A land surface temperature downscaling method and system based on feature interaction optimization and multi-model screening mechanism

Abstract

The application discloses a kind of based on feature interaction optimization and multi-model screening mechanism land surface temperature downscaling method and system, belong to land surface temperature data processing technical field, steps are as follows: filtering to spectral band, remote sensing index and topographic feature;And topographic feature parameter is resampled to with optical band and remote sensing index matching resolution;Significantly related variable is filtered by correlation analysis with land surface temperature, interactive feature is constructed, and high contribution feature is dynamically filtered based on SHAP value;Based on linear model and machine learning model double frame, regression model is constructed, and regression model is verified;SHAP optimized feature extracted from Sentinel-2 data is fused with SRTM topographic data, input verified regression model, and generate 10 meter resolution land surface temperature data;The application uses combined feature interaction to mine high-order nonlinear relationship, and the prediction RMSE of homogeneous land surface area is greatly reduced, the data scale effect is retained to make the prediction more in line with physical process, and spatial consistency is better.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of land surface temperature data processing, and in particular relates to a land surface temperature downscaling method and system based on feature interaction optimization and multi-model screening mechanism. BACKGROUND

[0002] Under the background of global climate change, accelerated urbanization and increasingly prominent ecological and environmental problems, it is crucial to accurately obtain high-resolution land surface temperature data. As a basic climate variable determined by the global climate observation system, land surface temperature plays an indispensable role in many fields. In climate change monitoring, accurate land surface temperature data can reflect the regional and even global heat change trend, helping scientists to deeply study the mechanism and influence of climate warming. In the assessment of urban heat island effect, land surface temperature can intuitively present the temperature distribution difference within the city, providing key basis for urban planning and strategy formulation to alleviate the heat island effect. In precision irrigation in agriculture, land surface temperature can assist in judging the water demand of crops, realizing efficient use of water resources. In the analysis of ecosystem health, it is helpful to evaluate the energy balance and stability of the ecosystem.

[0003] Currently, thermal infrared remote sensing is the main means to obtain regional scale land surface temperature data, but there are problems of long revisit period and high data acquisition cost. For example, in the case of traditional thermal infrared satellite data acquisition, the revisit period may be several weeks or even months. In some rapidly changing areas, such as urban expansion areas or ecologically fragile areas, it is difficult to capture the dynamic changes of land surface temperature in a timely manner. The high cost of data acquisition also limits its application in large-scale and long-term monitoring, making it difficult to meet the urgent demand for high spatio-temporal resolution land surface temperature data.

[0004] In existing land surface temperature downscaling methods, although traditional empirical or semi-empirical methods are computationally efficient, they are heavily dependent on regional experience. In the face of complex and diverse land cover, such as mixed areas of buildings, vegetation, water bodies and other ground objects in cities, and complex terrain areas in mountainous regions, it is difficult to accurately model, resulting in poor downscaling effect. Although geographically weighted regression can capture spatial heterogeneity to some extent, its accuracy is often limited to the order of hundreds of meters, which cannot meet the demand for high-precision land surface temperature data. Although deep learning models can achieve high-precision resolution and large-scale generalization by virtue of multi-layer nonlinear network architecture in mining complex spatial correlations among multi-source remote sensing data, there are still deficiencies in spatial feature representation ability and model computation efficiency optimization.

[0005] Meanwhile, the machine learning model is restricted by the selection of feature variables and the applicability in the research of land surface temperature downscaling: the existing methods directly apply common feature variables (spectral index, remote sensing vegetation index, urbanization index and terrain index) or rely on domain experience for screening, so that the complex nonlinear relationship and interaction effect between variables are not fully mined, resulting in insufficient utilization of multi-source data; at the same time, different models perform differently in different scenes, the random forest model has high precision, but in the complex heterogeneous surface environment, the capture ability of high-order interaction between variables is limited, and the prediction efficiency is limited; the deep learning model has strong ability to process nonlinear relationship, but its high computational complexity and "black box" characteristics become a bottleneck for its wide application in areas with limited computing resources or real-time monitoring scenes.

[0006] Therefore, a land surface temperature downscaling method with high precision and high stability is particularly important. SUMMARY

[0007] Problems to be solved

[0008] In view of the problems existing in the prior art, the present application provides a land surface temperature downscaling method and system based on feature interaction optimization and multi-model screening mechanism.

