Mountain land low temperature refined monitoring prediction method and system based on satellite-ground fusion and deep learning

Abstract

The invention discloses a mountain land low temperature refined monitoring prediction method and system based on satellite-ground fusion and deep learning, and belongs to the technical field of ecological meteorological monitoring and agricultural disaster prevention and reduction. The invention aims to solve the problems of insufficient mountain low-temperature monitoring space coverage, neglect of local factors, rough prediction resolution and poor terrain adaptability in the prior art. The method comprises the steps of multi-source data acquisition, high-density meteorological station, multi-source satellite and meteorological mode data acquisition, spatial-temporal feature analysis, correlation analysis, key variable screening and weather type association, satellite-ground fusion air temperature reconstruction, generation of refined monitoring data based on a CNN-ANN deep learning framework, and short-term and temporary prediction modeling. And LSTM is combined with a terrain attention mechanism to construct a prediction model, and two-dimensional reliability verification is carried out. According to the method, the mountain low-temperature monitoring precision and the predicted terrain adaptability can be improved, accurate support is provided for low-temperature disaster prevention and reduction of mountain economic crops, and the method has remarkable practical value.

Description

Technical Field

[0001] This invention relates to the field of ecological meteorological monitoring, specifically to a method and system for refined monitoring and prediction of low temperatures in mountainous areas based on satellite-ground fusion and deep learning. Background Technology

[0002] Mountainous areas have complex terrain and distinct local microclimates. The occurrence of low-temperature disasters is highly coupled with topographic relief, vegetation cover, and instantaneous meteorological conditions. Traditional monitoring and forecasting technologies are no longer sufficient to meet the needs of refined disaster prevention and mitigation.

[0003] Existing methods for monitoring low temperatures in mountainous areas have significant limitations: First, while ground-based meteorological stations are the primary means of temperature monitoring, most are concentrated in flat suburban areas due to construction standards, making it difficult to cover the complex core areas of mountains. This results in spatial gaps in mountain temperature data. Some studies use statistical interpolation methods to supplement mountain temperature data, but these methods do not fully consider the differences in local microclimates caused by slope, aspect, and vegetation cover, leading to insufficient monitoring accuracy and an inability to reflect the temperature heterogeneity in different mountain regions. Second, while satellite remote sensing technology offers advantages such as high spatiotemporal coverage and macroscopic observation, and can obtain continuous surface temperature data through thermal radiation inversion, high-resolution polar-orbiting satellites suffer from poor temporal continuity, making it difficult to meet the need for daily continuous monitoring of low temperatures in mountains. Furthermore, satellite remote sensing data has a time lag, making it unable to directly support real-time early warning of low-temperature disasters.

[0004] In terms of low temperature prediction in mountainous areas, current methods mainly rely on meteorological model simulations. However, most existing models do not dynamically adjust the prediction weights based on the complexity of mountainous terrain, resulting in significantly higher prediction errors in complex terrain areas compared to flat areas, making it impossible to achieve differentiated and accurate predictions of low temperatures in mountainous areas.

[0005] Therefore, in order to address the problems of insufficient spatial coverage and neglect of local factors in existing technologies for monitoring low temperatures in mountainous areas, as well as the defects of coarse resolution and poor terrain adaptability in prediction, this invention proposes a refined monitoring and prediction method and system for low temperatures in mountainous areas based on satellite-ground fusion and deep learning. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for refined monitoring and prediction of low temperatures in mountainous areas based on satellite-ground fusion and deep learning.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A refined monitoring and prediction method for low temperatures in mountainous areas based on satellite-to-ground fusion and deep learning includes: The multi-source data acquisition step involves acquiring multi-source input data of the target mountainous area through a data acquisition device. The multi-source input data includes data from high-density meteorological monitoring stations, multi-source satellite monitoring data, and meteorological model simulation data. The spatiotemporal feature analysis step involves performing mathematical statistics and analysis on the multi-source input data, screening low-temperature disaster events by combining statistical data on disasters to mountain economic crops, analyzing their spatiotemporal distribution range and occurrence time, and quantitatively analyzing the correlation between multi-source data and low temperatures in mountainous areas to form feature analysis results. The satellite-ground fusion temperature reconstruction step involves selecting key feature variables based on the feature analysis results and constructing a satellite-ground fusion reconstruction model through a deep learning network framework, and outputting refined monitoring data of low temperature in mountainous areas. The short-term prediction modeling step involves constructing a short-term prediction model based on the refined monitoring data of low temperatures in the mountainous area, by learning the time series variation pattern of low temperatures in the mountainous area and dynamically allocating spatial prediction weights based on terrain complexity, and outputting the short-term prediction results of low temperatures in the mountainous area. The results output step evaluates the reliability of the refined monitoring data and short-term prediction results for low temperature in mountainous areas and outputs the final monitoring and prediction results.

[0008] Preferably, the spatiotemporal feature analysis step includes calculating the correlation coefficient between high-density meteorological monitoring station temperature data and satellite surface temperature data, topographic elevation data and vegetation index data through correlation analysis, selecting variables whose correlation coefficients meet a preset threshold as key feature variables, analyzing the changes in low temperature characteristics under different weather conditions based on the key feature variables and weather type data, and outputting the low temperature change driving results.

