Deep learning-based extended-period temperature downscaling method and system under complex terrain

Abstract

The invention discloses an extension period temperature downscaling method and system based on deep learning in a complex terrain, and belongs to the technical field of temperature prediction. The method comprises the following steps: aiming at the problem of defects in an existing temperature downscaling task, acquiring a back-calculation data set and a reanalysis data set, preprocessing the back-calculation data set and the reanalysis data set to obtain a multivariable training data set, acquiring terrain data, and extracting multi-scale features of a terrain by using a convolutional neural network; constructing an extended-period temperature downscaling model based on deep learning, training the model, fusing topographic data, determining an objective function, and obtaining a trained model; and inputting the low-resolution air temperature forecast data into the trained model to obtain a high-resolution downscaling result. According to the method, complex topographic data and extended period forecast data are adopted, a deep learning method is utilized to extract nonlinear features of air temperature data, respective advantages of multi-scale features are mined and integrated, the forecast air temperature is downscaled, and a high-resolution downscaling result is output.

Description

Technical Field

[0001] This invention belongs to the field of temperature prediction technology, specifically relating to a deep learning-based extended-range temperature downscaling method and system for complex terrain. Background Technology

[0002] Against the backdrop of global warming, extreme events such as high temperatures, droughts, and torrential rains are becoming more frequent and severe in my country, seriously impacting sustainable economic and social development. The risks are exacerbated by the overlap of extreme events with the socio-economic system. Among numerous meteorological indicators, 2-meter temperature is closely related to human production and daily life. To better cope with extreme events, people have begun to focus on forecasting meteorological indicators. Weather forecasting is a practical and widely used example. By integrating relevant dynamic theories and established observation systems, the forecast data has been greatly improved. Although we can make relatively accurate forecasts of meteorological data on a daily basis, long-term forecasts on a weekly or even monthly basis remain challenging.

[0003] Subseason-to-seasonal (S2S) forecasts, with timescales ranging from 2 weeks to 2 months (15-60 days), bridge the gap between medium-range weather forecasts and seasonal forecasts, playing a crucial role in multi-timescale forecasting from weather to climate. However, the increased nonlinearity across multiple meteorological timescales leads to greater model complexity, making it challenging to extract features and build models from large-scale meteorological data using traditional methods.

[0004] Furthermore, due to the gap between the demand for high-resolution data in practical applications and the low-resolution forecasts obtained from ensemble dynamical models, downscaling methods have emerged. In recent years, downscaling methods have made significant progress, mainly falling into two categories: dynamic downscaling and statistical downscaling. Dynamic downscaling applies a high-resolution numerical model to a finite subdomain of the region of interest, with the output of the coarser-scale numerical model providing boundary conditions for finer-scale simulations. However, this method is limited by computational resource intensity, potential errors in regional models, and non-transferability between regions. Statistical downscaling, on the other hand, establishes statistical relationships between coarse-scale (low-resolution) and high-resolution climate variables, and then applies these relationships to coarse-scale data to obtain local high-resolution data at different times or locations. With the advent of the big data era, the scale of meteorological data is increasing daily, and traditional statistical downscaling methods have certain limitations for large-scale data. Therefore, it is particularly important to fully explore and utilize the hidden climate change patterns and trends within the data to improve the level of climate prediction.

[0005] It is well known that topography has a significant impact on the evolution of atmospheric circulation. Topography alters the trajectory and thermal state of atmospheric circulation through both dynamic and thermal effects. Therefore, at sub-seasonal to seasonal timescales, statistical downscaling is particularly important for integrating topographic data, observational data, and model outputs in complex terrain, handling topographically induced local variability, and finally outputting high-resolution actual forecasts to improve prediction resolution and reliability. Therefore, in the task of temperature downscaling, the following issues are considered to remain to be addressed:

[0006] How can we use deep learning methods to extract nonlinear features from temperature data, explore and integrate the advantages of multi-scale features, and improve forecast accuracy?

[0007] Topographic data has a significant impact on temperature, and the incorporation of complex topographic features into the model needs further consideration.

[0008] When downscaling low-resolution data, the choice of downscaling method also needs to be considered. Summary of the Invention

[0009] To address the shortcomings of existing temperature downscaling tasks, this invention provides a deep learning-based extended-range temperature downscaling method and system for complex terrain. The method uses deep learning to correct model forecasts, incorporates terrain data, and finally adds a downscaling component, thereby improving model accuracy while meeting the requirements for high-resolution output.

[0010] To achieve the above objectives, the present invention employs the following technical solutions:

[0011] A deep learning-based extended-range temperature downscaling method for complex terrain, comprising the following steps:

[0012] Step 1: Obtain the extended S2S back-calculation dataset and reanalysis dataset and preprocess them to obtain a multivariate training dataset. Use deep learning methods to fuse the multivariate training dataset and extract multivariate data features. Obtain the ETOPO1 dataset as terrain data and use a convolutional neural network to extract multi-scale features of the terrain.

