Full-coverage daily minimum air temperature remote sensing estimation method and system

Abstract

The invention discloses a full-coverage daily minimum air temperature remote sensing estimation method and system, and the system comprises the steps that a remote sensing collection module is connected with a satellite sensor and a public data set, and obtains related remote sensing information data; the data processing module preprocesses related remote sensing information data to obtain a remote sensing index set, a data set of altitude information, surface temperature and air temperature; an algorithm analysis module uses XGBoost for training according to the data preprocessed by the data processing module, and air temperature data with 1km spatial resolution per hour is calculated; and the air temperature estimation module performs day-by-day iteration processing on the air temperature data hour by hour every day according to the air temperature data, outputs day-by-day daily minimum air temperature data and calculates the daily minimum air temperature. According to the method, data are rapidly processed and analyzed, daily minimum temperature information updated in real time is provided, the estimation result is high in accuracy, and weather changes are responded in time.

Description

Technical Field

[0001] The present invention relates to the technical field of air temperature remote sensing, and in particular to constructing a full-coverage daily minimum temperature remote sensing estimation method and system. Background Art

[0002] Accurate measurement of daily minimum temperature is crucial for agriculture, post-disaster management and public health. Although traditional ground-based meteorological stations play an important role in temperature monitoring, their distribution is limited by geographical location and cost. With the advancement of remote sensing technology, thermal infrared remote sensing technology has become a widely covered and cost-effective alternative to obtain daily minimum temperature.

[0003] However, thermal infrared remote sensing technology also has some shortcomings. First, thermal infrared remote sensing data is greatly affected by clouds and rain, which makes it impossible to obtain full coverage of temperature data under severe weather conditions, and the temporal resolution is limited, making it difficult to provide high-frequency temperature change monitoring. On the other hand, although numerical weather forecast models can provide temperature data with high temporal resolution, their spatial resolution is low, which is particularly unfavorable in areas with complex terrain or densely populated cities. At the same time, existing meteorological models usually operate on a larger spatial scale and have difficulty capturing microclimate changes caused by terrain or local climate conditions. Although traditional temperature downscaling methods can improve spatial resolution, they have high computational costs and cannot provide very detailed temperature change information, and the temperature prediction accuracy for local areas is not enough. These problems limit our ability to monitor and predict temperature changes in more detail. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a method for estimating the daily minimum temperature based on multi-source remote sensing data to improve the accuracy and resolution of temperature estimation. The air temperature, surface temperature LST and other environmental and remote sensing indices of ERA5 are used to estimate the daily minimum temperature with a full coverage of 1km spatial resolution by introducing the XGBoost algorithm. This method is suitable for meteorological estimation, agricultural management and environmental monitoring, etc., which require high-resolution daily minimum temperature data.

[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a full-coverage daily minimum temperature remote sensing estimation system, including a remote sensing acquisition module, a data processing module, an algorithm analysis module and a temperature estimation module.

[0006] a. Remote sensing acquisition module, which obtains relevant remote sensing information data by connecting satellite sensors and public data sets, including remote sensing image units, climate acquisition units and digital elevation units.

[0007] Preferably, the remote sensing image unit is connected to the sensor on the satellite through the network to collect the bands required for the surface temperature LST and the bands required for the remote sensing index in a certain area. The climate acquisition unit is connected to the fifth-generation atmospheric reanalysis data set ERA5 data set of the European Center for Medium-Term Weather Forecasts ECMWF in a certain area, and is transmitted to the data processing module after being numbered. The digital elevation unit is connected to the geospatial data cloud to obtain the DEM data in the digital elevation model data of the area and transmit it to the data processing module.

[0008] b. The data processing module pre-processes the relevant remote sensing information data to obtain a remote sensing index set, a data set of altitude information, surface temperature and air temperature.

[0009] Preferably, according to the surface temperature data products collected from the satellite, after format conversion, splicing and cropping, the four groups of LST images of Terra daytime, Terra nighttime, Aqua daytime and Aqua nighttime are fully covered by the XGBoost algorithm to achieve seamless estimation of the surface temperature LST data, and finally obtain the four groups of fully covered surface temperature LST of Terra daytime, Terra nighttime, Aqua daytime and Aqua nighttime in the area.

[0010] Preferably, the ERA5 dataset of the region is clipped, projected, format converted and data cleaned, and the abnormal temperature grid values ​​are excluded to obtain the hourly 2m air temperature T with a spatial resolution of 0.1°×0.1°. 10km .

[0011] Preferably, according to the vegetation index data products collected from the satellite, after format conversion, cropping and splicing, the SG filtering algorithm and linear time interpolation method are used to obtain the normalized vegetation index NDVI with full coverage and 1 km spatial resolution.

[0012] Preferably, the NDWI and NDBI of the region are obtained by calculation based on the bands required by the remote sensing index collected from the satellite. The calculation formula is as follows:

[0013]

[0014] In the formula, NIR represents the reflectance value of the near-infrared band, and SWIR represents the reflectance value of the short infrared band. After format conversion, splicing, cropping and resampling, the SG filtering algorithm and linear time interpolation method are used to obtain the normalized water index NDWI and the normalized building index NDBI with full coverage of 1km spatial resolution, which are merged with the normalized vegetation index NDVI to form a remote sensing index set.