[0009] Preferably, the satellite-to-ground fusion temperature reconstruction step further includes multi-source data preprocessing, performing geometric correction and radiometric calibration on the multi-source satellite monitoring data to eliminate spatial position deviations and radiometric errors, performing outlier removal processing on the high-density meteorological monitoring station data, removing monitoring values ​​that exceed the normal fluctuation range based on statistical criteria, and performing spatiotemporal interpolation processing on the meteorological model simulation data to obtain sample data that maintains the same spatial resolution and time scale as the satellite monitoring data.

[0010] Preferably, the satellite-ground fusion temperature reconstruction step further includes inputting the topographic data and satellite surface temperature data from the key feature variables into a CNN in the form of gridded images, capturing the spatial heterogeneity features of different mountain areas through alternating operations of at least three convolutional layers and pooling layers, and outputting a high-dimensional spatial feature vector containing topographic undulations and spatial differences in surface temperature. The high-dimensional spatial feature vector is dimensionally fused with measured temperature data from high-density meteorological monitoring stations through feature splicing, and the correspondence between spatial features and actual mountain temperatures is clarified through this dimensional fusion. Based on spatial heterogeneity information analysis, the temperature distribution pattern under different mountain scenarios is obtained through this correspondence.

[0011] Preferably, the short-term prediction modeling step includes dividing the mountainous area into high-complexity and low-complexity regions based on the topographic data in the key feature variables. Using slope as the core dividing criterion, and combining slope aspect, topographic unit features, vegetation cover, and crop distribution scenarios for comprehensive judgment, areas with slopes exceeding a preset threshold and shady slopes are classified as high-complexity regions, as are areas with lush vegetation and concentrated crop planting areas. Areas with slopes below a preset threshold, sunny slopes, and sparse vegetation areas without sensitive crops are classified as low-complexity regions. Higher prediction weights are assigned to high-complexity regions, and relatively lower weights are assigned to low-complexity regions. The output time series features are then adjusted through a weight matrix.

[0012] Preferably, the training of the short-term prediction model includes constructing a training set and a validation set by using historical mountain low temperature fine monitoring data and multi-source input data of the corresponding time period, using the mean absolute error as the loss function, and iteratively optimizing the model's time step, the number of hidden layer neurons, and the weight allocation parameters of the attention mechanism through gradient descent until the model's loss function converges on the validation set.

[0013] Preferably, the reliability evaluation includes performing spatial and temporal dimension verification. When performing spatial dimension verification, the root mean square error of the monitoring data, prediction results, and measured data of stations not involved in modeling is calculated. When the root mean square error is less than a preset spatial accuracy threshold, the spatial verification is output as qualified. When performing temporal dimension verification, the prediction results of the model for typical historical low temperature events are compared with the actual disaster occurrence data, and the prediction accuracy is calculated. When the prediction accuracy is greater than a preset temporal accuracy threshold, the temporal verification is output as qualified.

[0014] Preferably, the method also includes a model optimization step, in which the model training samples are randomly divided into multiple sub-samples through cross-validation, and each sub-sample is used as a validation set to evaluate the accuracy of the model. Based on the evaluation results, the size of the convolutional kernel of the model's convolutional layer and the number of hidden layers of the neural network are adjusted to maintain the reconstruction accuracy of the model in different mountainous sub-regions.

[0015] A refined monitoring and prediction system for low temperatures in mountainous areas based on satellite-to-ground fusion and deep learning includes: The multi-source data acquisition module acquires multi-source input data of the target mountainous area through a data acquisition device. The multi-source input data includes data from high-density meteorological monitoring stations, multi-source satellite monitoring data, and meteorological model simulation data. The spatiotemporal feature analysis module performs mathematical statistics and analysis on the multi-source input data. By combining the statistical data on disasters to mountain economic crops, it screens low-temperature disaster events and analyzes their spatiotemporal distribution range and occurrence time. It also quantitatively analyzes the correlation between multi-source data and low temperatures in mountainous areas, forming feature analysis results. The satellite-ground fusion temperature reconstruction module filters key feature variables based on the feature analysis results and constructs a satellite-ground fusion reconstruction model through a deep learning network framework, and outputs refined monitoring data of low temperature in mountainous areas. The short-term prediction modeling module, based on the refined monitoring data of low temperature in the mountainous area, constructs a short-term prediction model by learning the time series variation pattern of low temperature in the mountainous area and dynamically allocating spatial prediction weights based on terrain complexity, and outputs the short-term prediction results of low temperature in the mountainous area. The results output module evaluates the reliability of the refined monitoring data and short-term prediction results for low temperature in mountainous areas and outputs the final monitoring and prediction results.

[0016] The beneficial effects of this invention are: (1) By using the satellite-ground fusion technology of multi-data fusion, the pain points of insufficient meteorological station coverage, discontinuous satellite remote sensing time and coarse meteorological model resolution in traditional single data monitoring are effectively solved. By combining convolutional neural networks to extract the spatial heterogeneity features of mountain terrain and surface temperature, and then constructing a nonlinear mapping relationship between features and temperature through artificial neural networks, refined monitoring data of mountain low temperature can be generated. Furthermore, the reliability of the data is guaranteed by spatiotemporal dual-dimensional verification, which significantly improves the refinement and reliability of mountain low temperature monitoring.

