Extreme temperature event identification method based on multi-scale perception and double threshold clustering

Abstract

The present application relates to a kind of extreme temperature event identification method based on multiscale perception and double threshold clustering, it is related to meteorological monitoring and deep learning technical field, solve the high false negative rate of existing extreme temperature event identification method, existing model is difficult to consider different scale extreme temperature event, data sample is extremely unbalanced, small scale event feature is easy to lose technical problem.The method includes the following steps: data preparation;Based on the quantile-physical double threshold extreme temperature event labeling of ground temperature data;Multi-scale label pre-computation and balanced sampling;Scale perception model construction;Scale perception multi-task loss function design;Model training and extreme temperature event identification.The present application adopts quantile statistical threshold+physical reasonableness threshold, ensure that the labeled event meets statistical anomaly standard;DBSCAN density clustering algorithm is used to the spatial clustering of screened grid, only keep the grid point of clustering success, eliminate isolated noise point, significantly improve the labeling quality.

Description

Technical Field

[0001] This invention relates to the fields of meteorological monitoring and deep learning technology, and in particular to a method for identifying extreme temperature events based on multi-scale sensing and dual-threshold clustering. Background Technology

[0002] Extreme temperature events (such as extreme heat waves and extreme cold waves) can have a severe impact on agricultural production, energy supply, transportation, and people's lives. Accurate identification and prediction of extreme temperature events are of great significance for disaster prevention and mitigation.

[0003] Traditional meteorological event identification methods primarily rely on manually set thresholds and statistical models. These methods suffer from several problems in application. First, the threshold setting for extreme temperature events heavily depends on expert experience, resulting in poor applicability across different regions and a lack of objective, unified judgment criteria, making the identification results susceptible to subjective influence. Second, manual identification suffers from poor timeliness, hindering real-time detection and widespread application. Furthermore, traditional statistical models have limited feature extraction capabilities, failing to capture the complex spatial distribution characteristics and nonlinear relationships of extreme temperature events, thus leading to significant identification errors.

[0004] In recent years, deep learning methods have been applied to meteorological image analysis, but they still face technical challenges stemming from the data itself. First, extreme weather datasets are extremely imbalanced, with few extreme temperature event samples and a very high proportion of normal samples. This leads to a severe bias in model prediction towards normal samples, resulting in a high rate of missed predictions for extreme events. Second, extreme temperature events vary significantly in spatial scale, some appearing as large, continuous coverage areas, while others exhibit localized patchy distributions. Existing models, using fixed receptive fields, struggle to handle extreme temperature events at different spatial scales simultaneously. Furthermore, small-scale extreme temperature events have small areas and weak features, making them easily lost during multiple downsampling processes in deep networks. Finally, traditional single-threshold labeling methods generate a large number of isolated noise grid points. These noise grid points lack spatial continuity, severely impacting training quality.

[0005] Therefore, there is an urgent need for a new method for labeling extreme temperature events and a design for an identification model to solve the above-mentioned technical problems and achieve efficient and high-precision identification of extreme temperature events. Summary of the Invention

[0006] This invention aims to address the technical problems of high false negative rates, difficulty in acknowledging extreme temperature events at different scales, extreme imbalance in data samples, and easy loss of features of small-scale events in existing extreme temperature event identification methods. It provides an extreme temperature event identification method based on multi-scale perception and dual-threshold clustering.