[0015] Preferably, the DEM data of the region are merged, and the vector boundaries of the region are used for projection conversion, clipping and resampling to obtain a data set of altitude information.

[0016] The average aggregation method is used to upscale the data in the data sets of land surface temperature (LST), remote sensing index set and altitude information from 1 km spatial resolution to 10 km on average, so that its spatial resolution matches the temperature data in the ERA5 dataset.

[0017] c. The algorithm analysis module uses XGBoost for training based on the data preprocessed by the data processing module to calculate the hourly air temperature data with a spatial resolution of 1 km.

[0018] Preferably, XGBoost is used to 10km The training was carried out, and the feature variables involved in the training included the data set of LST, NDVI, NDBI, NDWI and altitude information after average aggregation; the LST, NDVI, NDBI, NDWI and altitude information with a spatial resolution of 1 km before average aggregation were input into the trained XGBoost model as explanatory variables to obtain the hourly temperature T with a spatial resolution of 1 km. 1km ; The algorithm analysis module is connected to the temperature estimation module through a network.

[0019] Preferably, in the algorithm analysis module, multiple linear regression is added as a comparative experiment, and T 10km After training with the averaged aggregated LST, NDVI, NDBI, NDWI, and DEM feature variables, the 1km resolution NDVI, NDBI, NDWI, LST, and DEM are used as input data, and the calculation formula is as follows:

[0020] T 1km =m 0 +m 1 LST+m 2 NDVI+m 3 NDBI+m 4 NDWI+m 5 DEM+ε

[0021] In the formula, m 1 …m 5 is the regression coefficient, ε is the error, m 0 is the intercept, T 1km is the estimated temperature for the multiple linear regression model.

[0022] Preferably, in the algorithm analysis module, a random forest is added as a comparative experiment. This algorithm integrates multiple decision trees to combine and classify their prediction results.

[0023] T 10km A nonlinear regression relationship is constructed with driving factors such as LST, NDVI, NDBI, NDWI, and DEM after average aggregation to achieve downscaling of multi-level spatial resolution.

[0024] The calculation formula is as follows:

[0025] Tree i =BuildTree(S i , F i )

[0026]

[0027] In the formula, the independent variable is F i ={f 1 ,f 2 ,....f m}, F i Represents the feature set of the i-th decision tree, Si = {x 1 ,x 2 ,....x n}, where S i Represents the training set of the i-th tree, BuildTree is the decision tree construction function, Tree i is the i-th decision tree. k i is the number of decision trees, Tree i (x) is the prediction result of the i-th tree for the sample, T 1km It is the estimated temperature of the random forest model and is transmitted to the temperature estimation module through the network.

[0028] Preferably, in the algorithm analysis module, an artificial neural network ANN is added as a comparative experiment. ANN learns and adjusts T 10km The weights of LST, NDVI, NDBI, NDWI, and DEM after average aggregation are used for information processing and pattern recognition, and the weight parameters of the network are constructed according to the training criteria; the calculation formula is as follows:

[0029]

[0030] In the formula, w k is the weight from the hidden layer to the output layer, h k is the output of the hidden layer neural unit, b is the bias constant used to adjust the result of the linear combination, T 1km It is the estimated temperature of the artificial neural network model and is transmitted to the air temperature estimation module through the network.

[0031] d. Temperature estimation module, based on the air temperature data, iterates the air temperature data every hour every day, outputs the daily minimum temperature data, and calculates the daily minimum temperature T min The data is outputted to the user end through the network, and the accuracy is verified by comparing the downscaled daily minimum temperature result with the measured minimum temperature at the meteorological station.

[0032] Preferably, the specific implementation process in the temperature estimation module is as follows: based on the XGBoost predicted temperature T 1km , create the daily minimum temperature function f min , perform dimensionality reduction operations and aggregate the time dimension.

[0033] Reduce the temperature data of each hour into a single grid data, extract the lowest temperature of each grid, and iterate the data objects of each hour day by day. The calculation formula is as follows:

[0034] T min (x,y)=f min (T 1km (x,y,k)

[0035] In the formula, x and y represent the row and column index of the grid, respectively; k represents the time index in hours, with a value range of (1-24); 1km (x,y,k) represents the temperature of the grid cell in the xth row and yth column at the kth hour, T min (x, y) is the daily minimum temperature of each grid, and the daily minimum temperature of each grid is integrated into the daily minimum temperature T min After that, the data is transmitted to the user end through the network.

[0036] Preferably, a full coverage daily minimum temperature remote sensing estimation method comprises the following steps:

[0037] Step 1: Collect the surface temperature data products, vegetation index data products and the bands required for remote sensing index of a certain area through satellites; obtain the ERA5 data set of the area through the Weather Forecast Center; and obtain the data DEM in the digital elevation model data of the area through the geospatial data cloud.