[0013] The extended S2S back-calculation dataset obtained in step 1 is used as the forecast data, and the reanalysis dataset is used as the target data. First, correlation analysis is performed on the forecast data to obtain a four-channel multivariate dataset, where the four channels are 2m air temperature, 2m dew point temperature, 10mV wind speed, and net shortwave radiation at the surface. Second, since there are missing values ​​in the ocean region of the target data, a masking operation is needed to align the ocean region in the forecast data with the study area. Finally, the dates of the forecast data and the target data are aligned to obtain the climate factor training dataset. Specifically:

[0014] The preprocessing in step 1 includes data alignment, variable selection, data cleaning, and standardization; wherein:

[0015] Data alignment: There are three types of forecast data: daily average data, cumulative data, and high-frequency data. The cumulative data and high-frequency data are unified into daily average data to ensure the consistency of the dataset in the time dimension.

[0016] Cumulative data is processed using a differencing method, and high-frequency data is aligned to the daily average time scale by calculating the mean. The specific processing method is as follows:

[0017] The daily increment is obtained by calculating the difference between consecutive cumulative values, for a given time period. Recorded cumulative variables , its in arrive Daily average within the interval The calculation formula is as follows:

[0018]

[0019] High-frequency data is sampled at daily time intervals (6 hours), and aggregated into daily average data through time grouping and mean calculation. This allows for multiple time points within a single day. Observed variables Its daily average The calculation formula is as follows: ,in The number of samples taken within a day;

[0020] Meanwhile, since the ocean portion of the target dataset is missing, the ocean portion of the forecast data is masked to align with the study area and adapt it to the study area of ​​the target dataset.

[0021] Variable selection: The correlation between various meteorological variables and temperature, as well as the cross-correlation among characteristics, are assessed. The correlation coefficient r is calculated using the following formula:

[0022]

[0023] In the formula, and These are the observed values ​​of the two variables being calculated; and Its mean;

[0024] The correlation coefficients between each meteorological element and the 2-meter air temperature were calculated. We selected the 2-meter air temperature (t2m), the 2-meter dew point temperature (d2m), the 10-meter radial wind (v10), and the net shortwave radiation (ssrd) from the eight meteorological variables as core features for subsequent temperature prediction analysis.

[0025] Data Cleaning: Since the ERA5-land reanalysis dataset lacks data on ocean regions, it is necessary to process the ocean regions in the forecast data. Based on the ocean mask generated from the ERA5-Land data, the values ​​for ocean regions are set to NaN. The ocean mask is generated by marking consecutive NaN regions in the ERA5-Land data and applying an area ratio threshold (proportion of total grid points > 0.1). Subsequently, missing NaN values ​​in land regions are filled using linear interpolation, while retaining NaN values ​​in ocean regions. The linear interpolation formula is as follows:

[0026]

[0027] In the formula, These are interpolation parameters; and For neighboring grid points;

[0028] Standardization: To accelerate model convergence and improve model accuracy, training data needs to be standardized. The Z-Score standardization method is used to standardize the training dataset, which standardizes the data based on the mean and standard deviation of the training set. The formula is:

[0029]

[0030] In the formula, For data before standardization, For standardized data, , Data respectively The mean and variance are calculated on the training dataset.

[0031] Step 2: Construct a deep learning-based extended-range temperature descaling model, which includes a ResNet34-based encoder, a terrain feature extraction part, and a decoder; input the multivariate training dataset into the model for training, incorporate terrain data, determine the objective function, and iteratively update the parameters to update the deep learning-based extended-range temperature descaling model;

[0032] The terrain feature extraction part in step 2 consists of multiple repeated convolutional blocks. The convolutional blocks gradually reduce the resolution of the high-resolution terrain data through 3×3 convolution operations, while increasing the number of channels. Adaptive average pooling is used to adjust the output of each layer to match the spatial dimension of the encoder. Then, the number of channels is further matched through convolution so that the outputs of each layer of the encoder can be added and fused element-wise. This method not only adjusts the size of the feature map of the high-resolution terrain data for easy input, but also preserves the high-resolution information of the terrain data (such as terrain slope, valley and mountain details), and controls the proportion of terrain information in the feature extraction and post-processing of the network, thereby better capturing the influence of terrain on temperature distribution.

[0033] The ResNet34-based encoder consists of a preprocessing module and an encoder module. The preprocessing module adjusts the model to adapt to subsequent input scales through 7×7 convolutions and pooling operations. The encoder module is based on different feature extraction layers of the ResNet34 residual network. ResNet34 is a 34-layer residual network consisting of an initial convolutional layer, batch normalization, ReLU activation, and max pooling layer. Its residual structure is divided into four layers: layer1, layer2, layer3, and layer4. Layer1 contains 3 BasicBlocks, each consisting of two 3×3 convolutions, batch normalization, and ReLU. Layer2 contains 4 BasicBlocks. Layer3 contains 6 BasicBlocks. Layer4 contains 3 BasicBlocks.

