Temperature reconstruction method fusing data of micrometeorological device

Abstract

The invention relates to a multi-source meteorological data fusion technology, and discloses an air temperature reconstruction method fusing data of a micro-meteorological device, which improves the temporal-spatial resolution and precision of ground temperature under a complex terrain. The method comprises the following steps: densely deploying micrometeorological devices in a complex terrain area to obtain high-frequency observation data, carrying out quality control, and carrying out hierarchical processing in two dimensions of space and time by taking pattern forecast grid point data as an initial background field: in the spatial dimension, dynamically updating a fusion weight for grid points with observation stations by using geographical weighted regression, and carrying out data fusion; a residual machine learning model is combined with multi-topographic feature correction for grid points without observation stations, and a high-precision space fusion background field is generated; in the time dimension, errors after space fusion are decoupled into a trend term and a periodic term, an autoregressive integral moving average model is used for predicting a trend, a Fourier algorithm is used for correcting a periodic phase, and then time dimension machine learning correction is carried out on a grid point of an observation-free station. And finally, complete-process automatic, high-temporal-spatial-resolution and low-error complex terrain area air temperature reconstruction is realized.

Description

Technical Field

[0001] This invention relates to multi-source meteorological data fusion technology, specifically to a method for reconstructing temperature by fusing data from micro-meteorological devices. Background Technology

[0002] Land surface temperature, as a core parameter in the Earth's energy cycle system, is crucial foundational data for climate change research, urban thermal environment assessment, precision agricultural management, and ecosystem monitoring. Its data quality directly determines the accuracy and reliability of related research and applications. However, current land surface temperature acquisition technologies face significant challenges, failing to meet the precision requirements of complex scenarios. These challenges are mainly reflected in the following three aspects: (1) The limitations of point-scale monitoring are prominent: Traditional point-scale temperature observations relying on meteorological stations require dense station deployment to achieve wide-area coverage. However, in complex terrain areas such as mountains, valleys, and plateaus, the cost of setting up stations is extremely high and the maintenance is difficult due to factors such as terrain barriers, inconvenient transportation, and power supply difficulties. This results in sparse observation stations, many data coverage blind spots, and an inability to fully reflect the spatial heterogeneity of temperature caused by micro-topography in the region.

[0003] (2) Imbalance in the spatiotemporal resolution of remote sensing inversion products: While existing thermal infrared remote sensing technology can achieve large-scale temperature monitoring, there is an irreconcilable contradiction in spatiotemporal resolution among different remote sensing products: high spatial resolution products (such as the Landsat series) have high spatial accuracy, but their revisit period is as long as 16 days or more, making it impossible to capture short-term dynamic changes in temperature; high temporal resolution products (such as the MODIS series) can achieve multiple observations per day, but their spatial resolution is only at the kilometer level, making it difficult to characterize temperature differences in small-scale terrains (such as small watersheds and mountain depressions), and failing to meet the needs of urban microclimate regulation and precision irrigation of farmland for fine spatiotemporal scale data.

[0004] (3) Limitations of existing multi-source data fusion methods in temperature reconstruction in complex terrain areas: To address the limitations of single data sources, various spatiotemporal fusion models have been proposed (such as STARFM, ESTARFM, and FSDAF), but these methods still face key technical bottlenecks in temperature reconstruction in complex terrain areas. Relying on the assumption of homogeneity of the land surface, it cannot adapt to complex terrain: mainstream models such as STARFM and FSDAF all assume that the land cover type is uniform and the spatial distribution of temperature is continuous within the fusion area. However, in mountainous and valley areas with abrupt elevation changes and strong thermal heterogeneity, the local thermal differences caused by micro-topography will break this assumption, resulting in systematic bias in the fusion results and failing to accurately reflect the actual temperature distribution.

[0005] Ignoring the dynamic coupling of time-dimensional errors leads to significant long-term error accumulation: Existing models mostly focus on spatial dimension error correction, failing to consider the superposition of diurnal temperature variation cycles (such as diurnal temperature range) and topographic thermal effects. When performing continuous fusion of multiple time periods, periodic time errors (such as underestimation of morning temperature and overestimation of evening temperature) accumulate continuously, amplifying monthly and quarterly average temperature errors and affecting the accuracy of long-term climate change analysis.