[0017] (2) To achieve terrain adaptability and timeliness of short-term low temperature forecasts in mountainous areas and meet the needs of agricultural meteorological services, this application significantly reduces the forecast error in complex areas by using an attention mechanism based on terrain complexity; at the same time, it learns the low temperature time series pattern by relying on long short-term memory networks to achieve short-term forecast response, which solves the problems of insufficient forecast resolution of meteorological models and insufficient consideration of local microclimates, and meets the business needs of tea and citrus as economic crops in mountainous areas for accurate and timely low temperature early warning. Attached Figure Description

[0018] Figure 1 This is a flowchart of the refined monitoring and prediction method for low temperature in mountainous areas based on satellite-ground fusion and deep learning, as proposed in this invention. Figure 2This is a deep learning network framework diagram of a mountain low temperature fine monitoring and prediction based on satellite-ground fusion and deep learning in an embodiment of the present invention; Figure 3 This is a block diagram of the mountain low temperature fine monitoring and prediction system based on satellite-ground fusion and deep learning of the present invention; Figure 4 This is a flowchart of the model construction method of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms front, back, left, right, top, and bottom used in the following description refer to directions in the accompanying drawings, and the terms bottom surface and top surface, inside and outside refer to directions toward or away from the geometric center of a specific component, respectively.

[0020] This invention proposes a refined monitoring and prediction method for low-temperature conditions in mountainous areas based on satellite-to-ground fusion and deep learning, such as... Figure 1 As shown, it includes: The multi-source data acquisition step involves acquiring multi-source input data of the target mountainous area through a data acquisition device. The multi-source input data includes data from high-density meteorological monitoring stations, multi-source satellite monitoring data, and meteorological model simulation data. The high-density meteorological monitoring station data comes from the existing ground meteorological monitoring network deployed in the target area, such as more than 5,000 regional meteorological stations. The data collected includes 2-meter temperature, relative humidity, sunshine duration, and precipitation as meteorological elements, with a temporal resolution of 1 hour / time and a spatial coverage density of no less than 1 station / 50km. 2 After collection, invalid signals are initially filtered out through a data cleaning terminal to ensure data integrity > 95%.

[0021] Multi-source satellite monitoring data is received through a satellite receiving platform, specifically including: Surface temperature data, including LST products from the MODIS series satellites and surface temperature data from the Fengyun 3D satellite, are used to invert the spatial distribution of surface temperature in mountainous areas. Vegetation and topography data are used to extract vegetation cover, slope, and aspect as topographic and vegetation parameters. The data format is standardized as GeoTIFF to facilitate subsequent spatial matching and modeling.

[0022] The meteorological model simulation data is obtained through the output of the regional climate model. The content is the simulated daily average temperature data of the target area, with a spatial resolution of 1km and a temporal resolution of 1 day. After collection, it needs to be consistent with the time scale of satellite data and meteorological station data to provide a regional-scale temperature background reference for satellite-ground fusion.

[0023] During the data acquisition process, the data acquisition device uses the TCP / IP protocol to achieve real-time transmission and storage of multi-source data. The storage medium uses a high-performance server to ensure that the data is not lost and can be traced back.

[0024] The spatiotemporal feature analysis step involves performing mathematical statistics and analysis on the multi-source input data, screening low-temperature disaster events by combining statistical data on disasters to mountain economic crops, analyzing their spatiotemporal distribution range and occurrence time, and quantitatively analyzing the correlation between multi-source data and low temperatures in mountainous areas to form feature analysis results. Screening for low-temperature disaster events: Based on the statistical yearbook of disasters affecting mountain economic crops, and using frost disasters with a low temperature threshold of ≤5℃ during the budding stage for tea and ≤3℃ during the flowering stage for citrus as the criteria, typical low temperature disaster events in spring over the past 3-5 years were selected. By overlaying satellite surface temperature data with meteorological station temperature data, the spatiotemporal distribution range and duration of each event are analyzed to form a spatiotemporal distribution map of low-temperature disasters.

[0025] The multi-source data correlation quantitative analysis adopted the Pierce correlation coefficient method to calculate the correlation coefficients between 2-meter air temperature at high-density meteorological stations and satellite LST, DEM slope, DEM aspect, and NDVI. Variables with |r|≥0.6 were selected as key feature variables, such as the correlation coefficient between satellite LST and station temperature r=0.72 and the correlation coefficient between slope and station temperature r=-0.65. By combining weather type data on cold waves, radiative cooling, and stationary fronts, the differences in low-temperature characteristics under different weather conditions are analyzed. For example, cold waves cause low temperatures over a wide range and with a large drop in temperature, while radiative cooling causes low temperatures to be concentrated in valleys and shady slopes. Finally, the results of low-temperature change drivers are output, providing a basis for variables in subsequent modeling.

[0026] The satellite-ground fusion temperature reconstruction step involves selecting key feature variables based on the feature analysis results and constructing a satellite-ground fusion reconstruction model through a deep learning network framework, and outputting refined monitoring data of low temperature in mountainous areas. Key feature variables were selected from multi-source data, including satellite LST, DEM slope, DEM aspect, NDVI, and simulated temperature from meteorological models, totaling five categories of variables; The variables were preprocessed. Satellite data underwent geometric correction based on ground control points, with a correction error of ≤1 pixel. Radiometric calibration was performed, converting DN values ​​into radiance using satellite calibration coefficients. Outlier removal was performed on meteorological station data. Spatiotemporal interpolation was performed on meteorological model data, using bilinear interpolation to downscale the 1km resolution to 250m to match the satellite data resolution. Training samples were created based on spatial grid matching and time series alignment principles. Each 250m grid was used as a unit, and the sample in each grid contained 5 types of feature variables and the measured temperature of the corresponding meteorological station within the grid. The time span of the samples was the spring of the past 3 years, and a total of about 100,000 valid samples were generated.