[0038] Step 2: Preprocess the land surface temperature data products and use the XGBoost algorithm to seamlessly estimate four sets of fully covered land surface temperatures LST; preprocess the data in the ERA5 dataset to obtain the air temperature T 10km; Preprocess the bands of vegetation index data products and remote sensing index, and use the SG filtering algorithm and linear time interpolation method to obtain full coverage of NDVI, NDBI, and NDWI; obtain altitude information data by preprocessing DEM data; use the average aggregation method to calculate the data set of LST, NDVI, NDBI, NDWI and altitude information.

[0039] Step 3: Use the XGBoost algorithm based on the obtained air temperature T 10km The characteristic relationship between the average and aggregated data sets is used to calculate the hourly air temperature data with a spatial resolution of 1 km.

[0040] Step 4: Based on the hourly root air temperature data obtained in step 3, construct the daily minimum temperature function f min Perform iterative processing day by day to calculate the daily minimum temperature T min .

[0041] Compared with the prior art, the present invention provides a method for estimating the daily minimum temperature by combining multi-source remote sensing data, which has the following beneficial effects:

[0042] 1. The present invention sets a remote sensing image unit, a climate acquisition unit and a digital elevation unit through a remote sensing acquisition module. The data processing module realizes data preprocessing and full coverage. The algorithm analysis module reveals the performance of machine learning in the spatial downscaling process by using XGBoost, and the high and low values ​​of the downscaling results are better restored.

[0043] 2. The estimation output module of the present invention calculates the minimum temperature of the full coverage day T min , analyzed multiple groups of interference variables such as climate conditions, geographical altitude characteristics, vegetation coverage and urban buildings, and used remote sensing data as features to effectively compensate for the errors of reanalyzed temperature data. The estimation results are highly accurate. The estimation output module network connects to the user end to arrange and output multiple groups of daily minimum temperature data, which is convenient for monitoring personnel to obtain spatial temperature information more intuitively, and can quickly process and analyze data, provide real-time updated daily minimum temperature information, respond to weather changes in a timely manner, and have higher reference value. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the structure of the present invention;

[0045] Figure 2 This is a schematic diagram of the study area of ​​the present invention;

[0046] Figure 3 It is a step diagram of the method of the present invention;

[0047] Figure 4 It is a schematic diagram of the accuracy of the present invention;

[0048] Figure 5 This is a full coverage daily minimum temperature result map;

[0049] Figure 6 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0050] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0051] This study used machine learning combined with multi-source remote sensing data to estimate the daily minimum temperature, and conducted a comprehensive analysis using multiple data sources such as meteorological data, geographic information, and satellite data, which improved the accuracy and reliability of the predicted daily minimum temperature.

[0052] For specific implementation measures, please refer to Figure 1 ,A method for estimating daily minimum temperature by remote sensing includes a remote sensing acquisition module, a data processing module, an algorithm analysis module and a temperature estimation module.

[0053] The remote sensing acquisition module includes a remote sensing image unit, a climate acquisition unit and a digital elevation unit.

[0054] The remote sensing acquisition module is connected to the products developed by the medium-resolution imaging spectrometers on NASA's Terra and Aqua satellites (https: / / earthdata.nasa.gov) through the network. The land surface temperature LST comes from MOD11A1 and MYD11A1, NDVI comes from MOD13A2 and MYD13A2, NDBI and NDWI come from the short infrared band and near infrared band of MOD09A1 and MYD09A1; the time span of the data in the remote sensing image unit is from 2002 to 2023, the temporal resolution is daily, and the spatial resolution is 1km×1km.

[0055] The climate collection unit is connected to the ERA5 dataset through the network, and collects and numbers the ERA5 data provided by ECMWF (https: / / www.ecmwf.int), which contains up to 137 layers of atmospheric data covering the world and provides temperature information at different altitudes. The data used in the climate collection unit is hourly air temperature data 2m above the ground, with a time span of 2002 to 2022 and a spatial resolution of 10km×10km.

[0056] The digital elevation units were numbered through network connections to the DEM datasets of the three northeastern provinces, which were sourced from the Geospatial Data Cloud (https: / / www.gscloud.cn).

[0057] Preferably, the remote sensing acquisition module is connected to the data processing module via a network.

[0058] The data processing module includes the preprocessing and full coverage of the remote sensing index set, as well as data spatial resolution matching through the average aggregation method.

[0059] Firstly, the surface temperature data products provided in MOD11A1 and MYD11A1 were converted into different formats and processed by splicing and cropping to finally obtain four sets of surface temperatures (LST) of Terra daytime, Terra nighttime, Aqua daytime and Aqua nighttime in the three northeastern provinces with a spatial resolution of 1 km.

[0060] Preferably, the XGBoost algorithm is used to fully cover the four groups of LST images of Terra daytime, Terra nighttime, Aqua daytime, and Aqua nighttime to achieve full coverage of the surface temperature LST data. The results are shown in Table 1.