[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: An extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering includes the following steps: Step 1: Data Preparation; Acquire meteorological grid data for the target area and process the acquired meteorological data; Step 2: Annotation of extreme temperature events based on quantile-physical dual threshold of surface temperature data; By using surface temperature data, we can obtain labels for extreme temperature events, providing labels and training objectives for the subsequent training process. Step 3: Multi-scale label pre-computation and balanced sampling; For each valid sample, based on the extreme temperature event labels obtained in step two, calculate whether it contains extreme temperature events, and assign a scale label to it by calculating the proportion of extreme temperature event occurrence points to the total number of sample points. Construct a multi-scale balanced sampler and perform weighted random sampling based on joint weights; Step 4: Scale-aware model construction; Construct a scale-aware deep learning model based on the U-Net architecture; Step 5: Design of scale-aware multi-task loss function; Design a scale-aware joint loss function; Step Six: Model Training and Extreme Temperature Event Recognition; The multi-scale balanced sampler data constructed in step three is used for sampling, which is then input into the deep learning model constructed in step four. The model is trained using the joint loss function in step five, and the extreme temperature event probability and segmentation mask for each image are output. The final recognition results output by the deep learning model are compared with the true values ​​to evaluate the performance level of the deep learning model.

[0008] In the above technical solution, the processing of meteorological data in step one includes: numerical cleaning, outlier handling, and normalization.

[0009] In the above technical solution, step two specifically involves: First, based on historical surface temperature channel data in meteorological data, the specified quantile threshold for each grid point is calculated. Subsequently, the DBSCAN spatial clustering algorithm was used to cluster grid points that meet the dual threshold conditions. By setting an appropriate neighborhood radius and minimum number of samples, the cluster core points, boundary points and noise points were identified, the successfully clustered points were retained and isolated noise points were removed. Finally, the retained cluster points are converted into binary masks, and the same cluster is labeled as the same event, generating an extreme temperature event mask and obtaining extreme temperature event labels.

[0010] In the above technical solution, the method for calculating the joint sampling weight of each sample in step three is as follows: Sample weight = Event weight × Scale weight × Positive sample weight.

[0011] In the above technical solution, the deep learning model in step four includes: a multi-scale feature extraction module, a scale-aware residual block, and a multi-scale attention mechanism module; The multi-scale feature extraction module uses three parallel branches with scale factors of [1, 2, 3]. Each branch independently extracts meteorological features at different scales. Finally, after upsampling to unify the size, the features are fused to achieve comprehensive extraction of data features at different scales by the deep learning model. Scale-aware residual blocks are used to sample small-scale branches and large-scale branches respectively. Small-scale branches use small-scale convolutional kernels to fully learn their features. Large-scale branches use dilated convolutions with an inflation rate of 2 to actively encourage the deep learning model to expand its receptive field and identify the occurrence of large-scale events as a whole. The multi-scale attention mechanism module is used to learn channel attention and spatial attention; learning channel attention is used to obtain the importance weights between different channels, and learning spatial attention is used to obtain the importance weights of different spatial locations.

[0012] In the above technical solution, step five: the joint loss function is specifically as follows: in, For Tversky's loss, For Focal loss, For BCE loss, , , These are the scaling parameters for Tversky loss, Focal loss, and BCE loss, respectively.

[0013] In the above technical solution, the formula for calculating Tversky loss is as follows: in, , and These are all scale parameters. , , These represent true positive, false positive, and false negative, respectively.

[0014] In the above technical solution, the formula for calculating Focal loss is as follows: in, For real labels It is the probability that this point is classified as an extreme point; and These are all scale parameters.

[0015] In the above technical solution, the formula for calculating BCE loss is: in, For predicted values, This is the actual value.

[0016] In the above technical solution, when evaluating the performance level of the deep learning model in step six: For classification tasks that determine whether extreme temperature events have occurred, the metrics used include: precision, recall, and F1 score. For image segmentation tasks in the region where the event occurred, the metrics used include the Dice coefficient and IoU.

[0017] The present invention has the following beneficial effects: 1. Addressing the annotation noise issue for extreme temperature events: The extreme temperature event identification method based on multi-scale perception and dual-threshold clustering of the present invention adopts a dual screening of "quantile statistical threshold + physical rationality threshold" to ensure that the labeled events meet both statistical anomaly criteria and physical interpretability; the DBSCAN density clustering algorithm is used to spatially cluster the screened grid points, retaining only the successfully clustered grid points and eliminating isolated noise points, which significantly improves the labeling quality.