[0009] Technical scheme

[0010] In order to solve the above problems, the present application adopts the following technical scheme.

[0011] A land surface temperature downscaling method based on feature interaction optimization and multi-model screening mechanism, the steps are as follows:

[0012] S1, multi-source remote sensing data preprocessing: selecting spectral bands, remote sensing indexes and terrain features, including preprocessing of optical bands, remote sensing indexes and terrain feature parameters of 100-meter resolution data of Landsat9 and 10-meter resolution data of Sentinel-2; the optical bands include red, green and blue primary color bands, the remote sensing indexes include enhanced vegetation index, normalized building index and normalized water index, and the terrain feature parameters include elevation, slope and slope direction, and the terrain feature parameters are resampled to the resolution matched with the optical bands and remote sensing indexes;

[0013] S2, variable screening: screening variables significantly related to land surface temperature through correlation analysis, constructing interaction features, and dynamically screening high-contribution features based on SHAP value;

[0014] S3, establishing a regression model: based on the double framework of linear model and machine learning model, constructing a regression model, and verifying the regression model;

[0015] S4, High-resolution land surface temperature generation: The SHAP optimized features extracted from Sentinel-2 data are fused with SRTM terrain data, input into the verified regression model, and 10-meter resolution land surface temperature data is generated.

[0016] Preferably, the method for constructing interaction features in step S2 is:

[0017] The selected variables are divided into four categories: spectral bands, remote sensing indices, terrain features, and nonlinear transformation features. Through non-repeating combination of combinatorics, intra-class interaction features and cross-class interaction features are generated, including interactions between spectral bands, between bands and remote sensing indices, between bands and terrain features, between bands and nonlinear transformation features, between remote sensing indices, between remote sensing indices and terrain features, between remote sensing indices and nonlinear transformation features, between terrain features, between terrain features and nonlinear transformation features, and between nonlinear transformation features. Finally, 105 independent interaction features are formed, and the interaction operation preserves the original data scale effect.

[0018] Further, the specific method for SHAP dynamic screening in step S2 is:

[0019] Based on the contribution of SHAP value to the interaction feature, the top 10 high-contribution features are selected as model inputs. Through the model independence and interpretability mechanism of SHAP, the core driving factors in nonlinear relationships are identified, and redundant variables are removed.

[0020] Preferably, the training framework of the machine learning model in step S3 is:

[0021] The model is trained at 100-meter resolution and directly migrated to 10-meter resolution Sentinel-2 data for prediction. Through post-processing, outliers are removed to ensure the spatial consistency of high-resolution land surface temperature data.

[0022] Further, when the regression model is verified in step S3, the specific method is:

[0023] 10-fold cross-validation is used to evaluate the stability and generalization ability of the regression model: each validation performs triple evaluation—coefficient of determination (R²) quantifies the explanation of the predicted features to the variation of land surface temperature, root mean square error (RMSE) measures the deviation between predicted and measured temperature, and mean absolute error (MAE) directly reflects the prediction error level; models that meet all three criteria are considered effective.

[0024] Further, the model effectiveness evaluation criteria in step S3 include:

[0025] The coefficient of determination (R2) quantifies the degree of explanation of the predicted features to the variation of the land surface temperature, the root mean square error (RMSE) measures the deviation between the predicted and measured temperature, the mean absolute error (MAE) reflects the level of prediction error, and the model satisfying R2>0.4, RMSE<2°C and MAE<1.5°C is determined as effective.

[0026] A land surface temperature downscaling system based on feature interaction optimization and multi-model screening mechanism, comprising:

[0027] A multi-source remote sensing data preprocessing module: used for acquiring multi-source remote sensing data and terrain data, and performing preprocessing operations such as extraction and resampling on the data to ensure spatial consistency of the data.

[0028] A variable screening module: including a correlation analysis unit, an interaction feature construction unit and a SHAP score screening unit, used for screening high-contribution features.

[0029] A model establishment module: used for constructing a linear model and a machine learning model double framework, training and verifying the model, and selecting an effective model.

[0030] A high-resolution land surface temperature generation module: used for fusing the optimized features and terrain data, inputting the effective model, and generating high-resolution land surface temperature data.

[0031] Further, the data preprocessing module further includes a data quality control unit for performing radiation calibration, atmospheric correction and other processing on the multi-source remote sensing data to improve data quality.