[0061] Table 1 Comparison of the accuracy of XGBoost algorithm for full coverage of LST during daytime and nighttime in the three northeastern provinces

[0062] <![CDATA[R 2 ]]> RMSE MAE Terra Daytime 0.97384 0.7583 0.4932 Terra Night 0.96903 0.6163 0.4243 Aqua Day 0.95797 0.9732 0.6385 Aqua Night 0.96704 0.6543 0.4532

[0063] Based on the vegetation index data products collected from satellites, after preprocessing, the three northeastern provinces were taken as the study area, and the SG filtering algorithm and linear time interpolation method were used to obtain the normalized vegetation index NDVI with full coverage and 1km spatial resolution.

[0064] The calculation is carried out with the three northeastern provinces as the study area, and the calculation formulas for NDWI and NDBI of the study area are as follows:

[0065]

[0066] In the formula, NIR represents the reflectance value of the near infrared band, and SWIR represents the reflectance value of the short infrared band. The range of NDVI, NDBI, and NDWI values ​​is between (-1, 1). NDVI reflects the degree of vegetation coverage, NDBI represents the degree of urban building coverage, and NDWI represents the distribution of water bodies. Then, the SG filtering algorithm and linear time interpolation method are used to obtain the full coverage NDVI, NDBI, and NDWI. By combining these explanatory variables, this method can analyze the spatial variability of air temperature at a more detailed scale.

[0067] The preprocessing of the climate collection unit data involves the extraction of the air temperature at a height of 2m from ERA5. After downloading the data, it is clipped, projected, and format converted for the three northeastern provinces. Then, data cleaning is performed to exclude abnormal temperature grid values, and the hourly 2m air temperature T with a spatial resolution of 10km from 2002 to 2022 is obtained.10km .

[0068] The preprocessing of digital elevation unit data involves extracting the altitude, projecting the DEM data, and using the vector boundaries of the three northeastern provinces for projection transformation, clipping, and resampling, and then obtaining DEM data with a spatial resolution of 1 km above sea level, such as Figure 2 shown.

[0069] Before entering the algorithm training module, the average aggregation method is used for the five high-resolution feature variables to upscale the 1km spatial resolution of LST, NDVI, NDWI, NDBI and DEM to 10km spatial resolution, which is consistent with the 2m spatial resolution of ERA5 air temperature; the calculation formula is as follows:

[0070] X={LST,NDVI,NDBI,NDWI,DEM}

[0071]

[0072] Among them, X i is the value of the ith 1km spatial resolution pixel, X i,avg It is the average value of the 10km resolution grid; each 10km×10km pixel contains 10×10 1km×1km pixels. For each 10km×10km pixel, the average values ​​of LST, NDVI, NDWI, NDBI and DEM of the corresponding 100 pixels are calculated to achieve data upscaling.

[0073] Preferably, the data processing module is connected to the algorithm analysis module via a network.

[0074] The algorithm analysis module uses the XGBoost method to provide temperature estimation results with higher spatial resolution and higher accuracy through learning and comprehensive analysis of historical meteorological data and remote sensing data from 2002 to 2022; three methods, namely multivariate linear regression, random forest and artificial neural network, are introduced for comparative experiments. According to the training accuracy results, the XGBoost algorithm used in this study is superior to the other three algorithms. The year 2023 is selected for verification, and the model output temperature in 2023 is compared with the measured temperature of 198 meteorological stations in the three northeastern provinces; the algorithm analysis module is connected to the temperature estimation module through the network.

[0075] XGBoost is an efficient machine learning algorithm for gradient boosting decision trees, which is good at processing large-scale data and complex features. By gradually building trees and optimizing prediction errors, XGBoost can capture the nonlinear relationship between features and target variables, and control model complexity through regularization to effectively prevent overfitting. It is very suitable for downscaling temperature data with complex nonlinear features.

[0076] Specifically, the XGBoost model selects T with a resolution of 10 km in the three provinces of Northeast China from 2002 to 2022. 10km The driving factors such as LST, NDVI, NDBI, NDWI, and DEM with a resolution of 10 km that have been averaged and aggregated are used as training sets, and decision trees and nonlinear relationships are constructed through multiple rounds of iterations.

[0077] In each iteration, a new tree is constructed to further refine the temperature prediction at 1 km spatial resolution by fitting the residuals. The split gain is calculated as follows:

[0078]

[0079] In the formula, G split is the split gain, g i and h i are the first and second order derivatives of the loss function, L and R represent the left and right child node datasets, λ is the regularization parameter, and γ is the split penalty term. The XGBoost model selects the best feature split point by maximizing the split gain, thereby gradually enhancing the spatial resolution of air temperature estimation using features such as LST, DEM, NDVI, NDBI, and NDWI.

[0080] The prediction value of each round is updated by accumulating the output of the new tree, so that the model gradually approaches the expected spatial detail structure at a resolution of 1km. The calculation update formula is as follows:

[0081]

[0082] In the formula, η is the learning rate, which is used to control the contribution of the new tree to the total estimate, and f t It is the output of the t-th tree based on the fitting of the characteristic variables (LST, DEM, NDVI, NDBI, NDWI). Through continuous updating, the XGBoost model gradually refines the temperature estimate of each 1km grid to better reflect the spatial characteristics.