[0018] 2. Addressing the issue of balancing multiple scales of extreme temperature events: The extreme temperature event identification method based on multi-scale perception and dual-threshold clustering of the present invention extracts features of different scales in parallel through a multi-scale feature extraction module according to the spatial distribution characteristics of extreme temperature events at different scales.

[0019] The extreme temperature event recognition method based on multi-scale perception and dual-threshold clustering of the present invention adopts a multi-scale perception deep learning model. Through specially designed scale-aware residual blocks, the receptive field is adaptively adjusted according to the characteristics of events at different scales: small-scale events use small convolutional kernels and more network layers to preserve details, while large-scale events use dilated convolutions to increase the receptive field to capture the overall distribution, which significantly improves the segmentation accuracy and classification performance of extreme temperature events.

[0020] 3. Regarding the problem of extremely imbalanced data samples: The extreme temperature event identification method based on multi-scale perception and dual-threshold clustering of the present invention constructs a scale-aware joint weighted sampler, namely a multi-scale balanced sampler, which assigns higher sampling weights to small-scale events, thereby alleviating the problems of sample imbalance and loss of small-scale events from the data level.

[0021] 4. Regarding the issue of feature loss in small-scale events: This invention presents an extreme temperature event identification method based on multi-scale perception and dual-threshold clustering. It designs a scale-aware loss function and assigns differentiated weights to extreme temperature events of different scales by designing scale-weighted Tversky and Focal losses, making the model more focused on small-scale events that are difficult to classify. This invention can effectively identify both large-area continuous extreme temperature events (large-scale) and accurately locate localized patchy extreme temperature events (small-scale), solving the problem that traditional models struggle to handle events of different scales simultaneously. Attached Figure Description

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0023] Figure 1 This is a flowchart illustrating the steps of the extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering of the present invention.

[0024] Figure 2 This is a schematic diagram of a scale-aware deep learning model structure based on the U-Net architecture.

[0025] Figure 3 A visualization of extreme temperature detection results. Detailed Implementation

[0026] The inventive concept of this invention is as follows: The extreme temperature event identification method based on multi-scale perception and dual-threshold clustering of the present invention mainly includes steps such as data preparation, quantile-physical dual-threshold extreme temperature event labeling, multi-scale balanced sampling, scale-aware model construction, and multi-scale loss function construction. It solves the problem of low accuracy of extreme temperature identification caused by the difficulty in labeling extreme weather datasets and sample imbalance in the prior art, and improves the accuracy of extreme temperature detection and identification.

[0027] The present invention will now be described in detail with reference to the accompanying drawings.

[0028] like Figure 1 As shown, the extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering of the present invention specifically includes the following steps: Step 1: Data Preparation; Acquire gridded meteorological data for the target area, including atmospheric variable data and surface variable data (including a 2-meter temperature data channel). Perform numerical cleaning, outlier removal, and normalization on all meteorological data.

[0029] Step 2: Annotation of extreme temperature events based on quantile-physical dual threshold of surface temperature data; This step aims to obtain labels for extreme temperature events using surface temperature data, providing labels and training targets for subsequent training processes. To obtain more reliable labels, this invention employs a labeling method combining quantile thresholds and physical thresholds.

[0030] First, based on historical surface temperature channel data from meteorological data, a specified quantile threshold is calculated for each grid point. For extreme high-temperature events, the 95th percentile is used; for extreme low-temperature events, the 5th percentile is used.

[0031] Based on this, in order to ensure that event annotations are physically interpretable, this invention uses physical thresholds to check the physical rationality of grid points that exceed quantile thresholds. For extreme high-temperature events, the temperature value should be higher than the upper limit of the climatological normal range (e.g., 35°C), and for extreme low-temperature events, the temperature value should be lower than the lower limit of the climatological normal range (e.g., -10°C). Finally, only grid points that simultaneously meet the quantile threshold and physical threshold are retained.

