight transformation and directly passes the original input features of this layer to the output without any computational processing.

[0046] Specifically, the residual connection mechanism for each residual block in step S4.3 above is as follows: S4.3.1: Input feature vector x into the main path and shortcut path respectively; S4.3.2: The main path consists of a first linear transformation layer, a batch normalization layer, a nonlinear activation layer, a regularized dropout layer, a second linear transformation layer, and a batch normalization layer, used to learn the residual mapping F(x). The shortcut path consists of a linear transformation layer and a batch normalization layer to perform linear projection when the input and output dimensions are different; when the dimensions are the same, it is an identity connection. The output of the shortcut path is shortcut(x). S4.3.3: The final output of the residual learning block is ReLU(F(x) and shortcut(x)). When the main path does not learn any effective new features, F(x) approaches 0, and the network at least maintains an identity mapping.

[0047] S4.4: The output of the residual block layer is input into the regression head. The regression head is composed of multiple residual learning blocks with successively decreasing dimensions connected together. By progressively compressing the feature dimensions, the input feature vector is reduced from the hidden dimension D. h Finally, it is mapped to a one-dimensional surface temperature estimate.

[0048] S4.5: Input the training set and validation set into the constructed ResMLP model for iterative training. In each training round, calculate the loss on the training set and update the model parameters, and calculate the loss on the validation set. If the accuracy threshold is not reached, return to S4.1 to start a new training round. If the accuracy threshold is reached, output the final model.

[0049] S4.6: Input the test set into the final model trained in S4.5 to evaluate the model accuracy.

[0050] S5: Abnormal Data Repair: For the abnormal areas marked by the abnormal mask, a spatial smoothing algorithm is used to fill and repair them based on the predicted surface temperature values ​​of each pixel, so as to obtain the repaired and reconstructed surface temperature data of the study area.

[0051] The technical solution disclosed in this invention not only integrates surface reflectance, spectral index, atmospheric conditions, and solar geometry data, but also innovatively incorporates land cover classification and spatial neighborhood data. This invention employs a multi-stage outlier identification algorithm, combining null value detection, absolute thresholding, land use category statistics, and dual-scale local window statistics, supplemented by morphological dilation operations. This enables comprehensive and robust labeling of temperature anomaly pixels, ensuring high confidence in the training sample set, thereby improving the model's robustness and generalization ability. This invention utilizes an improved residual multilayer perceptron to fully explore the complex nonlinear relationship between features and surface temperature. In the restoration stage, this invention uses a spatial smoothing algorithm to fill in the anomaly region, ensuring a smooth transition and seamless integration between the restored temperature and surrounding valid pixels, effectively avoiding splicing marks and edge abrupt changes, and guaranteeing the spatial continuity of surface temperature data.

[0052] Example 3 In a specific embodiment of the present invention, a method for repairing Landsat surface temperature anomalies based on residual networks is provided, comprising the following steps: The study area (i.e., the target area) is the southern bank of Hangzhou Bay, and the study data is Landsat 8 satellite imagery data from August 16, 2020, covering both land and water bodies in the study area.

[0053] In this embodiment of the invention, the spectral feature indices in the spectral features include: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Built-up Index (NDBI), Soil Adjusted Vegetation Index (SAVI), and Bare Soil Index (BSI).

[0054] The formulas for calculating these indices are as follows: ; ; ; ; ; Wherein, NIR is the surface reflectance in the near-infrared band, with a wavelength range of 0.85-0.88 μm; Red is the surface reflectance in the red band, with a wavelength range of 0.64-0.67 μm; Green is the surface reflectance in the green band, with a wavelength range of 0.53-0.59 μm; Blue is the surface reflectance in the blue band, with a wavelength range of 0.45 – 0.51 μm; SWIR1 is the surface reflectance in the short-wave infrared band, with a wavelength range of 1.57 – 1.65 μm; L is a parameter that varies with vegetation density, with a value range from 0 to 1, and is assigned 0.5 in this implementation.

[0055] In this embodiment of the invention, environmental features include solar zenith angle (SZA), solar azimuth angle (SAA), upward atmospheric radiation (URAD), downward atmospheric radiation (DRAD), and atmospheric transmittance (ATRAN).

[0056] The formula for calculating the solar angle index is as follows: ; ; ; ; ; ; ; Where N is the cumulative number of days from January 1st of the current year to the imaging date; t is the solar angle in radians; δ is the solar declination angle in radians; AST is the local true solar time in hours; BT is Beijing time, obtained by adding 8 hours to UTC time in hours; Lon is the geographic longitude corresponding to the pixel in degrees; Lon BJ The longitude is Beijing standard, taken as 120°E; ω is the solar hour angle, in degrees, negative in the morning and positive in the afternoon; φ is the geographic latitude corresponding to the pixel, in radians.