[0027] Model structure as follows Figure 2 The deep learning network framework consists of two parts: feature extraction and nonlinear mapping. The feature extraction layer uses a 3-layer convolutional neural network (CNN), with a kernel size of 3×3 and the activation function being ReLU. The pooling layer uses max pooling. Five key feature variables are input into the CNN in the form of gridded images. Through alternating convolution and pooling operations, the spatial heterogeneity features of different areas of the mountain are captured, such as the temperature gradient between steep and gentle slopes and the temperature difference between vegetated areas and bare land. The output is a high-dimensional spatial feature vector with a dimension of 256. The nonlinear mapping layer introduces an artificial neural network (ANN). The number of neurons in the input layer is consistent with the dimension of the feature vector in the high-dimensional space. Two hidden layers are set with 64 and 32 neurons respectively. The activation function is Sigmoid. The output layer is a single neuron with the activation function Linear. The output is the reconstructed value of the daily average temperature in a 250m grid. The model was trained using the root mean square error (RMSE) as the loss function, with the Adam optimizer employed, a learning rate of 0.001, and 1000 iterations. The samples were divided into training and validation sets in a 7:3 ratio, and training was stopped when the validation set RMSE ≤ 0.8℃.

[0028] The preprocessed real-time multi-source data is input into the trained model, which outputs refined monitoring data of spring low temperature at a resolution of 250m in mountainous areas. The data is in gridded dataset format and can be directly used for subsequent prediction modeling and visualization.

[0029] The short-term prediction modeling step involves constructing a short-term prediction model based on the refined monitoring data of low temperatures in the mountainous area, by learning the time series variation pattern of low temperatures in the mountainous area and dynamically allocating spatial prediction weights based on terrain complexity, and outputting the short-term prediction results of low temperatures in the mountainous area. The prediction sample was constructed using the time window sliding method, with 250m resolution fine monitoring data for 7 consecutive days, including daily average temperature, slope, and NDVI, as the input sample X, and the daily average temperature of the 8th day as the output sample Y. The input sample X is normalized to min-max and mapped to the [0,1] interval to eliminate the difference in units; the sample time span is the spring of the past 3 years, and a total of about 5000 training samples are generated.

[0030] Short-term prediction model construction, the model structure integrates LSTM and spatial attention mechanism, such as Figure 2The network framework has an LSTM layer with 3 input dimensions, including daily average temperature, slope, and NDVI, a time step of 7, and 128 hidden layer neurons, and is used to learn the time series patterns of low temperature changes. The spatial attention layer uses the slope of the DEM as the basis for weight allocation, dividing the mountainous area into high-complexity regions (slope > 25°, shady slopes, concentrated tea-growing areas) and low-complexity regions (slope ≤ 25°, sunny slopes, sparse vegetation areas). High-complexity regions are assigned an attention weight w1 = 0.6, while low-complexity regions are assigned a weight w2 = 0.4. The time-series features of the LSTM output are adjusted using a weight matrix to enhance the prediction accuracy of complex regions. The output layer activation function is Linear, and the output is a predicted daily average temperature for the next day at a resolution of 250m.

[0031] Using the mean absolute error (MAE) as the loss function, the gradient descent method was used to iteratively optimize the model parameters, including the LSTM time step, attention weights, number of hidden layer neurons, and the ratio of training set to validation set of 7:3, until the validation set MAE ≤ 0.9℃ converged. Based on monitoring data of historical low-temperature events over the past three years, the model is retrospectively predicted. If the prediction error in a certain area is greater than 1.2℃, the attention weight of that area is dynamically adjusted. For example, w1 in areas with a slope greater than 25° is increased to 0.7, and the model is retrained to ensure the model's adaptability in complex terrain areas.

[0032] The model outputs 250m resolution low temperature prediction data for the next day, including the daily average temperature, daily minimum temperature, and low temperature risk level: low risk: daily minimum temperature > 5℃, medium risk: 3℃ < daily minimum temperature ≤ 5℃, high risk: daily minimum temperature ≤ 3℃, providing accurate early warning basis for agricultural disaster prevention.

[0033] The results output step evaluates the reliability of the refined monitoring data and short-term prediction results for low temperature in mountainous areas and outputs the final monitoring and prediction results.

[0034] The reliability evaluation and application of results are completed through verification and result output devices. The specific process is as follows: The reliability evaluation is carried out in two dimensions. Spatial dimension verification selects measured temperature data from 10% of high-density meteorological stations that are not involved in the modeling, and calculates the root mean square error (RMSE) between the monitoring data, prediction results and measured data. Spatial verification is deemed qualified when RMSE ≤ 1.0℃. Temporal dimension verification selects historical data of typical low-temperature events in the past 3 years, compares the model prediction results with the actual disaster occurrence data, and calculates the prediction accuracy. Temporal verification is deemed qualified when the accuracy is ≥ 85%. The monitoring and prediction results are displayed in the form of a grid map through the result output device, clearly marking the high-risk areas of low temperature and the duration of low temperature, and supporting interactive operations such as zooming and querying. The final monitoring and prediction data will be archived to provide data support for subsequent research on the patterns of low-temperature disasters in mountainous areas and for iterative optimization of models.