[0034] The core design of these residual blocks is the introduction of residual connections, which aims to alleviate the degradation problem of deep networks, ensure effective gradient propagation, and capture the nonlinear relationships of meteorological inputs through multi-scale feature extraction. The BasicBlock consists of two 3×3 convolutional layers and a residual connection. The encoder extracts features (e1 to e4) from the meteorological input and adds terrain features layer by layer to achieve element-level fusion, thereby incorporating terrain constraints and improving the ability to capture temperature patterns in terrain-sensitive areas (such as mountains or basins).

[0035] The decoder section employs a DecoderBlock module, which restores resolution through upsampling operations using transposed convolution. Each DecoderBlock includes a 3×3 convolution, batch normalization and ReLU activation, a transposed convolution upsampling layer, and a spatial attention mechanism. The spatial attention mechanism is used to highlight important spatial regions, suppress irrelevant noise, and improve the model's sensitivity to spatial heterogeneity of temperature. Each DecoderBlock ultimately achieves the task of downscaling meteorological data by using a series of operations such as transposed convolution and corresponding activation and batch normalization.

[0036] The decoder starts from the center block, adjusts its size through bilinear interpolation, and connects encoder features layer by layer to achieve skip connections to preserve high-resolution details; the final upsampling module adjusts to the specified output size through a series of convolutions and transposed convolutions, combined with activation and adaptive average pooling, to ensure the generation of a high-resolution temperature prediction field, while incorporating Dropout to prevent overfitting.

[0037] The training dataset was input into the extended-range temperature correction model for complex terrain based on ResNet34_Unet for training, and the mean squared error was set as the objective function; its expression is:

[0038]

[0039] in, This represents the actual 2-meter temperature value from the target dataset. This represents the downscaling result of the model's output forecast temperature. The number of samples;

[0040] The Adam parameter optimizer is used to train for a fixed number of rounds, and the model with the minimum validation loss is saved as the optimal model.

[0041] Step 3: Input the low-resolution temperature forecast data into the already trained deep learning-based extended-range temperature downscaling model to obtain high-resolution downscaling results.

[0042] Step 3 is the application of the trained model. The core operation is to first align the low-resolution extended-range temperature forecast data to be processed with high-resolution topographic data of the same spatial range and specifications, and then input it into the trained model. The model will rely on the learned topographic and temperature features to simultaneously correct the systematic bias and local random error of the forecast data, as well as downscale the low-resolution data to the level of the topographic data. Finally, the model output is restored to the actual temperature unit, and simple fine-tuning is performed in combination with the ground observation data of the same period to obtain the high-resolution extended-range temperature downscaled result with rich spatial details and reliable forecast accuracy.

[0043] A deep learning-based extended-range temperature downscaling system for complex terrain, the system comprising the following modules:

[0044] The data preprocessing module is used to acquire the dataset and preprocess it to obtain a multivariate training dataset. Deep learning methods are then used to fuse and model the multivariate training dataset and extract multivariate data features.

[0045] The training module is used to train the deep learning-based extended-range temperature descaling model. By setting the objective function, the module iteratively tunes the parameters to update the deep learning-based extended-range temperature descaling model.

[0046] The downscaling module is used to correct biases in the data to be downscaled; for the low-resolution forecast data used in the training of the main part of the model, multi-layer transposed convolution is used to upsample the data so that the model outputs high-resolution forecast data.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] Multi-scale terrain fusion improves accuracy in complex areas. This invention preprocesses S2S data to obtain multivariate data, which is then input into a ResNet34-based extended-range temperature downscaling model. This data is fused with the extracted multi-scale terrain data to jointly extract different feature elements. The 2m temperature forecast data is then comprehensively corrected and downscaled, thereby reducing correction errors and improving model accuracy. This method is particularly suitable for meteorological element correction and downscaling in complex terrain.

[0049] Transposed convolution combined with skip connections overcomes the shortcomings of traditional interpolation. In the downscaling stage, transposed convolution and skip connections across layers are used to replace traditional interpolation methods. The former dynamically generates high-resolution details that conform to physical laws through learnable parameters, while the latter passes fine-grained features from the encoder to the decoder, supplementing the local information lost during scale-up. This maintains spatial continuity while improving resolution, avoiding the over-smoothing and physical information loss problems of traditional methods.

[0050] Precise spatiotemporal modeling meets practical operational needs. The model is iteratively optimized using MSE as the loss function to fully extract spatiotemporal features, ultimately outputting high-resolution forecast data that is perfectly adapted to the terrain grid of the study area. This method not only maintains forecast stability under different weather conditions but also provides accurate and reliable data support for low-temperature early warning in agriculture in complex terrain areas and heat island assessment in urban agglomerations. Attached Figure Description

[0051] Figure 1 This is a flowchart of the extended-period temperature downscaling method based on deep learning for complex terrain according to the present invention;

[0052] Figure 2 This is a schematic diagram of the extended-period temperature downscaling method of the present invention;

[0053] Figure 3 This is a schematic diagram of the Basicblock residual block of the present invention;

[0054] Figure 4 This is a graph showing the correlation coefficient results of selecting multiple forecast variables in this invention;

[0055] Figure 5 This is a schematic diagram comparing the results of the model of the present invention with those of the traditional linear model and the U-Net model;

[0056] Figure 6 This is an architecture diagram of the extended-period temperature descaledation system based on deep learning for complex terrain, as described in this invention. Detailed Implementation

[0057] To gain a deeper understanding of this invention, we will provide a comprehensive and detailed description. However, this invention has various implementations and is not limited to the specific examples listed herein. These examples are presented to enhance a full understanding of the disclosure of this invention.