[0006] The lack of physical constraints in areas without observation sites can easily lead to non-physical results: In areas without meteorological station coverage, fusion methods based on machine learning (such as random forests and neural networks) rely solely on static terrain features (such as elevation and slope) for interpolation prediction. Without incorporating the physical constraints of numerical weather prediction models, they are prone to results that violate physical laws (such as misjudging temperature inversion layers in high-altitude areas as higher than in low-altitude areas, and abnormal high temperatures in winter).

[0007] Low level of automation, making it difficult to adapt to business applications: The existing fusion process requires manual setting of key parameters, but the non-stationary thermal characteristics of complex terrain areas make it impossible for manual parameters to be globally adapted, requiring repeated debugging, which is not only inefficient, but also prone to affecting the fusion accuracy due to human operation errors.

[0008] In addition, although micro-meteorological devices that have emerged in recent years can achieve high-frequency temperature observation at single points in complex terrain areas through low-cost and dense deployment, breaking through the limitations of traditional station deployment, such devices are still essentially point-scale observations and cannot be directly converted into continuous spatial temperature fields. They need to be combined with areal data such as remote sensing and numerical models through effective fusion methods to achieve high-precision temperature reconstruction of point-area collaboration.

[0009] In summary, current surface temperature monitoring in complex terrain areas faces multiple challenges, including insufficient point-scale data coverage, spatiotemporal imbalance of remote sensing data, and poor adaptability of fusion methods. There is an urgent need for an efficient fusion method that can integrate high-frequency observation data from micro-meteorological devices, take into account both spatiotemporal error correction, and incorporate physical constraints, in order to overcome existing technical bottlenecks and achieve high-precision, high spatiotemporal resolution temperature reconstruction in complex terrain areas. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide a method for reconstructing air temperature by integrating micro-meteorological device data, so as to improve the spatiotemporal resolution and accuracy of ground temperature under complex terrain.

[0011] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows: A method for reconstructing temperature by fusing data from micro-meteorological devices includes the following steps: S1. Deploy micro-meteorological devices in complex terrain areas to acquire near-surface temperature observation data, and use the median absolute deviation (MAD) method to identify outliers and control the quality of the observation data; S2. Using the model forecast fusion grid data as the initial background field, count and record the number of observation stations around each grid point after quality control, and mark the grid points with the observation stations. S3. For grid points with a number of surrounding observation stations greater than 0, calculate the error between the temperature of the grid point and the observed temperature of each observation station, and dynamically update the weights through geographic weighted regression to achieve temperature fusion of grid points with observation stations. S4. Statistically analyze the errors of the grid points with observation stations, construct a residual machine learning model, and use this model to correct the grid points without observation stations to generate a spatially fused background field. S5. Based on the spatial fusion background field, the error between the observation station grid points and the observation data of the micro-meteorological device is statistically analyzed, and the error is decoupled into a trend term and a periodic term using the seasonal trend decomposition (STL) algorithm; S6. The error trend term is predicted using an autoregressive integral moving average (ARIMA) model, and the phase of the error periodic term is dynamically corrected using a Fourier algorithm. S7. Statistically analyze the errors of grid points with observation stations, construct a residual machine learning model, and use this model to correct grid points without observation stations in the time dimension; S8. Output ground temperature reconstruction data.

[0012] Furthermore, in step S1, the criteria for outlier identification are: ; in, This represents the absolute deviation of the median.

[0013] Furthermore, in step S2, when counting the number of observation stations around each grid point, the statistical range is the area centered on that grid point with a radius not exceeding 10km.

[0014] Furthermore, in step S2, the grid point marked with an observation station refers to a grid point in which the number of observation stations is greater than 0 within an area centered on the grid point and with a radius not exceeding 10km.