[0032] Further, the interaction feature construction unit of the variable screening module can generate intra-class interaction features and cross-class interaction features according to feature classification rules, and the feature classification rules divide the features into spectral bands, remote sensing indexes, terrain features and nonlinear transformation features.

[0033] Advantages

[0034] Compared with the prior art, the present application has the following advantages:

[0035] Improved accuracy and stability, suitable for complex scenarios: using combined feature interaction to mine high-order nonlinear relationships, the RMSE in homogeneous surface area prediction is greatly reduced, the data scale effect is preserved to make the prediction more consistent with the physical process, and the spatial consistency is better;

[0036] Optimized calculation and generalization ability, eliminate redundant features: filter features with the help of SHAP framework, reduce model training calculation amount, improve efficiency, enhance model generalization in different scenarios, and avoid overfitting;

[0037] High-resolution products are highly practical: the generated 10-meter resolution surface temperature data meets the accuracy requirements, providing strong data support for refined scenarios such as urban heat island management and agricultural drought early warning;

[0038] Excellent technical compatibility and scalability: The multi-model library allows for the addition of new models and feature types, enabling the integration of new technologies to expand research dimensions and adapt to ever-changing research and application needs. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments or examples of this application, the accompanying drawings used in the embodiments or examples will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0040] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments. Generally, the components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0042] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. Example 1

[0043] like Figure 1 As shown, a land surface temperature downscaling method based on feature interaction optimization and multi-model screening mechanism includes:

[0044] Data acquisition and preprocessing:

[0045] 100-meter resolution data of Landsat 9 for the study area were obtained from the U.S. Geological Survey website. The data includes optical bands, enhanced vegetation index, normalized building index, normalized water index, and topographic feature parameters such as elevation, slope, and aspect. 10-meter resolution data of Sentinel-2 for the corresponding area were obtained from the European Space Agency's Copernicus Open Access Centre. The data also covers the above-mentioned optical bands and remote sensing indices.

[0046] Resample the terrain feature parameters using geographic information processing software such as ENVI or ArcGIS.

[0047] Take Landsat9 data as an example, resample the elevation, slope, and aspect to 100-meter resolution to match the optical band and remote sensing index resolution of Landsat9. For Sentinel-2 data, resample the terrain feature parameters to 10-meter resolution. At the same time, perform radiometric calibration and atmospheric correction on all data to eliminate the effects of atmospheric scattering, absorption, and other factors on data quality.

[0048] Variable selection:

[0049] Use Python's pandas and scipy libraries to calculate the Spearman correlation coefficient between all satellite bands, vegetation index, terrain index, and land surface temperature. Assuming there are 100 sample points in the study area, calculate the correlation coefficient between each variable and LST, and select variables with a significance level P<0.001, such as red band, EVI, slope, etc.

[0050] Divide the selected variables into four categories: spectral bands, remote sensing indices, terrain features, and nonlinear transformation features. Use Python's itertools library to perform non-repetitive combination calculations to generate intra-class interaction features and cross-class interaction features, ultimately forming 105 independent interaction features. In the calculation process, directly use the original data for multiplication to preserve the data scale effect.

[0051] Use the SHAP library to calculate the SHAP value of each interaction feature, and sort the 105 interaction features based on their contribution values. Select the top 10 high-contribution features. For example, after calculation, it is found that the SHAP value of EVI×elevation and NDBI×slope is high, so they are used as model inputs and redundant variables are removed.

[0052] Establish a regression model:

[0053] Construct a linear model using 11 basic features, and a machine learning model using high-contribution features optimized by SHAP. Take the XGBoost model as an example, train it under the 100-meter resolution Landsat9 data training framework.

[0054] 10-fold cross-validation is used to evaluate the stability and generalization ability of the model: 100 sample point data are randomly divided into 10 parts, 9 of which are taken as the training set and 1 as the test set each time; in each validation, the coefficient of determination, root mean square error and mean absolute error are calculated; after multiple cross-validation, if a certain model meets R²>0.4, RMSE<2°C and MAE<1.5°C, it is determined that the model is effective.

[0055] High-resolution land surface temperature generation:

[0056] SHAP-optimized features are extracted from Sentinel-2 data and fused with 30-meter resolution SRTM terrain data; for example, the selected EVI, NDBI and other optimized features are combined with SRTM elevation data.