[0083] XGBoost gradually builds multiple trees through multiple rounds of iterations, and each tree fits the residual of the previous round to continuously optimize the prediction ability of the model. Then, NDVI, NDBI, NDWI, LST, and DEM with a spatial resolution of 1km in 2023 are used as explanatory variables and input into the trained XGBoost model. The final prediction result of the model is the sum of the outputs of all trees.

[0084] This method enables the XGBoost model to gradually improve the estimation accuracy of the temperature at a spatial resolution of 1 km, while maintaining the spatial refinement effect of feature information such as LST, DEM, NDVI, NDBI and NDWI. The downscaled temperature result Y of the XGBoost model with a spatial resolution of 1 km per hour in 2023 is estimated. 1km back.

[0085] Three methods, namely multivariate linear regression, random forest and artificial neural network, were introduced for comparative experiments.

[0086] As an algorithm under supervised machine learning, multiple linear regression constructs a multiple linear regression downscaling model by constructing the changing relationship between the only continuous dependent variable and multiple numerical independent variables to estimate the high-resolution 2m air temperature.

[0087] Specifically, the 10km resolution LST, NDVI, NDBI, NDWI, and DEM were selected as independent variables after average aggregation in remote sensing image units and compared with the 10km resolution T 10km Construct a multiple linear regression equation. The calculation formula is as follows:

[0088] The calculation formula is as follows:

[0089] T 10km =m 1 LST+m 2 NDVI+m 3 NDBI+m 4 NDWI+m 5 DEM+ε

[0090] In the formula, m 1 …m 5 is the regression coefficient, ε is the error, T 10km is the 2m air temperature of ERA5; the average aggregated NDVI, NDBI, NDWI, LST, DEM and 2m air temperature of ERA5 at 10km resolution in the three northeastern provinces from 2002 to 2022 were selected as sample data for regression model training, and then the 1km high-resolution NDVI, NDBI, NDWI, LST and DEM in 2023 were used as input data.

[0091] The calculation formula is as follows:

[0092] T 1km =m 1 LST+m 2 NDVI+m 3 NDBI+m 4 NDWI+m 5 DEM+ε

[0093] After inputting the trained model, the multivariate regression downscaling temperature result T at an hourly spatial resolution of 1 km in 2023 is estimated. 1km .

[0094] As a machine learning idea of ​​nonlinear ensemble learning, the core idea of ​​random forest is to combine and classify the prediction results of multiple decision trees by integrating them. It can show its superiority when the number of variables is much larger than the number of observations.

[0095] Specifically, the 10km resolution T 10km A nonlinear regression relationship is constructed with the averaged aggregated 10km resolution LST, NDVI, NDBI, NDWI, DEM and other driving factors to achieve multi-level spatial resolution downscaling. The calculation formula is as follows:

[0096] Tree i =BuildTree(S i ,F i )

[0097] In the formula, the independent variable is F i ={f 1 ,f 2 ,....f m}, F i Represents the feature set of the i-th decision tree, Si = {x 1 ,x 2 ,....x n}, where S i Represents the training set of the i-th tree, BuildTree is the decision tree construction function, Tree i is the i-th decision tree. k i is the number of decision trees;

[0098]

[0099] H(x) is the entropy value of each decision tree node, p i is the proportion of samples of the i-th category in the data set;

[0100]

[0101] Gini is the Gini coefficient of each decision tree node;

[0102]

[0103] Tree i (x) is the prediction result of the i-th tree for the sample, T 1km Estimated temperatures after downscaling for the random forest model.

[0104] The NDVI, NDBI, NDWI, LST, DEM and T at a resolution of 10 km in the three provinces of Northeast China from 2002 to 2022 were selected. 10km as the training set for each tree.

[0105] Random forest extracts sample data from the training set at random with replacement. This sampling method ensures that the training data set for each decision tree is randomly selected, and each decision tree is independently established. The diversity and stability of the model are increased by random sampling of training data and random selection of features.

[0106] Selecting the land surface temperature LST with the smallest Gini coefficient as the main feature for division can improve the purity of the sub-dataset, limit the number of features of each root node, find the optimal parameter setting, reduce the imbalance between data sets, and thus reduce the entropy value. The lower the entropy value, the more ordered the data features are and the higher the purity of the decision tree model is.

[0107] All the features in the remote sensing acquisition unit are selected as explanatory variables, and a random forest downscaling model is constructed through the nonlinear mapping relationship between the explanatory variable set and the air temperature.

[0108] Subsequently, NDVI, NDBI, NDWI, LST, and DEM with a spatial resolution of 1 km in 2023 were used as explanatory variables and input into the trained random forest model, and feature extraction and discrimination were performed based on the principle of “scale invariance”.

[0109] Thus, the random forest downscaling temperature result T with a spatial resolution of 1 km per hour in 2023 is estimated. 1km .

[0110] Artificial neural network is a learning model that calculates nonlinear data sets. It is suitable for solving complex causal relationships and non-deterministic reasoning problems. By simulating the structure and function of the human brain nervous system, the connections between neurons have different weights.