[0057] In this embodiment of the invention, the land features in the study area are divided into seven categories: water bodies, woodlands, grasslands, construction land, unused land, cultivated land, and fishponds, i.e., M=7. The neighborhood window is a 3*3 window, and the number of pixels of each land feature category within the 3*3 neighborhood window surrounding each pixel is... .

[0058] The obtained data was overlaid with bands to construct a multidimensional feature dataset containing 24 features. Then, spatial registration was performed on all feature layers to ensure that the pixel positions of each band were strictly aligned in the same geographic coordinate system. Subsequently, normalization was performed on each feature to linearly map each feature value to the [0,1] interval, so as to eliminate the bias effect caused by the difference in the units of different features on the model training.

[0059] In this embodiment of the invention, the normalization process used is Min-Max normalization, and its general formula is: ; in, and These are the minimum and maximum values ​​of the feature across all valid pixels, respectively.

[0060] The baseline statistic A is the median, and the discrete statistic B is the median absolute deviation (MAD). The formula for calculating MAD is: ; in, For the i-th value in the dataset, This is the median value in the dataset.

[0061] In this embodiment of the invention, the sensitivity coefficients are set as follows: Se1=2.5, Se2=3.0, Se3=2.5. Finally, the high-confidence normal pixels in the study area are randomly divided into training set, validation set and test set in a ratio of 7:2:1.

[0062] The input feature dimension Din=24, the hidden layer dimension Dh=128, and the residual block stack contains 6 residual blocks (i.e., L=6). The regularization dropout rate is set to 0.15. The model uses mean squared error (MSE) as the loss function, employs the AdamW optimizer, and has an initial learning rate of 1×10⁻⁶. - ³, weight decay 1×10 -5 The training batch size is 1024, and the maximum number of training epochs is 400. The learning rate scheduler uses the ReduceLROnPlateau strategy, with monitoring and validation loss, a decay factor of 0.5, and a patience epoch of 8. To prevent overfitting, an early stopping mechanism is introduced, monitoring and validation loss, with a patience value set to 60 epochs and a minimum improvement threshold of 1×10⁻⁶. -5 Automatic Mixed Precision (AMP) was enabled during training to accelerate computation, and gradients were clipped. The maximum norm was set to 1.0, and the model that performed best on the validation set was selected as the final land surface temperature estimation model.

[0063] In this embodiment of the invention, the Poisson algorithm based on the spatial gradient domain is used for filling and repair. The specific process includes: First, the four-neighborhood of the pixels within the anomaly region Ω is used as the basic computational unit, with the spatial gradient of the temperature value estimated by the model serving as the guiding field. For each pixel p to be repaired, a discrete Poisson equation is established: ; Where U(p) is the four-neighborhood of p, deg(p) is the number of effective neighbors of p in the image, g is the model-estimated temperature, and I is the original effective temperature value. The left side of the equation requires that the Laplace operator of the repaired temperature f within Ω be consistent with the Laplace operator of the estimated value, and the right side uses the original effective temperature I(q) as the Dirichlet Boundary Condition, with the estimated value g(q) used to fill in invalid pixels; Secondly, the discrete Poisson equations of all pixels within Ω are combined, and the sparse direct solver is called to solve the linear system in one go, so as to obtain the restored surface temperature value that is seamlessly integrated with the surrounding area in the abnormal region.

[0064] Please see Figure 4 Figure 4(a) shows the original surface temperature image to be restored, with obvious anomalies and missing areas. Figure 4 Image (b) shows the complete surface temperature image restored using the method of this invention. The restored area is spatially continuous and transitions smoothly with surrounding pixels. The ResMLP proposed in this invention is compared with MLP and Tabular ResNet in terms of accuracy and absolute prediction difference.

[0065] Evaluation indicators include the coefficient of determination. Mean absolute error Root mean square error and average deviation The calculation formula is as follows: ; ; ; ; in, This represents the actual surface temperature value for a pixel. For model estimates, The total number of samples.

[0066] Table 1. Results of different models on the test set

[0067] In Table 1, MLP stands for Multilayer Perceptron, and Tabular ResNet stands for Tabular Residual Network. As shown in Table 1, the ResMLP model proposed in this invention outperforms the two comparative models in all four metrics. Compared to the standard MLP, It increased by approximately 0.021. It decreased by approximately 0.196 °C. The temperature decreased by approximately 0.180 °C, indicating that the model's ability to nonlinearly map temperature was significantly enhanced after introducing feature attention and residual connections.