[0057] The fused data is input into the verified effective regression model to generate 10-meter resolution land surface temperature data; the generated data is post-processed by setting temperature threshold range to remove outliers and ensure spatial consistency of the data. Example 2

[0058] A land surface temperature downscaling system based on feature interaction optimization and multi-model screening mechanism, comprising:

[0059] Multi-source remote sensing data preprocessing module:

[0060] Data acquisition unit: develop a special data acquisition program to automatically download Landsat9, Sentinel-2 data and SRTM terrain data in the study area by calling the API interface of USGS and ESA.

[0061] Resampling unit: use GDAL library to write resampling code to match the resolution of terrain feature parameters and optical bands, remote sensing index.

[0062] Data quality control unit: integrate radiation calibration and atmospheric correction algorithms such as 6S model to process multi-source remote sensing data and improve data quality.

[0063] Variable selection module:

[0064] Correlation analysis unit: based on Python data analysis library, develop correlation calculation program to calculate Spearman correlation coefficient of all variables with LST and select variables that meet the conditions.

[0065] Interaction feature construction unit: according to feature classification rules, use Python combination calculation function to generate intra-class and cross-class interaction features, forming 105 independent interaction features.

[0066] SHAP score screening unit: call SHAP library, calculate SHAP value of interaction feature, sort contribution degree, and screen top 10 high contribution features.

[0067] Model building module:

[0068] Model building unit: use machine learning libraries such as Scikit-learn and XGBoost to build linear models and machine learning models.

[0069] Model training unit: train the model on 100-meter resolution data and set appropriate training parameters.

[0070] Model verification unit: perform 10-fold cross-validation, calculate R², RMSE and MAE, and determine whether the model is effective.

[0071] High-resolution land surface temperature generation module:

[0072] Feature fusion unit: fuse SHAP optimized features of Sentinel-2 with SRTM terrain data.

[0073] Temperature generation unit: input the fused data into the effective model to generate 10-meter resolution land surface temperature data.

[0074] Post-processing unit: write data post-processing program, remove outliers, and output high-quality 10-meter resolution land surface temperature data.

[0075] In summary, the present application has the following beneficial effects:

[0076] Significantly improve accuracy and stability, enhance adaptability in complex scenarios: the present application uses innovative combined feature interaction method, such as constructing EVI×elevation, NDBI×slope, etc. interaction features, to fully exploit the high-order nonlinear relationship between different types of features. In the test of homogeneous surface area such as vegetation coverage area, compared with traditional machine learning method, the root mean square error of prediction is significantly reduced; at the same time, the original data scale effect is preserved, making the model more consistent with the real physical process of land surface energy exchange, and the spatial consistency of the prediction result is obviously improved, which can more truly reflect the distribution of land surface temperature.

[0077] Optimize computing efficiency and generalization, remove redundant features: introduce SHAP explainability framework for dynamic feature screening, based on quantitative feature contribution, accurately identify and remove redundant features; when processing large-scale multi-source remote sensing data, the amount of model training is reduced, and the computing efficiency is improved; after SHAP screening, the number of model input features is reduced by about 30%-50%, but the model accuracy is not reduced, on the contrary, the generalization ability in different scenarios is enhanced;

[0078] In specific implementation, the stability and adaptability of the model are significantly improved through verification in multiple test areas in different climate zones and different topographies, overfitting phenomenon is effectively avoided, and stable and reliable operation in complex and variable actual application scenarios is ensured.

[0079] Improve the practicability of high-resolution products: successfully generate 10-meter resolution land surface temperature data, and meet the index requirements of determination coefficient > 0.4 and root mean square error < 2°C.

[0080] Possess technical compatibility and expansibility: the multi-model library constructed by the application has good expansibility, allows to add machine learning models or feature types, and can flexibly integrate new models and data, further improving the performance and application range of land surface temperature downscaling.

[0081] The above-described embodiments only express the preferred embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the patent of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications, improvements and substitutions can be made, which are all within the protection scope of the present application.