[0111] ANN performs information processing and pattern recognition by learning and adjusting the weights of the 2m air temperature of ERA5 with a spatial resolution of 10km and the LST, NDVI, NDBI, NDWI, and DEM with a resolution of 10km after average aggregation, and constructs the weight parameters of the network according to the training criteria.

[0112] The calculation formula is as follows:

[0113]

[0114] In the formula, H j is the output of the jth neuron, w ijis the weight of the ith input to the jth neuron, x i is the ith input, b j is the bias, f is the activation function;

[0115]

[0116] f′ is the derivative of the activation function, is the output layer error, is the hidden layer error, J is the loss function;

[0117]

[0118] is the weight connecting the i-th neuron and the j-th neuron in the m-th layer and the m+1-th layer, is the input of the i-th neuron in the m-th layer, is the error of the jth neuron in the m+1th layer;

[0119]

[0120] is the output of the i-th neuron in the m-th layer;

[0121]

[0122] is the bias of the jth neuron in the m+1th layer, γ is the learning rate, is the updated deviation;

[0123]

[0124] is the gradient of the weight, is the updated weight;

[0125]

[0126] is the updated bias, w is the weight of the initialized model;

[0127] The choice of activation function has an important impact on the performance and training effect of the ANN network. This study tried three extreme functions, Relu, Sigmoid, and Tanh, for downscaling prediction. The results show that the Relu function works best.

[0128] In terms of weight optimization algorithms, this study tried four weight optimization algorithms, namely, backpropagation (BP), stochastic gradient descent (SGD), adaptive learning rate (ALP) and regularized learning algorithm (Elastic Net, EN), and used them separately and in combination.

[0129] It was found that after combining BP and SGD, by observing the training error and loss function of each iteration step, this BP-SGD combination algorithm can quickly converge the model to the optimal solution.

[0130] By inputting the training data into the ANN network through forward propagation, the BP-SGD algorithm can provide better convergence performance while balancing computational efficiency and stability.

[0131] Based on the simulation results of sample data, the present invention determines the basic structure of the downscaling neural network and initializes the connection weights and biases between neurons in the network. The input variables include NDVI, NDBI, NDWI, LST, DEM and T at a resolution of 10 km from 2002 to 2022. 10km , features used to train and test the network.

[0132] The downscaling network model designed in the present invention has three layers, and the parameters of each layer include the input channel, that is, the collected feature data set (such as LST, NDVI, etc.), the output features after the bias function and neuron signal processing, and the feature information of the raster data and longitude and latitude.

[0133] The number of nodes in the first two layers of the hidden layer was finally set to 75 and 35, because after increasing the number of hidden nodes to (75, 35), the correlation coefficient score of the hidden layer was calculated to be 0.785, and after quantitatively increasing or decreasing the number of nodes, the correlation coefficient score no longer increased.

[0134] In addition, the regularization process is controlled by adjusting the alpha value, so that the network can quickly reduce the weights of relatively unimportant node paths during training to avoid overfitting. The calculation formula is as follows:

[0135]

[0136] w k is the weight from the hidden layer to the output layer, h k is the output of the hidden layer neural unit, b is the bias constant used to adjust the result of the linear combination, m is the number of neural units in the hidden layer, and k here ranges from 1 to m, indicating that all neural units in the hidden layer are traversed.

[0137] Using the 1km high-resolution NDVI, NDBI, NDWI, LST, and DEM in 2023 as input data, the trained ANN model was input to estimate the 2023 hourly 1km spatial resolution artificial neural network downscaling temperature result T 1km , the experimental results are shown in Table 2.

[0138] Table 2 Comparison of accuracy evaluation under four methods

[0139] <![CDATA[R 2 ]]> RMSE The present invention 0.953 0.632 Multiple Linear Regression 0.629 1.486 Random Forest 0.911 0.702 Artificial Neural Networks 0.863 0.924

[0140] According to the training results in Table 2, the accuracy comparison table of the four models shows that the XGBoost method used in this study is more accurate than multiple linear regression, artificial neural network and random forest. Then the temperature results T of the XGBoost model with a spatial resolution of 1 km per hour in 2023 were downscaled. 1km Transmitted to the estimation output module through the network.

[0141] The algorithm analysis module is connected to the estimation output module through a network.

[0142] Temperature estimation module predicts temperature T based on XGBoost 1km , create the daily minimum temperature function f min , perform dimensionality reduction operations and aggregate the time dimension.

[0143] The temperature data of each hour is reduced to a single grid data, the lowest temperature of each grid is extracted, and the data object of each hour is iterated day by day. The calculation formula is as follows:

[0144] T min (x,y)=f min (T 1km (x,y,k)

[0145] In the formula, x and y represent the row and column index of the grid, respectively; k represents the time index in hours, with a value range of (1-24); 1km (x,y,k) represents the temperature of the grid cell in the xth row and yth column at the kth hour, T min (x, y) is the daily minimum temperature of each grid. The daily minimum temperature of each grid is integrated into the daily minimum temperature T min After that, the data is transmitted to the user end through the network.