[0058] A deep learning-based extended-range temperature downscaling method for complex terrain, comprising the following steps:

[0059] Step 1: Obtain the extended S2S back-calculation dataset and reanalysis dataset and preprocess them to obtain a multivariate training dataset. Use deep learning methods to fuse the multivariate training dataset and extract multivariate data features. Obtain the ETOPO1 dataset as terrain data and use a convolutional neural network to extract multi-scale features of the terrain.

[0060] The forecast area in this embodiment covers longitudes from 107°E to 123°E and latitudes from 31°N to 47°N. The forecast data resolution is 0.5°, and the total number of grid points is 33×33. The forecast data uses S2S scale forecast data from 2004 to 2023 with start dates in summer (June-August), provided by the European Centre for Medium-Range Weather Forecasts, including 27 start dates.

[0061] In this embodiment, the reanalysis data is the real temperature data, i.e., the real surface temperature data, which comes from the ERA5-land reanalysis dataset. The surface temperature data with the same latitude, longitude and time range as the forecast data are selected from the ERA5-land dataset as the real values ​​output by the subsequent gridded forecast correction model.

[0062] The terrain data used in this embodiment comes from the ETOPO1 dataset released by the National Oceanic and Atmospheric Administration (NOAA). The range of the terrain data dataset is consistent with that of the climate dataset.

[0063] Meteorological data selection and preprocessing: The ECMWF-released S2S back-analysis dataset was used as the model input, and the ERA5-Land reanalysis dataset (0.1° resolution) was used as the ground truth for training. For terrain data, the NOAA-released ETOPO1 high-resolution global terrain dataset was used. Compared to low-resolution data, ETOPO1 can more clearly depict micro-topographic details such as mountains and basins, providing high-precision basic data support for the model to capture the temperature decrease or warming effect caused by terrain.

[0064] Variable selection: Spearman correlation analysis was used, with the 2m air temperature of ERA5-Land as the baseline, to select a highly correlated four-channel variable combination from the S2S dataset: 2m air temperature, 2m dew point temperature, 10m V-direction wind speed, and net shortwave radiation at the surface. This combination eliminated redundant variables and covered the key physical processes of heat transfer and exchange.

[0065] Consistency Correction: To address the characteristic that ERA5-Land only covers land areas, a "mask alignment" strategy is adopted to mask out the ocean areas in the S2S data. Based on date labels, the input and target data are aligned step by step to construct a spatiotemporally consistent training dataset.

[0066] The extended S2S back-calculation dataset obtained in step 1 is used as the forecast data, and the reanalysis dataset is used as the target data. First, correlation analysis is performed on the forecast data to obtain a four-channel multivariate dataset, where the four channels are 2m air temperature, 2m dew point temperature, 10mV wind speed, and net shortwave radiation at the surface. Second, since there are missing values ​​in the ocean region of the target data, a masking operation is needed to align the ocean region in the forecast data with the study area. Finally, the dates of the forecast data and the target data are aligned to obtain the climate factor training dataset. Specifically:

[0067] The preprocessing in step 1 includes data alignment, variable selection, data cleaning, and standardization; wherein:

[0068] Data alignment: There are three types of forecast data: daily average data, cumulative data, and high-frequency data. The cumulative data and high-frequency data are unified into daily average data to ensure the consistency of the dataset in the time dimension.

[0069] Cumulative data is processed using a differencing method, and high-frequency data is aligned to the daily average time scale by calculating the mean. The specific processing method is as follows:

[0070] The daily increment is obtained by calculating the difference between consecutive cumulative values, for a given time period. Recorded cumulative variables , its in arrive Daily average within the interval The calculation formula is as follows:

[0071]

[0072] High-frequency data is sampled at daily time intervals (6 hours), and aggregated into daily average data through time grouping and mean calculation. This allows for multiple time points within a single day. Observed variables Its daily average The calculation formula is as follows: ,in The number of samples taken within a day;

[0073] Meanwhile, since the ocean portion of the target dataset is missing, the ocean portion of the forecast data is masked to align with the study area and adapt it to the study area of ​​the target dataset.