[0015] Furthermore, in step S3, the weighting function used in the geographically weighted regression is a Gaussian kernel function: ; in, grid points to be corrected To the observation point The Euclidean distance, where b is the bandwidth parameter.

[0016] Furthermore, the residual machine learning model used in steps S4 and S7 is the Extreme Gradient Boosting Tree (XGBoost) model.

[0017] Furthermore, in steps S4 and S7, during the correction of grid points without observation stations, the features input to the residual machine learning model include: grid point temperature, terrain elevation, slope, aspect, longitude, and latitude; the output label is the error between grid point and station temperature.

[0018] Furthermore, in step S5, the periodic term parameter of the seasonal trend decomposition algorithm is dynamically set according to the daily and seasonal temperature variation characteristics.

[0019] Furthermore, in step S6, when using the Fourier algorithm to dynamically correct the phase of the error periodic term, the fundamental frequency used is the daily period and its harmonic components.

[0020] Furthermore, in step S8, the spatial resolution of the output ground temperature reconstruction data does not exceed 1km×1km, and the temporal resolution does not exceed 1 hour.

[0021] The beneficial effects of this invention are: (1) Improve data reliability: By densely deploying a network of micro-meteorological devices in complex terrain areas, high-density near-surface temperature observation data is obtained. The median absolute deviation method is used to identify and control outliers in the observation data, effectively eliminating outlier interference and providing a high-quality data foundation for subsequent temperature reconstruction. This solves the problems of insufficient coverage and low data reliability of traditional meteorological stations in complex areas such as mountains and valleys.

[0022] (2) Optimize the accuracy of spatial dimension reconstruction: In the spatial dimension, by combining the dynamic weight allocation mechanism of geographic weighted regression with the residual machine learning model, for grid points with observation stations, the fusion weights are dynamically updated based on the temperature error between the grid points and the observation stations to accurately capture local micro-topographic temperature features; for grid points without observation stations, multi-source features such as grid point temperature, topographic elevation, slope, aspect, latitude and longitude are integrated for fusion correction to generate a high-precision spatial fusion background field with a spatial resolution of no more than 1km, correcting the systematic bias of the topographic system in the model forecast.

[0023] (3) Improve the continuity and accuracy of time dimension reconstruction: In the time dimension, based on the spatial fusion background field, the error between the grid points of the observation station and the micro-meteorological observation is decoupled into a trend term and a periodic term using a seasonal trend decomposition algorithm. Then, the error trend term is predicted by an autoregressive integral moving average model, and the phase of the periodic term is dynamically corrected by a Fourier algorithm, so as to achieve continuous temperature reconstruction with a time resolution of no more than 1 hour. At the same time, for the grid points without observation stations, the time dimension correction is completed by a residual machine learning model, which effectively suppresses the long-term accumulation of time dimension error and avoids the amplification of temperature error on a long time scale such as monthly average.

[0024] (4) Achieve full-process automation and business adaptation: The entire temperature reconstruction process does not require manual setting of key parameters. From data quality control and spatiotemporal fusion to final data output, all processes are automated. It can adapt to the non-stationary characteristics of complex terrain, solve the problem of low automation and difficulty in adapting to business applications of traditional fusion methods, and provide an efficient and scalable technical solution for temperature monitoring in complex terrain areas. Attached Figure Description

[0025] Figure 1 This is a flowchart of the temperature reconstruction method based on the fusion of micrometeorological device data in an embodiment of the present invention. Detailed Implementation

[0026] This invention aims to provide a method for reconstructing temperature by fusing data from micro-meteorological devices, thereby improving the spatiotemporal resolution and accuracy of ground temperature in complex terrain. The core idea is to achieve fully automated reconstruction of high-precision temperature fields in complex terrain areas through a densely deployed network of micro-meteorological devices and a spatiotemporal dual-dimensional fusion algorithm. Specifically, the method involves first densely deploying micro-meteorological devices in complex terrain areas to acquire high-frequency observation data. After quality control, model forecast grid data is used as the initial background field, and processed in two layers: spatial and temporal. Spatially, for grid points with observation stations, geographic weighted regression is used to dynamically update the fusion weights; for grid points without observation stations, a residual machine learning model combined with multi-terrain features is used for correction, generating a high-precision spatially fused background field. Temporally, the error after spatial fusion is decoupled into trend and periodic terms. An autoregressive integral moving average model is used to predict the trend, and a Fourier algorithm is used to correct the periodic phase. Finally, temporal machine learning correction is applied to grid points without observation stations, ultimately achieving fully automated, high spatiotemporal resolution, and low-error temperature reconstruction in complex terrain areas.