[0146] In order to evaluate the accuracy of the 2m daily minimum temperature results under the XGBoost method, the daily minimum temperature data of 198 meteorological stations in the three northeastern provinces in 2023 were selected to test the credibility of the downscaled daily minimum temperature estimates in 2023.

[0147] The evaluation criteria mainly include root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), where n represents the number of samples and T represents the number of samples. min is the downscaled high-resolution daily minimum air temperature data, T min_obs The air temperature data of the real meteorological station at the corresponding location at the longitude and latitude. The calculation formulas of each indicator are as follows:

[0148] (1) Root mean square error

[0149] RMSE is an indicator to measure the estimation accuracy of the 2m air temperature downscaling model, which can reflect the daily minimum temperature data T after downscaling. min and the actual daily minimum observation value T of the meteorological station min_obs The average difference between the two groups of data can reflect the degree to which the two groups of data deviate from each other.

[0150]

[0151] (2) Mean absolute error

[0152] MAE is used to measure the downscaled daily minimum temperature T min and the lowest measured value of the station day T min_obs The deviation between them refers to the average value of the absolute deviations of the absolute deviations of each measurement. The mean absolute error can avoid the problem of errors offsetting each other and accurately reflect the actual situation of the error in the downscaling results of the reanalysis of temperature data.

[0153]

[0154] (3) Mean absolute percentage error

[0155] MAPE is also a commonly used indicator to measure the estimation results. The mean absolute percentage error is very suitable for the accuracy of the estimation results of temperature downscaling time series. MAPE expresses the error in percentage form, which can intuitively reflect the daily minimum temperature T after downscaling. min and the lowest measured value of the station day T min_obs The relative error between them is particularly suitable for evaluating the performance of the temperature downscaling estimation model at different levels of scale and unit data sets in this study.

[0156]

[0157] On the other hand, the present invention provides a full coverage daily minimum temperature remote sensing estimation method, such as Figure 3 and Figure 6 As shown, the following steps are included:

[0158] Step 1: Collect the surface temperature data products, vegetation index data products and the bands required for remote sensing index of a certain area through satellites; obtain the ERA5 data set of the area through the Weather Forecast Center; and obtain the data DEM in the digital elevation model data of the area through the geospatial data cloud.

[0159] Step 2: Preprocess the land surface temperature data products and use the XGBoost algorithm to seamlessly estimate four sets of fully covered land surface temperatures LST; preprocess the data in the ERA5 dataset to obtain the air temperature T 10km ; Preprocess the bands of vegetation index data products and remote sensing index, and use the SG filtering algorithm and linear time interpolation method to obtain full coverage of NDVI, NDBI, and NDWI; obtain altitude information data by preprocessing DEM data; use the average aggregation method to calculate the data set of LST, NDVI, NDBI, NDWI and altitude information.

[0160] Step 3: Use the XGBoost algorithm based on the obtained air temperature T 10km The characteristic relationship between the average and aggregated data sets is used to calculate the hourly air temperature data with a spatial resolution of 1 km.

[0161] Step 4: Based on the hourly root air temperature data obtained in step 3, construct the daily minimum temperature function f min Perform iterative processing day by day to calculate the daily minimum temperature T min .

[0162] The present invention provides a method for estimating the daily minimum temperature based on machine learning combined with multi-source remote sensing data. The algorithm effectively analyzes multiple groups of interference variables such as climate conditions, geographical features, vegetation coverage, water body coverage, urban buildings, etc., and realizes the estimation of the daily minimum temperature. The estimation result is highly accurate. The temperature estimation module network connects the user end to arrange and output the daily minimum data under the XGBoost algorithm, which is convenient for monitoring personnel to obtain surface information more intuitively, and the reference value is higher.

[0163] The present invention effectively matches the lowest temperature T observed on the day of observation at the stations in the three northeastern provinces. min_obs The estimated daily minimum temperature T under the XGBoost model min , both have a good regression relationship, and the accuracy results are as follows Figure 4 The results of the daily minimum temperature after full coverage are as follows: Figure 5 , which makes it easier for monitoring personnel to obtain surface information more intuitively and has higher reference value.

[0164] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A full coverage daily minimum temperature remote sensing estimation system, characterized in that: It includes remote sensing acquisition module, data processing module, algorithm analysis module and temperature estimation module; The remote sensing acquisition module acquires relevant remote sensing information data by connecting to satellite sensors and public data sets; The data processing module pre-processes the relevant remote sensing information data to obtain a remote sensing index set, a data set of altitude information, a surface temperature and an air temperature; The algorithm analysis module uses XGBoost for training based on the data preprocessed by the data processing module to calculate the hourly air temperature data with a spatial resolution of 1 km; The temperature estimation module performs iterative processing on the air temperature data hour by hour every day according to the air temperature data, outputs the daily minimum temperature data every day, and calculates the daily minimum temperature T min , and transmit the output data to the user end through the network.