[0074] Variable selection: The correlation between various meteorological variables and temperature, as well as the cross-correlation among characteristics, are assessed. The correlation coefficient r is calculated using the following formula:

[0075]

[0076] In the formula, and These are the observed values ​​of the two variables being calculated; and Its mean;

[0077] The correlation coefficients between each meteorological element and the 2-meter air temperature were calculated. We selected the 2-meter air temperature (t2m), the 2-meter dew point temperature (d2m), the 10-meter radial wind (v10), and the net shortwave radiation (ssrd) from the eight meteorological variables as core features for subsequent temperature prediction analysis.

[0078] Data Cleaning: Since the ERA5-Land observation data lacks data on ocean regions, it is necessary to process the ocean regions in the forecast data. Based on the ocean mask generated from the ERA5-Land data, the values ​​for ocean regions are set to NaN. The ocean mask is generated by marking consecutive NaN regions in the ERA5-Land data and applying an area ratio threshold (proportion of total grid points > 0.1). Subsequently, missing NaN values ​​in land regions are filled using linear interpolation, while retaining NaN values ​​in ocean regions. The linear interpolation formula is as follows:

[0079]

[0080] In the formula, These are interpolation parameters; and For neighboring grid points;

[0081] Standardization: To accelerate model convergence and improve model accuracy, training data needs to be standardized. The Z-Score standardization method is used to standardize the training dataset, which standardizes the data based on the mean and standard deviation of the training set. The formula is:

[0082]

[0083] In the formula, For data before standardization, For standardized data, , Data respectively The mean and variance are calculated on the training dataset.

[0084] Step 2: Construct a deep learning-based extended-range temperature descaling model, which includes a ResNet34-based encoder, a terrain feature extraction part, and a decoder; input the multivariate training dataset into the model for training, incorporate terrain data, determine the objective function, and iteratively update the parameters to update the deep learning-based extended-range temperature descaling model;

[0085] The terrain feature extraction part in step 2 consists of multiple repeated convolutional blocks. The convolutional blocks gradually reduce the resolution of the high-resolution terrain data through 3×3 convolution operations, while increasing the number of channels. Adaptive average pooling is used to adjust the output of each layer to match the spatial dimension of the encoder. Then, the number of channels is further matched through convolution so as to perform element-wise addition and fusion with the output of each layer of the encoder.

[0086] The terrain feature extraction process involves the following steps: First, high-resolution terrain data is input. Then, a series of convolutional operations are used to progressively reduce the resolution while increasing the number of channels. Next, adaptive average pooling is applied to adjust the output of each layer to match the spatial dimension of the corresponding layer in the encoder, and additional convolutions are used to match the number of channels. Finally, the processed terrain features are element-wise added and fused with the outputs of each layer in the encoder. This process preserves high-resolution details of the terrain (such as slope, valleys, and mountains) and controls their proportion in feature extraction, helping the model capture the influence of terrain on temperature distribution, especially in complex terrain areas.

[0087] The ResNet34-based encoder consists of a preprocessing module and an encoder module. The preprocessing module adjusts the model to adapt to subsequent input scales through 7×7 convolutions and pooling operations. The encoder module is based on different feature extraction layers of the ResNet34 residual network. ResNet34 is a 34-layer residual network consisting of an initial convolutional layer, batch normalization, ReLU activation, and max pooling layers. Its residual structure is divided into four layer groups: layer1, layer2, layer3, and layer4. Layer1 contains three BasicBlocks, each consisting of two 3×3 convolutions, batch normalization, and ReLU; layer2 contains four BasicBlocks; layer3 contains six BasicBlocks; and layer4 contains three BasicBlocks. After each layer group, the corresponding output of the terrain feature extraction part is added for element-wise fusion. Its function is to extract the nonlinear relationship of meteorological input, incorporate terrain constraints, and improve the ability to capture temperature patterns in terrain-sensitive areas. Residual connections ensure gradient propagation in deep networks and avoid degradation.

[0088] The decoder part uses the DecoderBlock module, which restores the resolution by performing upsampling operation through transposed convolution. Each DecoderBlock includes a 3×3 convolution, batch normalization and ReLU activation, transposed convolution upsampling layer and spatial attention mechanism. The spatial attention mechanism is used to highlight important spatial regions, suppress irrelevant noise, and improve the model's sensitivity to spatial heterogeneity of temperature.

[0089] The operation steps are as follows: Starting with the lowest resolution feature (center block) of the encoder, the size is adjusted using bilinear interpolation. In each DecoderBlock, transposed convolutional upsampling is applied, combined with batch normalization and ReLU activation to restore resolution. An integrated spatial attention mechanism is used: the channel average and maximum values ​​of the feature map are calculated, concatenated to generate an attention map, which is multiplied with the original features to highlight spatially important regions. Corresponding features from the encoder (e4 to e1) are fused layer by layer through skip connections to preserve high-resolution details. In the final upsampling module, a series of convolutions and transposed convolutions are applied, combined with activation, adaptive average pooling, and Dropout, to adjust to the target output size and generate a high-resolution temperature prediction field. Its function is to progressively restore spatial resolution, enhance sensitivity to spatial heterogeneity of temperature, suppress noise, prevent overfitting, and ensure a fine temperature distribution in the output.