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] This embodiment provides a method for reconstructing air temperature by fusing data from micrometeorological devices. (See also...) Figure 1 It includes the following implementation process: I. Data Acquisition and Quality Control: 1. Deployment and data acquisition of micro-meteorological devices: In complex terrain areas such as mountains, valleys, and plateaus, micro-meteorological devices are deployed at high density to continuously acquire near-surface temperature observation data.

[0029] 2. Quality control based on median absolute deviation (MAD): Because the raw observation data may contain outliers due to equipment malfunctions or extreme weather interference, directly using it for data fusion would severely impact accuracy. Therefore, the median absolute deviation method is used for outlier identification and removal. Specifically, the median of continuous 24-hour observation data from a single micrometeorological device is calculated, and then the median absolute deviation (MAD) is calculated using a formula. Observations outside the range of "median - 3 × MAD to median + 3 × MAD" are identified as outliers and removed. This effectively eliminates outliers caused by equipment errors or sudden interference while retaining normal temperature fluctuation data, providing a reliable point-scale data foundation for subsequent fusion.

[0030] II. Initial Background Field Construction and Grid Classification: 1. Initial background field selection: Using model forecast grid data as the initial background field, such data (such as grid temperature output by numerical weather prediction models) has continuous spatial coverage characteristics, which can solve the limitation of micro-meteorological devices that can only observe points and cannot cover areas.

[0031] 2. Statistics and classification of observation stations around the grid points: First, it is determined whether each grid point is supported by micrometeorological observation data: A statistical range with a radius ≤ 10km is defined, centered on a single model grid point (this distance covers the effective observation area of ​​the micrometeorological device while avoiding interference from terrain differences on temperature correlation due to an excessively large range). The number of observation stations within this range after quality control is counted. Based on the statistical results, all grid points are divided into two categories: grid points with observation stations (number of stations within a 10km radius > 0) and grid points without observation stations (number of stations within a 10km radius = 0). The classification result will directly determine the technical path of subsequent spatial fusion, ensuring that resources are concentrated on the correction of grid points supported by observation data, while providing a clear target for interpolation of grid points without observation data.

[0032] III. Spatial Dimension Integration: The idea behind spatial fusion is to first correct grid points with observational support, and then use these as a basis to correct grid points without observational support. This layered processing improves spatial accuracy, as detailed below: 1. Temperature fusion with grid points at observation stations: Calculate the error between the grid point temperature and the temperature of each surrounding observation station: Using the model forecast temperature of a single grid point with observation stations as a benchmark, calculate the difference between the temperature of that grid point and the observed temperatures of each surrounding micrometeorological station to obtain a set of error data.

[0033] Then, the weights are dynamically updated based on Geographically Weighted Regression (GWR): a Gaussian kernel function is used as the weight function for GWR, i.e.: ;in, grid points to be corrected To the observation point The Euclidean distance is given by , and b is the bandwidth parameter. The closer the observation station is to the grid point, the greater its weight, and vice versa.

[0034] Finally, the grid point temperature is calculated by fusion: the temperature of the surrounding observation stations is weighted according to the dynamically updated weights, and then combined with the initial grid point error to obtain the preliminary fusion temperature of the grid points of the observation stations.

[0035] 2. Temperature correction without observation station grid points: Since grid points without observation stations lack direct observational data, traditional interpolation methods that rely solely on spatial distance would overlook the crucial impact of terrain on temperature. Therefore, a residual machine learning model is constructed for correction, such as the Extreme Gradient Boosting Tree (XGBoost) model, which leverages its strong ability to fit nonlinear relationships.