2. A full coverage daily minimum temperature remote sensing estimation system according to claim 1, characterized in that: The remote sensing acquisition module includes a remote sensing image unit, a climate acquisition unit and a digital elevation unit; The remote sensing image unit is connected to the sensor on the satellite through the network to collect the bands required for the land surface temperature LST and the bands required for the remote sensing index in a certain area; The climate acquisition unit is connected to the ERA5 data set of ECMWF in a certain area, and is transmitted to the data processing module after being numbered; the digital elevation unit is connected to the geospatial data cloud, obtains the DEM data in the digital elevation model data of the area and transmits it to the data processing module.

3. A full coverage daily minimum temperature remote sensing estimation system according to claim 2, characterized in that: The preprocessing operations in the data processing module are specifically as follows: Based on the surface temperature data products collected from satellites, after format conversion, splicing and cropping, the four groups of LST images of Terra daytime, Terra nighttime, Aqua daytime and Aqua nighttime are fully covered using the XGBoost algorithm to achieve seamless estimation of surface temperature LST data, and obtain four groups of fully covered surface temperature LST of Terra daytime, Terra nighttime, Aqua daytime and Aqua nighttime in the area; The ERA5 dataset of the region was clipped, projected, format converted and cleaned, and the abnormal temperature grid values ​​were excluded to obtain the hourly 2m air temperature T with a spatial resolution of 0.1°×0.1°. 10km ; According to the vegetation index data products collected from satellites, after format conversion, cropping and splicing, the SG filtering algorithm and linear time interpolation method are used to obtain the normalized vegetation index NDVI with a spatial resolution of 1km covering the entire area; According to the bands required by the remote sensing index collected from the satellite, calculations are performed to obtain the NDWI and NDBI of the area. The calculation formula is as follows: In the formula, NIR represents the reflectance value of the near infrared band, and SWIR represents the reflectance value of the short infrared band. After format conversion, splicing, cropping and resampling, the SG filtering algorithm and linear time interpolation method are used to obtain the normalized water index NDWI and the normalized building index NDBI with full coverage of 1km spatial resolution, which are combined with the normalized vegetation index NDVI to form a remote sensing index set. Merge the obtained DEM data, and use the vector boundary of the area to perform projection conversion, clipping and resampling to obtain a data set of altitude information; The data in the data set of land surface temperature (LST), remote sensing index set and altitude information are upscaled from 1 km spatial resolution to 10 km using the average aggregation method to match the air temperature in the ERA5 dataset.

4. A full coverage daily minimum temperature remote sensing estimation system according to claim 3, characterized in that: The specific implementation process of the algorithm analysis module is as follows: Use XGBoost to analyze the air temperature T 10km The training was carried out, and the feature variables involved in the training included the data set of LST, NDVI, NDBI, NDWI and altitude information after average aggregation; the LST, NDVI, NDBI, NDWI and altitude information with a spatial resolution of 1 km before average aggregation were input into the trained XGBoost model as explanatory variables to obtain the hourly temperature T with a spatial resolution of 1 km. 1km .

5. A full coverage daily minimum temperature remote sensing estimation system according to claim 4, characterized in that: The specific implementation process of the temperature estimation module is as follows: Based on XGBoost, the predicted temperature T 1km , create the daily minimum temperature function f min , perform dimensionality reduction operations and aggregate the time dimension; reduce the temperature data of each hour into a single grid data, extract the lowest temperature of each grid, and iterate the data objects of each hour day by day. The calculation formula is as follows: T min (x,y)=f min (T 1km (x,y,k)) In the formula, x and y represent the row and column index of the grid respectively, k represents the time index, the unit is hour, and the value range is 1-24, T 1km (x,y,k) represents the temperature of the grid cell in the xth row and yth column at the kth hour, T min (x, y) is the daily minimum temperature of each grid; the daily minimum temperature of each grid is integrated into the daily minimum temperature T min After that, the data is transmitted to the user end through the network.

6. A full coverage daily minimum temperature remote sensing estimation method, used to implement the daily minimum temperature remote sensing estimation system according to any one of claims 1 to 5, characterized in that: The following steps are involved: Step 1: Collect the surface temperature data products, vegetation index data products and the bands required for remote sensing index in a certain area through satellites; obtain the ERA5 data set of the area through the Weather Forecast Center; and obtain the data DEM in the digital elevation model data of the area through the geospatial data cloud; Step 2: Preprocess the land surface temperature data products and use the XGBoost algorithm to seamlessly estimate four sets of fully covered land surface temperatures LST; preprocess the data in the ERA5 dataset to obtain the air temperature T 10km ; Preprocess the bands of vegetation index data products and remote sensing index, and use SG filtering algorithm and linear time interpolation method to obtain full coverage of NDVI, NDBI, and NDWI; obtain altitude information data by preprocessing DEM data; use the average aggregation method to calculate the data set of LST, NDVI, NDBI, NDWI and altitude information; Step 3: Use the XGBoost algorithm based on the obtained air temperature T 10km The characteristic relationship between the averaged and aggregated data sets is used to calculate the hourly air temperature data with a spatial resolution of 1 km; Step 4: Based on the hourly root air temperature data obtained in step 3, construct the daily minimum temperature function f min Perform iterative processing day by day to calculate the daily minimum temperature T min .

Images