[0090] The final upsampling module, as the core link connecting low-resolution feature extraction and high-resolution temperature prediction field generation, is designed to perfectly meet the dual requirements of detail restoration and scale matching in downscaling tasks. This module is not simply an interpolation or scaling operation, but rather systematically addresses the key issues of detail blurring and poor spatial consistency that easily occur in low-resolution data during scaling through transposed convolutional structures, ultimately achieving a precise mapping from the low-resolution feature space to the high-resolution temperature prediction field.

[0091] The transposed convolution, as a core scaling technique for downscaling, differs from the simple pixel weighting of traditional interpolation. It achieves detail generation driven by physical laws through learnable convolution kernel parameters. In temperature downscaling scenarios, it can automatically learn and generate microscopic details in high-resolution grids based on the local correlations of low-resolution data, achieving a natural transition from macroscopic trends to microscopic features. Simultaneously, the ReLU activation function applied after the transposed convolution enhances the expression of nonlinear features in temperature data (such as diurnal temperature range and seasonal temperature abrupt changes) through nonlinear transformation, avoiding overly smoothing of the downscaling results and ensuring the realism of the high-resolution prediction field.

[0092] Step 3: Input the low-resolution temperature forecast data into the already trained deep learning-based extended-range temperature downscaling model to obtain high-resolution downscaling results.

[0093] Step 3 is the application of the trained model. The core operation is to first align the low-resolution extended-range temperature forecast data to be processed with high-resolution topographic data of the same spatial range and specifications, and then input it into the trained model. The model will rely on the learned topographic and temperature features to simultaneously correct the systematic bias and local random error of the forecast data, as well as downscale the low-resolution data to the level of the topographic data. Finally, the model output is restored to the actual temperature unit, and simple fine-tuning is performed in combination with the ground observation data of the same period to obtain the high-resolution extended-range temperature downscaled result with rich spatial details and reliable forecast accuracy.

[0094] In this embodiment, to comprehensively evaluate the performance of the downscaling method in predicting 2-meter temperatures in a certain region, four evaluation metrics were selected to measure the model's prediction accuracy and reliability from different perspectives. These four evaluation metrics are root mean square error (RMSE), Pearson correlation coefficient (PCC), coefficient of determination (R²), and accuracy.

[0095] In this embodiment, the forecast data used is the summer retrospective data of a certain region from 2004 to 2023. The first 17 years are divided into a training set and the last 3 years are divided into a test set. To prevent the dates of the training set and the test set from overlapping, the last 3 years of the training set are divided into a validation set for adjusting the hyperparameters of the model. The test set is used to evaluate the forecast performance of the corrected model.

[0096] The initial learning rate was set to 0.0005, and a plateau learning rate decay strategy was adopted. When the validation loss no longer improved, the learning rate decayed according to the set factor.

[0097] The loss function used is based on the mean squared error (MSE) of the studied land area, and its expression is as follows:

[0098]

[0099] in, This represents the actual 2-meter temperature value from the target dataset. This represents the downscaling result of the model's output forecast temperature. The number of samples is denoted as 'sample size'. In this embodiment, the deep learning model is implemented using PyTorch, and the training device is a server equipped with an Nvidia 4090 processor.

[0100] Training is performed for a fixed number of rounds using the Adam parameter optimizer. To avoid overfitting, an early stopping mechanism is implemented: training stops when the validation loss does not improve for 15 consecutive rounds, and the model with the minimum validation loss is saved as the optimal model.

[0101] For example, for a deep learning-based extended-range temperature drop-scale model, by comparing the original forecast, linear regression (LR), and the traditional U-Net model, the following results are obtained: Figure 5 The diagram shows the evaluation indicators for the downscaling results of each model.

[0102] Figure 5 Each section represents a schematic diagram of a different indicator, with the evaluation indicators being (a) RMSE, (b) PCC, and (c) (d) Accuracy: The results show that lower RMSE is better, as is lower PCC. The higher the accuracy, the better.

[0103] From the perspective of systematic analysis of experimental results, this invention compares the traditional linear regression correction model and the nonlinear deep learning model based on U-Net to address the error problem existing in traditional raw forecast data. The comparative experimental results show that both linear and nonlinear modeling approaches can produce positive correction and downscaling effects on traditional raw forecast data, but in terms of the degree of improvement, the correction effect of this invention is better.

[0104] The deep learning-based extended-period temperature downscaling method for complex terrain proposed in this invention demonstrates superior overall performance in comparative experiments with several mainstream downscaling methods (including traditional regression models and other deep learning models). In core evaluation metrics (such as root mean square error RMSE, coefficient of determination R², etc.), the correction error of this model is significantly lower than that of other comparative models, and the spatial matching and numerical consistency between the downscaled data and the measured data are also better.