[0035] Specifically, such as Figures 1 to 4 As shown, the spatiotemporal feature analysis steps include: calculating the correlation coefficient between high-density meteorological monitoring station temperature data and satellite surface temperature data, terrain elevation data and vegetation index data using correlation analysis methods; selecting variables whose correlation coefficients meet preset thresholds as key feature variables; analyzing the changes in low temperature characteristics under different weather conditions based on the key feature variables and weather type data; and outputting the low temperature change driving results.

[0036] To ensure the validity of the correlation calculation between high-density meteorological monitoring station temperature data and satellite surface temperature data, topographic elevation data, and vegetation index data, it is necessary to first complete the spatiotemporal consistency calibration of the data.

[0037] The data time scale is uniformly based on the 1-hour 2-meter temperature data from high-density meteorological monitoring stations. For satellite surface temperature data, such as MODISLST data (1 time per day) and Fengyun 3DLST data (6 times per hour), linear interpolation is used to interpolate to the 1-hour level. For vegetation index data, Landsat-8NDVI data (16 times per day) is selected from the spring, March-April period, and the period of high incidence of low temperature in mountainous areas, and the time phases that overlap with low temperature events are selected. The nearest neighbor method is used to match to the daily average time scale, because the analysis of low temperature disasters uses the daily average / daily minimum temperature as the core indicator. The data spatial grid alignment uses 250m×250m as the unified spatial grid unit. For topographic elevation data, the 30m resolution of SRTMDEM is upscaled to 250m using mean resampling, and the average elevation, slope, and aspect values ​​within each grid are extracted. For high-density meteorological monitoring station data, the inverse distance weighted interpolation method is used to interpolate the 1-hour temperature data of the stations to the 250m grid, ensuring that each grid has corresponding station temperature interpolation results. Satellite surface temperature data is directly adjusted to the 250m grid using bilinear interpolation, ultimately forming a multi-source data matrix with 1-hour temporal resolution and 250m spatial resolution, providing a unified computational basis for correlation analysis.

[0038] The Pearson correlation analysis method was used to calculate the correlation coefficients between each variable and the temperature data from high-density weather stations. This process requires setting the calculation dimensions according to the characteristics of different data types. Satellite surface temperature data selects the hourly average surface temperature within the grid, calculates its correlation coefficient with the corresponding hourly station temperature interpolation data of the grid, and reflects the degree of linear correlation between surface temperature and near-surface air temperature; The topographic elevation data selected were the average elevation and average slope values ​​within the grid. The correlation coefficients between these values ​​and the daily average station temperature data of the grid were calculated. Elevation was negatively correlated with temperature, with the temperature dropping by about 0.6℃ for every 100m increase in elevation. Slope was also negatively correlated with temperature, with cold air tending to accumulate in steep slope areas. The vegetation index data selected the daily average NDVI value within the grid and calculated its correlation coefficient with the daily average station temperature data of the grid. In areas with high vegetation coverage, such as tea gardens, NDVI > 0.6. Due to vegetation transpiration, the temperature is usually lower than that of bare land, so the two are negatively correlated. The calculation process and results are recorded using the Python function scipy.stats.pearsonr to calculate the correlation coefficient and output the corresponding P-value and significance level. Only correlation coefficients with P < 0.05 that are significantly correlated are retained, forming a statistical table of variable type, correlation coefficient, and significance, which provides a basis for subsequent threshold selection.

[0039] The determination of the preset correlation coefficient threshold needs to be verified in conjunction with historical low-temperature events. The threshold setting is based on the measured data of low-temperature events in the mountains in the spring of 2021-2023 over the past three years, such as the frost in March 2022 and the low temperature in April 2021. The modeling effect of variables corresponding to the three thresholds |r|=0.5, 0.6, and 0.7 was tested respectively. When |r|=0.5, some weakly correlated variables, such as the aspect data with r=-0.52, will cause the model RMSE>1.2℃; when |r|=0.7, the number of effective variables is insufficient, and only satellite LST and elevation data are retained, resulting in poor model generalization. Finally, the preset threshold was determined to be |r|≥0.6. At this time, the selected variables, satellite LST: r=0.72, elevation: r=-0.68, slope: r=-0.65, NDVI: r=-0.63, can ensure the correlation strength and cover the three core influencing factors of topography, vegetation, and satellite. After variable screening and processing, the key feature variables selected are standardized to eliminate differences in units, such as elevation (unit: m) and NDVI (unitless). This ensures that the weights of each variable are balanced during subsequent modeling. The standardized variable set will be used as the input data for CNN, directly affecting the spatial feature extraction effect.