Claims

1. A method for downscaling land surface temperature based on feature interaction optimization and multi-model screening mechanism, characterized in that, The steps are as follows: S1. Multi-source remote sensing data preprocessing: Screening of spectral bands, remote sensing indices, and topographic features, including preprocessing of optical bands, remote sensing indices, and topographic feature parameters of Landsat 9 100-meter resolution data and Sentinel-2 10-meter resolution data; the optical bands include red, green, and blue primary color bands, the remote sensing indices include enhanced vegetation index, normalized building index, and normalized water index, and the topographic feature parameters include elevation, slope, and aspect, and the topographic feature parameters are resampled to a spatial resolution that matches the optical bands and remote sensing indices; S2. Variable selection: Variables that are significantly related to surface temperature are selected through correlation analysis, interaction features are constructed, and high-contribution features are dynamically selected based on SHAP values. S3. Establish a regression model: Based on the dual framework of linear model and machine learning model, construct a regression model and validate the regression model; S4. High-resolution surface temperature generation: The SHAP optimized features extracted from Sentinel-2 data are fused with SRTM terrain data and input into a validated regression model to generate surface temperature data with a resolution of 10 meters. The method for constructing the interaction features in step S2 is as follows: The selected variables were divided into four categories: spectral bands, remote sensing indices, terrain features, and nonlinear transformation features. Through combinatorial mathematics, non-repeating combinations were used to generate intra-class and cross-class interactive features, including interactions between spectral bands, between bands and remote sensing indices, between bands and terrain features, between bands and nonlinear exchange features, between remote sensing indices, between remote sensing indices and terrain features, between remote sensing indices and nonlinear exchange features, between terrain features, between terrain features and nonlinear exchange features, and between nonlinear exchange features. Finally, 105 sets of independent interactive features were formed, and the interaction operation preserved the original data scaling effect. The specific method for dynamic SHAP filtering in step S2 is as follows: The contribution of interaction features is ranked based on SHAP values. The top 10 features with the highest contribution are selected as model inputs. The core driving factors in nonlinear relationships are identified and redundant variables are eliminated through the model independence and interpretability mechanism of SHAP.

2. The surface temperature downscaling method based on feature interaction optimization and multi-model screening mechanism according to claim 1, characterized in that: The training framework for the machine learning model in step S3 is as follows: The model was trained at a 100-meter resolution and then directly transferred to Sentinel-2 10-meter resolution data for prediction. Outliers were removed through post-processing to ensure the spatial consistency of the high-resolution surface temperature data.

3. The surface temperature downscaling method based on feature interaction optimization and multi-model screening mechanism according to claim 1, characterized in that: The specific method for step S3 in validating the regression model is as follows: The stability and generalization ability of the regression model were evaluated using 10-fold cross-validation: each validation performed a triple evaluation—the coefficient of determination R² quantifies the explanatory power of the predicted features for surface temperature variation, the root mean square error RMSE measures the deviation between the predicted and measured temperatures, and the mean absolute error MAE directly reflects the level of prediction error; models that meet all three criteria are deemed valid.

4. The surface temperature downscaling method based on feature interaction optimization and multi-model screening mechanism according to claim 3, characterized in that: The model validity evaluation criteria in step S3 include: The coefficient of determination R² quantifies the explanatory power of the predicted features for surface temperature variation, the root mean square error RMSE measures the deviation between the predicted and measured temperatures, and the mean absolute error MAE reflects the level of prediction error. Models that satisfy R² > 0.4, RMSE < 2°C, and MAE < 1.5°C are considered valid.

5. A system for the land surface temperature downscaling method based on feature interaction optimization and multi-model screening mechanism as described in claim 1, characterized in that, include: Multi-source remote sensing data preprocessing module: used to acquire multi-source remote sensing data and terrain data, and perform data extraction and resampling preprocessing operations to ensure spatial consistency of the data; Variable screening module: includes correlation analysis unit, interaction feature construction unit and SHAP scoring screening unit, used to screen high contribution features; Model building module: used to build a dual framework of linear models and machine learning models, train and validate models, and select effective models; High-resolution surface temperature generation module: used to fuse optimized features with terrain data, input into an effective model, and generate high-resolution surface temperature data; The data preprocessing module also includes a data quality control unit, which is used to perform radiometric calibration and atmospheric correction on multi-source remote sensing data to improve data quality. The interactive feature construction unit of the variable screening module can generate intra-class interactive features and cross-class interactive features according to the feature classification rules. The feature classification rules divide the features into spectral bands, remote sensing indices, terrain features and nonlinear transformation features.

Images