[0105] This result not only verifies the applicability and superiority of the proposed deep learning-based temperature correction and downscaling method in complex terrain scenarios, but also demonstrates the research value of the downscaling model from a practical perspective: compared with traditional methods, the deep learning model can more fully integrate multi-source information and more accurately depict the spatial distribution and variation characteristics of temperature under complex underlying surface conditions, providing a more effective technical path for high-resolution and high-precision temperature forecasting, and possessing important theoretical research significance and practical application prospects.

[0106] Further observation Figure 5 The diagram shows the evaluation indexes for the downscaling results of each model. This diagram not only presents the overall performance differences between different models, but also includes the control group data of the model fitting effect after removing terrain information, providing a direct basis for analyzing the impact of terrain factors on model performance.

[0107] By comparing the results of the two sets of models with and without terrain information, it is clear that the model with terrain information exhibits significantly better fit across all evaluation metrics. Specifically, it demonstrates a lower corrected error rate (RMSE) and a higher data correlation index (R²). This phenomenon fully illustrates that terrain information, as a key factor influencing the spatial distribution of temperature, can provide the model with richer feature inputs, helping it to more accurately learn the temperature variation patterns under different terrain conditions. This effectively improves the model's correction accuracy and downscaling effect, and further verifies the rationality and necessity of incorporating terrain information into the model design in this invention.

[0108] like Figure 6 The diagram shown is an architecture diagram of the extended-period temperature downscaling system based on deep learning for complex terrain provided in this embodiment.

[0109] This embodiment establishes a deep learning-based extended-range temperature downscaling system for complex terrain. Based on the actual production requirements for high-resolution forecast data, it can effectively utilize terrain elevation information to correct the extended-range forecast data released by numerical weather prediction models, further reducing the error of the original forecast data. At the same time, compared with traditional linear and neural network models, it improves the model forecast accuracy and provides more stable forecast information, which has certain practical significance.

[0110] A deep learning-based extended-range temperature downscaling system for complex terrain, the system comprising the following modules:

[0111] The data preprocessing module is used to acquire the dataset and preprocess it to obtain a multivariate training dataset. Deep learning methods are then used to fuse and model the multivariate training dataset and extract multivariate data features.

[0112] The training module is used to train the deep learning-based extended-range temperature descaling model. By setting the objective function, the module iteratively tunes the parameters to update the deep learning-based extended-range temperature descaling model.

[0113] The downscaling module is used to correct biases in the data to be downscaled; for the low-resolution forecast data used in the training of the main part of the model, multi-layer transposed convolution is used to upsample the data so that the model outputs high-resolution forecast data.

[0114] Contents not described in detail in this specification are prior art known to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications are readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of this invention are protected.

Claims

1. A deep learning-based extended-period temperature downscaling method for complex terrain, characterized in that, The method includes the following steps: Step 1: Obtain the extended S2S back-calculation dataset and reanalysis dataset and preprocess them to obtain a multivariate training dataset. Use deep learning methods to fuse the multivariate training dataset and extract multivariate data features. Obtain the ETOPO1 dataset as terrain data and use a convolutional neural network to extract multi-scale features of the terrain. Step 2: Construct a deep learning-based extended-range temperature descaling model, which includes a ResNet34-based encoder, a terrain feature extraction part, and a decoder; input the multivariate training dataset into the model for training, incorporate terrain data, determine the objective function, and update the parameters to update the deep learning-based extended-range temperature descaling model; Step 3: Input the low-resolution temperature forecast data into the already trained deep learning-based extended-range temperature downscaling model to obtain high-resolution downscaling results.

2. The deep learning-based extended-range temperature downscaling method for complex terrain as described in claim 1, characterized in that, Using the extended-range S2S back-calculation dataset obtained in step 1 as the forecast data, and the reanalysis dataset as the target data, firstly, correlation analysis is performed on the forecast data to obtain a four-channel multivariate dataset, where the four channels are 2m air temperature, 2m dew point temperature, 10mV wind speed, and surface shortwave radiation; secondly, the ocean region in the forecast data is masked to align the study area; finally, the dates of the forecast data and the target data are aligned to obtain the climate factor training dataset; specifically: The preprocessing in step 1 includes data alignment, variable selection, data cleaning, and standardization; wherein: Data alignment: There are three types of forecast data: daily average data, cumulative data, and high-frequency data. The cumulative data and high-frequency data are unified into daily average data to ensure the consistency of the dataset in the time dimension. Cumulative data is processed using a differencing method, and high-frequency data is aligned to the daily average time scale by calculating the mean. The specific processing method is as follows: The daily increment is obtained by calculating the difference between consecutive cumulative values, for a given time period. Recorded cumulative variables , its in arrive Daily average within the interval The calculation formula is as follows: High-frequency data is sampled at daily time intervals and aggregated into daily average data through time grouping and mean calculation. For multiple time points within a day... Observed variables Its daily average The calculation formula is as follows: ,in The number of samples taken within a day; Variable selection: The correlation between various meteorological variables and temperature, as well as the cross-correlation among characteristics, are assessed. The correlation coefficient r is calculated using the following formula: In the formula, and These are the observed values ​​of the two variables being calculated; and Its mean; Data cleaning: The ocean regions in the forecast data are processed. Based on the ocean mask generated from the reanalysis dataset, the values ​​of the ocean regions are set to NaN. The ocean mask is generated by marking consecutive NaN regions in the reanalysis dataset and applying an area ratio threshold. Subsequently, NaN values ​​in land regions are filled using a linear interpolation method, while retaining NaN values ​​in ocean regions. The linear interpolation formula is as follows: In the formula, These are interpolation parameters; and For neighboring grid points; Standardization: The training dataset is standardized using the Z-Score standardization method, which standardizes the data based on the mean and standard deviation of the training set. The formula is: In the formula, For data before standardization, For standardized data, , Data respectively The mean and variance are calculated on the training dataset.