[0068] The formula for calculating the absolute difference between the predicted temperature and the actual temperature is as follows: ; in, This is the absolute difference.

[0069] Table 2. Percentage of Absolute Temperature Difference Ranges for Different Models (%)

[0070] As shown in Table 2, the ResMLP model has the highest percentage of pixels in the low error range, indicating that the model's predictions are closer to the actual temperature overall, with less dispersion and stronger fine estimation ability.

[0071] Based on the two sets of experiments above, the ResMLP model proposed in this invention performs excellently in terms of standard regression indicators, error spatial distribution, and module necessity verification. It can provide high-precision initial estimates for the repair of surface temperature anomalies and ensure the reliability of subsequent spatial repair.

[0072] Example 4 The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not disclosed in the apparatus embodiments, please refer to the embodiments of the method of the present invention.

[0073] Please see Figure 5 In this embodiment of the invention, a remote sensing image surface temperature anomaly repair system is provided, comprising: The data acquisition module is used to acquire standard land satellite remote sensing data covering the target area, and extract multi-band surface reflectance data and surface temperature data to be repaired in the target area based on the standard land satellite remote sensing data. The multidimensional feature construction module is used to acquire the land cover classification data of the target area. Based on the multi-band land reflectance data of the target area, the standard data of land satellite remote sensing and the land cover classification data, a standardized multidimensional feature dataset containing spectral features, environmental features and land cover features is constructed and spatially registered and normalized. The anomaly identification and sample construction module is used to identify and mark temperature anomaly pixels and high-confidence normal pixels based on the surface temperature data to be repaired, and to construct a sample dataset based on each high-confidence normal pixel and the standardized multidimensional feature dataset; wherein, the high-confidence normal pixels are the remaining pixels other than the identified and marked temperature anomaly pixels; each sample data in the sample dataset includes: the standardized multidimensional feature vector corresponding to the high-confidence normal pixel and the ground truth label of the surface temperature; The model training and temperature prediction module is used to iteratively train the pre-built surface temperature estimation model based on the sample dataset until the model accuracy converges or the preset training termination condition is met, so as to obtain the trained surface temperature estimation model; based on the standardized multidimensional feature dataset, the standardized multidimensional feature vector of each pixel is input into the trained surface temperature estimation model to perform temperature estimation, so as to obtain the predicted surface temperature value of each pixel. The abnormal area filling and repair module is used to locate the abnormal area to be repaired based on the temperature abnormality pixels identified by the identification marker, and fill and repair the abnormal area based on the predicted surface temperature values ​​of each pixel to obtain the surface temperature data of the target area after repair and reconstruction. The pre-built land surface temperature estimation model is an improved residual multilayer perceptron, which sequentially includes an input projection layer, an adaptive channel attention layer, a residual block stack, and a regression head. The input projection layer is used to input standardized multidimensional feature vectors, map features from the original input dimensions to preset hidden dimensions, and output hidden dimension feature vectors. The adaptive channel attention layer is used to input hidden dimension feature vectors, calculate element-wise gating weights for each dimension of features, and adjust the original features element-wise based on the gating weights, outputting a filtered and enhanced weighted feature sequence. The residual block stack is used to input the weighted feature sequence. Each residual block is configured with an independent fully connected transformation main path and a residual cross-layer direct connection shortcut path. The transformation results of the main path are superimposed and fused with the original input features transmitted by the shortcut path to output deep fused features. The regression head is used to input deep fused features, and after multi-level dimensional compression linear mapping operations, outputs a single-dimensional predicted land surface temperature value.

[0074] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions from a computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used to perform the operation of a remote sensing image surface temperature anomaly repair method.

[0075] In one embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of the terminal. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM (Random Access Memory) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the remote sensing image surface temperature anomaly repair method in the above embodiments.