[0032] Subsequently, this invention employs the DBSCAN spatial clustering algorithm to cluster grid points that meet the dual threshold conditions. By setting an appropriate neighborhood radius and minimum number of samples, it identifies cluster core points, boundary points, and noise points, retaining only successfully clustered points (core points and boundary points) and eliminating isolated noise points to remove the impact of events such as localized wildfires on the accuracy of annotation.

[0033] Finally, the retained cluster points are converted into binary masks, and the same cluster is labeled as the same event, generating an extreme temperature event mask and obtaining extreme temperature event labels.

[0034] Step 3: Multi-scale label pre-computation and balanced sampling; This step aims to construct multi-scale labels for the samples while mitigating the problem of sample imbalance through a balanced sampling mechanism.

[0035] For each valid sample, calculate whether it contains extreme temperature events based on the extreme temperature event mask generated in step two, and assign a scale label (small scale or large scale) to it by calculating the proportion of extreme temperature event occurrence points to the total number of sample points.

[0036] A multi-scale balanced sampler is constructed, performing weighted random sampling based on joint weights. The joint sampling weight for each sample is calculated as follows: Sample weight = Event weight × Scale weight × Positive sample weight For example, in this invention, the weight of small-scale events can be set to 5.0, and the weight of large-scale events can be set to 0.5. The event weights can be set according to the difficulty of identifying various events (extreme low temperature and extreme high temperature).

[0037] Step 4: Scale-aware model construction; This step aims to build a scale-aware deep learning model based on the U-Net architecture. Its main modules include a multi-scale feature extraction module, scale-aware residual blocks, and a multi-scale attention mechanism module, as shown in the overall structure diagram below. Figure 2 As shown.

[0038] The multi-scale feature extraction module uses three parallel branches with scale factors of [1, 2, 3]. Each branch independently extracts meteorological features at different scales. Finally, after upsampling to unify the size, the features are fused to achieve comprehensive extraction of data features at different scales. Scale-aware residual blocks are used to sample different convolution forms and network stacking methods for small-scale and large-scale branches respectively. The small-scale branch uses 3×3 small-scale convolution kernels to fully learn its features, while the large-scale branch uses dilated convolution with an inflation rate of 2 to actively encourage the model to expand its receptive field and identify the occurrence of large-scale events as a whole.

[0039] The multi-scale attention mechanism module learns both channel attention and spatial attention simultaneously. Channel attention learns the importance weights between different channels, while spatial attention employs a multi-scale design to learn the importance weights of different spatial locations. Ultimately, these attention weights are adaptively fused to construct a comprehensive multi-scale attention mechanism.

[0040] Step 5: Design of scale-aware multi-task loss function; This step aims to design a scale-aware joint loss function. The joint loss function used in this invention mainly includes three parts: Tversky loss, Focal loss, and BCE loss.

[0041] The formula for calculating Tversky loss is as follows: in, , and These are all scale parameters. , , These represent true positives, false positives, and false negatives, respectively. The loss function can be adjusted... and balance and This makes the model more focused on reducing false negatives, thus making it more sensitive to small targets and allowing for targeted optimization of small target detection.

[0042] The formula for calculating Focal loss is as follows: in, For real labels It is the probability that this point is classified as an extreme point; and These are all scale parameters. By adjusting the scale parameters, the weights can be adjusted according to the event scale, thereby reducing the weight of easily classified samples and making the model focus on difficult-to-classify samples.

[0043] The formula for calculating BCE loss is: in, For predicted values, The true value is represented. The BCE loss forms the base loss of the joint loss function.

[0044] Ultimately, the training loss of the model is defined as a weighted sum of the three loss functions.

[0045] in, For Tversky's loss, For Focal loss, For BCE loss, , , These are the scaling parameters for Tversky loss, Focal loss, and BCE loss, respectively.