[0036] The specific process includes: First, constructing a model training dataset: using grid points with observation stations as samples, the input features include grid point temperature (temperature after preliminary fusion), terrain elevation, slope, aspect, longitude, and latitude, and the output label is the error between the preliminary fused grid point temperature and the micro-meteorological observation temperature. The XGBoost model is trained and the parameters are optimized in this way to ensure that the model can accurately fit the mapping relationship between multiple terrain features and temperature errors.

[0037] Based on the trained model, the above input features of the grid points without observation stations are substituted into the trained model to obtain the temperature error prediction value of this type of grid point. Then, the temperature of the grid points in the initial background field is subtracted from the error prediction value to complete the spatial correction of the grid points without observation stations.

[0038] Through the above steps, a spatially fused background field covering all grid points is generated, achieving high-precision spatial temperature distribution characterization with a spatial resolution of ≤1km.

[0039] IV. Integration of the Time Dimension: While spatially fused data resolves the issue of uneven spatial distribution, temporal errors still exist, which can amplify monthly and quarterly average errors over the long term. Therefore, the purpose of temporal fusion is to decouple these errors into trend and periodic components, correct them separately, and then integrate them to achieve continuous reconstruction with a time resolution of ≤1 hour. Specifically: 1. Error decomposition: Based on the spatially fused background field, the errors between the spatially fused temperature and the micrometeorological observation temperature at gridded observation stations are statistically analyzed, and this error sequence is used as the object of time-dimensional correction. The Seasonal Trend Decomposition (STL) algorithm is employed to decouple the error sequence into a trend term (reflecting the long-term trend of error over time) and a periodic term (reflecting the periodic fluctuations of error). When setting the periodic term parameter of the STL, it is necessary to dynamically adjust it in conjunction with the temperature change characteristics: the daily periodic parameter is set to 24 hours (matching daily temperature variations), and the seasonal periodic parameter is set to 30 days (matching monthly seasonal variations), ensuring that the decomposition results accurately reflect the temporal distribution pattern of the error, laying the foundation for subsequent targeted correction.

[0040] 2. Error trend term and periodic term correction: Trend term correction: The Autoregressive Integral Moving Average (ARIMA) model is used to predict the error trend term. The ARIMA model can effectively capture the autocorrelation of time series. By analyzing the changing patterns of the historical error trend term, the trend term value at future times is predicted. Then, the predicted trend term is subtracted from the spatially fused temperature to eliminate the long-term cumulative effect of the error.

[0041] Periodic term correction: The Fourier algorithm is used to dynamically correct the phase of the error periodic term. The Fourier algorithm can decompose a periodic signal into harmonic components of different frequencies. Taking into account the diurnal periodic characteristics of temperature error, the fluctuation pattern of the error periodic term is accurately fitted based on the 24-hour fundamental frequency and its harmonic components. Then, the phase of the periodic term is adjusted according to the fitting results. Finally, the corrected periodic term is subtracted from the temperature after trend term correction to obtain a preliminary time fusion product with gridded observation stations.

[0042] 3. Time correction without observation station grid points: Referring to the correction logic for grid points without observation stations in the spatial dimension, another set of XGBoost residual machine learning models is constructed: taking grid points with observation stations as samples, the input features include the initial time fusion temperature, terrain elevation, slope, aspect, longitude, latitude, and time factor, and the output label is the error between the initial time fusion temperature and the micro-meteorological observation temperature; after training the model, the input features of grid points without observation stations are substituted into the model to obtain the predicted time error value, thus completing the time dimension correction for this type of grid point.

[0043] V. Final Data Output: Integrate all gridded temperature data after spatial and temporal corrections to output the final ground temperature reconstruction data. This data must meet the following requirements: spatial resolution ≤ 1km × 1km (adapting to small-scale temperature differences in complex terrain), temporal resolution ≤ 1 hour (capturing short-term dynamic changes in temperature), and, as verified in practice, the reconstruction error in complex terrain areas is reduced by an average of more than 20%.