[0040] This step requires relying on historical weather classification data, derived from weather logs of the target area, and combining it with the spatial heterogeneity analysis approach of drought monitoring diagrams from satellite remote sensing, to analyze the spatiotemporal differences in low temperature characteristics under different weather types: In response to the low temperatures in mountainous areas during spring, three typical weather types are mainly identified: cold wave weather, radiative cooling weather, and weather influenced by stationary fronts. Taking the cold wave event on March 15, 2022 as an example, this type of weather is characterized by strong cold air and a wide range of influence. Key characteristic variables include a daily temperature drop of ≥8℃ on satellite LST, a significantly enhanced negative correlation between elevation and temperature (r=-0.75), uniform spatial distribution of low temperature with no obvious local differences, short duration (1-2 days), and no significant topographic dependence. Taking the radiative cooling weather event in the mountainous area on April 8, 2021 as an example, under this weather type, the night sky is clear and there are few clouds, and the ground radiates heat away quickly. The key characteristic variables are that the slope and temperature are significantly negatively correlated (r=-0.71), the NDVI is high, the low temperature in the tea garden lasts for ≥4 hours, the low temperature is concentrated in the valley and the low-lying areas of the shady slope, the spatial heterogeneity is strong, and the duration is relatively long, 3-4 days. Taking the stationary front event in the mountainous area on March 20, 2023 as an example, the weather under this type of weather is characterized by the convergence of cold and warm air masses, cloudy and rainy weather with little sunshine. The key characteristic variable is that the negative correlation between NDVI and temperature has weakened, with r=-0.58. Due to the cloud cover weakening vegetation transpiration and the small fluctuation of satellite LST, the daily variation is ≤2℃. The low temperature distribution is patchy, controlled by the position of the front, and has the longest duration, lasting 5-6 days. The analysis results were recorded in tabular form, including weather type, correlation strength of key variables, and spatiotemporal characteristics of low temperature. The core influencing factors under different weather types were identified, such as satellite LST for cold wave weather and slope for radiative cooling weather, which provided a basis for subsequent model parameter adjustments.

[0041] Specifically, such as Figures 1 to 4 As shown, the satellite-to-ground fusion temperature reconstruction step also includes multi-source data preprocessing, which involves geometric correction and radiometric calibration of the multi-source satellite monitoring data to eliminate spatial position deviations and radiometric errors, outlier removal of the high-density meteorological monitoring station data, removal of monitoring values ​​that exceed the normal fluctuation range based on statistical criteria, and spatiotemporal interpolation of the meteorological model simulation data to obtain sample data with consistent spatial resolution and time scale with the satellite monitoring data.

[0042] This step uses the conventional and repeatable 3σ criterion in meteorological data processing as the core statistical criterion, supplemented by consistency checks on data from adjacent time periods, to ensure the scientific and practical nature of outlier removal. The specific principles and implementation steps are as follows: Based on the normal distribution characteristics of temperature data from high-density meteorological monitoring stations, in a large amount of continuous monitoring data, such as hourly temperature data in spring, 99.73% of the normal data will be distributed within the range of the data mean μ ± 3 times the standard deviation σ. Data that exceeds this range are considered outliers because the probability of occurrence is extremely low, only 0.27%, and are mostly caused by equipment failure, transmission interference, or extreme environmental interference, such as sensor obstruction or false triggering by instantaneous strong winds. First, the normality test should be performed on the high-density meteorological monitoring station data of the target mountainous area, such as the mountainous area in spring, March-April, when low temperature disasters are frequent. The Shapiro-Wilk test method is used. When the test result P value > 0.05, the data conforms to the normal distribution and the 3σ criterion can be directly applied. If the P value ≤ 0.05, the data has a slight skewness. Then, the data should be logarithmically transformed first, y = ln(x+1) to make it approximately normal, and then the 3σ criterion can be applied. For suspected outliers identified by the 3σ criterion, further comparison is made between two adjacent monitoring periods, such as temperature data within one hour before and after. If the temperatures in adjacent periods are both within the normal range and change steadily, with a temperature difference of ≤2℃ between adjacent periods, which is consistent with the natural rate of temperature change in mountainous areas during spring, and only the data in the middle period exceeds the range, then the data is confirmed as an outlier. If the data in adjacent periods also show drastic fluctuations, such as a temperature difference of >5℃ within one hour, then it is necessary to combine the data with the satellite remote sensing data from the same period, such as whether the surface temperature changes drastically in sync, to determine whether it is a real extreme weather event, and to avoid mistakenly deleting valid data.

[0043] The monitoring value within the normal fluctuation range is not a fixed value, but a dynamic range calculated based on the characteristics of historical spring temperature data in the target mountain area. It needs to be determined in combination with the monitoring dimensions of high-density meteorological monitoring stations and the characteristics of mountain climate.

[0044] Specifically, such as Figures 1 to 4 As shown, the satellite-ground fusion temperature reconstruction step further includes inputting the topographic data and satellite surface temperature data from the key feature variables into a CNN in the form of gridded images. The CNN captures the spatial heterogeneity features of different mountain areas through alternating operations of at least three convolutional layers and pooling layers, and outputs a high-dimensional spatial feature vector containing topographic relief and spatial differences in surface temperature. The high-dimensional spatial feature vector is then fused with measured temperature data from high-density meteorological monitoring stations through feature splicing. The correspondence between spatial features and actual mountain temperatures is clarified through this dimensional fusion. Based on spatial heterogeneity information analysis, the temperature distribution patterns under different mountain scenarios are obtained through this correspondence.

[0045] Specifically, such as Figures 1 to 4 As shown, the short-term prediction modeling steps include dividing the mountainous area into high-complexity and low-complexity regions based on the topographic data in the key feature variables. Using slope as the core dividing criterion, and combining slope aspect, topographic unit features, vegetation cover, and crop distribution scenarios for comprehensive judgment, areas with slopes exceeding a preset threshold and shady slopes are classified as high-complexity regions, as are areas with lush vegetation and concentrated crop planting areas. Areas with slopes below a preset threshold, sunny slopes, and sparse vegetation areas without sensitive crops are classified as low-complexity regions. Higher prediction weights are assigned to high-complexity regions, and relatively lower weights are assigned to low-complexity regions. The output time series features are then adjusted through a weight matrix.