[0076] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0077] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for repairing surface temperature anomalies in remote sensing images, characterized in that, include: Acquire standard land satellite remote sensing data covering the target area, and extract multi-band surface reflectance data and surface temperature data to be repaired for the target area based on the standard land satellite remote sensing data. Obtain land cover classification data of the target area, and construct a standardized multidimensional feature dataset based on the multi-band land reflectance data of the target area, the standard data of the land satellite remote sensing and the land cover classification data, which includes spectral features, environmental features and land cover features, and has undergone spatial registration and normalization. Based on the surface temperature data to be restored, abnormal temperature pixels and high-confidence normal pixels are identified and marked. A sample dataset is constructed based on each high-confidence normal pixel and the standardized multidimensional feature dataset. The high-confidence normal pixels are the pixels other than the identified abnormal temperature pixels. Each sample data in the sample dataset includes: the standardized multidimensional feature vector corresponding to the high-confidence normal pixel and the ground truth label of the surface temperature. Based on the sample dataset, the pre-constructed surface temperature estimation model is iteratively trained until the model accuracy converges or the preset training termination condition is met, thus obtaining the trained surface temperature estimation model. Based on the standardized multidimensional feature dataset, the standardized multidimensional feature vector of each pixel is input into the trained surface temperature estimation model to estimate the temperature, thus obtaining the predicted surface temperature value of each pixel. The abnormal areas to be repaired are located based on the temperature anomaly pixels identified by the identification markers, and the abnormal areas are filled and repaired based on the predicted surface temperature values ​​of each pixel, so as to obtain the surface temperature data of the target area after repair and reconstruction. The pre-built land surface temperature estimation model is an improved residual multilayer perceptron, which sequentially includes an input projection layer, an adaptive channel attention layer, a residual block stack, and a regression head. The input projection layer is used to input standardized multidimensional feature vectors, map features from the original input dimensions to preset hidden dimensions, and output hidden dimension feature vectors. The adaptive channel attention layer is used to input hidden dimension feature vectors, calculate element-wise gating weights for each dimension of features, and adjust the original features element-wise based on the gating weights, outputting a filtered and enhanced weighted feature sequence. The residual block stack is used to input the weighted feature sequence. Each residual block is configured with an independent fully connected transformation main path and a residual cross-layer direct connection shortcut path. The transformation results of the main path are superimposed and fused with the original input features transmitted by the shortcut path to output deep fused features. The regression head is used to input deep fused features, and after multi-level dimensional compression linear mapping operations, outputs a single-dimensional predicted land surface temperature value.

2. The method for repairing surface temperature anomalies in remote sensing images according to claim 1, characterized in that, Based on the multi-band surface reflectance data of the target area, the standard data of the Landsat remote sensing system, and the classification data of land cover, a standardized multidimensional feature dataset is constructed, including spectral features, environmental features, and land cover features, and which has undergone spatial registration and normalization processing. This dataset includes: Spectral feature indices are calculated based on multi-band surface reflectance data of the target area, and spectral features are constructed based on the spectral feature indices and the multi-band surface reflectance data of the target area. Extract the header information and atmospheric parameters from the Landsat remote sensing standard data, calculate the solar angle parameters based on the header information, and construct environmental features based on the solar angle parameters and atmospheric parameters. Based on the surface feature classification data, the statistical characteristics of the feature classification of each pixel neighborhood are calculated, and surface cover characteristics are constructed together with the surface feature classification data. By integrating the spectral features, environmental features, and land cover features, a preliminary multidimensional heterogeneous feature dataset is obtained. Spatial registration is then uniformly performed on the preliminary multidimensional heterogeneous feature dataset, and normalization is applied to all features to obtain a standardized multidimensional feature dataset.

3. The method for repairing surface temperature anomalies in remote sensing images according to claim 2, characterized in that, The spectral characteristic indices include vegetation-related indices, water-related indices, soil-related indices, and building-related indices; the atmospheric-related parameters include upward and downward propagating thermal radiation-related parameters and atmospheric transmittance-related parameters; the solar angle parameters include solar zenith angle and solar azimuth angle; the land cover characteristics include land cover classification categories and spatial neighborhood statistical characteristics, wherein the land cover classification categories take values ​​from 1 to M integers, corresponding to M types of land cover; and the spatial neighborhood statistical characteristics are the number of pixels of each land cover category within a preset-size neighborhood window around each pixel.

4. The method for repairing surface temperature anomalies in remote sensing images according to claim 1, characterized in that, Based on the surface temperature data to be restored, in the process of identifying and marking temperature anomaly pixels and high-confidence normal pixels, surface temperature anomalies are identified from two dimensions: data value range and pixel geospatial correlation, including: Pixels with null or non-finite surface temperatures are marked as temperature anomalous pixels; Pixels with surface temperatures below a preset low temperature threshold and above a preset high temperature threshold are marked as temperature anomalous pixels; Obtain the central tendency statistics of surface temperature of different land types in the target area. And the dispersion statistic D will deviate Pixels within a certain range are labeled as temperature anomaly pixels; among them, 1 represents the preset sensitivity coefficient; The remote sensing image is traversed pixel by pixel using two preset sliding windows. The temperature central tendency and dispersion statistics within the sliding windows of the two scales are calculated respectively. The local normal temperature range is constructed by combining the preset sensitivity coefficients of each scale. Pixels that exceed the local normal temperature range are marked as temperature anomalous pixels.