3. The deep learning-based extended-range temperature downscaling method for complex terrain as described in claim 2, characterized in that, Step 2: Constructing a deep learning-based extended-range temperature descaling model, which includes a ResNet34-based encoder, a terrain feature extraction component, and a decoder; inputting a multivariate training dataset into the model for training, incorporating terrain data, determining the objective function, and iteratively updating the parameters to update the deep learning-based extended-range temperature descaling model are the specific operations as follows: The terrain feature extraction part in step 2 consists of multiple repeated convolutional blocks. The convolutional blocks gradually reduce the resolution of the high-resolution terrain data through 3×3 convolution operations, while increasing the number of channels. Adaptive average pooling is used to adjust the output of each layer to match the spatial dimension of the encoder. Then, the number of channels is further matched through convolution so as to perform element-wise addition and fusion with the output of each layer of the encoder. The ResNet34-based encoder consists of a preprocessing module and an encoder module. The preprocessing module adjusts the model to adapt to subsequent input scales through 7×7 convolutions and pooling operations. The encoder module is based on different feature extraction layers of the ResNet34 residual network. ResNet34 is a 34-layer residual network consisting of an initial convolutional layer, batch normalization, ReLU activation, and max pooling layer. Its residual structure is divided into four layers: layer1, layer2, layer3, and layer4. Layer1 contains 3 BasicBlocks, each consisting of two 3×3 convolutions, batch normalization, and ReLU. Layer2 contains 4 BasicBlocks. Layer3 contains 6 BasicBlocks. Layer4 contains 3 BasicBlocks. The decoder part uses the DecoderBlock module, which restores the resolution by performing upsampling operation through transposed convolution. Each DecoderBlock includes a 3×3 convolution, batch normalization and ReLU activation, transposed convolution upsampling layer and spatial attention mechanism. The spatial attention mechanism is used to highlight important spatial regions, suppress irrelevant noise, and improve the model's sensitivity to spatial heterogeneity of temperature. The decoder starts from the center block, adjusts its size through bilinear interpolation, and connects encoder features layer by layer to achieve skip connections to preserve high-resolution details; the final upsampling module adjusts to the specified output size through a series of convolutions and transposed convolutions, combined with activation and adaptive average pooling, to ensure the generation of a high-resolution temperature prediction field, while incorporating Dropout to prevent overfitting. The training dataset was input into the extended-range temperature correction model for complex terrain based on ResNet34_Unet for training, and the mean squared error was set as the objective function; its expression is: ,in, This represents the actual 2-meter temperature value from the target dataset. This represents the downscaling result of the model's output forecast temperature. The number of samples; The Adam parameter optimizer is used to train for a fixed number of rounds, and the model with the minimum validation loss is saved as the optimal model.

4. The deep learning-based extended-range temperature downscaling method for complex terrain as described in claim 3, characterized in that, Step 3: Inputting low-resolution temperature forecast data into the trained deep learning-based extended-range temperature downscaling model to obtain high-resolution downscaling results involves the following steps: First, aligning the low-resolution extended-range temperature forecast data to be processed with high-resolution terrain data of the same spatial range and specifications. Then, inputting the terrain data into the trained model. Based on the learned terrain and temperature features, the model simultaneously corrects the systematic bias and local random errors of the forecast data, and performs downscaling by scaling the low-resolution data to the level of the terrain data. Finally, the model output is restored to the actual temperature units, and simple fine-tuning is performed in conjunction with the concurrent ground observation data to obtain the high-resolution extended-range temperature downscaling results.

5. A deep learning-based extended-range temperature downscaling system for complex terrain, characterized in that, The system is used to implement a deep learning-based extended-range temperature downscaling method for complex terrain. The system includes the following modules: The data preprocessing module is used to acquire the dataset and preprocess it to obtain a multivariate training dataset. Deep learning methods are then used to fuse and model the multivariate training dataset and extract multivariate data features. The training module is used to train the deep learning-based extended-range temperature descaling model. By setting the objective function, the module iteratively tunes the parameters to update the deep learning-based extended-range temperature descaling model. The downscaling module is used to correct biases in the data to be downscaled; for the low-resolution forecast data used in the training of the main part of the model, multi-layer transposed convolution is used to upsample the data so that the model outputs high-resolution forecast data.

Images