[0046] Specifically, such as Figures 1 to 4 As shown, the training of the short-term prediction model includes constructing a training set and a validation set using recent mountain low temperature fine monitoring data and corresponding multi-source input data. Historical mountain low temperature fine monitoring data and corresponding multi-source input data refer to data from the past five years. The average absolute error is used as the loss function, and the model's time step, the number of hidden layer neurons, and the weight allocation parameters of the attention mechanism are iteratively optimized using the gradient descent method until the model's loss function converges on the validation set.

[0047] Specifically, such as Figures 1 to 4 As shown, the reliability evaluation includes spatial and temporal dimension verification. When performing spatial dimension verification, the root mean square error (RMSE) of the monitoring data, prediction results, and measured data from stations not involved in modeling is calculated. If the RMSE is less than a preset spatial accuracy threshold, the spatial verification is output as qualified. When performing temporal dimension verification, the prediction results of the model for typical historical low-temperature events are compared with the actual disaster occurrence data, and the prediction accuracy is calculated. If the prediction accuracy is greater than a preset temporal accuracy threshold, the temporal verification is output as qualified.

[0048] Specifically, such as Figures 1 to 4 As shown, it also includes a model optimization step, in which the model training samples are randomly divided into multiple sub-samples through cross-validation, and each sub-sample is used as a validation set to evaluate the accuracy of the model. Based on the evaluation results, the size of the convolutional kernel of the model's convolutional layer and the number of hidden layers of the neural network are adjusted to maintain the reconstruction accuracy of the model in different mountainous sub-regions.

[0049] A refined monitoring and prediction system for low temperatures in mountainous areas based on satellite-to-ground fusion and deep learning includes: The multi-source data acquisition module acquires multi-source input data of the target mountainous area through a data acquisition device. The multi-source input data includes data from high-density meteorological monitoring stations, multi-source satellite monitoring data, and meteorological model simulation data. The spatiotemporal feature analysis module performs mathematical statistics and analysis on the multi-source input data. By combining the statistical data on disasters to mountain economic crops, it screens low-temperature disaster events and analyzes their spatiotemporal distribution range and occurrence time. It also quantitatively analyzes the correlation between multi-source data and low temperatures in mountainous areas, forming feature analysis results. The satellite-ground fusion temperature reconstruction module filters key feature variables based on the feature analysis results and constructs a satellite-ground fusion reconstruction model through a deep learning network framework, and outputs refined monitoring data of low temperature in mountainous areas. The short-term prediction modeling module, based on the refined monitoring data of low temperature in the mountainous area, constructs a short-term prediction model by learning the time series variation pattern of low temperature in the mountainous area and dynamically allocating spatial prediction weights based on terrain complexity, and outputs the short-term prediction results of low temperature in the mountainous area. The results output module evaluates the reliability of the refined monitoring data and short-term prediction results for low temperature in mountainous areas and outputs the final monitoring and prediction results.

[0050] The foregoing has illustrated and described the basic features, principles, and advantages of the present invention. It should be noted that the present invention is not limited to the above embodiments, but only to some embodiments. Any improvements and additions made without departing from the spirit and scope of the present invention are considered to be within the scope of protection of the present invention.

Claims

1. A refined monitoring and prediction method for low temperature in mountainous areas based on satellite-ground fusion and deep learning, characterized in that, include: The multi-source data acquisition step involves acquiring multi-source input data of the target mountainous area through a data acquisition device. The multi-source input data includes data from high-density meteorological monitoring stations, multi-source satellite monitoring data, and meteorological model simulation data. The spatiotemporal feature analysis step involves performing mathematical statistics and analysis on the multi-source input data, screening low-temperature disaster events by combining statistical data on disasters to mountain economic crops, analyzing their spatiotemporal distribution range and occurrence time, and quantitatively analyzing the correlation between multi-source data and low temperatures in mountainous areas to form feature analysis results. The satellite-ground fusion temperature reconstruction step involves selecting key feature variables based on the feature analysis results and constructing a satellite-ground fusion reconstruction model through a deep learning network framework, and outputting refined monitoring data of low temperature in mountainous areas. The short-term prediction modeling step involves constructing a short-term prediction model based on the refined monitoring data of low temperatures in the mountainous area, by learning the time series variation pattern of low temperatures in the mountainous area and dynamically allocating spatial prediction weights based on terrain complexity, and outputting the short-term prediction results of low temperatures in the mountainous area. The results output step evaluates the reliability of the refined monitoring data and short-term prediction results for low temperature in mountainous areas and outputs the final monitoring and prediction results.

2. The method for refined monitoring and prediction of low temperature in mountainous areas based on satellite-ground fusion and deep learning according to claim 1, characterized in that, The spatiotemporal feature analysis steps include: calculating the correlation coefficient between high-density meteorological monitoring station temperature data and satellite surface temperature data, topographic elevation data and vegetation index data using correlation analysis methods; selecting variables whose correlation coefficients meet preset thresholds as key feature variables; analyzing the changes in low temperature characteristics under different weather conditions based on the key feature variables and weather type data; and outputting the low temperature change driving results.