5. The method for repairing surface temperature anomalies in remote sensing images according to claim 1, characterized in that, In the pre-constructed surface temperature estimation model The input projection layer consists of a single linear transformation layer; The adaptive channel attention layer consists of a linear transformation layer, a ReLU nonlinear activation layer, another linear transformation layer, and a Sigmoid activation layer arranged in sequence. The residual block stack consists of L residual learning blocks with the same structure stacked together. Each residual learning block contains a fully connected transformation main path and a residual cross-layer direct connection shortcut path set in parallel. The regression head is composed of multiple residual learning blocks with progressively decreasing dimensions, strung together.

6. The method for repairing surface temperature anomalies in remote sensing images according to claim 1, characterized in that, The fully connected transformation main path consists of a first linear transformation layer, a batch normalization layer, a nonlinear activation layer, a regularized dropout layer, a second linear transformation layer, and a batch normalization layer, used to learn the residual mapping; the residual cross-layer direct connection shortcut path is composed of a linear transformation layer and a batch normalization layer to perform linear projection when the input and output dimensions are different, and is an identity connection when the input and output dimensions are the same.

7. The method for repairing surface temperature anomalies in remote sensing images according to claim 1, characterized in that, Based on the identification of temperature anomaly pixels, the abnormal areas to be repaired are located, including: The temperature anomaly pixels identified by the merging and identification are expanded using a morphological dilation operator to absorb isolated anomaly pixels in the neighborhood, resulting in an outlier binary image; wherein, in the outlier binary image, the abnormal region is 0 and the normal region is 1; Based on the binary image of the outlier, the standard land satellite remote sensing data of the target area is masked to complete the location of the outlier area to be repaired.

8. A system for repairing surface temperature anomalies in remote sensing images, characterized in that, include: The data acquisition module is used to acquire standard land satellite remote sensing data covering the target area, and extract multi-band surface reflectance data and surface temperature data to be repaired in the target area based on the standard land satellite remote sensing data. The multidimensional feature construction module is used to acquire the land cover classification data of the target area. Based on the multi-band land reflectance data of the target area, the standard data of land satellite remote sensing and the land cover classification data, a standardized multidimensional feature dataset containing spectral features, environmental features and land cover features is constructed and spatially registered and normalized. The anomaly identification and sample construction module is used to identify and mark temperature anomaly pixels and high-confidence normal pixels based on the surface temperature data to be repaired, and to construct a sample dataset based on each high-confidence normal pixel and the standardized multidimensional feature dataset; wherein, the high-confidence normal pixels are the remaining pixels other than the identified and marked temperature anomaly pixels; each sample data in the sample dataset includes: the standardized multidimensional feature vector corresponding to the high-confidence normal pixel and the ground truth label of the surface temperature; The model training and temperature prediction module is used to iteratively train the pre-built surface temperature estimation model based on the sample dataset until the model accuracy converges or the preset training termination condition is met, so as to obtain the trained surface temperature estimation model; based on the standardized multidimensional feature dataset, the standardized multidimensional feature vector of each pixel is input into the trained surface temperature estimation model to perform temperature estimation, so as to obtain the predicted surface temperature value of each pixel. The abnormal area filling and repair module is used to locate the abnormal area to be repaired based on the temperature abnormality pixels identified by the identification marker, and fill and repair the abnormal area based on the predicted surface temperature values ​​of each pixel to obtain the surface temperature data of the target area after repair and reconstruction. The pre-built land surface temperature estimation model is an improved residual multilayer perceptron, which sequentially includes an input projection layer, an adaptive channel attention layer, a residual block stack, and a regression head. The input projection layer is used to input standardized multidimensional feature vectors, map features from the original input dimensions to preset hidden dimensions, and output hidden dimension feature vectors. The adaptive channel attention layer is used to input hidden dimension feature vectors, calculate element-wise gating weights for each dimension of features, and adjust the original features element-wise based on the gating weights, outputting a filtered and enhanced weighted feature sequence. The residual block stack is used to input the weighted feature sequence. Each residual block is configured with an independent fully connected transformation main path and a residual cross-layer direct connection shortcut path. The transformation results of the main path are superimposed and fused with the original input features transmitted by the shortcut path to output deep fused features. The regression head is used to input deep fused features, and after multi-level dimensional compression linear mapping operations, outputs a single-dimensional predicted land surface temperature value.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the remote sensing image surface temperature anomaly repair method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the remote sensing image surface temperature anomaly repair method as described in any one of claims 1 to 7.

Images