[0044] The entire process described above, from data acquisition and quality control to spatiotemporal fusion, requires no human intervention and achieves fully automated processing. It can adapt to the non-stationary characteristics of complex terrain and meet the operational application needs of fields such as climate change research and urban thermal environment assessment.

[0045] Although embodiments of the present invention have been described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and alterations shall not depart from the protection scope of the present invention.

Claims

1. A method for reconstructing air temperature by integrating data from a micrometeorological device, characterized in that, Includes the following steps: S1. Deploy micro-meteorological devices in complex terrain areas to acquire near-surface temperature observation data, and use the median absolute deviation method to identify outliers and control the quality of the observation data; S2. Using the model forecast fusion grid data as the initial background field, count and record the number of observation stations around each grid point after quality control, and mark the grid points with the observation stations. S3. For grid points with a number of surrounding observation stations greater than 0, calculate the error between the temperature of the grid point and the observed temperature of each observation station, and dynamically update the weights through geographic weighted regression to achieve temperature fusion of grid points with observation stations. S4. Statistically analyze the errors of the grid points with observation stations, construct a residual machine learning model, and use this model to correct the grid points without observation stations to generate a spatially fused background field. S5. Based on the spatial fusion background field, the error between the observation station grid points and the micro-meteorological device observation data is statistically analyzed, and the error is decoupled into a trend term and a periodic term using a seasonal trend decomposition algorithm. S6. The error trend term is predicted using an autoregressive integral moving average model, and the phase of the error period term is dynamically corrected using a Fourier algorithm. S7. Statistically analyze the errors of grid points with observation stations, construct a residual machine learning model, and use this model to correct grid points without observation stations in the time dimension; S8. Output ground temperature reconstruction data.

2. The method for reconstructing air temperature by fusing micrometeorological device data as described in claim 1, characterized in that, In step S1, the criteria for outlier identification are: ； in, This represents the absolute deviation of the median.

3. The method for reconstructing temperature by fusing micrometeorological device data as described in claim 1, characterized in that, In step S2, when counting the number of observation stations around each grid point, the statistical range is the area with a radius of no more than 10km centered on that grid point.

4. The method for reconstructing air temperature by fusing micrometeorological device data as described in claim 3, characterized in that, In step S2, the grid point marked with an observation station refers to a grid point in which the number of observation stations is greater than 0 within an area centered on the grid point and with a radius not exceeding 10km.

5. The method for reconstructing air temperature by fusing micrometeorological device data as described in claim 1, characterized in that, In step S3, the weighting function used in the geographic weighted regression is the Gaussian kernel function: ； in, grid points to be corrected To the observation point The Euclidean distance, where b is the bandwidth parameter.

6. The method for reconstructing air temperature by fusing micrometeorological device data as described in claim 1, characterized in that, The residual machine learning models used in steps S4 and S7 are both extreme gradient boosting tree models.

7. The method for reconstructing air temperature by fusing micrometeorological device data as described in claim 1, characterized in that, In steps S4 and S7, during the correction of grid points without observation stations, the features input to the residual machine learning model include: grid point temperature, terrain elevation, slope, aspect, longitude, and latitude; the output label is the error between grid point and station temperature.

8. The method for reconstructing air temperature by fusing micrometeorological device data as described in claim 1, characterized in that, In step S5, the periodic term parameter of the seasonal trend decomposition algorithm is dynamically set according to the daily and seasonal temperature variation characteristics.

9. The method for reconstructing air temperature by fusing micrometeorological device data as described in claim 1, characterized in that, In step S6, when the phase of the error periodic term is dynamically corrected using the Fourier algorithm, the fundamental frequency used is the daily period and its harmonic components.

10. The method for reconstructing air temperature by fusing micrometeorological device data as described in claim 1, characterized in that, In step S8, the spatial resolution of the output ground temperature reconstruction data does not exceed 1km×1km, and the temporal resolution does not exceed 1 hour.

Images