3. The method for refined monitoring and prediction of low temperature in mountainous areas based on satellite-ground fusion and deep learning according to claim 1, characterized in that, The satellite-to-ground fusion temperature reconstruction step also includes multi-source data preprocessing, which involves geometric correction and radiometric calibration of the multi-source satellite monitoring data to eliminate spatial position deviations and radiometric errors, outlier removal of the high-density meteorological monitoring station data, removal of monitoring values ​​exceeding the normal fluctuation range based on statistical criteria, and spatiotemporal interpolation of the meteorological model simulation data to obtain sample data with consistent spatial resolution and time scale with the satellite monitoring data.

4. The method for refined monitoring and prediction of low temperature in mountainous areas based on satellite-ground fusion and deep learning as described in claim 1, characterized in that, The satellite-ground fusion temperature reconstruction step further includes inputting the topographic data and satellite surface temperature data from the key feature variables into a CNN in the form of gridded images. The CNN captures the spatial heterogeneity features of different mountain areas through alternating operations of at least three convolutional layers and pooling layers, and outputs a high-dimensional spatial feature vector containing topographic relief and spatial differences in surface temperature. The high-dimensional spatial feature vector is then dimensionally fused with measured temperature data from high-density meteorological monitoring stations through feature splicing. The dimensional fusion clarifies the correspondence between spatial features and actual mountain temperatures. Based on spatial heterogeneity information, the temperature distribution patterns under different mountain scenarios are obtained through this correspondence analysis.

5. The method for refined monitoring and prediction of low temperature in mountainous areas based on satellite-ground fusion and deep learning according to claim 1, characterized in that, The short-term prediction modeling steps include dividing the mountainous area into high-complexity and low-complexity regions based on topographic data in key feature variables. Using slope as the core dividing criterion, and combining slope aspect, topographic unit features, vegetation cover, and crop distribution scenarios for comprehensive judgment, areas with slopes exceeding a preset threshold and shady slopes are classified as high-complexity regions, as are areas with lush vegetation and concentrated crop planting areas. Areas with slopes below a preset threshold, sunny slopes, and sparse vegetation areas without sensitive crops are classified as low-complexity regions. High prediction weights are assigned to high-complexity regions, and low weights are assigned to low-complexity regions. The output time series features are then adjusted through a weight matrix.

6. The method for refined monitoring and prediction of low temperature in mountainous areas based on satellite-ground fusion and deep learning according to claim 1, characterized in that, The training of the short-term prediction model includes constructing a training set and a validation set using historical mountain low temperature fine monitoring data and multi-source input data for the corresponding time period. The model's time step, the number of hidden layer neurons, and the weight allocation parameters of the attention mechanism are iteratively optimized using the gradient descent method, with the mean absolute error as the loss function, until the model's loss function converges on the validation set.

7. The method for refined monitoring and prediction of low temperature in mountainous areas based on satellite-ground fusion and deep learning according to claim 1, characterized in that, The reliability evaluation includes spatial and temporal dimension verification. When performing spatial dimension verification, the root mean square error (RMSE) of the monitoring data, prediction results, and measured data from sites not involved in modeling is calculated. If the RMSE is less than a preset spatial accuracy threshold, the spatial verification is output as qualified. When performing temporal dimension verification, the prediction results of the model for typical historical low-temperature events are compared with actual disaster occurrence data, and the prediction accuracy is calculated. If the prediction accuracy is greater than a preset temporal accuracy threshold, the temporal verification is output as qualified.

8. The method for refined monitoring and prediction of low temperature in mountainous areas based on satellite-ground fusion and deep learning according to claim 1, characterized in that, It also includes a model optimization step, which randomly divides the model training samples into multiple sub-samples through cross-validation. Each sub-sample is used as a validation set to evaluate the accuracy of the model. Based on the evaluation results, the kernel size of the model's convolutional layers and the number of hidden layers in the neural network are adjusted to maintain the model's reconstruction accuracy in different mountainous sub-regions.

9. A refined monitoring and prediction system for low temperatures in mountainous areas based on satellite-ground fusion and deep learning, applicable to the refined monitoring and prediction method for low temperatures in mountainous areas based on satellite-ground fusion and deep learning as described in any one of claims 1 to 8, characterized in that, include: The multi-source data acquisition module acquires multi-source input data of the target mountainous area through a data acquisition device. The multi-source input data includes data from high-density meteorological monitoring stations, multi-source satellite monitoring data, and meteorological model simulation data. The spatiotemporal feature analysis module performs mathematical statistics and analysis on the multi-source input data. By combining the statistical data on disasters to mountain economic crops, it screens low-temperature disaster events and analyzes their spatiotemporal distribution range and occurrence time. It also quantitatively analyzes the correlation between multi-source data and low temperatures in mountainous areas, forming feature analysis results. The satellite-ground fusion temperature reconstruction module filters key feature variables based on the feature analysis results and constructs a satellite-ground fusion reconstruction model through a deep learning network framework, and outputs refined monitoring data of low temperature in mountainous areas. The short-term prediction modeling module, based on the refined monitoring data of low temperature in the mountainous area, constructs a short-term prediction model by learning the time series variation pattern of low temperature in the mountainous area and dynamically allocating spatial prediction weights based on terrain complexity, and outputs the short-term prediction results of low temperature in the mountainous area. The results output module evaluates the reliability of the refined monitoring data and short-term prediction results for low temperature in mountainous areas and outputs the final monitoring and prediction results.

Images