[0046] Step Six: Model Training and Extreme Temperature Event Recognition; The data from the multi-scale balanced sampler constructed in step three is used for sampling and input into the model constructed in step four. The model is trained using the joint loss function from step five, and the outputs are the extreme temperature event probabilities and segmentation masks for each image. The final recognition results output by the model are compared with the true values ​​to evaluate the model's performance.

[0047] This invention uses the following five parameters as evaluation criteria. For the classification task of determining whether an extreme temperature event has occurred, this invention uses three metrics: precision, recall, and F1 score. For the image segmentation task of the event occurrence region, this invention uses two metrics: Dice coefficient and IoU.

[0048] The following examples illustrate the extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering of the present invention.

[0049] 1. Dataset and preprocessing; ERA5 meteorological reanalysis data was used for model performance testing. The data has a temporal resolution of 6 hours, a spatial resolution of 0.25°×0.25°, and covers the eastern part of China (longitude 80°-140°, latitude 60°-0°).

[0050] The data includes 13 pressure layers (50-1000 hPa), with 5 variables (t, u, v, z, q) per layer, totaling 65 upper-level channels, as well as 6 surface channels for surface temperature, surface dew point temperature, radial wind, zonal wind, precipitation, and mean sea level pressure, for a total of 71 meteorological data channels.

[0051] Subsequently, z-score normalization was performed on the meteorological reanalysis data, and the normalization parameters were retained to restore the true output.

[0052] 2. Model Training Strategy For labeling extreme low-temperature events, a quantile threshold of 5% and a physical threshold of -10℃ were used. The neighborhood radius of the DBSCAN algorithm was set to 2 grid points, and the minimum number of samples was 10. For sampling weights, the base weight for positive samples was set to 2.0, the base weight for negative samples was 1.0, the scale weight for large-scale samples was set to 0.5, and the scale weight for small-scale samples was set to 5.0. In the joint loss function, the weights of the Tversky loss, Focal loss, and BCE loss were set to 0.4, 0.4, and 0.2, respectively.

[0053] Data from 2019 to 2023 was used as the training set, data from 2024 as the validation set, and data from 2025 as the independent test set. An NVIDIA GeForce RTX 3090 GPU was used for model training and testing. The Adam optimizer was employed, with an initial learning rate set to 5e. -4 The batch size was set to 32, and the total number of training epochs was 100. After each training epoch, the loss was calculated on the validation set, and an "early stopping" strategy was adopted. Finally, the model with the smallest loss on the validation set was saved for testing.

[0054] 3. Performance Comparison and Verification of Beneficial Effects On independent test sets, the performance of the proposed model and the standard U-Net network were tested respectively, and the key performance indicators are compared in Table 1. As shown in Table 1, compared to the standard U-Net architecture, the proposed method achieves superior performance in both segmentation and classification tasks. High precision indicates that the method has an extremely low false positive rate for extreme low-temperature events, demonstrating the effectiveness of dual-threshold clustering annotation and multi-scale balanced sampling. High recall indicates that the method can effectively identify most images containing real extreme low-temperature events, with a low false negative rate, demonstrating the scale-aware network's excellent ability to capture events at different scales.

[0055] Table 1 Performance Indicators for Extreme Low Temperature Detection

[0056] Figure 3 The model of this invention is shown for two samples in the test set, one belonging to a large-scale event and the other to a small-scale event. The first row shows the range of the extreme low-temperature event predicted by the model, the second row shows the binarized result of the prediction, and the third row shows the true label. It can be seen that the model proposed in this invention accurately segments the extreme low-temperature occurrence region with clear boundaries, indicating that the method of this invention can achieve accurate extreme temperature identification.

[0057] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for identifying extreme temperature events based on multi-scale sensing and dual-threshold clustering, characterized in that, Includes the following steps: Step 1: Data Preparation; Acquire meteorological grid data for the target area and process the acquired meteorological data; Step 2: Annotation of extreme temperature events based on quantile-physical dual threshold of surface temperature data; By using surface temperature data, we can obtain labels for extreme temperature events, providing labels and training objectives for the subsequent training process. Step 3: Multi-scale label pre-computation and balanced sampling; For each valid sample, based on the extreme temperature event labels obtained in step two, calculate whether it contains extreme temperature events, and assign a scale label to it by calculating the proportion of extreme temperature event occurrence points to the total number of sample points. Construct a multi-scale balanced sampler and perform weighted random sampling based on joint weights; Step 4: Scale-aware model construction; Construct a scale-aware deep learning model based on the U-Net architecture; Step 5: Design of scale-aware multi-task loss function; Design a scale-aware joint loss function; Step Six: Model Training and Extreme Temperature Event Recognition; The multi-scale balanced sampler data constructed in step three is used for sampling, which is then input into the deep learning model constructed in step four. The model is trained using the joint loss function in step five, and the extreme temperature event probability and segmentation mask for each image are output. The final recognition results output by the deep learning model are compared with the true values ​​to evaluate the performance level of the deep learning model.

2. The extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering according to claim 1, characterized in that, The processing of meteorological data in step one includes: numerical cleaning, outlier handling, and normalization.

3. The extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering according to claim 1, characterized in that, Step two is as follows: First, based on historical surface temperature channel data in meteorological data, the specified quantile threshold for each grid point is calculated. Subsequently, the DBSCAN spatial clustering algorithm was used to cluster grid points that meet the dual threshold conditions. By setting an appropriate neighborhood radius and minimum number of samples, the cluster core points, boundary points and noise points were identified, the successfully clustered points were retained and isolated noise points were removed. Finally, the retained cluster points are converted into binary masks, and the same cluster is labeled as the same event, generating an extreme temperature event mask and obtaining extreme temperature event labels.

4. The extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering according to claim 1, characterized in that, In step three, the joint sampling weight for each sample is calculated as follows: Sample weight = Event weight × Scale weight × Positive sample weight.

5. The extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering according to claim 1, characterized in that, The deep learning model in step four includes: a multi-scale feature extraction module, a scale-aware residual block, and a multi-scale attention mechanism module; The multi-scale feature extraction module uses three parallel branches with scale factors of [1, 2, 3]. Each branch independently extracts meteorological features at different scales. Finally, after upsampling to unify the size, the features are fused to achieve comprehensive extraction of data features at different scales by the deep learning model. Scale-aware residual blocks are used to sample small-scale branches and large-scale branches respectively. Small-scale branches use small-scale convolutional kernels to fully learn their features. Large-scale branches use dilated convolutions with an inflation rate of 2 to actively encourage the deep learning model to expand its receptive field and identify the occurrence of large-scale events as a whole. The multi-scale attention mechanism module is used to learn channel attention and spatial attention; learning channel attention is used to obtain the importance weights between different channels, and learning spatial attention is used to obtain the importance weights of different spatial locations.

6. The extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering according to claim 1, characterized in that, In step five: the joint loss function is specifically as follows: in, For Tversky's loss, For Focal loss, For BCE loss, , , These are the scaling parameters for Tversky loss, Focal loss, and BCE loss, respectively.

7. The extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering according to claim 6, characterized in that, The formula for calculating Tversky loss is as follows: in, , and These are all scale parameters. , , These represent true positive, false positive, and false negative, respectively.

8. The extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering according to claim 6, characterized in that, The formula for calculating Focal loss is as follows: in, For real labels, It is the probability that this point is classified as an extreme point; and These are all scale parameters.

9. The extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering according to claim 6, characterized in that, The formula for calculating BCE loss is: in, For predicted values, This is the actual value.

10. The extreme temperature event identification method based on multi-scale sensing and dual-threshold clustering according to claim 1, characterized in that, When evaluating the performance level of the deep learning model in step six: For classification tasks that determine whether extreme temperature events have occurred, the metrics used include: precision, recall, and F1 score. For image segmentation tasks in the region where the event occurred, the metrics used include the Dice coefficient and IoU.

Images